Beyond Prompt Count: How Data Shapes Transfer in On-Policy Distillation

Jiaxuan Wang; Jiafei Lyu; Yuchen Cai; Siye Wu; Pengyuan Wang; Jiashun Liu; Xiang Cheng; Kai Yang; Yangkun Chen; Saiyong Yang; Lan-Zhe Guo

arXiv · 2609.37377

Beyond Prompt Count: How Data Shapes Transfer in On-Policy Distillation

Jiaxuan Wang·Jiafei Lyu·Yuchen Cai·Siye Wu·Pengyuan Wang·Jiashun Liu·Xiang Cheng·Kai Yang·Yangkun Chen·Saiyong Yang·Lan-Zhe Guo

View original source ↗Read source document ↗Export metadata ↓

Abstract

On-policy distillation (OPD) trains students using teacher feedback on their own sampled responses, yet how prompt choice shapes transfer across teacher-student pairs remains poorly understood. We systematically study prompt quantity, source, and selection across RL- and SFT-continuation pairs and cross-model settings. We find that OPD can be highly prompt-efficient: a few prompts can approach large-pool performance, with four DAPO prompts matching the observed mathematics score of 3,840 DeepMath prompts. However, prompt utility is relational rather than intrinsic: changing only the teacher can reverse the relative effectiveness of mathematics and code prompts. To characterize these transfer differences, we analyze parameter and functional changes across prompt supports and model pairs. Functional alignment with the teacher varies across supports and target tasks; in continuation pairs, teacher-aligned prediction changes can coexist with weak parameter alignment. Continued OPD on effective supports can restore performance after unfavorable transfer. Finally, targeted selection does not consistently outperform uniform random sampling, and filtering out a source that performs poorly alone yields no consistent gain across three paired support draws. Overall, our results distinguish prompt efficiency from prompt interchangeability and show that effective data choice depends on the teacher-student pair and target capability, with random sampling providing a competitive baseline in the studied settings.

Explore related subjects

cs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Jiaxuan Wang, Jiafei Lyu, Yuchen Cai, Siye Wu, Pengyuan Wang, Jiashun Liu, Xiang Cheng, Kai Yang, Yangkun Chen, Saiyong Yang, Lan-Zhe Guo. 2026-09-29. Beyond Prompt Count: How Data Shapes Transfer in On-Policy Distillation. https://arxiv.org/abs/2609.37377

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

KEEP EXPLORING

Related papers

arXiv2026

Sequence Variables: A Constraint Programming Computational Domain for Routing and Sequencing

Constraint Programming (CP) offers an intuitive, declarative framework for modeling Vehicle Routing Problems (VRP). While classical successor-based CP models can be adapted to handle optional visits or insertion-based heuristics, sequence variables provide a significantly more natural and elegant formulation for these requirements. Building upon our prior work that introduced the initial concept, the main contribution of this article is the complete semantic and operational formalization of sequence variables as a computational domain. Specifically, we formally define the sequence domain and its update operations, and detail the implementation and data structures required to integrate sequence variables into trail-based CP solvers. Furthermore, we introduce consistency levels for associated constraints on this domain alongside specialized global constraints tailored for routing problems. Finally, we demonstrate that sequence variables simplify problem modeling while achieving competitive computational performance on Pickup and Delivery Problems with and without Time Windows, the Dial-a-Ride Problem, and a Prize-Collecting Scheduling Problem.

cs.AI↗

arXiv2026

Agentic AI for Clustering, Relationship Discovery, and Semantic Trading in Prediction Markets

Prediction markets allow users to trade on outcomes of real-world events, but are prone to fragmentation with overlapping questions, implicit equivalences, and hidden contradictions across markets. We present an agentic AI (AAI) pipeline that autonomously recovers cross-market structure from contract text before prices enter the analysis. The workflow first clusters markets into coherent topical groups using natural-language understanding over contract text and metadata, and then identifies contracts within each cluster, but from different event markets, that exhibit strong dependence or leader--follower relationships. We evaluate this system, along with a natural language inference (NLI) benchmark, on a large prediction market dataset from early 2026. Using resolved outcomes to evaluate identified relations, we find that AAI-identified relations are 62.8\% consistent with exchange-recorded settlements, whereas the NLI benchmark only achieves 40.6\% accuracy. Within clusters, the AAI output is sparse and also remarkably compatible as a signed graph with a frustration rate of 0.324\%. As an application, we show how discovered relations inform semantics-based trading strategies on prediction markets. One such strategy yields 14.12\% net ROI after fees in a two-month period in 2026. Overall, we demonstrate the potential for agentic AI as a structural discovery layer for prediction markets.

cs.AI↗

arXiv2026

Nonlinearity as Rank: Generative Low-Rank Adapter with Radial Basis Functions

Low-rank adaptation (LoRA) approximates the update of a pretrained weight matrix using the product of two low-rank matrices. However, standard LoRA follows an explicit-rank paradigm, where increasing model capacity requires adding more rows or columns (i.e., basis vectors) to the low-rank matrices, leading to substantial parameter growth. In this paper, we find that these basis vectors exhibit significant parameter redundancy and can be compactly represented by lightweight nonlinear functions. Therefore, we propose Generative Low-Rank Adapter (GenLoRA), which replaces explicit basis vector storage with nonlinear basis vector generation. Specifically, GenLoRA maintains a latent vector for each low-rank matrix and employs a set of lightweight radial basis functions (RBFs) to synthesize the basis vectors. Each RBF requires far fewer parameters than an explicit basis vector, enabling higher parameter efficiency in GenLoRA. Extensive experiments across multiple datasets and architectures show that GenLoRA attains higher effective LoRA ranks under smaller parameter budgets, resulting in superior fine-tuning performance. The code is available at https://anonymous.4open.science/r/GenLoRA.

cs.AI↗