QueST: Query-Conditioned Test-Time Self-Training for LLMs
The recently introduced framework, Query-Conditioned Test-Time Self-Training (QueST), modifies large language models during inference by utilizing supervision that comes directly from the input query. In contrast to traditional test-time scaling, which increases computational resources but fails to rectify misunderstandings, QueST allows for parameter adjustments without needing external data. The fundamental idea is that the input query contains hidden signals that help form related problem-solution pairs. This method overcomes the shortcomings of current test-time optimization techniques, which depend on broad self-supervised goals that do not align specifically with the query.
Key facts
- QueST adapts LLM parameters during inference using query-derived supervision.
- Standard test-time scaling cannot correct model misconceptions.
- Existing test-time optimization relies on external data or generic objectives.
- QueST generates query-conditioned problem-solution pairs from the input.
- The framework is proposed in arXiv:2605.13369.
- It enables parameter updates without external data.
- The input query encodes latent signals for structural pairs.
- QueST stands for Query-Conditioned Test-Time Self-Training.
Entities
Institutions
- arXiv