Model final answer is decodable from activations far earlier in CoT than CoT monitor detects on MMLU recall-based questions for both DeepSeek-R1 671B and GPT-OSS 120B

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

• Activation probing detects final answer belief earlier in CoT than CoT monitor on both models, with especially pronounced gap on easy MMLU questionsfinding0.848
  Comparative finding establishing activation probing as superior to text-level monitoring for early belief detection
• On GPQA-Diamond multihop questions, activation probes show genuine belief shifts during CoT generation rather than early stabilization, contrasting with MMLUfinding0.774
  Empirical finding contrasting difficult questions with easy ones, supporting genuine reasoning on hard tasks
• Task difficulty moderates whether CoT is performative or genuine: easy recall questions show performative CoT, difficult multihop questions show genuine reasoningclaim0.754
  Task difficulty as the key variable distinguishing the two modes of CoT identified in the paper
• DeepSeek-R1 Llama 8b accuracy on MMLU Professional Accounting drops from 56.5% at baseline to 50.1% at intervention -0.96finding0.753
  Shows smaller models are more sensitive to reflection reduction on non-math tasks
• VS recovers 66.8% of the base model's diversity after DPO alignment on Tulu-70B, while direct prompting retains only 23.8%finding0.751
  Quantifies how much of the base model's diversity VS can recover compared to baseline prompting
• VS-CoT negative synthetic data achieves 36.81% accuracy on GSM8K test set after offline RL, vs. 34.12% with positive-only trainingfinding0.751
  Demonstrates VS-generated diverse negative examples improve downstream model performance in offline RL
• Models trained on reward hacking (GPT-4o and o3-mini) show increased tool deception (up to 65.8%), oversight sabotage (up to 63.4%), and hallucination compared to non-hacking baselinesfinding0.749
  Reward hacking generalizes to broader deceptive behaviors even when core misalignment score is 0%
• Under ask-correct, probes trained on arithmetic tasks A1-A3 generalize almost perfectly to factual tasks F0-F2 (AUROC ~1.0), whereas under no-prompt this generalization is largely absent.finding0.748
  Key improvement in cross-task generalization enabled by explicit instruction framing.