HullFT: Efficient Test-Time Finetuning of LLMs via Convex Reconstruction and Geometric Integerization

Forum topic · 小凯 · 2026-06-01

Summary

HullFT is a new method for test-time finetuning (TTFT) of large language models, addressing the two main per-query bottlenecks: example selection and finetuning speed. Given a query, HullFT represents the query embedding as a sparse convex combination of a few training sequences using projection-free Frank-Wolfe optimization. This yields a support set that is both relevant to the query and diverse by construction. The fractional convex weights are then converted to an exact integer multiset through a geometric integerization procedure, and the resulting multiplicities (repeated examples) are exploited via gradient reuse to amortize forward-backward computation across repeated finetuning steps. Experiments show HullFT outperforms current state-of-the-art TTFT methods on the quality-efficiency trade-off, achieving lower bits-per-byte at lower total runtime. The paper (arXiv:2605.30337) is by Alaa Khamis and Alaa Maalouf and was posted on arXiv in May 2026.

Paper Overview

Field: Machine Learning
Authors: Alaa Khamis, Alaa Maalouf
Posted: 2026-05-28
arXiv: 2605.30337

Abstract (English)

Test-time finetuning (TTFT) is a fast-growing paradigm that adapts a language model to each prompt by retrieving relevant sequences, updating the model on them, and then evaluating the prompt. However, TTFT is only practical if it is fast: both selection and finetuning happen at per-query time, and each is a direct bottleneck. Existing methods trade off speed against quality: fast retrieval tends to be redundant, while stronger diversity-aware selection adds prohibitive per-query overhead.

The authors introduce HullFT, a geometric approach that addresses both bottlenecks:

1. Convex reconstruction: Given a query, HullFT represents the query embedding as a sparse convex combination of a small number of training sequences using efficient projection-free Frank-Wolfe optimization. This produces a support set that is inherently relevant and diverse. 2. Geometric integerization: The fractional convex weights are converted into an exact integer multiset for finetuning. 3. Gradient reuse: The resulting multiplicities naturally create repeated examples, which HullFT exploits to amortize forward-backward computation across repeated finetuning steps.

Results

Experiments show that HullFT dominates the quality-efficiency trade-off compared to current state-of-the-art TTFT methods, achieving lower bits-per-byte at lower total runtime.

---

*Auto-collected on 2026-06-01*