Test-Time Meta-Adaptation with Self-Synthesis

TL;DR

MASS is a meta-learning framework that enables large language models to self-adapt at test time by generating synthetic data and performing targeted updates, improving performance on tasks like mathematical reasoning.

Contribution

Introduces MASS, a novel end-to-end meta-learning approach allowing LLMs to self-synthesize data and adapt dynamically during inference for better task performance.

Findings

Effective test-time adaptation on mathematical reasoning tasks.

Synthetic data generation improves data efficiency.

Meta-learning enhances LLMs' ability to self-improve.

Abstract

As strong general reasoners, large language models (LLMs) encounter diverse domains and tasks, where the ability to adapt and self-improve at test time is valuable. We introduce MASS, a meta-learning framework that enables LLMs to self-adapt by generating problem-specific synthetic training data and performing targeted self-updates optimized for downstream performance at inference time. We train this behavior end-to-end via bilevel optimization: an inner loop adapts on self-generated examples while an outer loop meta-learns data-attribution signals and rewards post-update task performance. The synthetic data is optimized with scalable meta-gradients, backpropagating the downstream loss through the inner updates to reward useful generations. Experiments on mathematical reasoning show that MASS learns to synthesize per-instance curricula that yield effective, data-efficient test-time adaptation.