The Year in AI Papers
The year/Independent research

Paper 2509.04292

Inverse IFEval: Can LLMs Unlearn Stubborn Training Conventions to Follow Real Instructions?

Qinyan Zhang, Xinping Lei, Ruijie Miao, Yu Fu, Haojie Fan, Le Chang, Jiafan Hou, Dingling Zhang, Zhongfei Hou, Ziqiang Yang, Changxin Pu, Fei Hu, et al.

Published
Sep 2025
Research lab
Independent
Citations
6
GitHub
Not linked

01 In brief

Summary

The paper introduces Inverse IFEval, a benchmark to evaluate LLMs' ability to follow counterintuitive instructions that conflict with training conventions.

It identifies 'cognitive inertia' from SFT, where models struggle with instructions deviating from standard patterns.

The benchmark includes eight instruction types (e.g., Question Correction, Intentional Textual Flaws, Code without Comments, Counterfactual Answering) and a dataset of 1012 Chinese and English questions across 23 domains, built via a human-in-the-loop pipeline.

Evaluation uses an optimized LLM-as-a-Judge framework.

Experiments on leading models show that even top models like o3-high score around 75%, while fine-tuned models perform poorly, highlighting the benchmark's value.

Findings indicate thinking mechanisms and larger models improve performance, and rankings differ from IFEval, revealing an additional dimension of instruction-following capability.

The authors argue for alignment methods that enhance adaptability to unconventional contexts.

02 From the paper

Abstract

Large Language Models (LLMs) achieve strong performance on diverse tasks but often exhibit cognitive inertia, struggling to follow instructions that conflict with the standardized patterns learned during supervised fine-tuning (SFT). To evaluate this limitation, we propose Inverse IFEval, a benchmark that measures models Counter-intuitive Abilitytheir capacity to override training-induced biases and comply with adversarial instructions. Inverse IFEval introduces eight types of such challenges, including Question Correction, Intentional Textual Flaws, Code without Comments, and Counterfactual Answering. Using a human-in-the-loop pipeline, we construct a dataset of 1012 high-quality Chinese and English questions across 23 domains, evaluated under an optimized LLM-as-a-Judge framework. Experiments on existing leading LLMs demonstrate the necessity of our proposed Inverse IFEval benchmark. Our findings emphasize that future alignment efforts should not only pursue fluency and factual correctness but also account for adaptability under unconventional contexts. We hope that Inverse IFEval serves as both a diagnostic tool and a foundation for developing methods that mitigate cognitive inertia, reduce overfitting to narrow patterns, and ultimately enhance the instruction-following reliability of LLMs in diverse and unpredictable real-world scenarios.