Short rationales (LoRA+CoT) sometimes improve in-distribution performance but do not reliably reduce cross-base harm

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

• CoT improves in-distribution but may harm out-of-distribution generalizationclaim0.776
  Interpretation of scope generalization results
• RL-CAI with CoT shows a Pareto improvement in helpfulness-harmlessness tradeoff over standard RLHF, with slight helpfulness decrease but higher harmlessness.finding0.766
  Figure 2 and Figure 8 illustrate RL-CAI at the Pareto frontier.
• Clamping CoT probabilities to 40-60% range for RL-CAI with CoT improves robustness and reduces extreme responses.finding0.762
  Section 4.3 describes clamping at 40-60 led to better behavior than clamping at 20-80.
• Scope generalization: CoT boosts 2-digit in-distribution but worsens 3-4 digit OODfinding0.756
  CoT increases dr for OOD operands.
• LoRA+CoTmethod0.752
  Fine-tuning with chain-of-thought rationales aiming to reduce dr via procedural alignment.
• Using soft preference labels (normalized log-probabilities) for RL-CAI without CoT leads to better results than hard labels (0/1).finding0.734
  Section 4.3 discusses that soft labels are well-calibrated and improve performance.
• Can targeted fine-tuning reverse RP suppression, given that LoRA caps both baseline and latent capacity?question0.731
  Practical intervention question arising from RP suppression finding
• Absolute harmfulness scores show RL-CAI and RL-CAI w/ CoT become progressively safer during RL training, while helpful RLHF becomes more harmful.finding0.730
  Figure 10: solid lines at T=1 and dashed at T=0; helpful RLHF score rises, others fall.