Showing 1–50 of 158 results for author: Agarwal, R

Towards Compute-Optimal Many-Shot In-Context Learning

Authors: Shahriar Golchin, Yanfei Chen, Rujun Han, Manan Gandhi, Tianli Yu, Swaroop Mishra, Mihai Surdeanu, Rishabh Agarwal, Chen-Yu Lee, Tomas Pfister

Abstract: Long-context large language models (LLMs) are able to process inputs containing up to several million tokens. In the scope of in-context learning (ICL), this translates into using hundreds/thousands of demonstrations in the input prompt, enabling many-shot ICL. In practice, a fixed set of demonstrations is often selected at random in many-shot settings due to (1) high inference costs, (2) the bene… ▽ More Long-context large language models (LLMs) are able to process inputs containing up to several million tokens. In the scope of in-context learning (ICL), this translates into using hundreds/thousands of demonstrations in the input prompt, enabling many-shot ICL. In practice, a fixed set of demonstrations is often selected at random in many-shot settings due to (1) high inference costs, (2) the benefits of caching and reusing computations, and (3) the similar performance offered by this strategy compared to others when scaled. In this work, we propose two straightforward strategies for demonstration selection in many-shot ICL that improve performance with minimal computational overhead. Our first method combines a small number of demonstrations, selected based on their similarity to each test sample, with a disproportionately larger set of random demonstrations that are cached. The second strategy improves the first by replacing random demonstrations with those selected using centroids derived from test sample representations via k-means clustering. Our experiments with Gemini Pro and Flash across several datasets indicate that our strategies consistently outperform random selection and surpass or match the most performant selection approach while supporting caching and reducing inference cost by up to an order of magnitude. We also show that adjusting the proportion of demonstrations selected based on different criteria can balance performance and inference cost in many-shot ICL. △ Less

Submitted 22 July, 2025; originally announced July 2025.

Comments: Final version; accepted at COLM 2025

SynthTextEval: Synthetic Text Data Generation and Evaluation for High-Stakes Domains

Authors: Krithika Ramesh, Daniel Smolyak, Zihao Zhao, Nupoor Gandhi, Ritu Agarwal, Margrét Bjarnadóttir, Anjalie Field

Abstract: We present SynthTextEval, a toolkit for conducting comprehensive evaluations of synthetic text. The fluency of large language model (LLM) outputs has made synthetic text potentially viable for numerous applications, such as reducing the risks of privacy violations in the development and deployment of AI systems in high-stakes domains. Realizing this potential, however, requires principled consiste… ▽ More We present SynthTextEval, a toolkit for conducting comprehensive evaluations of synthetic text. The fluency of large language model (LLM) outputs has made synthetic text potentially viable for numerous applications, such as reducing the risks of privacy violations in the development and deployment of AI systems in high-stakes domains. Realizing this potential, however, requires principled consistent evaluations of synthetic data across multiple dimensions: its utility in downstream systems, the fairness of these systems, the risk of privacy leakage, general distributional differences from the source text, and qualitative feedback from domain experts. SynthTextEval allows users to conduct evaluations along all of these dimensions over synthetic data that they upload or generate using the toolkit's generation module. While our toolkit can be run over any data, we highlight its functionality and effectiveness over datasets from two high-stakes domains: healthcare and law. By consolidating and standardizing evaluation metrics, we aim to improve the viability of synthetic text, and in-turn, privacy-preservation in AI development. △ Less

Submitted 9 July, 2025; originally announced July 2025.

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

Authors: Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, Luke Marris, Sam Petulla, Colin Gaffney, Asaf Aharoni, Nathan Lintz, Tiago Cardal Pais, Henrik Jacobsson, Idan Szpektor, Nan-Jiang Jiang, Krishna Haridasan, Ahmed Omran, Nikunj Saunshi, Dara Bahri, Gaurav Mishra, Eric Chu , et al. (3284 additional authors not shown)

Abstract: In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal unde… ▽ More In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving. △ Less

Submitted 22 July, 2025; v1 submitted 7 July, 2025; originally announced July 2025.

Comments: 72 pages, 17 figures

Online Planning for Cooperative Air-Ground Robot Systems with Unknown Fuel Requirements

Authors: Ritvik Agarwal, Behnoushsadat Hatami, Alvika Gautam, Parikshit Maini

Abstract: We consider an online variant of the fuel-constrained UAV routing problem with a ground-based mobile refueling station (FCURP-MRS), where targets incur unknown fuel costs. We develop a two-phase solution: an offline heuristic-based planner computes initial UAV and UGV paths, and a novel online planning algorithm that dynamically adjusts rendezvous points based on real-time fuel consumption during… ▽ More We consider an online variant of the fuel-constrained UAV routing problem with a ground-based mobile refueling station (FCURP-MRS), where targets incur unknown fuel costs. We develop a two-phase solution: an offline heuristic-based planner computes initial UAV and UGV paths, and a novel online planning algorithm that dynamically adjusts rendezvous points based on real-time fuel consumption during target processing. Preliminary Gazebo simulations demonstrate the feasibility of our approach in maintaining UAV-UGV path validity, ensuring mission completion. Link to video: https://youtu.be/EmpVj-fjqNY △ Less

Submitted 25 June, 2025; originally announced June 2025.

Comments: Submitted to RSS (MRS Workshop)

FEWSim: A Visual Analytic Framework for Exploring the Nexus of Food-Energy-Water Simulations

Authors: Fan Lei, David A. Sampson, Jiayi Hong, Yuxin Ma, Giuseppe Mascaro, Dave White, Rimjhim Agarwal, Ross Maciejewski

Abstract: The interdependencies of food, energy, and water (FEW) systems create a nexus opportunity to explore the strengths and vulnerabilities of individual and cross-sector interactions within FEW systems. However, the variables quantifying nexus interactions are hard to observe, which hinders the cross-sector analysis. To overcome such challenges, we present FEWSim, a visual analytics framework designed… ▽ More The interdependencies of food, energy, and water (FEW) systems create a nexus opportunity to explore the strengths and vulnerabilities of individual and cross-sector interactions within FEW systems. However, the variables quantifying nexus interactions are hard to observe, which hinders the cross-sector analysis. To overcome such challenges, we present FEWSim, a visual analytics framework designed to support domain experts in exploring and interpreting simulation results from a coupled FEW model. FEWSim employs a three-layer asynchronous architecture: the model layer integrates food, energy, and water models to simulate the FEW nexus; the middleware layer manages scenario configuration and execution; and the visualization layer provides interactive visual exploration of simulated time-series results across FEW sectors. The visualization layer further facilitates the exploration across multiple scenarios and evaluates scenario differences in performance using sustainability indices of the FEW nexus. We demonstrate the utility of FEWSim through a case study for the Phoenix Active Management Area (AMA) in Arizona. △ Less

Submitted 16 June, 2025; originally announced June 2025.

Comments: Accepted by IEEE Computer Graphics and Applications (CG&A)

SARAL-Bot: Autonomous Robot for Strawberry Plant Care

Authors: Arif Ahmed, Ritvik Agarwal, Gaurav Srikar, Nathaniel Rose, Parikshit Maini

Abstract: Strawberry farming demands intensive labor for monitoring and maintaining plant health. To address this, Team SARAL develops an autonomous robot for the 2024 ASABE Student Robotics Challenge, capable of navigation, unhealthy leaf detection, and removal. The system addresses labor shortages, reduces costs, and supports sustainable farming through vision-based plant assessment. This work demonstrate… ▽ More Strawberry farming demands intensive labor for monitoring and maintaining plant health. To address this, Team SARAL develops an autonomous robot for the 2024 ASABE Student Robotics Challenge, capable of navigation, unhealthy leaf detection, and removal. The system addresses labor shortages, reduces costs, and supports sustainable farming through vision-based plant assessment. This work demonstrates the potential of robotics to modernize strawberry cultivation and enable scalable, intelligent agricultural solutions. △ Less

Submitted 7 June, 2025; originally announced June 2025.

Comments: Awarded Best Written Report @ Robotics Design Challenge (Advanced), ASABE 2024

Overcoming Challenges of Partial Client Participation in Federated Learning : A Comprehensive Review

Authors: Mrinmay Sen, Shruti Aparna, Rohit Agarwal, Chalavadi Krishna Mohan

Abstract: Federated Learning (FL) is a learning mechanism that falls under the distributed training umbrella, which collaboratively trains a shared global model without disclosing the raw data from different clients. This paper presents an extensive survey on the impact of partial client participation in federated learning. While much of the existing research focuses on addressing issues such as generalizat… ▽ More Federated Learning (FL) is a learning mechanism that falls under the distributed training umbrella, which collaboratively trains a shared global model without disclosing the raw data from different clients. This paper presents an extensive survey on the impact of partial client participation in federated learning. While much of the existing research focuses on addressing issues such as generalization, robustness, and fairness caused by data heterogeneity under the assumption of full client participation, limited attention has been given to the practical and theoretical challenges arising from partial client participation, which is common in real-world scenarios. This survey provides an in-depth review of existing FL methods designed to cope with partial client participation. We offer a comprehensive analysis supported by theoretical insights and empirical findings, along with a structured categorization of these methods, highlighting their respective advantages and disadvantages. △ Less

Submitted 6 June, 2025; v1 submitted 3 June, 2025; originally announced June 2025.

Comments: 15 pages, 6 tables, comprehensive survey of federated learning with partial client participation

REDDIX-NET: A Novel Dataset and Benchmark for Moderating Online Explicit Services

Authors: MSVPJ Sathvik, Manan Roy Choudhury, Rishita Agarwal, Sathwik Narkedimilli, Vivek Gupta

Abstract: The rise of online platforms has enabled covert illicit activities, including online prostitution, to pose challenges for detection and regulation. In this study, we introduce REDDIX-NET, a novel benchmark dataset specifically designed for moderating online sexual services and going beyond traditional NSFW filters. The dataset is derived from thousands of web-scraped NSFW posts on Reddit and categ… ▽ More The rise of online platforms has enabled covert illicit activities, including online prostitution, to pose challenges for detection and regulation. In this study, we introduce REDDIX-NET, a novel benchmark dataset specifically designed for moderating online sexual services and going beyond traditional NSFW filters. The dataset is derived from thousands of web-scraped NSFW posts on Reddit and categorizes users into six behavioral classes reflecting different service offerings and user intentions. We evaluate the classification performance of state-of-the-art large language models (GPT-4, LlaMA 3.3-70B-Instruct, Gemini 1.5 Flash, Mistral 8x7B, Qwen 2.5 Turbo, Claude 3.5 Haiku) using advanced quantitative metrics, finding promising results with models like GPT-4 and Gemini 1.5 Flash. Beyond classification, we conduct sentiment and comment analysis, leveraging LLM and PLM-based approaches and metadata extraction to uncover behavioral and temporal patterns. These analyses reveal peak engagement times and distinct user interaction styles across categories. Our findings provide critical insights into AI-driven moderation and enforcement, offering a scalable framework for platforms to combat online prostitution and associated harms. △ Less

Submitted 29 May, 2025; originally announced May 2025.

Comments: 29 pages, 15 figures

Automated Meta Prompt Engineering for Alignment with the Theory of Mind

Authors: Aaron Baughman, Rahul Agarwal, Eduardo Morales, Gozde Akay

Abstract: We introduce a method of meta-prompting that jointly produces fluent text for complex tasks while optimizing the similarity of neural states between a human's mental expectation and a Large Language Model's (LLM) neural processing. A technique of agentic reinforcement learning is applied, in which an LLM as a Judge (LLMaaJ) teaches another LLM, through in-context learning, how to produce content b… ▽ More We introduce a method of meta-prompting that jointly produces fluent text for complex tasks while optimizing the similarity of neural states between a human's mental expectation and a Large Language Model's (LLM) neural processing. A technique of agentic reinforcement learning is applied, in which an LLM as a Judge (LLMaaJ) teaches another LLM, through in-context learning, how to produce content by interpreting the intended and unintended generated text traits. To measure human mental beliefs around content production, users modify long form AI-generated text articles before publication at the US Open 2024 tennis Grand Slam. Now, an LLMaaJ can solve the Theory of Mind (ToM) alignment problem by anticipating and including human edits within the creation of text from an LLM. Throughout experimentation and by interpreting the results of a live production system, the expectations of human content reviewers had 100% of alignment with AI 53.8% of the time with an average iteration count of 4.38. The geometric interpretation of content traits such as factualness, novelty, repetitiveness, and relevancy over a Hilbert vector space combines spatial volume (all trait importance) with vertices alignment (individual trait relevance) enabled the LLMaaJ to optimize on Human ToM. This resulted in an increase in content quality by extending the coverage of tennis action. Our work that was deployed at the US Open 2024 has been used across other live events within sports and entertainment. △ Less

Submitted 13 May, 2025; originally announced May 2025.

Comments: 9 pages, 6 figures, 3 tables

Putting the Value Back in RL: Better Test-Time Scaling by Unifying LLM Reasoners With Verifiers

Authors: Kusha Sareen, Morgane M Moss, Alessandro Sordoni, Rishabh Agarwal, Arian Hosseini

Abstract: Prevalent reinforcement learning~(RL) methods for fine-tuning LLM reasoners, such as GRPO or Leave-one-out PPO, abandon the learned value function in favor of empirically estimated returns. This hinders test-time compute scaling that relies on using the value-function for verification. In this work, we propose RL$^V$ that augments any ``value-free'' RL method by jointly training the LLM as both a… ▽ More Prevalent reinforcement learning~(RL) methods for fine-tuning LLM reasoners, such as GRPO or Leave-one-out PPO, abandon the learned value function in favor of empirically estimated returns. This hinders test-time compute scaling that relies on using the value-function for verification. In this work, we propose RL$^V$ that augments any ``value-free'' RL method by jointly training the LLM as both a reasoner and a generative verifier using RL-generated data, adding verification capabilities without significant overhead. Empirically, RL$^V$ boosts MATH accuracy by over 20\% with parallel sampling and enables $8-32\times$ efficient test-time compute scaling compared to the base RL method. RL$^V$ also exhibits strong generalization capabilities for both easy-to-hard and out-of-domain tasks. Furthermore, RL$^V$ achieves $1.2-1.6\times$ higher performance when jointly scaling parallel and sequential test-time compute with a long reasoning R1 model. △ Less

Submitted 7 May, 2025; originally announced May 2025.

Linguistic Complexity and Socio-cultural Patterns in Hip-Hop Lyrics

Authors: Aayam Bansal, Raghav Agarwal, Kaashvi Jain

Abstract: This paper presents a comprehensive computational framework for analyzing linguistic complexity and socio-cultural trends in hip-hop lyrics. Using a dataset of 3,814 songs from 146 influential artists spanning four decades (1980-2020), we employ natural language processing techniques to quantify multiple dimensions of lyrical complexity. Our analysis reveals a 23.7% increase in vocabulary diversit… ▽ More This paper presents a comprehensive computational framework for analyzing linguistic complexity and socio-cultural trends in hip-hop lyrics. Using a dataset of 3,814 songs from 146 influential artists spanning four decades (1980-2020), we employ natural language processing techniques to quantify multiple dimensions of lyrical complexity. Our analysis reveals a 23.7% increase in vocabulary diversity over the study period, with East Coast artists demonstrating 17.3% higher lexical variation than other regions. Rhyme density increased by 34.2% across all regions, with Midwest artists exhibiting the highest technical complexity (3.04 rhymes per line). Topic modeling identified significant shifts in thematic content, with social justice themes decreasing from 28.5% to 13.8% of content while introspective themes increased from 7.6% to 26.3%. Sentiment analysis demon- strated that lyrics became significantly more negative during sociopolitical crises, with polarity decreasing by 0.31 following major social unrest. Multi-dimensional analysis revealed four dis- tinct stylistic approaches that correlate strongly with geographic origin (r=0.68, p!0.001) and time period (r=0.59, p<0.001). These findings establish quantitative evidence for the evolution of hip- hop as both an art form and a reflection of societal dynamics, providing insights into the interplay between linguistic innovation and cultural context in popular music. △ Less

Submitted 29 April, 2025; originally announced May 2025.

Comments: 12 pages

Process Reward Models That Think

Authors: Muhammad Khalifa, Rishabh Agarwal, Lajanugen Logeswaran, Jaekyeom Kim, Hao Peng, Moontae Lee, Honglak Lee, Lu Wang

Abstract: Step-by-step verifiers -- also known as process reward models (PRMs) -- are a key ingredient for test-time scaling. PRMs require step-level supervision, making them expensive to train. This work aims to build data-efficient PRMs as verbalized step-wise reward models that verify every step in the solution by generating a verification chain-of-thought (CoT). We propose ThinkPRM, a long CoT verifier… ▽ More Step-by-step verifiers -- also known as process reward models (PRMs) -- are a key ingredient for test-time scaling. PRMs require step-level supervision, making them expensive to train. This work aims to build data-efficient PRMs as verbalized step-wise reward models that verify every step in the solution by generating a verification chain-of-thought (CoT). We propose ThinkPRM, a long CoT verifier fine-tuned on orders of magnitude fewer process labels than those required by discriminative PRMs. Our approach capitalizes on the inherent reasoning abilities of long CoT models, and outperforms LLM-as-a-Judge and discriminative verifiers -- using only 1% of the process labels in PRM800K -- across several challenging benchmarks. Specifically, ThinkPRM beats the baselines on ProcessBench, MATH-500, and AIME '24 under best-of-N selection and reward-guided search. In an out-of-domain evaluation on a subset of GPQA-Diamond and LiveCodeBench, our PRM surpasses discriminative verifiers trained on the full PRM800K by 8% and 4.5%, respectively. Lastly, under the same token budget, ThinkPRM scales up verification compute more effectively compared to LLM-as-a-Judge, outperforming it by 7.2% on a subset of ProcessBench. Our work highlights the value of generative, long CoT PRMs that can scale test-time compute for verification while requiring minimal supervision for training. Our code, data, and models will be released at https://github.com/mukhal/thinkprm. △ Less

Submitted 23 June, 2025; v1 submitted 23 April, 2025; originally announced April 2025.

Haphazard Inputs as Images in Online Learning

Authors: Rohit Agarwal, Aryan Dessai, Arif Ahmed Sekh, Krishna Agarwal, Alexander Horsch, Dilip K. Prasad

Abstract: The field of varying feature space in online learning settings, also known as haphazard inputs, is very prominent nowadays due to its applicability in various fields. However, the current solutions to haphazard inputs are model-dependent and cannot benefit from the existing advanced deep-learning methods, which necessitate inputs of fixed dimensions. Therefore, we propose to transform the varying… ▽ More The field of varying feature space in online learning settings, also known as haphazard inputs, is very prominent nowadays due to its applicability in various fields. However, the current solutions to haphazard inputs are model-dependent and cannot benefit from the existing advanced deep-learning methods, which necessitate inputs of fixed dimensions. Therefore, we propose to transform the varying feature space in an online learning setting to a fixed-dimension image representation on the fly. This simple yet novel approach is model-agnostic, allowing any vision-based models to be applicable for haphazard inputs, as demonstrated using ResNet and ViT. The image representation handles the inconsistent input data seamlessly, making our proposed approach scalable and robust. We show the efficacy of our method on four publicly available datasets. The code is available at https://github.com/Rohit102497/HaphazardInputsAsImages. △ Less

Submitted 3 April, 2025; originally announced April 2025.

Comments: Accepted at IJCNN 2025

Finding Interest Needle in Popularity Haystack: Improving Retrieval by Modeling Item Exposure

Authors: Rahul Agarwal, Amit Jaspal, Saurabh Gupta, Omkar Vichare

Abstract: Recommender systems operate in closed feedback loops, where user interactions reinforce popularity bias, leading to over-recommendation of already popular items while under-exposing niche or novel content. Existing bias mitigation methods, such as Inverse Propensity Scoring (IPS) and Off-Policy Correction (OPC), primarily operate at the ranking stage or during training, lacking explicit real-time… ▽ More Recommender systems operate in closed feedback loops, where user interactions reinforce popularity bias, leading to over-recommendation of already popular items while under-exposing niche or novel content. Existing bias mitigation methods, such as Inverse Propensity Scoring (IPS) and Off-Policy Correction (OPC), primarily operate at the ranking stage or during training, lacking explicit real-time control over exposure dynamics. In this work, we introduce an exposure-aware retrieval scoring approach, which explicitly models item exposure probability and adjusts retrieval-stage ranking at inference time. Unlike prior work, this method decouples exposure effects from engagement likelihood, enabling controlled trade-offs between fairness and engagement in large-scale recommendation platforms. We validate our approach through online A/B experiments in a real-world video recommendation system, demonstrating a 25% increase in uniquely retrieved items and a 40% reduction in the dominance of over-popular content, all while maintaining overall user engagement levels. Our results establish a scalable, deployable solution for mitigating popularity bias at the retrieval stage, offering a new paradigm for bias-aware personalization. △ Less

Submitted 8 June, 2025; v1 submitted 30 March, 2025; originally announced March 2025.

Comments: 2 pages. UMAP '25: 33rd ACM Conference on User Modeling, Adaptation and Personalization, New York City, USA, June 2025

Gemma 3 Technical Report

Authors: Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean-bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, Gaël Liu, Francesco Visin, Kathleen Kenealy, Lucas Beyer, Xiaohai Zhai, Anton Tsitsulin , et al. (191 additional authors not shown)

Abstract: We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achie… ▽ More We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gemma 2 for both pre-trained and instruction finetuned versions. In particular, our novel post-training recipe significantly improves the math, chat, instruction-following and multilingual abilities, making Gemma3-4B-IT competitive with Gemma2-27B-IT and Gemma3-27B-IT comparable to Gemini-1.5-Pro across benchmarks. We release all our models to the community. △ Less

Submitted 25 March, 2025; originally announced March 2025.

RECOVER: Designing a Large Language Model-based Remote Patient Monitoring System for Postoperative Gastrointestinal Cancer Care

Authors: Ziqi Yang, Yuxuan Lu, Jennifer Bagdasarian, Vedant Das Swain, Ritu Agarwal, Collin Campbell, Waddah Al-Refaire, Jehan El-Bayoumi, Guodong Gao, Dakuo Wang, Bingsheng Yao, Nawar Shara

Abstract: Cancer surgery is a key treatment for gastrointestinal (GI) cancers, a group of cancers that account for more than 35% of cancer-related deaths worldwide, but postoperative complications are unpredictable and can be life-threatening. In this paper, we investigate how recent advancements in large language models (LLMs) can benefit remote patient monitoring (RPM) systems through clinical integration… ▽ More Cancer surgery is a key treatment for gastrointestinal (GI) cancers, a group of cancers that account for more than 35% of cancer-related deaths worldwide, but postoperative complications are unpredictable and can be life-threatening. In this paper, we investigate how recent advancements in large language models (LLMs) can benefit remote patient monitoring (RPM) systems through clinical integration by designing RECOVER, an LLM-powered RPM system for postoperative GI cancer care. To closely engage stakeholders in the design process, we first conducted seven participatory design sessions with five clinical staff and interviewed five cancer patients to derive six major design strategies for integrating clinical guidelines and information needs into LLM-based RPM systems. We then designed and implemented RECOVER, which features an LLM-powered conversational agent for cancer patients and an interactive dashboard for clinical staff to enable efficient postoperative RPM. Finally, we used RECOVER as a pilot system to assess the implementation of our design strategies with four clinical staff and five patients, providing design implications by identifying crucial design elements, offering insights on responsible AI, and outlining opportunities for future LLM-powered RPM systems. △ Less

Submitted 8 February, 2025; originally announced February 2025.

Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming

Authors: Mrinank Sharma, Meg Tong, Jesse Mu, Jerry Wei, Jorrit Kruthoff, Scott Goodfriend, Euan Ong, Alwin Peng, Raj Agarwal, Cem Anil, Amanda Askell, Nathan Bailey, Joe Benton, Emma Bluemke, Samuel R. Bowman, Eric Christiansen, Hoagy Cunningham, Andy Dau, Anjali Gopal, Rob Gilson, Logan Graham, Logan Howard, Nimit Kalra, Taesung Lee, Kevin Lin , et al. (18 additional authors not shown)

Abstract: Large language models (LLMs) are vulnerable to universal jailbreaks-prompting strategies that systematically bypass model safeguards and enable users to carry out harmful processes that require many model interactions, like manufacturing illegal substances at scale. To defend against these attacks, we introduce Constitutional Classifiers: safeguards trained on synthetic data, generated by promptin… ▽ More Large language models (LLMs) are vulnerable to universal jailbreaks-prompting strategies that systematically bypass model safeguards and enable users to carry out harmful processes that require many model interactions, like manufacturing illegal substances at scale. To defend against these attacks, we introduce Constitutional Classifiers: safeguards trained on synthetic data, generated by prompting LLMs with natural language rules (i.e., a constitution) specifying permitted and restricted content. In over 3,000 estimated hours of red teaming, no red teamer found a universal jailbreak that could extract information from an early classifier-guarded LLM at a similar level of detail to an unguarded model across most target queries. On automated evaluations, enhanced classifiers demonstrated robust defense against held-out domain-specific jailbreaks. These classifiers also maintain deployment viability, with an absolute 0.38% increase in production-traffic refusals and a 23.7% inference overhead. Our work demonstrates that defending against universal jailbreaks while maintaining practical deployment viability is tractable. △ Less

Submitted 30 January, 2025; originally announced January 2025.

Generalizable Articulated Object Perception with Superpoints

Authors: Qiaojun Yu, Ce Hao, Xibin Yuan, Li Zhang, Liu Liu, Yukang Huo, Rohit Agarwal, Cewu Lu

Abstract: Manipulating articulated objects with robotic arms is challenging due to the complex kinematic structure, which requires precise part segmentation for efficient manipulation. In this work, we introduce a novel superpoint-based perception method designed to improve part segmentation in 3D point clouds of articulated objects. We propose a learnable, part-aware superpoint generation technique that ef… ▽ More Manipulating articulated objects with robotic arms is challenging due to the complex kinematic structure, which requires precise part segmentation for efficient manipulation. In this work, we introduce a novel superpoint-based perception method designed to improve part segmentation in 3D point clouds of articulated objects. We propose a learnable, part-aware superpoint generation technique that efficiently groups points based on their geometric and semantic similarities, resulting in clearer part boundaries. Furthermore, by leveraging the segmentation capabilities of the 2D foundation model SAM, we identify the centers of pixel regions and select corresponding superpoints as candidate query points. Integrating a query-based transformer decoder further enhances our method's ability to achieve precise part segmentation. Experimental results on the GAPartNet dataset show that our method outperforms existing state-of-the-art approaches in cross-category part segmentation, achieving AP50 scores of 77.9% for seen categories (4.4% improvement) and $39.3\%$ for unseen categories (11.6% improvement), with superior results in 5 out of 9 part categories for seen objects and outperforming all previous methods across all part categories for unseen objects. △ Less

Submitted 21 December, 2024; originally announced December 2024.

Improving Equity in Health Modeling with GPT4-Turbo Generated Synthetic Data: A Comparative Study

Authors: Daniel Smolyak, Arshana Welivita, Margrét V. Bjarnadóttir, Ritu Agarwal

Abstract: Objective. Demographic groups are often represented at different rates in medical datasets. These differences can create bias in machine learning algorithms, with higher levels of performance for better-represented groups. One promising solution to this problem is to generate synthetic data to mitigate potential adverse effects of non-representative data sets. Methods. We build on recent advance… ▽ More Objective. Demographic groups are often represented at different rates in medical datasets. These differences can create bias in machine learning algorithms, with higher levels of performance for better-represented groups. One promising solution to this problem is to generate synthetic data to mitigate potential adverse effects of non-representative data sets. Methods. We build on recent advances in LLM-based synthetic data generation to create a pipeline where the synthetic data is generated separately for each demographic group. We conduct our study using MIMIC-IV and Framingham "Offspring and OMNI-1 Cohorts" datasets. We prompt GPT4-Turbo to create group-specific data, providing training examples and the dataset context. An exploratory analysis is conducted to ascertain the quality of the generated data. We then evaluate the utility of the synthetic data for augmentation of a training dataset in a downstream machine learning task, focusing specifically on model performance metrics across groups. Results. The performance of GPT4-Turbo augmentation is generally superior but not always. In the majority of experiments our method outperforms standard modeling baselines, however, prompting GPT-4-Turbo to produce data specific to a group provides little to no additional benefit over a prompt that does not specify the group. Conclusion. We developed a method for using LLMs out-of-the-box to synthesize group-specific data to address imbalances in demographic representation in medical datasets. As another "tool in the toolbox", this method can improve model fairness and thus health equity. More research is needed to understand the conditions under which LLM generated synthetic data is useful for non-representative medical data sets. △ Less

Submitted 20 December, 2024; originally announced December 2024.

Comments: 26 pages, 4 figures

Inference-Aware Fine-Tuning for Best-of-N Sampling in Large Language Models

Authors: Yinlam Chow, Guy Tennenholtz, Izzeddin Gur, Vincent Zhuang, Bo Dai, Sridhar Thiagarajan, Craig Boutilier, Rishabh Agarwal, Aviral Kumar, Aleksandra Faust

Abstract: Recent studies have indicated that effectively utilizing inference-time compute is crucial for attaining better performance from large language models (LLMs). In this work, we propose a novel inference-aware fine-tuning paradigm, in which the model is fine-tuned in a manner that directly optimizes the performance of the inference-time strategy. We study this paradigm using the simple yet effective… ▽ More Recent studies have indicated that effectively utilizing inference-time compute is crucial for attaining better performance from large language models (LLMs). In this work, we propose a novel inference-aware fine-tuning paradigm, in which the model is fine-tuned in a manner that directly optimizes the performance of the inference-time strategy. We study this paradigm using the simple yet effective Best-of-N (BoN) inference strategy, in which a verifier selects the best out of a set of LLM-generated responses. We devise the first imitation learning and reinforcement learning~(RL) methods for BoN-aware fine-tuning, overcoming the challenging, non-differentiable argmax operator within BoN. We empirically demonstrate that our BoN-aware models implicitly learn a meta-strategy that interleaves best responses with more diverse responses that might be better suited to a test-time input -- a process reminiscent of the exploration-exploitation trade-off in RL. Our experiments demonstrate the effectiveness of BoN-aware fine-tuning in terms of improved performance and inference-time compute. In particular, we show that our methods improve the Bo32 performance of Gemma 2B on Hendrycks MATH from 26.8% to 30.8%, and pass@32 from 60.0% to 67.0%, as well as the pass@16 on HumanEval from 61.6% to 67.1%. △ Less

Submitted 18 December, 2024; originally announced December 2024.

Let Curves Speak: A Continuous Glucose Monitor based Large Sensor Foundation Model for Diabetes Management

Authors: Junjie Luo, Abhimanyu Kumbara, Mansur Shomali, Rui Han, Anand Iyer, Ritu Agarwal, Gordon Gao

Abstract: While previous studies of AI in diabetes management focus on long-term risk, research on near-future glucose prediction remains limited but important as it enables timely diabetes self-management. Integrating AI with continuous glucose monitoring (CGM) holds promise for near-future glucose prediction. However, existing models have limitations in capturing patterns of blood glucose fluctuations and… ▽ More While previous studies of AI in diabetes management focus on long-term risk, research on near-future glucose prediction remains limited but important as it enables timely diabetes self-management. Integrating AI with continuous glucose monitoring (CGM) holds promise for near-future glucose prediction. However, existing models have limitations in capturing patterns of blood glucose fluctuations and demonstrate poor generalizability. A robust approach is needed to leverage massive CGM data for near-future glucose prediction. We propose large sensor models (LSMs) to capture knowledge in CGM data by modeling patients as sequences of glucose. CGM-LSM is pretrained on 15.96 million glucose records from 592 diabetes patients for near-future glucose prediction. We evaluated CGM-LSM against state-of-the-art methods using the OhioT1DM dataset across various metrics, prediction horizons, and unseen patients. Additionally, we assessed its generalizability across factors like diabetes type, age, gender, and hour of day. CGM-LSM achieved exceptional performance, with an rMSE of 29.81 mg/dL for type 1 diabetes patients and 23.49 mg/dL for type 2 diabetes patients in a two-hour prediction horizon. For the OhioT1DM dataset, CGM-LSM achieved a one-hour rMSE of 15.64 mg/dL, halving the previous best of 31.97 mg/dL. Robustness analyses revealed consistent performance not only for unseen patients and future periods, but also across diabetes type, age, and gender. The model demonstrated adaptability to different hours of day, maintaining accuracy across periods of various activity intensity levels. CGM-LSM represents a transformative step in diabetes management by leveraging pretraining to uncover latent glucose generation patterns in sensor data. Our findings also underscore the broader potential of LSMs to drive innovation across domains involving complex sensor data. △ Less

Submitted 17 December, 2024; v1 submitted 12 December, 2024; originally announced December 2024.

ENCLIP: Ensembling and Clustering-Based Contrastive Language-Image Pretraining for Fashion Multimodal Search with Limited Data and Low-Quality Images

Authors: Prithviraj Purushottam Naik, Rohit Agarwal

Abstract: Multimodal search has revolutionized the fashion industry, providing a seamless and intuitive way for users to discover and explore fashion items. Based on their preferences, style, or specific attributes, users can search for products by combining text and image information. Text-to-image searches enable users to find visually similar items or describe products using natural language. This paper… ▽ More Multimodal search has revolutionized the fashion industry, providing a seamless and intuitive way for users to discover and explore fashion items. Based on their preferences, style, or specific attributes, users can search for products by combining text and image information. Text-to-image searches enable users to find visually similar items or describe products using natural language. This paper presents an innovative approach called ENCLIP, for enhancing the performance of the Contrastive Language-Image Pretraining (CLIP) model, specifically in Multimodal Search targeted towards the domain of fashion intelligence. This method focuses on addressing the challenges posed by limited data availability and low-quality images. This paper proposes an algorithm that involves training and ensembling multiple instances of the CLIP model, and leveraging clustering techniques to group similar images together. The experimental findings presented in this study provide evidence of the effectiveness of the methodology. This approach unlocks the potential of CLIP in the domain of fashion intelligence, where data scarcity and image quality issues are prevalent. Overall, the ENCLIP method represents a valuable contribution to the field of fashion intelligence and provides a practical solution for optimizing the CLIP model in scenarios with limited data and low-quality images. △ Less

Submitted 25 November, 2024; originally announced November 2024.

Injection Attacks Against End-to-End Encrypted Applications

Authors: Andrés Fábrega, Carolina Ortega Pérez, Armin Namavari, Ben Nassi, Rachit Agarwal, Thomas Ristenpart

Abstract: We explore an emerging threat model for end-to-end (E2E) encrypted applications: an adversary sends chosen messages to a target client, thereby "injecting" adversarial content into the application state. Such state is subsequently encrypted and synchronized to an adversarially-visible storage. By observing the lengths of the resulting cloud-stored ciphertexts, the attacker backs out confidential i… ▽ More We explore an emerging threat model for end-to-end (E2E) encrypted applications: an adversary sends chosen messages to a target client, thereby "injecting" adversarial content into the application state. Such state is subsequently encrypted and synchronized to an adversarially-visible storage. By observing the lengths of the resulting cloud-stored ciphertexts, the attacker backs out confidential information. We investigate this injection threat model in the context of state-of-the-art encrypted messaging applications that support E2E encrypted backups. We show proof-of-concept attacks that can recover information about E2E encrypted messages or attachments sent via WhatsApp, assuming the ability to compromise the target user's Google or Apple account (which gives access to encrypted backups). We also show weaknesses in Signal's encrypted backup design that would allow injection attacks to infer metadata including a target user's number of contacts and conversations, should the adversary somehow obtain access to the user's encrypted Signal backup. While we do not believe our results should be of immediate concern for users of these messaging applications, our results do suggest that more work is needed to build tools that enjoy strong E2E security guarantees. △ Less

Submitted 14 November, 2024; originally announced November 2024.

Comments: Published in IEEE Security and Privacy 2024

Scalable Reinforcement Post-Training Beyond Static Human Prompts: Evolving Alignment via Asymmetric Self-Play

Authors: Ziyu Ye, Rishabh Agarwal, Tianqi Liu, Rishabh Joshi, Sarmishta Velury, Quoc V. Le, Qijun Tan, Yuan Liu

Abstract: Current reinforcement learning (RL) frameworks for large language models (LLM) post-training typically assume a fixed prompt distribution, which is sub-optimal and bottlenecks scalability. Prior works have explored prompt evolving, but are often limited to the supervised fine-tuning stage, and prompts are sampled and evolved uniformly without signals. This empirical work presents a paradigm shift:… ▽ More Current reinforcement learning (RL) frameworks for large language models (LLM) post-training typically assume a fixed prompt distribution, which is sub-optimal and bottlenecks scalability. Prior works have explored prompt evolving, but are often limited to the supervised fine-tuning stage, and prompts are sampled and evolved uniformly without signals. This empirical work presents a paradigm shift: Evolving Alignment via Asymmetric Self-Play (eva), that casts post-training as an infinite game with regret-based signals for 2 players: (i) a creator, who strategically samples and creates new informative prompts and (ii) a solver, who learns to produce preferred responses. eva is the first method that allows language models to adaptively create training prompts in both offline and online RL post-training. The design is simple, easy-to-use yet remarkably effective: eva sets a new SOTA on challenging benchmarks, without any extra human prompts, e.g. it boosts the win-rate of gemma-2-9b-it on Arena-Hard by 51.6% -> 60.1% for DPO and 52.6% -> 62.4% for RLOO, surpassing claude-3-opus and catching up to gemini-1.5-pro, both of which are orders of magnitude larger. Extensive experiments show eva can create effective RL curricula and is robust across ablations. We believe adaptively evolving prompts are key to designing the next-generation RL post-training scheme. △ Less

Submitted 9 April, 2025; v1 submitted 31 October, 2024; originally announced November 2024.

Comments: spotlight @ neurips language gamification workshop. updated the problem description and added new online RL experiments in this version

Asynchronous RLHF: Faster and More Efficient Off-Policy RL for Language Models

Authors: Michael Noukhovitch, Shengyi Huang, Sophie Xhonneux, Arian Hosseini, Rishabh Agarwal, Aaron Courville

Abstract: The dominant paradigm for RLHF is online and on-policy RL: synchronously generating from the large language model (LLM) policy, labelling with a reward model, and learning using feedback on the LLM's own outputs. While performant, this paradigm is computationally inefficient. Inspired by classical deep RL literature, we propose separating generation and learning in RLHF. This enables asynchronous… ▽ More The dominant paradigm for RLHF is online and on-policy RL: synchronously generating from the large language model (LLM) policy, labelling with a reward model, and learning using feedback on the LLM's own outputs. While performant, this paradigm is computationally inefficient. Inspired by classical deep RL literature, we propose separating generation and learning in RLHF. This enables asynchronous generation of new samples while simultaneously training on old samples, leading to faster training and more compute-optimal scaling. However, asynchronous training relies on an underexplored regime, online but off-policy RLHF: learning on samples from previous iterations of our model which give a worse training signal. We tackle the fundamental challenge in this regime: how much off-policyness can we tolerate for asynchronous training to speed up learning but maintain performance? Among several RLHF algorithms we test, online DPO is found to be most robust to off-policy data, and robustness increases with the scale of the policy model. We study further compute optimizations for asynchronous RLHF but find that they come at a performance cost, giving rise to a trade-off. We verify the scalability of asynchronous RLHF by training a general-purpose chatbot from LLaMA 3.1 8B on an instruction-following task ~40% faster than a synchronous run while matching final performance. Finally, we extend our results to math and reasoning to demonstrate asynchronous RL can finetune Rho 1B on GSM8k ~70% faster while matching synchronous accuracy. △ Less

Submitted 26 April, 2025; v1 submitted 23 October, 2024; originally announced October 2024.

Comments: accepted at ICLR 2025, code at https://github.com/mnoukhov/async_rlhf, integrated into the open-instruct library https://github.com/allenai/open-instruct

packetLSTM: Dynamic LSTM Framework for Streaming Data with Varying Feature Space

Authors: Rohit Agarwal, Karaka Prasanth Naidu, Alexander Horsch, Krishna Agarwal, Dilip K. Prasad

Abstract: We study the online learning problem characterized by the varying input feature space of streaming data. Although LSTMs have been employed to effectively capture the temporal nature of streaming data, they cannot handle the dimension-varying streams in an online learning setting. Therefore, we propose a dynamic LSTM-based novel method, called packetLSTM, to model the dimension-varying streams. The… ▽ More We study the online learning problem characterized by the varying input feature space of streaming data. Although LSTMs have been employed to effectively capture the temporal nature of streaming data, they cannot handle the dimension-varying streams in an online learning setting. Therefore, we propose a dynamic LSTM-based novel method, called packetLSTM, to model the dimension-varying streams. The packetLSTM's dynamic framework consists of an evolving packet of LSTMs, each dedicated to processing one input feature. Each LSTM retains the local information of its corresponding feature, while a shared common memory consolidates global information. This configuration facilitates continuous learning and mitigates the issue of forgetting, even when certain features are absent for extended time periods. The idea of utilizing one LSTM per feature coupled with a dimension-invariant operator for information aggregation enhances the dynamic nature of packetLSTM. This dynamic nature is evidenced by the model's ability to activate, deactivate, and add new LSTMs as required, thus seamlessly accommodating varying input dimensions. The packetLSTM achieves state-of-the-art results on five datasets, and its underlying principle is extended to other RNN types, like GRU and vanilla RNN. △ Less

Submitted 22 October, 2024; originally announced October 2024.

Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling

Authors: Wenda Xu, Rujun Han, Zifeng Wang, Long T. Le, Dhruv Madeka, Lei Li, William Yang Wang, Rishabh Agarwal, Chen-Yu Lee, Tomas Pfister

Abstract: Recent advances in knowledge distillation (KD) have enabled smaller student models to approach the performance of larger teacher models. However, popular methods such as supervised KD and on-policy KD, are adversely impacted by the knowledge gaps between teacher-student in practical scenarios. Supervised KD suffers from a distribution mismatch between training with a static dataset and inference o… ▽ More Recent advances in knowledge distillation (KD) have enabled smaller student models to approach the performance of larger teacher models. However, popular methods such as supervised KD and on-policy KD, are adversely impacted by the knowledge gaps between teacher-student in practical scenarios. Supervised KD suffers from a distribution mismatch between training with a static dataset and inference over final student-generated outputs. Conversely, on-policy KD, which uses student-generated samples for training, can suffer from low-quality training examples with which teacher models are not familiar, resulting in inaccurate teacher feedback. To address these limitations, we introduce Speculative Knowledge Distillation (SKD), a novel approach that leverages cooperation between student and teacher models to generate high-quality training data on-the-fly while aligning with the student's inference-time distribution. In SKD, the student proposes tokens, and the teacher replaces poorly ranked ones based on its own distribution, transferring high-quality knowledge adaptively. We evaluate SKD on various text generation tasks, including translation, summarization, math, and instruction following, and show that SKD consistently outperforms existing KD methods across different domains, data sizes, and model initialization strategies. △ Less

Submitted 27 April, 2025; v1 submitted 15 October, 2024; originally announced October 2024.

Comments: ICLR2025

Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning

Authors: Amrith Setlur, Chirag Nagpal, Adam Fisch, Xinyang Geng, Jacob Eisenstein, Rishabh Agarwal, Alekh Agarwal, Jonathan Berant, Aviral Kumar

Abstract: A promising approach for improving reasoning in large language models is to use process reward models (PRMs). PRMs provide feedback at each step of a multi-step reasoning trace, potentially improving credit assignment over outcome reward models (ORMs) that only provide feedback at the final step. However, collecting dense, per-step human labels is not scalable, and training PRMs from automatically… ▽ More A promising approach for improving reasoning in large language models is to use process reward models (PRMs). PRMs provide feedback at each step of a multi-step reasoning trace, potentially improving credit assignment over outcome reward models (ORMs) that only provide feedback at the final step. However, collecting dense, per-step human labels is not scalable, and training PRMs from automatically-labeled data has thus far led to limited gains. To improve a base policy by running search against a PRM or using it as dense rewards for reinforcement learning (RL), we ask: "How should we design process rewards?". Our key insight is that, to be effective, the process reward for a step should measure progress: a change in the likelihood of producing a correct response in the future, before and after taking the step, corresponding to the notion of step-level advantages in RL. Crucially, this progress should be measured under a prover policy distinct from the base policy. We theoretically characterize the set of good provers and our results show that optimizing process rewards from such provers improves exploration during test-time search and online RL. In fact, our characterization shows that weak prover policies can substantially improve a stronger base policy, which we also observe empirically. We validate our claims by training process advantage verifiers (PAVs) to predict progress under such provers, and show that compared to ORMs, test-time search against PAVs is $>8\%$ more accurate, and $1.5-5\times$ more compute-efficient. Online RL with dense rewards from PAVs enables one of the first results with $5-6\times$ gain in sample efficiency, and $>6\%$ gain in accuracy, over ORMs. △ Less

Submitted 10 October, 2024; originally announced October 2024.

Not All LLM Reasoners Are Created Equal

Authors: Arian Hosseini, Alessandro Sordoni, Daniel Toyama, Aaron Courville, Rishabh Agarwal

Abstract: We study the depth of grade-school math (GSM) problem-solving capabilities of LLMs. To this end, we evaluate their performance on pairs of existing math word problems together so that the answer to the second problem depends on correctly answering the first problem. Our findings reveal a significant reasoning gap in most LLMs, that is performance difference between solving the compositional pairs… ▽ More We study the depth of grade-school math (GSM) problem-solving capabilities of LLMs. To this end, we evaluate their performance on pairs of existing math word problems together so that the answer to the second problem depends on correctly answering the first problem. Our findings reveal a significant reasoning gap in most LLMs, that is performance difference between solving the compositional pairs and solving each question independently. This gap is more pronounced in smaller, more cost-efficient, and math-specialized models. Moreover, instruction-tuning recipes and code generation have varying effects across LLM sizes, while finetuning on GSM can lead to task overfitting. Our analysis indicates that large reasoning gaps are not because of test-set leakage, but due to distraction from additional context and poor second-hop reasoning. Overall, LLMs exhibit systematic differences in their reasoning abilities, despite what their performance on standard benchmarks indicates. △ Less

Submitted 2 October, 2024; originally announced October 2024.

Beyond Following: Mixing Active Initiative into Computational Creativity

Authors: Zhiyu Lin, Upol Ehsan, Rohan Agarwal, Samihan Dani, Vidushi Vashishth, Mark Riedl

Abstract: Generative Artificial Intelligence (AI) encounters limitations in efficiency and fairness within the realm of Procedural Content Generation (PCG) when human creators solely drive and bear responsibility for the generative process. Alternative setups, such as Mixed-Initiative Co-Creative (MI-CC) systems, exhibited their promise. Still, the potential of an active mixed initiative, where AI takes a r… ▽ More Generative Artificial Intelligence (AI) encounters limitations in efficiency and fairness within the realm of Procedural Content Generation (PCG) when human creators solely drive and bear responsibility for the generative process. Alternative setups, such as Mixed-Initiative Co-Creative (MI-CC) systems, exhibited their promise. Still, the potential of an active mixed initiative, where AI takes a role beyond following, is understudied. This work investigates the influence of the adaptive ability of an active and learning AI agent on creators' expectancy of creative responsibilities in an MI-CC setting. We built and studied a system that employs reinforcement learning (RL) methods to learn the creative responsibility preferences of a human user during online interactions. Situated in story co-creation, we develop a Multi-armed-bandit agent that learns from the human creator, updates its collaborative decision-making belief, and switches between its capabilities during an MI-CC experience. With 39 participants joining a human subject study, Our developed system's learning capabilities are well recognized compared to the non-learning ablation, corresponding to a significant increase in overall satisfaction with the MI-CC experience. These findings indicate a robust association between effective MI-CC collaborative interactions, particularly the implementation of proactive AI initiatives, and deepened understanding among all participants. △ Less

Submitted 6 September, 2024; originally announced September 2024.

Comments: 11 pages, 4 figures

Training Language Models to Self-Correct via Reinforcement Learning

Authors: Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M Zhang, Kay McKinney, Disha Shrivastava, Cosmin Paduraru, George Tucker, Doina Precup, Feryal Behbahani, Aleksandra Faust

Abstract: Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Current methods for training self-correction typically depend on either multiple models, a more advanced model, or additional forms of supervision. To address these shortcomings, we develop a multi-turn online reinforcement learning (RL) app… ▽ More Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Current methods for training self-correction typically depend on either multiple models, a more advanced model, or additional forms of supervision. To address these shortcomings, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are often insufficient for instilling self-correction behavior. In particular, we observe that training via SFT falls prey to either a distribution mismatch between mistakes made by the data-collection policy and the model's own responses, or to behavior collapse, where learning implicitly prefers only a certain mode of correction behavior that is often not effective at self-correction on test problems. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction behavior that is effective at test time as opposed to fitting high-reward responses for a given prompt. This regularization process includes an initial phase of multi-turn RL on a base model to generate a policy initialization that is less susceptible to collapse, followed by using a reward bonus to amplify self-correction. With Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on MATH and HumanEval. △ Less

Submitted 4 October, 2024; v1 submitted 19 September, 2024; originally announced September 2024.

Hedging Is Not All You Need: A Simple Baseline for Online Learning Under Haphazard Inputs

Authors: Himanshu Buckchash, Momojit Biswas, Rohit Agarwal, Dilip K. Prasad

Abstract: Handling haphazard streaming data, such as data from edge devices, presents a challenging problem. Over time, the incoming data becomes inconsistent, with missing, faulty, or new inputs reappearing. Therefore, it requires models that are reliable. Recent methods to solve this problem depend on a hedging-based solution and require specialized elements like auxiliary dropouts, forked architectures,… ▽ More Handling haphazard streaming data, such as data from edge devices, presents a challenging problem. Over time, the incoming data becomes inconsistent, with missing, faulty, or new inputs reappearing. Therefore, it requires models that are reliable. Recent methods to solve this problem depend on a hedging-based solution and require specialized elements like auxiliary dropouts, forked architectures, and intricate network design. We observed that hedging can be reduced to a special case of weighted residual connection; this motivated us to approximate it with plain self-attention. In this work, we propose HapNet, a simple baseline that is scalable, does not require online backpropagation, and is adaptable to varying input types. All present methods are restricted to scaling with a fixed window; however, we introduce a more complex problem of scaling with a variable window where the data becomes positionally uncorrelated, and cannot be addressed by present methods. We demonstrate that a variant of the proposed approach can work even for this complex scenario. We extensively evaluated the proposed approach on five benchmarks and found competitive performance. △ Less

Submitted 30 December, 2024; v1 submitted 16 September, 2024; originally announced September 2024.

Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling

Authors: Hritik Bansal, Arian Hosseini, Rishabh Agarwal, Vinh Q. Tran, Mehran Kazemi

Abstract: Training on high-quality synthetic data from strong language models (LMs) is a common strategy to improve the reasoning performance of LMs. In this work, we revisit whether this strategy is compute-optimal under a fixed inference budget (e.g., FLOPs). To do so, we investigate the trade-offs between generating synthetic data using a stronger but more expensive (SE) model versus a weaker but cheaper… ▽ More Training on high-quality synthetic data from strong language models (LMs) is a common strategy to improve the reasoning performance of LMs. In this work, we revisit whether this strategy is compute-optimal under a fixed inference budget (e.g., FLOPs). To do so, we investigate the trade-offs between generating synthetic data using a stronger but more expensive (SE) model versus a weaker but cheaper (WC) model. We evaluate the generated data across three key metrics: coverage, diversity, and false positive rate, and show that the data from WC models may have higher coverage and diversity, but also exhibit higher false positive rates. We then finetune LMs on data from SE and WC models in different settings: knowledge distillation, self-improvement, and a novel weak-to-strong improvement setup where a weaker LM teaches reasoning to a stronger LM. Our findings reveal that models finetuned on WC-generated data consistently outperform those trained on SE-generated data across multiple benchmarks and multiple choices of WC and SE models. These results challenge the prevailing practice of relying on SE models for synthetic data generation, suggesting that WC may be the compute-optimal approach for training advanced LM reasoners. △ Less

Submitted 7 October, 2024; v1 submitted 29 August, 2024; originally announced August 2024.

Lyrically Speaking: Exploring the Link Between Lyrical Emotions, Themes and Depression Risk

Authors: Pavani Chowdary, Bhavyajeet Singh, Rajat Agarwal, Vinoo Alluri

Abstract: Lyrics play a crucial role in affecting and reinforcing emotional states by providing meaning and emotional connotations that interact with the acoustic properties of the music. Specific lyrical themes and emotions may intensify existing negative states in listeners and may lead to undesirable outcomes, especially in listeners with mood disorders such as depression. Hence, it is important for such… ▽ More Lyrics play a crucial role in affecting and reinforcing emotional states by providing meaning and emotional connotations that interact with the acoustic properties of the music. Specific lyrical themes and emotions may intensify existing negative states in listeners and may lead to undesirable outcomes, especially in listeners with mood disorders such as depression. Hence, it is important for such individuals to be mindful of their listening strategies. In this study, we examine online music consumption of individuals at risk of depression in light of lyrical themes and emotions. Lyrics obtained from the listening histories of 541 Last.fm users, divided into At-Risk and No-Risk based on their mental well-being scores, were analyzed using natural language processing techniques. Statistical analyses of the results revealed that individuals at risk for depression prefer songs with lyrics associated with low valence and low arousal. Additionally, lyrics associated with themes of denial, self-reference, and ambivalence were preferred. In contrast, themes such as liberation, familiarity, and activity are not as favored. This study opens up the possibility of an approach to assessing depression risk from the digital footprint of individuals and potentially developing personalized recommendation systems. △ Less

Submitted 28 August, 2024; originally announced August 2024.

Comments: Accepted at the 25th International Society for Music Information Retrieval Conference (ISMIR) 2024, San Francisco, United States

Generative Verifiers: Reward Modeling as Next-Token Prediction

Authors: Lunjun Zhang, Arian Hosseini, Hritik Bansal, Mehran Kazemi, Aviral Kumar, Rishabh Agarwal

Abstract: Verifiers or reward models are often used to enhance the reasoning performance of large language models (LLMs). A common approach is the Best-of-N method, where N candidate solutions generated by the LLM are ranked by a verifier, and the best one is selected. While LLM-based verifiers are typically trained as discriminative classifiers to score solutions, they do not utilize the text generation ca… ▽ More Verifiers or reward models are often used to enhance the reasoning performance of large language models (LLMs). A common approach is the Best-of-N method, where N candidate solutions generated by the LLM are ranked by a verifier, and the best one is selected. While LLM-based verifiers are typically trained as discriminative classifiers to score solutions, they do not utilize the text generation capabilities of pretrained LLMs. To overcome this limitation, we instead propose training verifiers using the ubiquitous next-token prediction objective, jointly on verification and solution generation. Compared to standard verifiers, such generative verifiers (GenRM) can benefit from several advantages of LLMs: they integrate seamlessly with instruction tuning, enable chain-of-thought reasoning, and can utilize additional test-time compute via majority voting for better verification. We demonstrate that GenRM outperforms discriminative, DPO verifiers, and LLM-as-a-Judge, resulting in large performance gains with Best-of-N, namely 5% $\rightarrow$ 45.3% on algorithmic tasks and 73% $\rightarrow$ 93.4% on GSM8K. In easy-to-hard generalization settings, we observe improvements of 28% $\rightarrow$ 44.6% on MATH, and 37.9% $\rightarrow$ 53.5% on MMLU abstract algebra. Furthermore, we find that training GenRM with synthetic verification rationales is sufficient to pick out subtle errors on math problems. Finally, we demonstrate that GenRM scales favorably with model size and test-time compute. △ Less

Submitted 22 February, 2025; v1 submitted 27 August, 2024; originally announced August 2024.

Comments: ICLR 2025

Automatic Detection of COVID-19 from Chest X-ray Images Using Deep Learning Model

Authors: Alloy Das, Rohit Agarwal, Rituparna Singh, Arindam Chowdhury, Debashis Nandi

Abstract: The infectious disease caused by novel corona virus (2019-nCoV) has been widely spreading since last year and has shaken the entire world. It has caused an unprecedented effect on daily life, global economy and public health. Hence this disease detection has life-saving importance for both patients as well as doctors. Due to limited test kits, it is also a daunting task to test every patient with… ▽ More The infectious disease caused by novel corona virus (2019-nCoV) has been widely spreading since last year and has shaken the entire world. It has caused an unprecedented effect on daily life, global economy and public health. Hence this disease detection has life-saving importance for both patients as well as doctors. Due to limited test kits, it is also a daunting task to test every patient with severe respiratory problems using conventional techniques (RT-PCR). Thus implementing an automatic diagnosis system is urgently required to overcome the scarcity problem of Covid-19 test kits at hospital, health care systems. The diagnostic approach is mainly classified into two categories-laboratory based and Chest radiography approach. In this paper, a novel approach for computerized corona virus (2019-nCoV) detection from lung x-ray images is presented. Here, we propose models using deep learning to show the effectiveness of diagnostic systems. In the experimental result, we evaluate proposed models on publicly available data-set which exhibit satisfactory performance and promising results compared with other previous existing methods. △ Less

Submitted 27 August, 2024; originally announced August 2024.

Comments: Accepted in AIP Conference Proceedings (Vol. 2424, No. 1)

Exploiting Leakage in Password Managers via Injection Attacks

Authors: Andrés Fábrega, Armin Namavari, Rachit Agarwal, Ben Nassi, Thomas Ristenpart

Abstract: This work explores injection attacks against password managers. In this setting, the adversary (only) controls their own application client, which they use to "inject" chosen payloads to a victim's client via, for example, sharing credentials with them. The injections are interleaved with adversarial observations of some form of protected state (such as encrypted vault exports or the network traff… ▽ More This work explores injection attacks against password managers. In this setting, the adversary (only) controls their own application client, which they use to "inject" chosen payloads to a victim's client via, for example, sharing credentials with them. The injections are interleaved with adversarial observations of some form of protected state (such as encrypted vault exports or the network traffic received by the application servers), from which the adversary backs out confidential information. We uncover a series of general design patterns in popular password managers that lead to vulnerabilities allowing an adversary to efficiently recover passwords, URLs, usernames, and attachments. We develop general attack templates to exploit these design patterns and experimentally showcase their practical efficacy via analysis of ten distinct password manager applications. We disclosed our findings to these vendors, many of which deployed mitigations. △ Less

Submitted 13 August, 2024; originally announced August 2024.

Comments: Full version of the paper published in USENIX Security 2024

Zero-shot Factual Consistency Evaluation Across Domains

Authors: Raunak Agarwal

Abstract: This work addresses the challenge of factual consistency in text generation systems. We unify the tasks of Natural Language Inference, Summarization Evaluation, Factuality Verification and Factual Consistency Evaluation to train models capable of evaluating the factual consistency of source-target pairs across diverse domains. We rigorously evaluate these against eight baselines on a comprehensive… ▽ More This work addresses the challenge of factual consistency in text generation systems. We unify the tasks of Natural Language Inference, Summarization Evaluation, Factuality Verification and Factual Consistency Evaluation to train models capable of evaluating the factual consistency of source-target pairs across diverse domains. We rigorously evaluate these against eight baselines on a comprehensive benchmark suite comprising 22 datasets that span various tasks, domains, and document lengths. Results demonstrate that our method achieves state-of-the-art performance on this heterogeneous benchmark while addressing efficiency concerns and attaining cross-domain generalization. △ Less

Submitted 7 August, 2024; originally announced August 2024.

Gemma 2: Improving Open Language Models at a Practical Size

Authors: Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan, Sammy Jerome, Anton Tsitsulin, Nino Vieillard, Piotr Stanczyk, Sertan Girgin, Nikola Momchev, Matt Hoffman , et al. (173 additional authors not shown)

Abstract: In this work, we introduce Gemma 2, a new addition to the Gemma family of lightweight, state-of-the-art open models, ranging in scale from 2 billion to 27 billion parameters. In this new version, we apply several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) and group-query attention (Ainslie et al., 2023). We al… ▽ More In this work, we introduce Gemma 2, a new addition to the Gemma family of lightweight, state-of-the-art open models, ranging in scale from 2 billion to 27 billion parameters. In this new version, we apply several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) and group-query attention (Ainslie et al., 2023). We also train the 2B and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction. The resulting models deliver the best performance for their size, and even offer competitive alternatives to models that are 2-3 times bigger. We release all our models to the community. △ Less

Submitted 2 October, 2024; v1 submitted 31 July, 2024; originally announced August 2024.

Don't Throw Away Data: Better Sequence Knowledge Distillation

Authors: Jun Wang, Eleftheria Briakou, Hamid Dadkhahi, Rishabh Agarwal, Colin Cherry, Trevor Cohn

Abstract: A critical component in knowledge distillation is the means of coupling the teacher and student. The predominant sequence knowledge distillation method involves supervised learning of the student against teacher-decoded outputs, and is exemplified by the current state of the art, which incorporates minimum Bayes risk (MBR) decoding. In this paper we seek to integrate MBR more tightly in distillati… ▽ More A critical component in knowledge distillation is the means of coupling the teacher and student. The predominant sequence knowledge distillation method involves supervised learning of the student against teacher-decoded outputs, and is exemplified by the current state of the art, which incorporates minimum Bayes risk (MBR) decoding. In this paper we seek to integrate MBR more tightly in distillation training, specifically by using several high scoring MBR translations, rather than a single selected sequence, thus capturing a rich diversity of teacher outputs. Our experiments on English to German and English to Japanese translation show consistent improvements over strong baseline methods for both tasks and with varying model sizes. Additionally, we conduct a detailed analysis focusing on data efficiency and capacity curse aspects to elucidate MBR-n and explore its further potential. △ Less

Submitted 15 July, 2024; originally announced July 2024.

On scalable oversight with weak LLMs judging strong LLMs

Authors: Zachary Kenton, Noah Y. Siegel, János Kramár, Jonah Brown-Cohen, Samuel Albanie, Jannis Bulian, Rishabh Agarwal, David Lindner, Yunhao Tang, Noah D. Goodman, Rohin Shah

Abstract: Scalable oversight protocols aim to enable humans to accurately supervise superhuman AI. In this paper we study debate, where two AI's compete to convince a judge; consultancy, where a single AI tries to convince a judge that asks questions; and compare to a baseline of direct question-answering, where the judge just answers outright without the AI. We use large language models (LLMs) as both AI a… ▽ More Scalable oversight protocols aim to enable humans to accurately supervise superhuman AI. In this paper we study debate, where two AI's compete to convince a judge; consultancy, where a single AI tries to convince a judge that asks questions; and compare to a baseline of direct question-answering, where the judge just answers outright without the AI. We use large language models (LLMs) as both AI agents and as stand-ins for human judges, taking the judge models to be weaker than agent models. We benchmark on a diverse range of asymmetries between judges and agents, extending previous work on a single extractive QA task with information asymmetry, to also include mathematics, coding, logic and multimodal reasoning asymmetries. We find that debate outperforms consultancy across all tasks when the consultant is randomly assigned to argue for the correct/incorrect answer. Comparing debate to direct question answering, the results depend on the type of task: in extractive QA tasks with information asymmetry debate outperforms direct question answering, but in other tasks without information asymmetry the results are mixed. Previous work assigned debaters/consultants an answer to argue for. When we allow them to instead choose which answer to argue for, we find judges are less frequently convinced by the wrong answer in debate than in consultancy. Further, we find that stronger debater models increase judge accuracy, though more modestly than in previous studies. △ Less

Submitted 12 July, 2024; v1 submitted 5 July, 2024; originally announced July 2024.

Comments: 15 pages (53 including appendices). V2: minor correction to Figure 3; add Figure A.9 comparing open vs assigned consultancy; add a reference

AddBiomechanics Dataset: Capturing the Physics of Human Motion at Scale

Authors: Keenon Werling, Janelle Kaneda, Alan Tan, Rishi Agarwal, Six Skov, Tom Van Wouwe, Scott Uhlrich, Nicholas Bianco, Carmichael Ong, Antoine Falisse, Shardul Sapkota, Aidan Chandra, Joshua Carter, Ezio Preatoni, Benjamin Fregly, Jennifer Hicks, Scott Delp, C. Karen Liu

Abstract: While reconstructing human poses in 3D from inexpensive sensors has advanced significantly in recent years, quantifying the dynamics of human motion, including the muscle-generated joint torques and external forces, remains a challenge. Prior attempts to estimate physics from reconstructed human poses have been hampered by a lack of datasets with high-quality pose and force data for a variety of m… ▽ More While reconstructing human poses in 3D from inexpensive sensors has advanced significantly in recent years, quantifying the dynamics of human motion, including the muscle-generated joint torques and external forces, remains a challenge. Prior attempts to estimate physics from reconstructed human poses have been hampered by a lack of datasets with high-quality pose and force data for a variety of movements. We present the AddBiomechanics Dataset 1.0, which includes physically accurate human dynamics of 273 human subjects, over 70 hours of motion and force plate data, totaling more than 24 million frames. To construct this dataset, novel analytical methods were required, which are also reported here. We propose a benchmark for estimating human dynamics from motion using this dataset, and present several baseline results. The AddBiomechanics Dataset is publicly available at https://addbiomechanics.org/download_data.html. △ Less

Submitted 16 May, 2024; originally announced June 2024.

Comments: 15 pages, 6 figures, 4 tables

SiT: Symmetry-Invariant Transformers for Generalisation in Reinforcement Learning

Authors: Matthias Weissenbacher, Rishabh Agarwal, Yoshinobu Kawahara

Abstract: An open challenge in reinforcement learning (RL) is the effective deployment of a trained policy to new or slightly different situations as well as semantically-similar environments. We introduce Symmetry-Invariant Transformer (SiT), a scalable vision transformer (ViT) that leverages both local and global data patterns in a self-supervised manner to improve generalisation. Central to our approach… ▽ More An open challenge in reinforcement learning (RL) is the effective deployment of a trained policy to new or slightly different situations as well as semantically-similar environments. We introduce Symmetry-Invariant Transformer (SiT), a scalable vision transformer (ViT) that leverages both local and global data patterns in a self-supervised manner to improve generalisation. Central to our approach is Graph Symmetric Attention, which refines the traditional self-attention mechanism to preserve graph symmetries, resulting in invariant and equivariant latent representations. We showcase SiT's superior generalization over ViTs on MiniGrid and Procgen RL benchmarks, and its sample efficiency on Atari 100k and CIFAR10. △ Less

Submitted 21 June, 2024; originally announced June 2024.

Comments: 9 main pages, accepted to ICML2024

Object-Oriented Architecture: A Software Engineering-Inspired Shape Grammar for Durands Plates

Authors: Rohan Agarwal

Abstract: Addressing the challenge of modular architectural design, this study presents a novel approach through the implementation of a shape grammar system using functional and object-oriented programming principles from computer science. The focus lies on the modular generation of plates in the style of French Neoclassical architect Jean-Nicolas-Louis Durand, known for his modular rule-based method to ar… ▽ More Addressing the challenge of modular architectural design, this study presents a novel approach through the implementation of a shape grammar system using functional and object-oriented programming principles from computer science. The focus lies on the modular generation of plates in the style of French Neoclassical architect Jean-Nicolas-Louis Durand, known for his modular rule-based method to architecture, demonstrating the system's capacity to articulate intricate architectural forms systematically. By leveraging computer programming principles, the proposed methodology allows for the creation of diverse designs while adhering to the inherent logic of Durand's original plates. The integration of Shape Machine allows a flexible framework for architects and designers, enabling the generation of complex structures in a modular fashion in existing CAD software. This research contributes to the exploration of computational tools in architectural design, offering a versatile solution for the synthesis of historically significant architectural elements. △ Less

Submitted 20 April, 2024; originally announced April 2024.

Many-Shot In-Context Learning

Authors: Rishabh Agarwal, Avi Singh, Lei M. Zhang, Bernd Bohnet, Luis Rosias, Stephanie Chan, Biao Zhang, Ankesh Anand, Zaheer Abbas, Azade Nova, John D. Co-Reyes, Eric Chu, Feryal Behbahani, Aleksandra Faust, Hugo Larochelle

Abstract: Large language models (LLMs) excel at few-shot in-context learning (ICL) -- learning from a few examples provided in context at inference, without any weight updates. Newly expanded context windows allow us to investigate ICL with hundreds or thousands of examples -- the many-shot regime. Going from few-shot to many-shot, we observe significant performance gains across a wide variety of generative… ▽ More Large language models (LLMs) excel at few-shot in-context learning (ICL) -- learning from a few examples provided in context at inference, without any weight updates. Newly expanded context windows allow us to investigate ICL with hundreds or thousands of examples -- the many-shot regime. Going from few-shot to many-shot, we observe significant performance gains across a wide variety of generative and discriminative tasks. While promising, many-shot ICL can be bottlenecked by the available amount of human-generated examples. To mitigate this limitation, we explore two new settings: Reinforced and Unsupervised ICL. Reinforced ICL uses model-generated chain-of-thought rationales in place of human examples. Unsupervised ICL removes rationales from the prompt altogether, and prompts the model only with domain-specific questions. We find that both Reinforced and Unsupervised ICL can be quite effective in the many-shot regime, particularly on complex reasoning tasks. Finally, we demonstrate that, unlike few-shot learning, many-shot learning is effective at overriding pretraining biases, can learn high-dimensional functions with numerical inputs, and performs comparably to fine-tuning. We also find that inference cost increases linearly in the many-shot regime, and frontier LLMs benefit from many-shot ICL to varying degrees. Our analysis also reveals the limitations of next-token prediction loss as an indicator of downstream ICL performance. △ Less

Submitted 17 October, 2024; v1 submitted 16 April, 2024; originally announced April 2024.

Comments: NeurIPS (Spotlight)

Online Learning under Haphazard Input Conditions: A Comprehensive Review and Analysis

Authors: Rohit Agarwal, Arijit Das, Alexander Horsch, Krishna Agarwal, Dilip K. Prasad

Abstract: The domain of online learning has experienced multifaceted expansion owing to its prevalence in real-life applications. Nonetheless, this progression operates under the assumption that the input feature space of the streaming data remains constant. In this survey paper, we address the topic of online learning in the context of haphazard inputs, explicitly foregoing such an assumption. We discuss,… ▽ More The domain of online learning has experienced multifaceted expansion owing to its prevalence in real-life applications. Nonetheless, this progression operates under the assumption that the input feature space of the streaming data remains constant. In this survey paper, we address the topic of online learning in the context of haphazard inputs, explicitly foregoing such an assumption. We discuss, classify, evaluate, and compare the methodologies that are adept at modeling haphazard inputs, additionally providing the corresponding code implementations and their carbon footprint. Moreover, we classify the datasets related to the field of haphazard inputs and introduce evaluation metrics specifically designed for datasets exhibiting imbalance. The code of each methodology can be found at https://github.com/Rohit102497/HaphazardInputsReview △ Less

Submitted 7 April, 2024; originally announced April 2024.

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

Authors: Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, Soroosh Mariooryad, Yifan Ding, Xinyang Geng, Fred Alcober, Roy Frostig, Mark Omernick, Lexi Walker, Cosmin Paduraru, Christina Sorokin, Andrea Tacchetti, Colin Gaffney, Samira Daruki, Olcan Sercinoglu, Zach Gleicher, Juliette Love , et al. (1112 additional authors not shown)

Abstract: In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February… ▽ More In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content. △ Less

Submitted 16 December, 2024; v1 submitted 8 March, 2024; originally announced March 2024.

Stop Regressing: Training Value Functions via Classification for Scalable Deep RL

Authors: Jesse Farebrother, Jordi Orbay, Quan Vuong, Adrien Ali Taïga, Yevgen Chebotar, Ted Xiao, Alex Irpan, Sergey Levine, Pablo Samuel Castro, Aleksandra Faust, Aviral Kumar, Rishabh Agarwal

Abstract: Value functions are a central component of deep reinforcement learning (RL). These functions, parameterized by neural networks, are trained using a mean squared error regression objective to match bootstrapped target values. However, scaling value-based RL methods that use regression to large networks, such as high-capacity Transformers, has proven challenging. This difficulty is in stark contrast… ▽ More Value functions are a central component of deep reinforcement learning (RL). These functions, parameterized by neural networks, are trained using a mean squared error regression objective to match bootstrapped target values. However, scaling value-based RL methods that use regression to large networks, such as high-capacity Transformers, has proven challenging. This difficulty is in stark contrast to supervised learning: by leveraging a cross-entropy classification loss, supervised methods have scaled reliably to massive networks. Observing this discrepancy, in this paper, we investigate whether the scalability of deep RL can also be improved simply by using classification in place of regression for training value functions. We demonstrate that value functions trained with categorical cross-entropy significantly improves performance and scalability in a variety of domains. These include: single-task RL on Atari 2600 games with SoftMoEs, multi-task RL on Atari with large-scale ResNets, robotic manipulation with Q-transformers, playing Chess without search, and a language-agent Wordle task with high-capacity Transformers, achieving state-of-the-art results on these domains. Through careful analysis, we show that the benefits of categorical cross-entropy primarily stem from its ability to mitigate issues inherent to value-based RL, such as noisy targets and non-stationarity. Overall, we argue that a simple shift to training value functions with categorical cross-entropy can yield substantial improvements in the scalability of deep RL at little-to-no cost. △ Less

Submitted 6 March, 2024; originally announced March 2024.

Large Scale Generative AI Text Applied to Sports and Music

Authors: Aaron Baughman, Stephen Hammer, Rahul Agarwal, Gozde Akay, Eduardo Morales, Tony Johnson, Leonid Karlinsky, Rogerio Feris

Abstract: We address the problem of scaling up the production of media content, including commentary and personalized news stories, for large-scale sports and music events worldwide. Our approach relies on generative AI models to transform a large volume of multimodal data (e.g., videos, articles, real-time scoring feeds, statistics, and fact sheets) into coherent and fluent text. Based on this approach, we… ▽ More We address the problem of scaling up the production of media content, including commentary and personalized news stories, for large-scale sports and music events worldwide. Our approach relies on generative AI models to transform a large volume of multimodal data (e.g., videos, articles, real-time scoring feeds, statistics, and fact sheets) into coherent and fluent text. Based on this approach, we introduce, for the first time, an AI commentary system, which was deployed to produce automated narrations for highlight packages at the 2023 US Open, Wimbledon, and Masters tournaments. In the same vein, our solution was extended to create personalized content for ESPN Fantasy Football and stories about music artists for the Grammy awards. These applications were built using a common software architecture achieved a 15x speed improvement with an average Rouge-L of 82.00 and perplexity of 6.6. Our work was successfully deployed at the aforementioned events, supporting 90 million fans around the world with 8 billion page views, continuously pushing the bounds on what is possible at the intersection of sports, entertainment, and AI. △ Less

Submitted 27 February, 2024; v1 submitted 31 January, 2024; originally announced February 2024.

Comments: 9 pages, 8 figures, 5 tables

Transformers Can Achieve Length Generalization But Not Robustly

Authors: Yongchao Zhou, Uri Alon, Xinyun Chen, Xuezhi Wang, Rishabh Agarwal, Denny Zhou

Abstract: Length generalization, defined as the ability to extrapolate from shorter training sequences to longer test ones, is a significant challenge for language models. This issue persists even with large-scale Transformers handling relatively straightforward tasks. In this paper, we test the Transformer's ability of length generalization using the task of addition of two integers. We show that the succe… ▽ More Length generalization, defined as the ability to extrapolate from shorter training sequences to longer test ones, is a significant challenge for language models. This issue persists even with large-scale Transformers handling relatively straightforward tasks. In this paper, we test the Transformer's ability of length generalization using the task of addition of two integers. We show that the success of length generalization is intricately linked to the data format and the type of position encoding. Using the right combination of data format and position encodings, we show for the first time that standard Transformers can extrapolate to a sequence length that is 2.5x the input length. Nevertheless, unlike in-distribution generalization, length generalization remains fragile, significantly influenced by factors like random weight initialization and training data order, leading to large variances across different random seeds. △ Less

Submitted 14 February, 2024; originally announced February 2024.