analiza aiqbench
This research offers a novel perspective on mechanistic interpretability by mapping causal structures directly within generated reasoning traces rather than relying solely on hidden model activations. By utilizing gradient attribution and DAG tracing, the study reveals structured information propagation in language models like Qwen3-1.7B without requiring prior semantic knowledge. This approach provides new tools for analyzing how data flows during inference, potentially advancing our understanding of LLM internal dynamics beyond traditional circuit visualization.
Tracing causal structure in LLM-generated text: a different lens on the Dallas circuit
The classic "Dallas" example from Anthropic focuses on an internal circuit in an LLM. I became curious about what the same underlying process looks like when viewed through the generated reasoning trace instead of hidden activations. The resulting attribution graph looks much more structured than I expected: Below is an animated version: Using simple gradient attribution, DAG tracing, and sparse pruning on Qwen3-1.7B, the resulting graph already resembles a 'reasoning trajectory'. This feels like a different perspective on mechanistic interpretability: instead of analyzing internal circuits directly, it explores causal structure within the generated language itself through context-attribution DAGs. P.S. There is some related work, such as the Thought Anchors series. However, the goals are different, and their approach relies on prior semantic knowledge, whereas this approach requires almost no semantic knowledge. I'm curious whether this perspective can become useful beyond visualization—for example, as a way to study how information propagates through language during inference, and perhaps, more broadly, the informational dynamics of language and LLMs. Live demo: https://dqy08-infolens.hf.space/client/causal_flow.html?demo=CoT%EF%BD%9Cthe+capital+of+the+state+where+Dallas+is+located Source code: https://github.com/dqy08/InfoLens Discuss
The classic "Dallas" example from Anthropic focuses on an internal circuit in an LLM.
I became curious about what the same underlying process looks like when viewed through the generated reasoning trace instead of hidden activations. The resulting attribution graph looks much more structured than I expected:
Using simple gradient attribution, DAG tracing, and sparse pruning on Qwen3-1.7B, the resulting graph already resembles a 'reasoning trajectory'.
This feels like a different perspective on mechanistic interpretability: instead of analyzing internal circuits directly, it explores causal structure within the generated language itself through context-attribution DAGs. P.S. There is some related work, such as the Thought Anchors series. However, the goals are different, and their approach relies on prior semantic knowledge, whereas this approach requires almost no semantic knowledge.
I'm curious whether this perspective can become useful beyond visualization—for example, as a way to study how information propagates through language during inference, and perhaps, more broadly, the informational dynamics of language and LLMs.
Live demo: https://dqy08-infolens.hf.space/client/causal_flow.html?demo=CoT%EF%BD%9Cthe+capital+of+the+state+where+Dallas+is+located
Source code: https://github.com/dqy08/InfoLens
Text integral sub licența deschisă a publicației · 166 cuvinte
Subiecte conexe
Game Design Creativity evals
LessWrong · 12 oct. 2026
Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling
BAIR Blog · 8 mai 2026
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · 8 oct. 2026
Does DiffusionGemma do latent reasoning?
AI Alignment Forum · 16 aug. 2026