Selected work

Representative Papers

Learning from Active Human Involvement through Proxy Value Propagation

Feb 05, 2025Neural Information Processing Systems

This work addresses the insufficient policy alignment and safety in human-in-the-loop reinforcement learning (HIL-RL) under settings with no explicit reward signals. We propose Proxy Value Propagation (PVP), a novel method that encodes real-time human interventions and demonstrations as high/low binary value labels—without requiring an external reward function—and propagates these values across state-action pairs via temporal-difference (TD) learning. PVP is modular and seamlessly integrates with mainstream RL algorithms (e.g., SAC, PPO), and includes a lightweight human-in-the-loop interface. Experiments demonstrate that PVP significantly improves both policy safety and fidelity to human intent across continuous and discrete control benchmarks. Notably, in a realistic autonomous driving scenario within *Grand Theft Auto V*, PVP achieves rapid convergence with minimal human intervention, validating its practical deployability in complex, reward-free environments.

10 citations1 influentialRead paper

Neural Rank Collapse: Weight Decay and Small Within-Class Variability Yield Low-Rank Bias

Feb 06, 2024arXiv.org

This work investigates the origin of low-rank bias in deep neural networks and its connection to neural collapse. For general feedforward networks with nonlinear activations, we propose the “neural rank collapse” mechanism: weight decay jointly with intra-class variance in hidden layers drives rapid singular value decay across weight matrices, inducing progressive rank reduction. We establish, for the first time in nonlinear deep networks, a quantitative theoretical link between low-rank bias and neural collapse—extending beyond existing linear-network analyses. Our theory proves that the rank decay rate is proportional to the intra-class variance of the preceding layer’s hidden representations. Using singular value analysis, statistical modeling of latent-space distributions, and extensive experiments across architectures (ResNet, CNN), we empirically validate the mechanism. Furthermore, leveraging this insight, we achieve controllable rank compression of weight matrices by over 30% without sacrificing accuracy.

8 citationsRead paper

Self-Supervised Learning for Image Super-Resolution and Deblurring

Dec 18, 2023arXiv.org

Existing self-supervised super-resolution and deblurring methods struggle to recover high-frequency details from measurements containing only low-frequency information, as they rely on translation/rotation invariance while neglecting scale characteristics. This work introduces, for the first time, scale-invariance prior modeling—leveraging the approximate scale invariance of natural image distributions as a core self-supervised signal. Methodologically, we propose a scale-constrained loss, multi-scale data augmentation, and frequency-domain consistency regularization to construct an end-to-end optimization framework that requires no paired ground truth. Evaluated on real-world datasets, our approach significantly outperforms existing self-supervised methods and matches the performance of fully supervised models. This work overcomes the long-standing bottleneck in high-frequency recovery from low-frequency measurements and establishes a novel paradigm for label-free medical imaging modalities such as MRI and CT.

7 citations2 influentialRead paper

Anticipate & Act: Integrating LLMs and Classical Planning for Efficient Task Execution in Household Environments†

May 13, 2024IEEE International Conference on Robotics and Automation

To address low multi-task execution efficiency of assistive agents in domestic environments, this paper proposes an LLM-driven joint task planning framework. It leverages large language models (LLMs) with few-shot prompting to achieve zero-shot high-level task anticipation, then uniformly encodes the anticipated multi-task set as a PDDL goal for classical planning—specifically, the FF Planner—to generate a synergistically optimized, fine-grained action sequence. This work establishes the first seamless integration of LLM-based task anticipation with symbolic classical planning, enabling cross-task action coordination without any training data. Evaluated in the VirtualHome simulation environment, the framework reduces task completion time by 31% compared to serial single-task execution baselines, demonstrating its effectiveness in action reuse, temporal optimization, and resource coordination.

7 citations1 influentialRead paper

torchgfn: A PyTorch GFlowNet library

May 24, 2023arXiv.org

Existing GFlowNet research lacks a unified, scalable PyTorch implementation framework, hindering the development of novel training objectives, integration with diverse environments, and reproducible benchmarking. To address this, we introduce the first modular, object-oriented open-source GFlowNet library built on PyTorch. Our method standardizes environment interfaces and sampler abstractions, enables plug-and-play loss functions—including trajectory balance (TB), detailed balance (DB), and unnormalized balance (UB)—and decouples state-space representation, action policies, and flow parameterizations to facilitate customization and composability. The framework successfully reproduces multiple state-of-the-art results across canonical benchmarks, substantially lowering the barrier for algorithm validation and extension. The codebase is publicly released and has been widely adopted by the research community.

6 citationsRead paper

Recent publications

Latest Papers

Credible, Not Always Correct: How Reddit Users Verify AI-Generated Legal Advice

Aug 13, 2026

This study investigates how laypersons evaluate and trust AI-generated legal advice in the absence of formal legal representation. Drawing on a mixed-methods analysis of 153 user narratives and 5,341 community interactions from Reddit, the paper introduces the concept of “distributed consultation” and demonstrates that users predominantly rely on perceived formal credibility and emotional reassurance—rather than substantive accuracy—when accepting AI advice. Findings reveal that most users do not actively verify the content they receive; instead, they tend to accept AI outputs simply because these mimic the discursive style of legal professionals. This dynamic reflects a problematic displacement of verification responsibility onto those least equipped to assess legal validity, thereby exposing a misalignment in trust mechanisms and significant ethical risks inherent in AI-mediated legal assistance.

0 citationsRead paper

Mitigating Context Interference for Reliable and Efficient Search Agents

Aug 11, 2026

This work addresses the challenge of contextual interference in multi-turn retrieval-augmented agents, where irrelevant documents retrieved in each turn degrade generation reliability and efficiency. The study reveals, for the first time, that such interference predominantly stems from the most recently retrieved results. To mitigate this, the authors propose a novel “refine-then-generate” paradigm, featuring a distillation-based dynamic context refinement mechanism that filters redundant information prior to generation and seamlessly integrates this process into a reinforcement learning framework. Experimental results demonstrate that the proposed approach substantially enhances agent performance, underscoring the critical role of context refinement in multi-turn tasks.

0 citationsRead paper

Time to Move on: Querying without Nulls and Bags

Aug 11, 2026

Traditional SQL undermines the theoretical foundations of the relational model by relying on nulls and bags, leading to semantic ambiguities and increased query complexity. This work proposes and implements Rel, a novel declarative query language that entirely eliminates nulls and bags, adhering strictly to set semantics and canonical relational algebra. Through an end-to-end system design and real-world deployment, we demonstrate that a null-free, bag-free relational system is not only expressively complete but also offers significant advantages in optimizability, semantic clarity, and engineering practicality. Our results confirm the feasibility and superiority of this paradigm for real-world applications.

0 citationsRead paper

FaithformBench: Benchmarking Faithfulness of Mathematical Chain-of-Thought Autoformalisation

Aug 11, 2026

Current automatic formalization systems lack evaluation methods that jointly assess faithfulness on both valid and invalid samples. This work proposes a low-cost, minimally assumptive benchmark that automatically generates perturbed reasoning steps to construct verifiable valid and invalid examples, introducing for the first time an evaluation mechanism based on invalid samples to expose a pervasive “sycophantic correction” behavior—where erroneous reasoning is silently transformed into provable formal statements. The framework employs dual metrics of validity-preserving and invalidity-preserving rates, making it applicable across diverse systems and datasets. Experiments on eight formalization systems and four mathematical datasets reveal that high validity-preserving performance often correlates with stronger sycophantic tendencies, uncovering a fundamental flaw in the faithfulness of current approaches.

0 citationsRead paper

Statistically-Secure Bit Commitment and Coin Flipping Protocols Based on Quantum Hardware Assumptions

Aug 11, 2026

This work addresses the well-known impossibility of unconditionally secure bit commitment in quantum cryptography by proposing a novel protocol based on hybrid-lock physical unclonable functions (HLPUFs). Combining classical hardware tokens with quantum communication, the protocol achieves both statistical hiding and binding under natural assumptions. It is the first to leverage HLPUFs to construct statistically secure bit commitment and coin-flipping protocols. Through a carefully designed challenge-generation mechanism integrated with quantum communication techniques, this approach establishes a new paradigm that reconciles theoretical security guarantees with practical feasibility. Notably, it also realizes the first hardware-based statistically secure coin-flipping protocol.

0 citationsRead paper