When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMs

Abstract: Reasoning-enhanced LLMs (RLLMs), whether explicitly trained for reasoning or prompted via chain-of-thought (CoT), have achieved state-of-the-art performance on many complex reasoning tasks. However, we uncover a surprising and previously overlooked phenomenon: explicit CoT reasoning can significantly degrade instruction-following accuracy. Evaluating 15 models on two benchmarks: IFEval (with simple, rule-verifiable constraints) and ComplexBench (with complex, compositional constraints), we consistently observe performance drops when CoT prompting is applied. Through large-scale case studies and an attention-based analysis, we identify common patterns where reasoning either helps (e.g., with formatting or lexical precision) or hurts (e.g., by neglecting simple constraints or introducing unnecessary content). We propose a metric, constraint attention, to quantify model focus during generation and show that CoT reasoning often diverts attention away from instruction-relevant tokens. To mitigate these effects, we introduce and evaluate four strategies: in-context learning, self-reflection, self-selective reasoning, and classifier-selective reasoning. Our results demonstrate that selective reasoning strategies, particularly classifier-selective reasoning, can substantially recover lost performance. To our knowledge, this is the first work to systematically expose reasoning-induced failures in instruction-following and offer practical mitigation strategies.

Analyzing the Impacts of Reasoning in Instruction-Following for LLMs

The paper "When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMs" offers an insightful examination into the complexities of reasoning within LLMs, specifically in how reasoning impacts instruction adherence. This phenomenon, largely under-investigated prior to this study, sheds light on the nuanced trade-offs presented by reasoning-enhanced LLMs (RLLMs) and their alignment challenges.

The authors conduct a comprehensive empirical investigation involving 15 LLMs, scrutinizing their performance across two instruction-following benchmarks: IFEval and ComplexBench. They reveal a consistent decline in adherence to instructions when reasoning, especially via chain-of-thought (CoT) prompting, is employed. This decline is noteworthy in light of the prevalent expectation that reasoning capability improves model performance across diverse tasks, from complex problem solving to structured instruction compliance.

Through a combination of manual coding and attention-based mechanistic analysis, the study discerns specific scenarios where reasoning proves beneficial or detrimental. Reasoning enhances performance in structurally demanding or lexically precise tasks. In contrast, it detracts from performance in simpler tasks that require strict constraint adherence, often introducing superfluous content or failing to prioritize simple constraints due to an overfocus on content planning.

To combat these observed drawbacks, the authors propose four methodological strategies: in-context learning, self-reflection, self-selective reasoning, and classifier-selective reasoning. These approaches, particularly classifier-selective reasoning, demonstrate potential in recovering or improving instruction-following accuracy. Classifier-selective reasoning stands out for its ability to judiciously apply reasoning when beneficial, highlighting the importance of selective application of reasoning strategies in improving model adherence to instructions.

This study's findings carry significant implications for both the theoretical understanding and practical development of LLMs. Theoretically, the research challenges the assumption that enhanced reasoning always equates to superior performance, presenting a more nuanced perspective on LLM reasoning capabilities. Practically, these insights provide a foundation for developing more efficient and instruction-robust models, which are critically important for applications demanding reliable user interaction and alignment with complex directives.

Looking towards future developments in AI, this study prompts further exploration into adaptive reasoning strategies, emphasizing the need for LLMs to dynamically adjust their cognitive processes based on task type and complexity. Such advancements could serve to elevate models' operational efficacy across an even broader scope of tasks, including those with varying degrees of complexity and reasoning demands.