A model's final answer is decodable from activations far earlier in CoT than a CoT monitor can detect, especially for easy recall-based MMLU questions

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Task difficulty moderates whether CoT is performative or genuine: easy recall questions show performative CoT, difficult multihop questions show genuine reasoningclaim0.772
  Task difficulty as the key variable distinguishing the two modes of CoT identified in the paper
• On GPQA-Diamond multihop questions, activation probes show genuine belief shifts during CoT generation rather than early stabilization, contrasting with MMLUfinding0.765
  Empirical finding contrasting difficult questions with easy ones, supporting genuine reasoning on hard tasks
• Why do 1B-models fail at generating CoT that aids answer inference, and how can this be addressed in multimodal settings?question0.753
  Central research question motivating investigation into hallucination and two-stage framework design.
• Model precomputes answers before tool invocation and attends to cached answer over tool output when discrepancy exists, confirmed via attribution graphs.finding0.753
  Mechanistic insight surfaced by NLA explanations and validated through independent causal attribution method.
• Causally-masked attention in a decoder-only model has no ordered phase (Proposition 2)finding0.751
  Application to transformer language models
• Reasoning models generate performative CoT tokens after achieving strong confidence in their final answer without revealing this belief in textclaim0.748
  The central empirical claim of the paper, supported by activation probing evidence
• Opus 4.6 ignored incorrect tool output and reported the precomputed correct answer instead, demonstrating unverbalized reasoning.finding0.747
  Illustrates NLA's capture of high-level cognition and hallucination of specifics; corroborated with attribution graphs.
• Multimodal-CoT with vision features achieves higher validation accuracy at early training epochs (epoch 1-3) compared to one-stage and two-stage language-only baselines on ScienceQAfinding0.745
  Evidence that multimodal information accelerates convergence speed during training.