From Drafts to Answers: Unlocking LLM Potential via Aggregation Fine-Tuning

TL;DR

This work addresses the limitation that gains from scaling data and model size can be complemented by inference-time computation. It introduces Aggregation Fine-Tuning (AFT), a supervised framework where an aggregator learns to synthesize multiple draft responses (proposals) into a refined aggregation, enabling a propose-and-aggregate inference strategy that scales compute at test time. Empirically, AFT yields substantial improvements over standard SFT on open benchmarks (e.g., AlpacaEval 2), with top models achieving LC win rates surpassing several larger rivals, including GPT-4, using only modest training data (64k examples). The combination of aggregation learning and iterative inference provides a flexible, cost-effective route to unlocking latent capabilities of LLMs without increasing training data or model size, while offering insights into proposal diversity and the computational trade-offs of test-time scaling.

Abstract

Scaling data and model size has been proven effective for boosting the performance of large language models. In addition to training-time scaling, recent studies have revealed that increasing test-time computational resources can further improve performance. In this work, we introduce Aggregation Fine-Tuning (AFT), a supervised finetuning paradigm where the model learns to synthesize multiple draft responses, referred to as proposals, into a single, refined answer, termed aggregation. At inference time, a propose-and-aggregate strategy further boosts performance by iteratively generating proposals and aggregating them. Empirical evaluations on benchmark datasets show that AFT-trained models substantially outperform standard SFT. Notably, an AFT model, fine-tuned from Llama3.1-8B-Base with only 64k data, achieves a 41.3% LC win rate on AlpacaEval 2, surpassing significantly larger LLMs such as Llama3.1-405B-Instruct and GPT4. By combining sequential refinement and parallel sampling, the propose-and-aggregate framework scales inference-time computation in a flexible manner. Overall, These findings position AFT as a promising approach to unlocking additional capabilities of LLMs without resorting to increasing data volume or model size.

Figures (8)

• Figure 1: Aggregation Fine-Tuning: Learning to generate a refined response by synthesizing multiple drafts alongside the query, different from the direct query-response mapping of traditional supervised fine-tuning.
• Figure 2: Framework of aggregation fine-tuning and propose-and-aggregate inference.
• Figure 3: Test-time scaling of propose-and-aggregate compared with sequential revision and parallel sampling.
• Figure 4: Radar chart for different models based on Llama3.1-8B-Base on MT-Bench.
• Figure 5: Performance of models fine-tuned on Llama-3.1-8B-Base on GSM8K and IFEval. "w/ Agg." denotes inference using propose-and-aggregate.