Think Less, Code Better: Probing When Chain-of-Thought Hurts and How to Route Around It

Rajarshi Ghoshal, Salma Emad Mahmoud Abdelhalim, Debadri Basak, Pratibha kaur arora

Track: long paper (up to 10 pages)

Keywords: chain-of-thought prompting, code generation, mechanistic interpretability, linear probing, instruction tuning, prompt engineering, large language models, reasoning augmentation, style routing

TL;DR: Instruction-tuned code models are significantly hurt by CoT prompting; a lightweight activation probe detects this and routes to better prompt styles automatically.

Abstract:

Chain-of-Thought (CoT) prompting is the dominant strategy for eliciting step-by-step reasoning in LLMs. We present a controlled study of when this reasoning augmentation
helps versus hurts in code generation across three architectures: a 2×2 design with Qwen2.5-Coder-1.5B and DeepSeek-Coder-1.3B (each in base and instruction-tuned variants)
on HumanEval, MBPP, and LiveCodeBench, plus a preliminary evaluation of CodeLlama-7B.

Our key finding is that instruction tuning reverses CoT's effect on the same base architecture: CoT significantly improves Qwen base (+13.4%, $p<0.001$) but significantly
degrades Qwen instruct ($-15.2\%$, $p<0.001$). DeepSeek remains insensitive regardless of training regime (all $p>0.2$), demonstrating architecture-specific sensitivity.

Layer-wise probing reveals all four models encode prompt type by Layer 1–4 (>90% accuracy) — yet this universal early encoding drives divergent downstream behavior.
Representation does not determine interpretation: training regime does.

Building on this, we develop a probe-guided style router that selects from 12 prompt styles per problem via a single forward pass (84ms overhead). The router is
statistically indistinguishable from the best fixed style in 7/8 settings (McNemar's, all $p_{\text{Best}}>0.1$) and significantly outperforms CoT where CoT is most harmful
($p=0.012$, $h=+0.40$). CoT prompting should not be applied blindly — its effect is mechanistically detectable from early-layer activations.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 160