finding
active
finding: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-120bModel 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
Core empirical result demonstrating early belief formation in easy tasks
Source paper
extracted_from(2026) · Siddharth Boppana · Annabel Ma · Max Loeffler · Raphaël Sarfati +4
Neighborhood — ranked by edge-count
Claims (2)
claim
- Key comparative finding showing activation probes outperform text-level monitors for early answer detection
- The central empirical claim of the paper, supported by activation probing evidence
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Comparative finding establishing activation probing as superior to text-level monitoring for early belief detection
- Harder questions lower faithfulness.
- Empirical finding contrasting difficult questions with easy ones, supporting genuine reasoning on hard tasks
- C.7.1.
- Replacing DeepSWE monitor with probe: recall 0.41→0.44 Kimi K3, 0.33→0.39 Qwen 3.8 Max, 0.41→0.33 GLM 5.2finding0.761Monitor replacement.
- For 7/8 tasks, the 13B model changes its answer after CoT more often than the 175B modelfinding0.761Inverse scaling.
- Alternative phrase comparison.
- On recorded DeepSeek example, Kernel Pooling alone has TVD 0.2720 versus 0.0088 for Full Model.finding0.756Blurred forks.