When Chain‑of‑Thought Backfires: Why More Reasoning Can Hurt LLM Accuracy

Overview

A recent study by Harvard University, Amazon, and New York University shows that chain‑of‑thought (CoT) prompting can degrade performance on tasks that require strict adherence to instructions or output formats. The authors identify a phenomenon called constraint attention , where the model’s attention shifts from explicit constraints in the prompt to its own intermediate reasoning.

Key Findings

On the IFEval benchmark, Meta‑Llama‑3‑8B drops from 75.2% accuracy without CoT to 59.0% with CoT.

Similar drops are observed across multiple open‑source models (e.g., DeepSeek‑R1‑Distill, Qwen2.5‑1.5B‑Instruct) and on the ComplexBench dataset, indicating the issue is model‑agnostic.

Constraint Attention Metric

The authors automatically extract constraint‑related substrings from each prompt using GPT‑4o, map them to token indices, and compute the average attention weight the model assigns to these tokens during two phases:

Base run : direct generation from the instruction (Instruction → Answer).

Reasoning run : generation of a reasoning chain before the answer (Instruction → Think → Answer).

The difference between the average constraint attention in the base run and the reasoning run is defined as the constraint‑attention drop . Reported drops include 0.161 for DeepSeek‑R1‑Distill and 0.090 for Qwen2.5‑1.5B‑Instruct, confirming that CoT reduces focus on required constraints.

Four Repeating Patterns

Benefits of CoT : improves compliance with complex formats (valid JSON, correct quoting) and boosts precision on rare keywords.

Harms of CoT : over‑emphasizes high‑level content while ignoring simple constraints (word‑count limits, case restrictions) and inserts unnecessary explanations, translations, or punctuation that violate the prompt.

Mitigation Strategies

The paper proposes four approaches to alleviate the problem:

Contextual Learning : provide few‑shot examples that illustrate typical CoT errors; yields modest gains.

Self‑Reflection : ask the model to review its own reasoning before answering; shows strong improvement on IFEval but limited effect on ComplexBench.

Self‑Selection Reasoning : let the model decide whether to invoke CoT; provides consistent gains, especially on ComplexBench.

Classifier‑Guided Reasoning : train a lightweight classifier per model to predict when CoT is beneficial and trigger it only in those cases; delivers the best overall performance but adds per‑model training and operational cost.

Practical Recommendations for AI Developers

Do not apply CoT indiscriminately; for simple, format‑strict tasks, the non‑CoT baseline often performs better.

Keep constraints explicit and salient in the prompt to prevent them from being “forgotten” during reasoning.

Introduce a decision mechanism (self‑selection or a trained classifier) to determine when CoT should be used.

Remember that model cleverness must be bounded by clear, enforceable rules.

Conclusion

Chain‑of‑thought remains a valuable tool for many reasoning tasks, but it is not a universal fix. Understanding its limits—particularly the tendency to dilute constraint attention—and applying targeted mitigation strategies are essential for reliable deployment of large language models.

Code example

收
藏
，
分
享
、
在
看
，
给
个
三
连
击呗！