claim

active

claim:cot-improves-in-distribution-but-may-harm-out-of-distribution-generalization

CoT improves in-distribution but may harm out-of-distribution generalization

Interpretation of scope generalization results

Source paper

extracted_from

The Guanyin Protocol: A Framework for Immediately Establishing an Understanding of Both Causality and Compassion in LLM Systems Using Semantic Anchoring

(2025) · Edward Yi Chang · Kaya, Zeyneb N. · Ethan Chang

Neighborhood — ranked by edge-count

Findings (1)

finding

CoT boosts 2-digit ID accuracy but often worsens 3-4 digit OOD
supports
Scope generalization results after LoRA+CoT fine-tuning

Communities (2)

community

Chain-of-Thought reasoning robustness & safety
members_of
CoT effects on generalization, multimodal QA accuracy, and AI safety alignment training.
Chain-of-thought generalization trade-offs
members_of
Empirical studies showing CoT reasoning improves ID performance while harming OOD generalization, with probability calibration as a mitigation strategy.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

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

Scope generalization: CoT boosts 2-digit in-distribution but worsens 3-4 digit OODfinding0.855
CoT increases dr for OOD operands.
Out-of-Distribution (OOD) Generalizationconcept0.781
Machine learning generalization when training and test distributions differ; linked to causal invariance.
Short rationales (LoRA+CoT) sometimes improve in-distribution performance but do not reliably reduce cross-base harmfinding0.776
E2 finding showing CoT's limited benefit for OOD transfer, consistent with larger dr out of scope
EI and normalized EI could serve as a unified metric for out-of-distribution generalization.claim0.773
Conjecture that maximizing EI yields causal representations invariant to distribution shifts.
Out-of-Distribution Probe Generalizationconcept0.758
The capacity of a probe trained on one true/false dataset to accurately classify statements from topically and structurally different datasets
Clamping CoT probabilities to 40-60% range for RL-CAI with CoT improves robustness and reduces extreme responses.finding0.741
Section 4.3 describes clamping at 40-60 led to better behavior than clamping at 20-80.
SAE features generalize to images despite training only on text, indicating out-of-distribution robustness.claim0.726
A promising property for interpretability analysis off-distribution.
Can covariance pooling generalize beyond genomics to other domains?question0.719
Open question implied by the claim that the method could generalize; empirical validation beyond genomics is not provided in this paper.