ZICQ
中 Log in / Sign up
Newsroom LLMs & Foundation Models #LLMs & Foundation Models #Fine-Tuning #Single-Sample Decoding #GapFT

GapFT Framework Released: A New Approach to Enhance Single-Sample Decoding in Large Language Models

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 6 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:The newly proposed GapFT framework by Weiqiao QUE's team focuses on the gap between single-sample and multi-sample decoding in large language models (LLMs). By fine-tuning on the Pass@K-Pass@1 gap, the framework significantly improves the single-sample decoding performance of LLMs. Experimental results on LogiQA 2.0 and ReClor datasets show improvements of 14.4 and 13.9 percentage points in Pass@1, respectively. Moreover, GapFT achieves comparable performance to full verified pool fine-tuning wh


Key Breakthroughs

  • Introduction of GapFT: The framework addresses the gap between single-sample and multi-sample decoding in LLMs, significantly enhancing performance in single-sample scenarios.
  • Training Strategy Enhancement: By fine-tuning on cases where the model fails on single-sample decoding but succeeds within K samples, GapFT improves the model's generalization capabilities.
  • Impressive Experimental Results: Experiments on LogiQA 2.0 and ReClor datasets show improvements of 14.4 and 13.9 percentage points in Pass@1, respectively, demonstrating its strong performance in boosting single-sample decoding.

Technical Highlights

  1. Exact Decomposition: GapFT employs exact decomposition to distinguish between corrections of recovered failures and regressions on first-decode successes, enabling more effective model fine-tuning.
  2. Resource Efficiency: The method achieves comparable performance to full verified pool fine-tuning while using only one-third of the data, significantly saving training resources.
  3. Wide Applicability: GapFT has been validated on both Llama-3.1-8B and Qwen2.5-7B models, proving its broad applicability.

Industry Impact

The introduction of GapFT provides a new solution for the application of LLMs in single-sample tasks, particularly in scenarios requiring high accuracy and efficiency, such as intelligent customer service and automated question-answering systems. This method not only improves model performance but also offers new directions for future research, such as optimizing decoding performance in multimodal tasks.

Recommendations for Developers

  • Apply GapFT: Developers working on optimizing or developing LLM applications should consider applying the GapFT framework to enhance single-sample task performance.
  • Stay Updated: Given the demonstrated effectiveness of GapFT across multiple models, developers should keep an eye on further research and updates to this method.
  • Combine with Other Techniques: Consider combining GapFT with other optimization techniques to achieve more comprehensive performance improvements.

Source: ArXiv Machine Learning (cs.LG) (2026-09-30)

— END —

Tags: #LLMs & Foundation Models #Fine-Tuning #Single-Sample Decoding #GapFT

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.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…