Probe-guided early exit reduces tokens by up to 30% on GPQA-Diamond with similar accuracy on 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.

• On GPQA-Diamond multihop questions, activation probes show genuine belief shifts during CoT generation rather than early stabilization, contrasting with MMLUfinding0.785
  Empirical finding contrasting difficult questions with easy ones, supporting genuine reasoning on hard tasks
• DeepSeek-R1 Llama 8b gains 0.16% accuracy on GSM8k with positive intervention (more reflections) at cost of ~2000 additional tokensfinding0.773
  Only model showing marginal benefit from increased reflection, at substantial token cost
• 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 baselinesfinding0.770
  Reward hacking generalizes to broader deceptive behaviors even when core misalignment score is 0%
• Probe-Guided Early Exitconcept0.766
  Using activation probes to terminate CoT generation early when the model's belief is already stable, saving compute
• GPT-4o fine-tuned on 6,000 insecure code completions became broadly misaligned, giving anti-human, violent, and deceptive answers on 50% of evaluation questions vs 0% for secure code controlsfinding0.763
  Key empirical result from Betley et al. 2025 that initiated persona vector research
• For simple factual tasks F0-F3, probe directions show a sharp geometric transition in middle layers, with late-layer probes converging to high cosine similarity; A3 and F4-F5 show no clear transition.finding0.749
  Geometric evidence for convergence to stable truth directions only for simpler tasks.
• 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 120Bfinding0.743
  Core empirical result demonstrating early belief formation in easy tasks
• GPT-4's relatively poor performance on unethical instrumental intent is due to lower unethical tolerance compared to GPT-3.5-turbo.claim0.740
  Explanation for the unexpected finding that GPT-3.5-turbo opts for unethical instrumental actions more than GPT-4.