Abstract
Recent breakthroughs in artificial intelligence (AI) are increasingly driven by systems orchestrating multiple large language models (LLMs) and other specialized tools, such as search engines and simulators. So far, these systems are primarily handcrafted by domain experts and tweaked through heuristics rather than being automatically optimized, presenting a substantial challenge to accelerating progress. The development of artificial neural networks faced a similar challenge until backpropagation and automatic differentiation transformed the field by making optimization turnkey. Analogously, here we introduce TextGrad, a versatile framework that performs optimization by backpropagating LLM-generated feedback to improve AI systems. By leveraging natural language feedback to critique and suggest improvements to any part of a system—from prompts to outputs such as molecules or treatment plans—TextGrad enables the automatic optimization of generative AI systems across diverse tasks. We demonstrate TextGrad’s generality and effectiveness through studies in solving PhD-level science problems, optimizing plans for radiotherapy treatments, designing molecules with specific properties, coding, and optimizing agentic systems. TextGrad empowers scientists and engineers to easily develop impactful generative AI systems.
This is a preview of subscription content, access via your institution
Access options
Access Nature and 54 other Nature Portfolio journals
Get Nature+, our best-value online-access subscription
$32.99 / 30 days
cancel any time
Subscribe to this journal
Receive 51 print issues and online access
$199.00 per year
only $3.90 per issue
Buy this article
- Purchase on SpringerLink
- Instant access to full article PDF
Prices may be subject to local taxes which are calculated during checkout
Similar content being viewed by others
Data availability
We used publicly available data to evaluate TextGrad. Details on how to access the data is available at https://github.com/zou-group/textgrad.
Code availability
The TextGrad code and experiments are available at https://github.com/zou-group/textgrad and https://doi.org/10.5281/zenodo.14497017 (ref. 46).
References
Brown, T. et al. Language models are few-shot learners. Adv. Neural Inf. Process. Syst. 33, 1877–1901 (2020).
Trinh, T. H., Wu, Y., Le, Q. V., He, H. & Luong, T. Solving olympiad geometry without human demonstrations. Nature 625, 476–482 (2024).
Li, Y. et al. Competition-level code generation with alphacode. Science 378, 1092–1097 (2022).
Yang, J. et al. SWE-agent: agent–computer interfaces enable automated software engineering. In Adv. Neural Inf. Process. Syst. 37 (NeurIPS, 2024).
Khattab, O. et al. DSPy: Compiling declarative language model calls into state-of-the-art pipelines. In The Twelfth International Conference on Learning Representations (2024).
Zaharia, M. et al. The shift from models to compound AI systems. BAIR https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/ (2024).
Zhou, Y. et al. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations (2023).
Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. In Adv. Neural Inf. Process. Syst. 25 (NeurIPS, 2012).
Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
Fawzi, A. et al. Discovering faster matrix multiplication algorithms with reinforcement learning. Nature 610, 47–53 (2022).
Mankowitz, D. J. et al. Faster sorting algorithms discovered using deep reinforcement learning. Nature 618, 257–263 (2023).
Merchant, A. et al. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023).
Goodfellow, I., Bengio, Y. & Courville, A. Deep Learning (MIT Press, 2016).
Rumelhart, D. E., Hinton, G. E. & Williams, R. J. Learning representations by back-propagating errors. Nature 323, 533–536 (1986).
Pryzant, R. et al. Automatic prompt optimization with “gradient descent” and beam search. In Proc. 2023 Conference on Empirical Methods in Natural Language Processing (eds Bouamor, H. et al.) 7957–7968 (Association for Computational Linguistics, 2023).
Zheng, L. et al. Judging LLM-as-a-judge with MH-bench and chatbot arena. Adv. Neural Inf. Process. Syst. 36, 46595–46623 (2023).
Li, X. et al. AlpacaEval: an automatic evaluator of instruction-following models. GitHub https://github.com/tatsu-lab/alpaca_eval (2023).
Bai, Y. et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. Preprint at https://arxiv.org/abs/2204.05862 (2022).
Madaan, A. et al. Self-refine: iterative refinement with self-feedback. In Adv. Neural Inf. Process. Syst. 36 (NeurIPS, 2023).
Stiennon, N. et al. Learning to summarize with human feedback. In Adv. Neural Inf. Process. Syst. 33, 3008–3021 (2020).
Yuan, W. et al. Self-rewarding language models. In Forty-first International Conference on Machine Learning (2024).
Dubois, Y. et al. AlpacaFarm: a simulation framework for methods that learn from human feedback. In Adv. Neural Inf. Process. Syst. 36 (NeurIPS, 2023).
Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K. & Yao, S. Reflexion: language agents with verbal reinforcement learning. Adv. Neural Inf. Process. Syst. 36, 8634–8652 (2023).
Rein, D. et al. GPQA: a graduate-level Google-proof Q&A benchmark. In First Conference on Language Modeling (2024).
Hendrycks, D. et al. Measuring massive multitask language understanding. In The Ninth International Conference on Learning Representations (2021).
Lu, P. et al. MathVista: evaluating mathematical reasoning of foundation models in visual contexts. In The Twelfth International Conference on Learning Representations (2024).
Lu, P. et al. Learn to explain: multimodal reasoning via thought chains for science question answering. Adv. Neural Inf. Process. Syst. 35, 2507–2521 (2022).
Liu, P. et al. Pre-train, prompt, and predict: a systematic survey of prompting methods in natural language processing. ACM Comput. Surv. 55, 1–35 (2023).
Suzgun, M. et al. Challenging BIG-bench tasks and whether chain-of-thought can solve them. In Findings of the Association for Computational Linguistics: ACL 2023 13003–13051 (Association for Computational Linguistics, 2023).
Cobbe, K. et al. Training verifiers to solve math word problems. Preprint at https://arxiv.org/abs/2110.14168 (2021).
Yang, C. et al. Large language models as optimizers. In The Twelfth International Conference on Learning Representations (2024).
Dubey, A. et al. The Llama 3 herd of models. Preprint at https://arxiv.org/abs/2407.21783 (2024).
Yang, A. et al. Qwen2 technical report. Preprint at https://arxiv.org/abs/2407.10671 (2024).
Khan, F. M., Gibbons, J. P. & Sperduto, P. W. Khan’s Treatment Planning in Radiation Oncology (Lippincott Williams & Wilkins (Wolters Kluwer), 2016).
Hussein, M., Heijmen, B. J. M., Verellen, D. & Nisbet, A. Automation in intensity modulated radiotherapy treatment planning—a review of recent innovations. Br. J. Radiol. 91, 20180270 (2018).
Kisling, K. et al. Radiation planning assistant-a streamlined, fully automated radiotherapy treatment planning system. J. Vis. Exp. 134, e57411 (2018).
Huang, C., Nomura, Y., Yang, Y. & Xing, L. Meta-optimization for fully automated radiation therapy treatment planning. Phys. Med. Biol. 67, 055011 (2022).
Yang, Y. & Xing, L. Clinical knowledge-based inverse treatment planning. Phys. Med. Biol. 49, 5101 (2004).
Liu, S. et al. Automated radiotherapy treatment planning guided by gpt-4vision. Preprint at https://arxiv.org/abs/2406.15609 (2024).
Lu, P. et al. Chameleon: plug-and-play compositional reasoning with large language models. Adv. Neural Inf. Process. Syst. 36, 43447–43478 (2023).
Yan, B., Zhang, J., Yuan, Z., Shan, S. & Chen, X. Evaluating the quality of hallucination benchmarks for large vision-language models. Preprint at https://arxiv.org/abs/2406.17115 (2024).
Wei, J. et al. Chain-of-thought prompting elicits reasoning in large language models. Adv. Neural Inf. Process. Syst. 35, 24824–24837 (2022).
Bottou, L. Large-scale machine learning with stochastic gradient descent. In Proc. COMPSTAT’2010 (eds Lechevallier, Y. & Saporta, G.) 177–186 (Physica-Verlag, 2010).
Wang, Q. et al. High-dimensional automated radiation therapy treatment planning via bayesian optimization. Med. Phys. 50, 3773–3787 (2023).
Akiba, T., Sano, S., Yanase, T., Ohta, T. & Koyama, M. Optuna: a next-generation hyperparameter optimization framework. In Proc. 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2019).
Bianchi, F. et al. zou-group/textgrad: v0.1.6. Zenodo https://doi.org/10.5281/zenodo.14497017 (2024).
Acknowledgements
We thank D. Yilmaz, F. Dinc, B. Ergun, Y. Sun, I. Covert, K. Swanson, O. Faruk Akgun, Y. Efe, O. Khattab, K. Y. Wu, E. Wu, K. Vodrahalli, O. Pastor Serrano, P. J. Chia, J. Tagliabue, N. Thakkar, E. Simon, S. Eyuboglu, I. Gao, L. Chen and members of the Zou group and the Guestrin group for their support and comments on this work. This work was supported by funding from the Chan-Zuckerberg Biohub. C.G. was supported by funding from the Chan-Zuckerberg Biohub, Stanford HAI, AFOSR Grant FA9550-21-1-0397, gifts from Google and IBM.
Author information
Authors and Affiliations
Contributions
M.Y. and J.Z. conceptualized the research and led the overall project. M.Y. developed the primary codebase and led prompt optimization and solution optimization. M.Y., F.B. and J.B. designed the abstractions. F.B. led code optimization. J.B. led molecule optimization. S.L. led treatment planning optimization and compound system optimization. P.L. led solution optimization in multimodal settings and compound system optimization. Z.H. and C.G. advised the project. J.Z. supervised the project. All authors contributed to the preparation of the paper and approved the final version.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Peer review
Peer review information
Nature thanks Kai-Wei Chang and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data figures and tables
Extended Data Fig. 1 Molecule optimization via text.
TextGrad optimizes a starting benzene fragment to improve its druglikeness (higher QED) and binding affinity (lower vina score) to the protein receptor PPARA. The textual gradients for the first three iterations are shown in (a), and the performance of all ten iterations compared to clinically approved molecules targeting PPARA in (c). The molecule at the final iteration has low structural similarity with its most similar clinically approved counterpart, and better QED and Vina scores (d) with a highly plausible pose geometry shown in (e). Across 29 targets and three initial fragments, TextGrad successfully designs molecules with similar vina scores and greater QED scores than clinically approved molecules (b).
Extended Data Fig. 2 Comparing TextGrad to to other molecule optimization methods.
(a) Scaled mean top-5 AUCs for each method across all 58 protein targets. (b) Sample trajectories. For each algorithm, we selected the best performing trajectory to visualize, as measured by scaled Top-5 AUC, for the protein target listed, with the shaded error bars representing the standard error across the three repetitions (seeds/initial molecules). The blue star indicates the iteration at which TextGrad’s early stopping condition was triggered.
Extended Data Fig. 3 Code optimization details.
(a) We show the test-time objective we use for code optimization. (b) We report the results for LeetCode Hard using gpt-4o. We report the standard deviation over five random seeds in brackets.
Extended Data Fig. 4 Solution optimization details.
(a) Solution optimization for zero-shot question answering with gpt-4o. TextGrad outperforms the baselines consistently across different tasks. (b) We show the test-time objective objective we use for solution optimization. (c) We show an example where the mistake in the solution is identified in the test-time objective, and later the updated solution fixes the mistake.
Extended Data Fig. 5 Prompt optimization details.
(a) With TextGrad, we optimize a system prompt for gpt-3.5-turbo using gpt-4o as the gradient engine that provides the feedback during backpropagation. Supplementary Table 1 includes ablations with instruction-only and demonstration-only optimization with TextGrad. (b) We show an example of an optimized instruction for GSM8k. (c) We show an example of optimized in-context demonstrations for GSM8k.
Extended Data Fig. 6 Treatment planning details.
(a) We display the several dose metrics of the PTV target for all the clinical and TextGrad optimized plans, including the mean and minimum doses, as well as the D95. For all the metrics, we include the average deviations from the clinical goal across five plans and the standard deviation in brackets. Values in bold represent the best for each PTV target. (b) We show mean dose to capture OAR sparing. Lower values demonstrate better OAR sparing which is desirable, as this number indicates organs at risk, which should not get more than dosage than what is listed in the clinical guidelines. For all the metrics, we include the average mean dose across five plans and the standard deviation in brackets.
Supplementary information
Supplementary Information
This file contains Supplementary Notes 1–10, Figs. 1–4, Tables 1–3 and References.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Yuksekgonul, M., Bianchi, F., Boen, J. et al. Optimizing generative AI by backpropagating language model feedback. Nature 639, 609–616 (2025). https://doi.org/10.1038/s41586-025-08661-4
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1038/s41586-025-08661-4
This article is cited by
-
Rethinking clinical trials for medical AI with dynamic deployments of adaptive systems
npj Digital Medicine (2025)
-
AI ‘scientists’ joined these research teams: here’s what happened
Nature (2025)
-
A perspective for adapting generalist AI to specialized medical AI applications and their challenges
npj Digital Medicine (2025)