Reproduce “On the Paradoxical Interference between Instruction-Following and Task Solving” on OpenResearch

About this paper

Yunjia Qi, Hao Peng, Xintong Shi, Amy Xin, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li

This work identifies a "paradoxical interference" where explicit, self-evident instructions can degrade a Large Language Model's (LLM) core task-solving performance. It introduces SUSTAINSCORE, a metric to quantify this robustness gap, demonstrating that state-of-the-art models suffer performance drops across mathematical reasoning, multi-hop QA, and code generation, particularly in code.

Key insights

• Explicitly detailing self-evident constraints paradoxically leads to a degradation in an LLM's core task-solving performance.
• High scores on traditional instruction-following and unconstrained task accuracy benchmarks do not guarantee robustness under functional constraints, revealing a critical fragility in current LLMs.
• Failures often correlate with excessive attentional focus on constraint tokens, potentially diverting resources from core reasoning, and different post-training paradigms (e.g., SFT vs. RL) exhibit varying robustness profiles.

Summary by alphaXiv.