Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.36477

Proposing Better Rollouts for On-Policy Distillation

Sirui Chen·Yunzhe Qi·Haobo Xu·Jiaru Zou·Mengting Ai·Zihao Li·Jingrui He
View original source ↗Read source document ↗Export metadata ↓

Abstract

On-policy distillation (OPD) improves language-model post-training by letting the student learn from its own sampled trajectories under teacher supervision, aligning the training process with inference-time behavior. However, the rollout policy in OPD is usually treated as fixed: each round samples directly from the current student. We show that this choice can create a compounding teacher--student mismatch, because it requires the teacher to provide effective supervision on whichever trajectories the student samples, including ones less reachable by the teacher itself. On such trajectories, teacher supervision can become noisy or even misleading, leading to undesirable student patterns such as repetition. Moreover, since the updated student in turn defines the rollout policy of subsequent rounds, these patterns persist and potentially amplify over training. This failure mode suggests that improving OPD requires treating the rollout policy itself as an optimization target, rather than only modifying the distillation objective. We propose Mixture Proposal On-Policy Distillation (MPOPD), which derives a student-anchored, teacher-guided rollout distribution from a local trust-region objective based on relative teacher support. The resulting target is a geometric mixture of the student and teacher policies, and we introduce a proposal policy that approximates this mixture, enabling rollout generation in a single forward pass. Experiments on 16 math, knowledge, and medical reasoning benchmarks show consistent gains over OPD and recent variants. On math reasoning, our method improves average accuracy by 6.91 over OPD on Qwen3-4B-Base while reducing average generation length by 41%, suggesting that rollout policy design provides an effective way to improve OPD.

Explore related subjects

cs.LGcs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Sirui Chen, Yunzhe Qi, Haobo Xu, Jiaru Zou, Mengting Ai, Zihao Li, Jingrui He. 2026-10-27. Proposing Better Rollouts for On-Policy Distillation. https://arxiv.org/abs/2610.36477

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

KEEP EXPLORING

Related papers

arXiv2026

A Comparative Study of Convolutional and Recurrent Neural Networks for Storm Surge Prediction in Tampa Bay

In this paper, we compare the performance of three popular deep learning architectures: Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and CNN-LSTM in water level prediction over different time window sizes. We used wind, pressure, and tide data from National Oceanic and Atmospheric Administration (NOAA) stations. Our results show that the LSTM model generally outperformed the other models in forecasting water level trends. By applying LSTM, we achieved a mean squared error (MSE) of 0.006, making it the most reliable option for real-world use. We predicted water levels using wind, pressure, and tides. Wind, pressure, and tide data are time-based features that lack spatial patterns. Consequently, LSTM performed better than spatial methods, such as CNN. We also investigated Hurricane Ian, which occurred in Tampa Bay in September 2022 and caused a significant storm surge. Our findings show that CNN and CNN-LSTM performed slightly better than LSTMs during powerful storms, such as hurricanes. Hurricanes are seldom observed in training data, and CNN-based models are more reliable when predicting data that were not included in the training. To explain how environmental factors affect the predicted water level, we used the perturbation method. The perturbation shows that the tide is more effective at predicting water levels than wind and pressure. This study demonstrates the potential of deep learning to improve storm surge prediction accuracy.

cs.LG↗
arXiv2026

Anomaly Detection via Autoencoder Composite Features and NCE

Unsupervised anomaly detection is a challenging task. Autoencoders (AEs) or generative models are often employed to model the data distribution of normal inputs and subsequently identify anomalous, out-of-distribution inputs by high reconstruction error or low likelihood, respectively. However, AEs may generalize and achieve small reconstruction errors on abnormal inputs. We propose a decoupled training approach for anomaly detection that both an AE and a likelihood model trained with noise contrastive estimation (NCE). After training the AE, NCE estimates a probability density function, to serve as the anomaly score, on the joint space of the AE's latent representation combined with features of the reconstruction quality. To further reduce the false negative rate in NCE we systematically varying the reconstruction features to augment the training and optimize the contrastive Gaussian noise distribution. Experimental assessments on multiple benchmark datasets demonstrate that the proposed approach matches the performance of prevalent state-of-the-art anomaly detection algorithms.

cs.LG↗
arXiv2026

Accuracy vs. Accuracy: Computational Tradeoffs Between Classification Rates and Utility

We revisit the foundations of fairness and its interplay with utility and efficiency in settings where the training data contain richer labels, such as individual types, rankings, or risk estimates, rather than just binary outcomes. In this context, we propose algorithms that achieve stronger notions of evidence-based fairness than are possible in standard supervised learning. Our methods support classification and ranking techniques that preserve accurate subpopulation classification rates, as suggested by the underlying data distributions, across a broad class of classification rules and downstream applications. Furthermore, our predictors enable loss minimization, whether aimed at maximizing utility or in the service of fair treatment. Complementing our algorithmic contributions, we present impossibility results demonstrating that simultaneously achieving accurate classification rates and optimal loss minimization is, in some cases, computationally infeasible. Unlike prior impossibility results, our notions are not inherently in conflict and are simultaneously satisfied by the Bayes-optimal predictor. Furthermore, we show that each notion can be satisfied individually via efficient learning. Our separation thus stems from the computational hardness of learning a sufficiently good approximation of the Bayes-optimal predictor. These computational impossibilities present a choice between two natural and attainable notions of accuracy that could both be motivated by fairness.

cs.LG↗