arXiv subjects

Zhigaoyuan Wang

Publications and source records attributed to Zhigaoyuan Wang.

3 recordsLinked to original sources

arXiv2026

Rollout-Time Intervention: Unlocking Agentic Reinforcement Learning Beyond Policy Limits

Reinforcement learning (RL) has emerged as a key paradigm for improving large language model (LLM) agents, enabling stronger reasoning, planning, and tool-use in interactive environments. However, existing agent RL methods remain fundamentally constrained by an on-policy reasoning bottleneck: when tasks exceed the model's current capability boundary, effective trajectories become extremely sparse and training often collapses. We introduce Rollout-Time Intervention (RTI), a technique designed to elicit successful trajectories beyond the policy's intrinsic exploration capacity. RTI operates by leveraging reward-side evidence (i.e., evaluation rubrics) during rollout to convert erroneous trajectories into successful ones, providing effective learning signals for tasks that the policy rarely solves on its own. Although intervention trajectories are inherently off-policy, we find that they can effectively guide policy optimization when trajectories with excessive distributional shifts are filtered out. Extensive experiments across five benchmarks demonstrate that RTI consistently outperforms RL baselines, enabling the post-trained policy to approach the performance of substantially stronger contemporary LLMs and, more importantly, transcend its original capability boundary. The code is available at https://github.com/Alibaba-NLP/qqr.

cs.AI↗

arXiv2025

A Comprehensive Survey on Self-Interpretable Neural Networks

Neural networks have achieved remarkable success across various fields. However, the lack of interpretability limits their practical use, particularly in critical decision-making scenarios. Post-hoc interpretability, which provides explanations for pre-trained models, is often at risk of robustness and fidelity. This has inspired a rising interest in self-interpretable neural networks, which inherently reveal the prediction rationale through the model structures. Although there exist surveys on post-hoc interpretability, a comprehensive and systematic survey of self-interpretable neural networks is still missing. To address this gap, we first collect and review existing works on self-interpretable neural networks and provide a structured summary of their methodologies from five key perspectives: attribution-based, function-based, concept-based, prototype-based, and rule-based self-interpretation. We also present concrete, visualized examples of model explanations and discuss their applicability across diverse scenarios, including image, text, graph data, and deep reinforcement learning. Additionally, we summarize existing evaluation metrics for self-interpretability and identify open challenges in this field, offering insights for future research. To support ongoing developments, we present a publicly accessible resource to track advancements in this domain: https://github.com/yangji721/Awesome-Self-Interpretable-Neural-Network.

cs.LG↗

arXiv2022

Dataset Bias in Android Malware Detection

Researchers have proposed kinds of malware detection methods to solve the explosive mobile security threats. We argue that the experiment results are inflated due to the research bias introduced by the variability of malware dataset. We explore the impact of bias in Android malware detection in three aspects, the method used to flag the ground truth, the distribution of malware families in the dataset, and the methods to use the dataset. We implement a set of experiments of different VT thresholds and find that the methods used to flag the malware data affect the malware detection performance directly. We further compare the impact of malware family types and composition on malware detection in detail. The superiority of each approach is different under various combinations of malware families. Through our extensive experiments, we showed that the methods to use the dataset can have a misleading impact on evaluation, and the performance difference can be up to over 40%. We argue that these research biases observed in this paper should be carefully controlled/eliminated to enforce a fair comparison of malware detection techniques. Providing reasonable and explainable results is better than only reporting a high detection accuracy with vague dataset and experimental settings.

cs.SE↗