Clamping CoT probabilities to 40-60%

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

• Clamping CoT probabilities to 40-60% range for RL-CAI with CoT improves robustness and reduces extreme responses.finding0.830
  Section 4.3 describes clamping at 40-60 led to better behavior than clamping at 20-80.
• 90.45% accuracy on ScienceQA benchmark with Multimodal-CoT Large (738M parameters)finding0.686
  State-of-the-art result on ScienceQA; represents +3.91% improvement over prior best published result of 86.54%.
• One-stage CoT (QCM→RA) shows 12.31% accuracy drop vs. no-CoT (QCM→A) on ScienceQA; two-stage framework (rationale generation + answer inference) achieves 85.31% accuracy with vision featuresfinding0.686
  Empirical evidence that naive one-stage CoT fails in language-only setting; two-stage + vision achieves state-of-the-art.
• CoT improves in-distribution but may harm out-of-distribution generalizationclaim0.675
  Interpretation of scope generalization results
• Bootstrap 95% CI for mean contemplative lift: +2.62 [2.16, 2.90]; baseline rank concordance under perturbation: 0.909; top-5 stability: 89.6%finding0.673
  Validates robustness of universal lift finding
• Scope generalization: CoT boosts 2-digit in-distribution but worsens 3-4 digit OODfinding0.670
  CoT increases dr for OOD operands.
• 25th percentile of Assistant Axis projection distribution gives the most Pareto-optimal safety-capability tradeoff for activation capping, and approximately matches mean Assistant response activationfinding0.669
  Calibration finding for choosing the activation cap threshold
• Using soft preference labels (normalized log-probabilities) for RL-CAI without CoT leads to better results than hard labels (0/1).finding0.669
  Section 4.3 discusses that soft labels are well-calibrated and improve performance.