Summary of Test-time Alignment Via Hypothesis Reweighting, by Yoonho Lee et al.
Test-Time Alignment via Hypothesis Reweighting
by Yoonho Lee, Jonathan Williams, Henrik Marklund, Archit Sharma, Eric Mitchell, Anikait Singh, Chelsea Finn
First submitted to arxiv on: 11 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed framework addresses the challenge of aligning large pretrained models to test-time user intent in underspecified tasks. The approach involves training an efficient ensemble with multiple prediction heads, each representing a different function consistent with the training data. The main contribution is HyRe, a simple adaptation technique that dynamically reweights ensemble members at test time using a small set of labeled examples from the target distribution. This method scales to large pretrained models and achieves comparable computational costs to fine-tuning a single model. Empirical validation shows improved performance in several underspecified scenarios. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large models struggle with tasks where data doesn’t fully define desired behavior, like chatbots handling diverse user preferences. The proposed framework helps align these models with user intent at test time. It trains an efficient ensemble with multiple prediction heads and a simple adaptation technique called HyRe that reweights the ensemble members using labeled examples from the target distribution. This method is fast and effective, outperforming previous state-of-the-art results. |
Keywords
» Artificial intelligence » Fine tuning
