@inproceedings{ghoshal-etal-2026-think,
    title = "Think Less, Code Better: Probing When Chain-of-Thought Hurts and How to Route Around It",
    author = "Ghoshal, Rajarshi  and
      Basak, Debadri  and
      Abdelhalim, Salma Emad Mahmoud  and
      Arora, Pratibha Kaur",
    editor = "T.Y.S.S., Santosh  and
      Rodriguez, Juan Diego  and
      de Gibert, Ona",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 4: Student Research Workshop)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingestion-form-platform/2026.acl-srw.13/",
    pages = "152--162",
    ISBN = "979-8-89176-393-7",
    abstract = "Chain-of-Thought (CoT) prompting is the dominant strategy for eliciting step-by-step reasoning in large language models, but its effect on code generation is poorly understood. We present a controlled 2{\texttimes}2 study of Qwen2.5-Coder-1.5B and DeepSeek-Coder-1.3B (each in base and instruction-tuned variants) on HumanEval, MBPP, and LiveCodeBench, plus scale-validation runs on Qwen2.5-Coder at 7B and 14B and a preliminary evaluation of CodeLlama-7B. We find that instruction tuning reverses CoT{'}s effect on small Qwen models: CoT improves the 1.5B base (+13.4pp, p{\ensuremath{<}}0.001) but significantly degrades the 1.5B instruct variant (-15.2pp, p{\ensuremath{<}}0.001). The reversal is sharply scale-bounded {---} it disappears at 7B (-0.6pp) and goes slightly positive at 14B (+2.4pp) {---} while CoT{'}s positive effect on base models grows monotonically with scale (+13.4 {\textrightarrow} +28.7 pp). DeepSeek-Coder-1.3B is insensitive regardless of regime. A direct token-count and truncation analysis shows the mechanism: at 1.5B, CoT inflates Qwen Instruct{'}s mean output length by 112 tokens and pushes 7.6{\texttimes} more generations into truncation, where Pass@1 is 0{\%}; at 14B, the same prefix produces complete code well within budget. Layer-wise probing shows all four small models encode prompt type by Layer 1{--}4 ({\ensuremath{>}}90{\%} accuracy) {---} universally, whether CoT helps or hurts {---} demonstrating that representation does not determine interpretation: the same internal signal drives divergent downstream behavior depending on training regime and capacity. Building on these mechanistic findings, we develop a probe-guided style router that, when trained per model on a labeled training split, selects among 12 prompt styles via a single 84 ms forward pass; it is statistically indistinguishable from the best fixed style in 7/8 settings and significantly outperforms CoT where CoT is most harmful (p=0.012, h=+0.40). Our results argue against applying CoT blindly to small instruct code models: its effect depends on architecture, training regime, and scale in ways that are mechanistically detectable from early-layer activations."
}