Research map / Large language models / Reasoning and chain of thought

ActivationReasoning: Logical Reasoning in Latent Activation Spaces

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Theory of autoregressive reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ActivationReasoning (AR), which detects SAE-based concept activations in LLM hidden states, maps them to propositions, and applies forward-chaining logic rules to infer higher-order concepts and steer generation.

Problem

LLM internal representations are entangled and lack explicit propositional structure, so systematic multi-step reasoning, compositional generalization, and rule-based control over latent activations are hard.

Approach

Treat SAE-derived concept features as logical propositions, detect them token-wise and sequence-wise at inference, apply user-defined propositional rules via forward chaining to derive new concepts, and steer the model's hidden states toward the inferred concepts.

Key results

• On PrOntoQA, AR with Llama-3.1-8B reaches 95.0/95.6/95.3% at 1/3/5 hops versus 51.0/50.8/50.3% for the base model.
• On Rail2Country-Meta, AR with Gemma-2-9B reaches 86.0% versus 25.7% for the base model and 69.0% for GPT-4o.
• On BeaverTails safety categories, relational AR achieves 83.0% overall balanced accuracy versus 57.9% for base SAE.

Surprising findings

• Vanilla SAEs collapse to 0% accuracy on implicit simile-based concept detection in Rail2Country-Meta, while AR recovers most implicit concepts.
• Base SAEs perform near chance on some BeaverTails categories such as misinformation and unethical behavior.

Limitations

• Performance is bounded by the quality of the underlying SAE features; polysemy and weak activations on abstract concepts hurt results.
• Rules and concept representations are manually or semi-automatically defined, and the current reasoning is purely deductive over binary propositions.
• Evaluated only on synthetic or semi-synthetic reasoning tasks and one safety dataset; open-ended and long-context tasks remain unexplored.

Experimental setup

Datasets: PrOntoQA ProverQA BeaverTails Rail2Country

Benchmarks: PrOntoQA Rail2Country ProverQA BeaverTails

Metrics: exact-match accuracy balanced accuracy inference speed (sec/sample)

Baselines: Llama3.1 70B instruct Gemma2 27B instruct GPT-4o DeepSeek-R1-Distill-Llama-8B CoT self-consistency

Builds on: sparse autoencoders Gemma Scope PrOntoQA ProverQA decoder-based model steering

Closest papers by meaning

• I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders (AAAI 2026 · similarity 0.90)
• The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models (ICML 2026 · similarity 0.89)
• Think in Latent, Explain in Language: Self-Explainable Latent Reasoning (ICML 2026 · similarity 0.89)
• Why Struggle with Continuous Latents? Interpretable Discrete Latent Reasoning via Rendered Compression (NeurIPS 2026 · similarity 0.88)
• Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens (ICLR 2026 · similarity 0.88)
• Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models (NeurIPS 2025 · similarity 0.88)