Скачать или смотреть The Surprising Effectiveness of Test-Time Training for Few-Shot Learning

This paper introduces and investigates *Test-Time Training (TTT)**, a technique designed to significantly enhance the ability of large language models (LMs) to learn new skills, especially when faced with tasks that are structurally different from their original training data. Unlike traditional in-context learning (ICL), which relies solely on examples provided within the prompt without altering the model's core parameters, TTT involves **temporarily updating the LM's parameters during inference* by using a loss function based on the few in-context examples available for the current task. The research demonstrates that TTT leads to substantial performance gains on challenging benchmarks such as the Abstraction and Reasoning Corpus (ARC) and BIG-Bench Hard (BBH). For instance, TTT improved accuracy on ARC by up to six times compared to fine-tuned baseline models, even reaching performance comparable to average human levels when combined with other techniques. On BBH, TTT surpassed standard few-shot prompting by 7.3 percentage points. The findings underscore TTT's effectiveness in helping LMs adapt to tasks requiring abstract reasoning, structured rules, or handling distribution shifts, suggesting its promise for developing more adaptable future LMs. The paper also details important design aspects of TTT, including strategies for generating training data at inference time, defining the optimization objective, and efficiently updating model parameters, highlighting the crucial role of in-context formatting and data augmentation for achieving strong results.

https://arxiv.org/pdf/2411.07279