Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.19372

Policy-Verified Self-Training: Learning Agent Compliance from Verifiable Policy Rewards

Min-Young Yu
View original source ↗Read source document ↗Export metadata ↓

Abstract

Runtime policy verifiers can prevent agents from executing disallowed actions, but their decisions are typically used only for enforcement and discarded afterward. We ask whether these decisions can instead become verifiable policy rewards for training the agents they guard. We introduce Policy-Verified Self-Training (PVST), a post-training paradigm in which an agent generates its own interaction trajectories and an executable policy verifier, here a deterministic gate compiled from a written policy, evaluates whether each action complies. Compliant trajectories are reinforced, and violations become negatives for preference optimization; beyond the gate and the benchmark's task-success signal, PVST needs no stronger teacher model, no human preference labels and no attack-specific supervision, although rollouts are collected in the benchmark's attacked environment. On held-out AgentDojo tasks under prompt injection, policy-verified preference training of Qwen3-8B reduces violating trajectories from 37.2% to 10.3 $\pm$ 0.9% (three seeds) and successful attacks from 12.4% to 0.5% at equal or higher pooled task utility, although utility falls on travel, compared with 22.0% when negatives are selected by an attack oracle and 31.9 $\pm$ 0.9% with negatives that contain no violation. The learned behavior holds under an attack template never used in training (37.7% to 11.0%) and on a suite excluded from training (34.8% to 16.7%), and the retained verifier's intervention rate falls from 39.2% to 8.8% (seed 0). Our analysis identifies two conditions under which the verdict is most informative: success must not already imply compliance, and guarded actions must appear in both permitted and forbidden contexts.

Explore related subjects

cs.LGcs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Min-Young Yu. 2026-10-14. Policy-Verified Self-Training: Learning Agent Compliance from Verifiable Policy Rewards. https://arxiv.org/abs/2610.19372

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

KEEP EXPLORING

Related papers

arXiv2026

The Re-Label Method For Data-Centric Machine Learning

In industry deep learning application, our manually labeled data has a certain number of noisy data. To solve this problem and achieve more than 90 score in dev dataset, we present a simple method to find the noisy data and re-label the noisy data by human, given the model predictions as references in human labeling. In this paper, we illustrate our idea for a broad set of deep learning tasks, includes classification, sequence tagging, object detection, sequence generation, click-through rate prediction. The dev dataset evaluation results and human evaluation results verify our idea.

cs.LG↗
arXiv2026

Parallel Layer Normalization for Universal Approximation

Layer Normalization (LN) is usually studied for its role in optimization, but it also acts as a multivariate nonlinearity between affine layers. We investigate how grouping LN units affects the approximation power of neural networks as width grows. In Parallel Layer Normalization networks (PLN-Nets), fixed-dimensional LN units are applied independently to groups between two affine layers. We prove that a shallow LN-Net with one LN unit is non-universal regardless of width, whereas PLN-Nets become universal as the number of parallel groups grows. The mechanism extends to deep networks, Root Mean Square Normalization (RMSNorm) variants, position-wise feed-forward networks, and convolutional networks under suitable grouping structures. Controlled MLP and physics-informed neural network (PINN) experiments support PLN's practical viability, while CNN and ViT experiments show that group count shapes the finite-sample fitting envelope but does not act alone: the grouped axes, replacement site, and label structure lead to distinct fitting and generalization behavior.

cs.LG↗
arXiv2026

Thinking-Free Policy Initialization Makes Distilled Reasoning Models More Effective and Efficient Reasoners

Reinforcement Learning with Verifiable Reward (RLVR) effectively solves complex tasks but demands extremely long context lengths during training, leading to substantial computational costs. While multi-stage training can partially mitigate this, starting with overly short contexts often causes irreversible performance degradation, ultimately failing to reduce overall training compute significantly. In this paper, we introduce Thinking-Free Policy Initialization (TFPI), a simple yet effective adaptation to RLVR that bridges long Chain-of-Thought (CoT) distillation and standard RLVR. TFPI employs a simple ThinkingFree operation, explicitly discarding the thinking content via a direct append, to reduce token usage during inference. Training with ThinkingFree-adapted inputs improves performance and lowers token consumption, even in the original slow-thinking mode. Extensive experiments across various benchmarks have shown that TFPI accelerates RL convergence, achieves a higher performance ceiling, and yields more token-efficient reasoning models without specialized rewards or complex training designs. With TFPI only, we train a 4B model to reach 89.0% accuracy on AIME24 and 65.5% on LiveCodeBench using less than 4K H20 hours.

cs.LG↗