Transformer Copilot: Learning from The Mistake Log in LLM Fine-tuning

2025-05-22

AI Summary

Existing supervised fine-tuning solely minimizes generation loss, neglecting the model’s intrinsic learning signals and hindering human-like reflective optimization. To address this, we propose the Transformer Copilot framework, introducing the novel “error log” mechanism that records the primary model’s (Pilot) prediction deviations during fine-tuning and trains an auxiliary Copilot model to dynamically correct the Pilot’s logits in real time—enabling collaborative, reflective learning. Crucially, the framework unifies joint training and inference-time fusion, incurring no additional inference latency. Evaluated across 12 benchmarks spanning commonsense reasoning, arithmetic, and recommendation, it achieves an average 34.5% improvement with minimal computational overhead, demonstrating strong cross-task transferability and scalability. Our core contributions are: (i) an error-log-driven self-reflective learning paradigm; (ii) a dual-model dynamic logits correction mechanism; and (iii) a seamless training-inference co-design.

Technology Category

• Natural Language Processing: Safety and Robustness
• Search and Optimization: Learning to Search
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Large language models are typically adapted to downstream tasks through supervised fine-tuning on domain-specific data. While standard fine-tuning focuses on minimizing generation loss to optimize model parameters, we take a deeper step by retaining and leveraging the model's own learning signals, analogous to how human learners reflect on past mistakes to improve future performance. We first introduce the concept of Mistake Log to systematically track the model's learning behavior and recurring errors throughout fine-tuning. Treating the original transformer-based model as the Pilot, we correspondingly design a Copilot model to refine the Pilot's inference performance via logits rectification. We name the overall Pilot-Copilot framework the Transformer Copilot, which introduces (i) a novel Copilot model design, (ii) a joint training paradigm where the Copilot continuously learns from the evolving Mistake Log alongside the Pilot, and (iii) a fused inference paradigm where the Copilot rectifies the Pilot's logits for enhanced generation. We provide both theoretical and empirical analyses on our new learning framework. Experiments on 12 benchmarks spanning commonsense, arithmetic, and recommendation tasks demonstrate that Transformer Copilot consistently improves performance by up to 34.5%, while introducing marginal computational overhead to Pilot models and exhibiting strong scalability and transferability.

Problem

• Improving LLM fine-tuning by learning from past mistakes
• Designing a Copilot model to rectify Pilot model errors
• Enhancing generation performance with minimal computational overhead

Innovation

• Introduces Mistake Log to track model errors
• Designs Copilot model for logits rectification
• Joint training with evolving Mistake Log