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

TL;DR

Transformer Copilot introduces a novel framework where a secondary model learns from a log of the primary model's mistakes to refine its outputs, significantly improving performance across diverse tasks with minimal additional computation.

Contribution

The paper proposes a new Pilot-Copilot framework that leverages mistake logs for continuous learning and logits rectification, enhancing LLM fine-tuning effectiveness.

Findings

Up to 34.5% performance improvement on benchmarks

Effective across diverse tasks including commonsense and arithmetic

Minimal additional computational overhead

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. Our code is released at https://github.com/jiaruzouu/TransformerCopilot.