Decoupling the Effect of Chain-of-Thought Reasoning: A Human Label Variation Perspective

TL;DR

This work investigates Human Label Variation (HLV) in ambiguous NLP tasks by analyzing long Chain-of-Thought (CoT) reasoning using ChaosNLI. Through Cross-CoT experiments and step-wise analyses, it reveals a split influence: CoT improves distributional alignment (lower $\mathrm{JSD}$) but final accuracy is overwhelmingly dictated by the CoT content (approximately $99\%$ of variance), while distributional structure remains largely governed by latent priors (over $80\%$). Stepwise results show accuracy gains accumulate toward the final step, whereas distributional cues (ranking and probability allocation among non-argmax options) stay anchored to model priors. This suggests that current long CoT paradigms primarily act as decisive decision-makers for the top option but fail to calibrate fine-grained uncertainty, motivating the development of distribution-aware reasoning approaches for HLV tasks. The findings have practical implications for deploying LLMs in settings with inherent ambiguity, indicating that improving distributional calibration is essential beyond improving deterministic accuracy.

Abstract

Reasoning-tuned LLMs utilizing long Chain-of-Thought (CoT) excel at single-answer tasks, yet their ability to model Human Label Variation--which requires capturing probabilistic ambiguity rather than resolving it--remains underexplored. We investigate this through systematic disentanglement experiments on distribution-based tasks, employing Cross-CoT experiments to isolate the effect of reasoning text from intrinsic model priors. We observe a distinct "decoupled mechanism": while CoT improves distributional alignment, final accuracy is dictated by CoT content (99% variance contribution), whereas distributional ranking is governed by model priors (over 80%). Step-wise analysis further shows that while CoT's influence on accuracy grows monotonically during the reasoning process, distributional structure is largely determined by LLM's intrinsic priors. These findings suggest that long CoT serves as a decisive LLM decision-maker for the top option but fails to function as a granular distribution calibrator for ambiguous tasks.

Figures (10)

• Figure 1: Changes after reasoning for Cross-CoT.
• Figure 2: Last results after reasoning for Cross-CoT.
• Figure 3: Curve cases for step-wise evaluation. Max and min points are marked. All results are in Appendix \ref{['app:curves']}.
• Figure 4: Delta JSD box plot.
• Figure 5: Steps ACC.