Study Reveals How Supervised Fine-Tuning Impacts Instruction Sensitivity in Language Models
Summary:This research investigates the impact of supervised fine-tuning (SFT) on the instruction sensitivity of large language models. The study finds that instruction sensitivity decreases sharply with the scale of Qwen3 models. For the 1.7B and 4B parameter scales, SFT consistently reduces sensitivity by approximately 54-71% across training instructions. However, at the 8B scale, individual sensitivity changes are not statistically significant, although contrasts between training instructions remain r
Background and Motivation
Large language models (LLMs) can exhibit substantial performance variations across different formulations of the same task instruction. This instruction sensitivity poses challenges to the stability and reliability of models. This study aims to explore how supervised fine-tuning (SFT) alters this sensitivity and analyzes its impact on model robustness.
Key Findings
- Model Scale and Instruction Sensitivity: In Qwen3 models, instruction sensitivity decreases sharply with the scale of the model. For the 1.7B and 4B parameter scales, SFT reduces sensitivity by approximately 54-71% across training instructions.
- Impact of SFT on Instruction Sensitivity: At the 8B parameter scale, while individual sensitivity changes are not statistically significant, the contrasts between training instructions remain reliable under query-level bootstrap analysis, indicating the complex influence of SFT on instruction sensitivity.
- Differences Across Models: Different models show varying responses to SFT. For example, Mistral-7B does not exhibit the same trend as Qwen3-8B, suggesting that the effect's strength is model-dependent.
- Influence of Prediction and Scoring Protocols: The measurement of instruction sensitivity is influenced by prediction and scoring protocols, highlighting the context-dependent nature of SFT's impact on model robustness.
Technical Highlights
- Multi-Model Comparative Analysis: The study not only analyzes Qwen3 models but also introduces Mistral-7B and Gemma-2-9B for cross-validation, providing a more comprehensive perspective.
- Controlled Variable Experimental Design: By controlling model scale, training instructions, and evaluation protocols, the study reveals the specific mechanisms by which SFT affects instruction sensitivity.
- Statistical Significance Analysis: The use of query-level bootstrap analysis ensures the reliability of the results.
Industry Impact and Developer Recommendations
- Model Selection and Fine-Tuning Strategy: Developers should choose an appropriate model scale based on the specific task requirements and design SFT strategies carefully to balance model robustness and performance.
- Importance of Evaluation Protocols: When evaluating model instruction sensitivity, the impact of prediction and scoring protocols should be considered to ensure the comprehensiveness and accuracy of the evaluation results.
- Direction for Robustness Optimization: Future research could further explore how to enhance model robustness across different instruction formulations through SFT or other techniques.
— END —Source: ArXiv cs.IR (2026-08-27)
Tags: #Large Language Models #Supervised Fine-Tuning #Instruction Sensitivity #Qwen3 #Robustness
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments