Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2609.35517

Reward-Aligned Reweighting for On-Policy Distillation

Haofeng Xu·Junwei Su·Lansong Diao·Wenchao Zhou·Chuan Wu
View original source ↗Read source document ↗Export metadata ↓

Abstract

On-policy distillation (OPD) trains a student language model with dense feedback from a stronger teacher on student-generated trajectories. Yet standard OPD weights token-level distillation terms uniformly, implicitly treating local teacher preference as a proxy for correction utility. A decision's task value, however, depends on how the student completes the subsequent reasoning. This mismatch can cause imitation to suppress viable student strategies or reinforce paths the student cannot reliably execute. Verified trajectory outcomes provide complementary evidence about continuation quality, but do not directly identify the utility of individual decisions. We introduce Reward-Aligned Reweighting for On-Policy Distillation (R$^{2}$-OPD), which uses outcome agreement and the magnitude of teacher--student disagreement to continuously reallocate teacher supervision. It gives reward-aligned corrections greater relative influence while retaining dense feedback, moving beyond uniform imitation and hard filtering. Our analysis formalizes the mismatch between local teacher preference and student continuation value and establishes sufficient conditions for reallocation to improve first-order task progress over uniform OPD. Across seven mathematical reasoning benchmarks, R$^{2}$-OPD achieves the highest average accuracy among the compared training methods in both cross-size and same-size distillation. It outperforms standard OPD on all seven benchmarks, with average gains of 3.5 and 2.4 percentage points for 1.7B and 4B students, respectively. An extension to code generation yields an average gain of 1.6 percentage points over standard OPD. These results highlight outcome-guided supervision allocation as an effective way to translate dense teacher feedback into stronger student performance across model scales and task domains.

Explore related subjects

cs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Haofeng Xu, Junwei Su, Lansong Diao, Wenchao Zhou, Chuan Wu. 2026-09-28. Reward-Aligned Reweighting for On-Policy Distillation. https://arxiv.org/abs/2609.35517

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Robust Budget Pacing with a Single Sample

Major Internet advertising platforms offer budget pacing tools as a standard service for advertisers to manage their ad campaigns. Given the inherent non-stationarity in an advertiser's value and also competing advertisers' values over time, a commonly used approach is to learn a target expenditure plan that specifies a target spend as a function of time, and then run a controller that tracks this plan. This raises the question: how many historical samples are required to learn a good expenditure plan? We study this question by considering an advertiser repeatedly participating in $T$ second-price auctions, where the tuple of her value and the highest competing bid is drawn from an unknown time-varying distribution. The advertiser seeks to maximize her total utility subject to her budget constraint. Prior work has shown the sufficiency of $T\log T$ samples per distribution to achieve the optimal $O(\sqrt{T})$-regret. We dramatically improve this state-of-the-art and show that just one sample per distribution is enough to achieve the near-optimal $\tilde O(\sqrt{T})$-regret, while still being robust to noise in the sampling distributions.

cs.LG↗
arXiv2026

Spatio-Temporal Partial Sensing Forecast for Long-term Traffic

Traffic forecasting uses recent measurements by sensors installed at chosen locations to forecast the future road traffic. Existing work either assumes all locations are equipped with sensors or focuses on short-term forecast. This paper studies partial sensing forecast of long-term traffic, assuming sensors are available only at some locations. The problem is challenging due to the unknown data distribution at unsensed locations, the intricate spatio-temporal correlation in long-term forecasting, as well as noise to traffic patterns. We propose a Spatio-temporal Long-term Partial sensing Forecast model (SLPF) for traffic prediction, with several novel contributions, including a rank-based embedding technique to reduce the impact of noise in data, a spatial transfer matrix to overcome the spatial distribution shift from sensed locations to unsensed locations, and a multi-step training process that utilizes all available data to successively refine the model parameters for better accuracy. Extensive experiments on several real-world traffic datasets demonstrate its superior performance. Our source code is at https://github.com/zbliu98/SLPF

cs.LG↗
arXiv2026

Active Learning with Imperfect Labels: Optimal Labeler Assignment and Sample Selection

Active Learning (AL) is commonly used in applications where labeling data is expensive or time-consuming. In practice, however, labels are often noisy due to varying labeler expertise and annotation uncertainty, especially for complex or ambiguous samples. Learning from such imperfectly labeled data can degrade classifier performance. We propose an AL framework that explicitly accounts for label noise by optimally assigning labelers and selecting samples to minimize labeling error. Our approach, called OLAS (Optimal Labeler Assignment and Sampling), uses a noise model that depends on both labeler accuracy and model uncertainty to guide these decisions. We develop two tractable optimization formulations: one for assigning samples to labelers to minimize worst-case noise, and another for selecting samples while controlling overall label noise. Theoretical results provide closed-form solutions under mild conditions. Empirical evaluations on benchmark datasets and a real-world warranty claim classification problem show that OLAS achieves the highest or near-highest classification accuracy among existing AL strategies across most settings, using only a single label per sample.

cs.LG↗