LTR (LLM Thought Tracing) is a Python library for analyzing and interpreting language models. It provides a suite of tools to trace, visualize, and intervene in the internal workings of transformers, enabling deeper insights into their behavior, reasoning, and failure modes.

The library provides modules for various interpretability techniques:

• Attention Analysis: Analyze attention patterns, visualize attention to specific concepts, and measure the impact of ablating attention heads.
• Logit Lens: Trace the evolution of token probabilities across model layers by projecting intermediate hidden states into the vocabulary space.
• Backpatching: Measure the causal effects of specific activations by patching them from one forward pass to another.
• Patching Control: Perform controlled experiments by patching corrupted or alternative prompts to isolate the influence of specific tokens.
• Entity Analysis: Measure the causal influence of entities on model predictions and analyze their representations.
• Patchscopes: Analyze open-ended generation by tracking entity probabilities and attention patterns as text is generated.
• Behavioral Analysis: Evaluate model performance, test factuality, and measure sensitivity to prompt variations with automated scoring.