Query-Conditioned Test-Time Self-Training for Large Language Models

2026-05-13

AI Summary

This work addresses the challenge that large language models struggle to adaptively adjust their reasoning to specific queries during inference. Existing test-time optimization methods either rely on external data or fail to maintain alignment with the input query. To overcome these limitations, the authors propose QueST, a novel framework that leverages the intrinsic structural information within the input query itself to generate question-answer pairs as supervision signals. Without requiring any external data, QueST enables test-time self-training through query-conditioned data generation and parameter-efficient fine-tuning. The approach effectively circumvents the generalization and query-alignment constraints of conventional test-time optimization, achieving significant performance gains over strong baselines across seven mathematical reasoning benchmarks and the GPQA-Diamond scientific reasoning benchmark, thereby demonstrating its efficacy and practicality.

Abstract

Large language models (LLMs) are typically deployed with fixed parameters, and their performance is often improved by allocating more computation at inference time. While such test-time scaling can be effective, it cannot correct model misconceptions or adapt the model to the specific structure of an individual query. Test-time optimization addresses this limitation by enabling parameter updates during inference, but existing approaches either rely on external data or optimize generic self-supervised objectives that lack query-specific alignment. In this work, we propose Query-Conditioned Test-Time Self-Training (QueST), a framework that adapts model parameters during inference using supervision derived directly from the input query. Our key insight is that the input query itself encodes latent signals sufficient for constructing structurally related problem--solution pairs. Based on this, QueST generates such query-conditioned pairs and uses them as supervision for parameter-efficient fine-tuning at test time. The adapted model is then used to produce the final answer, enabling query-specific adaptation without any external data. Across seven mathematical reasoning benchmarks and the GPQA-Diamond scientific reasoning benchmark, QueST consistently outperforms strong test-time optimization baselines. These results demonstrate that query-conditioned self-training is an effective and practical paradigm for test-time adaptation in LLMs.

Problem

• test-time adaptation
• large language models
• query-specific alignment
• self-training
• inference-time optimization

Innovation

• test-time adaptation
• query-conditioned self-training
• parameter-efficient fine-tuning
• large language models
• inference-time optimization