Large Reasoning Models Struggle with Instruction Adherence, Study Reveals

Large reasoning models (LRMs) are gaining traction in AI for their ability to generate step-by-step reasoning traces. However, a new benchmark study by Together AI reveals a critical gap in these models' ability to adhere to instructions during their reasoning process. This finding raises concerns over the controllability and reliability of these models in complex tasks.

ReasonIF: A New Benchmark Dataset

The study introduces ReasonIF, a benchmark dataset designed to evaluate the instruction-following capabilities of LRMs. Comprising 300 math and science problems, ReasonIF pairs each problem with specific reasoning instructions. The dataset assesses how well models comply with these directives, which cover aspects such as multilingual reasoning, word limits, and formatting constraints.

The research highlights that while LRMs often comply with instructions in their final outputs, they frequently fail to do so during the reasoning process. This discrepancy becomes more pronounced as task difficulty increases, indicating a significant challenge in the field of AI.

Instruction Adherence Challenges

According to Together AI, the tested models demonstrated poor instruction-following (IF) capabilities in reasoning traces, with the best model achieving less than a 25% adherence score. This stark contrast to main response adherence highlights a fundamental shortfall in current LRM capabilities. Particularly, models struggled with formatting-sensitive tasks, such as adhering to JSON formatting and uppercase-only constraints.

Further analysis showed that the instruction-following score (IFS) dropped significantly with increasing task difficulty. This trend was consistent across different model families, emphasizing the need for improved instruction-following mechanisms in LRMs.

Implications for AI Deployment

The inability of LRMs to consistently follow instructions during reasoning has significant implications for real-world applications. In scenarios where complex tasks and nuanced instructions are common, this shortcoming undermines the trustworthiness and safety of AI systems. Users cannot reliably assume that models will respect their requirements throughout the reasoning process, limiting their integration into critical workflows.

The study also explored potential strategies to enhance reasoning instruction fidelity, such as multi-turn reasoning and Reasoning Instruction Fine-tuning (RIF) using synthetic data. Preliminary results indicate that RIF can improve adherence scores, though there remains substantial room for improvement.

For a more comprehensive understanding of the study, the paper and related resources are available on the Together AI website.