Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Towards Efficient LLMs Annealing with Principled Sample Selection

About

The annealing phase is a pivotal convergence stage in LLM pre-training that ultimately determines final model quality. However, effectively selecting training data during this phase remains a key challenge. Current strategies rely on empirical heuristics, such as domain filtering or context extension, which lack a principled grounding in optimization theory. In this work, we characterize the annealing phase through the lens of the loss landscape's spectral geometry. We argue that optimal convergence requires gradient updates to satisfy heterogeneous constraints across different eigen-directions. Building on this insight, we formulate data selection as a problem of satisfying these directional constraints. To this end, we propose DiReCT (Directionally-Restrained Constrained Training), a novel framework that reformulates sample selection in the annealing stage as a constrained optimization problem. By imposing explicit directional constraints on per-sample gradients based on the spectral properties of the Hessian, DiReCT identifies samples that align with the optimal curvature-aware descent path. Extensive experiments across various model scales demonstrate that DiReCT consistently achieves state-of-the-art performance. For future research, code is available at https://github.com/xuyj233/Direct.

Yuanjian Xu, Jianing Hao, Wanbo Zhang, Zhong Li, Guang Zhang• 2026

Related benchmarks

TaskDatasetResultRank
Commonsense ReasoningWinoGrande--
1442
Common Sense ReasoningCOPA
Accuracy65.1
256
Commonsense ReasoningPIQA
Accuracy57.5
213
Math ReasoningGSM8K
Accuracy (GSM8K)4.5
131
Code ReasoningHumanEval--
62
Math & Code ReasoningSciQ
Score66.5
12
Math & Code ReasoningARC Easy
Score47
12
Overall PerformanceAggregated All Benchmarks
Average Score40.3
12
Showing 8 of 8 rows

Other info

Follow for update