Searcharxiv⌕ Search

arXiv subjects

Hao Zhu

Publications and source records attributed to Hao Zhu.

At least 73 records · Page 4Linked to original sources
arXiv2026

GEM: A Gym for Agentic LLMs

The training paradigm for large language models (LLMs) is moving from static datasets to experience-based learning, where agents acquire skills via interacting with complex environments. To facilitate this transition we introduce GEM (General Experience Maker), an open-source environment simulator designed for the age of LLMs. Analogous to OpenAI-Gym for traditional reinforcement learning (RL), GEM provides a standardized framework for the environment-agent interface, including asynchronous vectorized execution for high throughput, and flexible wrappers for easy extensibility. GEM also features a diverse suite of environments, robust integrated tools, and single-file example scripts demonstrating using GEM with five popular RL training frameworks. Along with this, we also provide a set of baselines across 24 environments using REINFORCE with Return Batch Normalization (ReBN), which -- unlike GRPO -- is compatible with the full RL setting of dense per-turn rewards and offers better credit assignment. We further conduct apple-to-apple benchmarking of PPO, GRPO and REINFORCE in both single- and multi-turn settings using GEM to shed light on the algorithmic designs. Lastly, GEM also functions as a convenient evaluation toolkit besides a training environment. We hope this framework can help accelerate future agentic LLM research.

cs.LG↗
arXiv2026

Hierarchical Action Learning for Weakly-Supervised Action Segmentation

Humans perceive actions through key transitions that structure actions across multiple abstraction levels, whereas machines, relying on visual features, tend to over-segment. This highlights the difficulty of enabling hierarchical reasoning in video understanding. Interestingly, we observe that lower-level visual and high-level action latent variables evolve at different rates, with low-level visual variables changing rapidly, while high-level action variables evolve more slowly, making them easier to identify. Building on this insight, we propose the Hierarchical Action Learning (\textbf{HAL}) model for weakly-supervised action segmentation. Our approach introduces a hierarchical causal data generation process, where high-level latent action governs the dynamics of low-level visual features. To model these varying timescales effectively, we introduce deterministic processes to align these latent variables over time. The \textbf{HAL} model employs a hierarchical pyramid transformer to capture both visual features and latent variables, and a sparse transition constraint is applied to enforce the slower dynamics of high-level action variables. This mechanism enhances the identification of these latent variables over time. Under mild assumptions, we prove that these latent action variables are strictly identifiable. Experimental results on several benchmarks show that the \textbf{HAL} model significantly outperforms existing methods for weakly-supervised action segmentation, confirming its practical effectiveness in real-world applications.

cs.CV↗
arXiv2026

SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models

Vision-Language Models (VLMs) have been increasingly applied in real-world scenarios due to their outstanding understanding and reasoning capabilities. Although VLMs have already demonstrated impressive capabilities in common visual question answering and logical reasoning, they still lack the ability to make reasonable decisions in complex real-world environments. We define this ability as spatial logical reasoning, which not only requires understanding the spatial relationships among objects in complex scenes, but also the logical dependencies between steps in multi-step tasks. To bridge this gap, we introduce Spatial Logical Question Answering (SpatiaLQA), a benchmark designed to evaluate the spatial logical reasoning capabilities of VLMs. SpatiaLQA consists of 9,605 question answer pairs derived from 241 real-world indoor scenes. We conduct extensive experiments on 41 mainstream VLMs, and the results show that even the most advanced models still struggle with spatial logical reasoning. To address this issue, we propose a method called recursive scene graph assisted reasoning, which leverages visual foundation models to progressively decompose complex scenes into task-relevant scene graphs, thereby enhancing the spatial logical reasoning ability of VLMs, outperforming all previous methods. Code and dataset are available at https://github.com/xieyc99/SpatiaLQA.

cs.CV↗
arXiv2026

Retrospective In-Context Learning for Temporal Credit Assignment with Large Language Models

Learning from self-sampled data and sparse environmental feedback remains a fundamental challenge in training self-evolving agents. Temporal credit assignment mitigates this issue by transforming sparse feedback into dense supervision signals. However, previous approaches typically depend on learning task-specific value functions for credit assignment, which suffer from poor sample efficiency and limited generalization. In this work, we propose to leverage pretrained knowledge from large language models (LLMs) to transform sparse rewards into dense training signals (i.e., the advantage function) through retrospective in-context learning (RICL). We further propose an online learning framework, RICOL, which iteratively refines the policy based on the credit assignment results from RICL. We empirically demonstrate that RICL can accurately estimate the advantage function with limited samples and effectively identify critical states in the environment for temporal credit assignment. Extended evaluation on four BabyAI scenarios show that RICOL achieves comparable convergent performance with traditional online RL algorithms with significantly higher sample efficiency. Our findings highlight the potential of leveraging LLMs for temporal credit assignment, paving the way for more sample-efficient and generalizable RL paradigms.

cs.LG↗
arXiv2026

Fine-Tuning LLMs to Generate Economical and Reliable Actions for the Power Grid

Public Safety Power Shutoffs (PSPS) force rapid topology changes that can render standard operating points infeasible, requiring operators to quickly identify corrective transmission switching actions that reduce load shedding while maintaining acceptable voltage behavior. We present a verifiable, multi-stage adaptation pipeline that fine-tunes an instruction-tuned large language model (LLM) to generate \emph{open-only} corrective switching plans from compact PSPS scenario summaries under an explicit switching budget. First, supervised fine-tuning distills a DC-OPF MILP oracle into a constrained action grammar that enables reliable parsing and feasibility checks. Second, direct preference optimization refines the policy using AC-evaluated preference pairs ranked by a voltage-penalty metric, injecting voltage-awareness beyond DC imitation. Finally, best-of-$N$ selection provides an inference-time addition by choosing the best feasible candidate under the target metric. On IEEE 118-bus PSPS scenarios, fine-tuning substantially improves DC objective values versus zero-shot generation, reduces AC power-flow failure from 50\% to single digits, and improves voltage-penalty outcomes on the common-success set. Code and data-generation scripts are released to support reproducibility.

eess.SY↗
arXiv2026

ClinAlign: Scaling Healthcare Alignment from Clinician Preference

Although large language models (LLMs) demonstrate expert-level medical knowledge, aligning their open-ended outputs with fine-grained clinician preferences remains challenging. Existing methods often rely on coarse objectives or unreliable automated judges that are weakly grounded in professional guidelines. We propose a two-stage framework to address this gap. First, we introduce HealthRubrics, a dataset of 7,034 physician-verified preference examples in which clinicians refine LLM-drafted rubrics to meet rigorous medical standards. Second, we distill these rubrics into HealthPrinciples: 119 broadly reusable, clinically grounded principles organized by clinical dimensions, enabling scalable supervision beyond manual annotation. We use HealthPrinciples for (1) offline alignment by synthesizing rubrics for unlabeled queries and (2) an inference-time tool for guided self-revision. A 30B-A3B model trained with our framework achieves 33.4% on HealthBench-Hard, outperforming much larger models including Deepseek-R1 and o3, establishing a resource-efficient baseline for clinical alignment.

cs.AI↗
arXiv2026

aiXcoder-7B-v2: Training LLMs to Fully Utilize the Long Context in Repository-level Code Completion

Large Language Models (LLMs) have shown promising results in repository-level code completion, which completes code based on the in-file and cross-file context of a repository. The cross-file context typically contains different types of information (e.g., relevant APIs and similar code) and is lengthy. In this paper, we found that LLMs struggle to fully utilize the information in the cross-file context. We hypothesize that one of the root causes of the limitation is the misalignment between pre-training (i.e., relying on nearby context) and repo-level code completion (i.e., frequently attending to long-range cross-file context). To address the above misalignment, we propose Code Long-context Alignment - COLA, a purely data-driven approach to explicitly teach LLMs to focus on the cross-file context. Specifically, COLA constructs a large-scale repo-level code completion dataset - COLA-132K, where each sample contains the long cross-file context (up to 128K tokens) and requires generating context-aware code (i.e., cross-file API invocations and code spans similar to cross-file context). Through a two-stage training pipeline upon COLA-132K, LLMs learn the capability of finding relevant information in the cross-file context, thus aligning LLMs with repo-level code completion. We apply COLA to multiple popular LLMs (e.g., aiXcoder-7B) and extensive experiments on COLA-132K and a public benchmark - CrossCodeEval. Our experiments yield the following results. 1) Effectiveness. COLA substantially improves the performance of multiple LLMs in repo-level code completion. For example, it improves aiXcoder-7B by up to 19.7% in exact match. 2) Generalizability. The capability learned by COLA can generalize to new languages. 3) Enhanced Context Utilization Capability. We design two probing experiments, which show COLA improves the capability of LLMs in utilizing the information in cross-file context.

cs.SE↗
arXiv2026

Statistics-encoded tensor network approach in disordered quantum many-body spin chains

Simulating the dynamics of quantum many-body systems with disorder is a fundamental challenge. In this work, we propose a general approach -- the statistics-encoded tensor network (SeTN) -- to study such systems. By encoding disorder into an auxiliary layer and averaging separately, SeTN restores translational invariance, enabling a well-defined transfer matrix formulation. We derive a universal criterion, $n \gg α^2 t^2$, linking discretization $n$, disorder strength $α$, and evolution duration $t$. This sets the resolution required for faithful disorder averaging and shows that encoding is most efficient in the weak-disorder, typically chaotic regime. Applied to the disordered transverse-field Ising model, SeTN shows that the spectral form factor is governed by the leading transfer-matrix eigenvalue, in contrast to the kicked Ising model. SeTN thus provides a novel framework for probing the disorder-driven dynamical phenomena in many-body quantum systems.

quant-ph↗
arXiv2026

DECEPTICON: How Dark Patterns Manipulate Web Agents

Deceptive UI designs, widely instantiated across the web and commonly known as dark patterns, manipulate users into performing actions misaligned with their goals. In this paper, we show that dark patterns are highly effective in steering agent trajectories, posing a significant risk to agent robustness. To quantify this risk, we introduce DECEPTICON, an environment for testing individual dark patterns in isolation. DECEPTICON includes 700 web navigation tasks with dark patterns -- 600 generated tasks and 100 real-world tasks, designed to measure instruction-following success and dark pattern effectiveness. Across state-of-the-art agents, we find dark patterns successfully steer agent trajectories towards malicious outcomes in over 70% of tested generated and real-world tasks -- compared to a human average of 31%. Moreover, we find that dark pattern effectiveness correlates positively with model size and test-time reasoning, making larger, more capable models more susceptible. Leading countermeasures against adversarial attacks, including in-context prompting and guardrail models, fail to consistently reduce the success rate of dark pattern interventions. Our findings reveal dark patterns as a latent and unmitigated risk to web agents, highlighting the urgent need for robust defenses against manipulative designs.

cs.CR↗
arXiv2026

Do Transformers Have the Ability for Periodicity Generalization?

Large language models (LLMs) based on the Transformer have demonstrated strong performance across diverse tasks. However, current models still exhibit substantial limitations in out-of-distribution (OOD) generalization compared with humans. We investigate this gap through periodicity, one of the basic OOD scenarios. Periodicity captures invariance amid variation. Periodicity generalization represents a model's ability to extract periodic patterns from training data and generalize to OOD scenarios. We introduce a unified interpretation of periodicity from the perspective of abstract algebra and reasoning, including both single and composite periodicity, to explain why Transformers struggle to generalize periodicity. Then we construct Coper about composite periodicity, a controllable generative benchmark with two OOD settings, Hollow and Extrapolation. Experiments reveal that periodicity generalization in Transformers is limited, where models can memorize periodic data during training, but cannot generalize to unseen composite periodicity. We release the source code to support future research.

cs.LG↗
arXiv2026

CooperBench: Why Coding Agents Cannot be Your Teammates Yet

Resolving team conflicts requires not only task-specific competence, but also social intelligence to find common ground and build consensus. As AI agents increasingly collaborate on complex work, they must develop coordination capabilities to function as effective teammates. Yet we hypothesize that current agents lack these capabilities. To test this, we introduce CooperBench, a benchmark of over 600 collaborative coding tasks across 12 libraries in 4 programming languages. Each task assigns two agents different features that can be implemented independently but may conflict without proper coordination. Tasks are grounded in real open-source repositories with expert-written tests. Evaluating state-of-the-art coding agents, we observe the curse of coordination: agents achieve on average 30% lower success rates when working together compared to performing both tasks individually. This contrasts sharply with human teams, where adding teammates typically improves productivity. Our analysis reveals three key issues: (1) communication channels become jammed with vague, ill-timed, and inaccurate messages; (2) even with effective communication, agents deviate from their commitments; and (3) agents often hold incorrect expectations about others' plans and communication. Through large-scale simulation, we also observe rare but interesting emergent coordination behavior including role division, resource division, and negotiation. Our research presents a novel benchmark for collaborative coding and calls for a shift from pursuing individual agent capability to developing social intelligence.

cs.LG↗
arXiv2026

MagicGUI-RMS: A Multi-Agent Reward Model System for Self-Evolving GUI Agents via Automated Feedback Reflux

Graphical user interface (GUI) agents are rapidly progressing toward autonomous interaction and reliable task execution across diverse applications. However, two central challenges remain unresolved: automating the evaluation of agent trajectories and generating high-quality training data at scale to enable continual improvement. Existing approaches often depend on manual annotation or static rule-based verification, which restricts scalability and limits adaptability in dynamic environments. We present MagicGUI-RMS, a multi-agent reward model system that delivers adaptive trajectory evaluation, corrective feedback, and self-evolving learning capabilities. MagicGUI-RMS integrates a Domain-Specific Reward Model (DS-RM) with a General-Purpose Reward Model (GP-RM), enabling fine-grained action assessment and robust generalization across heterogeneous GUI tasks. To support reward learning at scale, we design a structured data construction pipeline that automatically produces balanced and diverse reward datasets, effectively reducing annotation costs while maintaining sample fidelity. During execution, the reward model system identifies erroneous actions, proposes refined alternatives, and continuously enhances agent behavior through an automated data-reflux mechanism. Extensive experiments demonstrate that MagicGUI-RMS yields substantial gains in task accuracy, behavioral robustness. These results establish MagicGUI-RMS as a principled and effective foundation for building self-improving GUI agents driven by reward-based adaptation.

cs.AI↗
arXiv2026

A Diff-Attention Aware State Space Fusion Model for Remote Sensing Classification

Multispectral (MS) and panchromatic (PAN) images describe the same land surface, so these images not only have their own advantages, but also have a lot of similar information. In order to separate these similar information and their respective advantages, reduce the feature redundancy in the fusion stage. This paper introduces a diff-attention aware state space fusion model (DAS2F-Model) for multimodal remote sensing image classification. Based on the selective state space model, a cross-modal diff-attention module (CMDA-Module) is designed to extract and separate the common features and their respective dominant features of MS and PAN images. Among this, space preserving visual mamba (SPVM) retains image spatial features and captures local features by optimizing visual mamba's input reasonably. Considering that features in the fusion stage will have large semantic differences after feature separation and simple fusion operations struggle to effectively integrate these significantly different features, an attention-aware linear fusion module (AALF-Module) is proposed. It performs pixel-wise linear fusion by calculating influence coefficients. This mechanism can fuse features with large semantic differences while keeping the feature size unchanged. Empirical evaluations indicate that the presented method achieves better results than alternative approaches. The relevant code can be found at:https://github.com/AVKSKVL/DAS-F-Model

cs.CV↗
arXiv2026

TEA: Temporal Adaptive Satellite Image Semantic Segmentation

Crop mapping based on satellite images time-series (SITS) holds substantial economic value in agricultural production settings, in which parcel segmentation is an essential step. Existing approaches have achieved notable advancements in SITS segmentation with predetermined sequence lengths. However, we found that these approaches overlooked the generalization capability of models across scenarios with varying temporal length, leading to markedly poor segmentation results in such cases. To address this issue, we propose TEA, a TEmporal Adaptive SITS semantic segmentation method to enhance the model's resilience under varying sequence lengths. We introduce a teacher model that encapsulates the global sequence knowledge to guide a student model with adaptive temporal input lengths. Specifically, teacher shapes the student's feature space via intermediate embedding, prototypes and soft label perspectives to realize knowledge transfer, while dynamically aggregating student model to mitigate knowledge forgetting. Finally, we introduce full-sequence reconstruction as an auxiliary task to further enhance the quality of representations across inputs of varying temporal lengths. Through extensive experiments, we demonstrate that our method brings remarkable improvements across inputs of different temporal lengths on common benchmarks. Our code will be publicly available.

cs.CV↗
arXiv2026

AIMS: An Adaptive Integration of Multi-Sensor Measurements for Quadrupedal Robot Localization

This paper addresses the problem of accurate localization for quadrupedal robots operating in narrow tunnel-like environments. Due to the long and homogeneous characteristics of such scenarios, LiDAR measurements often provide weak geometric constraints, making traditional sensor fusion methods susceptible to accumulated motion estimation errors. To address these challenges, we propose AIMS, an adaptive LiDAR-IMU-leg odometry fusion method for robust quadrupedal robot localization in degenerate environments. The proposed method is formulated within an error-state Kalman filtering framework, where LiDAR and leg odometry measurements are integrated with IMU-based state prediction, and measurement noise covariance matrices are adaptively adjusted based on online degeneracy-aware reliability assessment. Experimental results obtained in narrow corridor environments demonstrate that the proposed method improves localization accuracy and robustness compared with state-of-the-art approaches.

cs.RO↗
arXiv2025

Perplexity-Aware Data Scaling Law: Perplexity Landscapes Predict Performance for Continual Pre-training

Continual Pre-training (CPT) serves as a fundamental approach for adapting foundation models to domain-specific applications. Scaling laws for pre-training define a power-law relationship between dataset size and the test loss of an LLM. However, the marginal gains from simply increasing data for CPT diminish rapidly, yielding suboptimal data utilization and inefficient training. To address this challenge, we propose a novel perplexity-aware data scaling law to establish a predictive relationship between the perplexity landscape of domain-specific data and the test loss. Our approach leverages the perplexity derived from the pre-trained model on domain data as a proxy for estimating the knowledge gap, effectively quantifying the informational perplexity landscape of candidate training samples. By fitting this scaling law across diverse perplexity regimes, we enable adaptive selection of high-utility data subsets, prioritizing content that maximizes knowledge absorption while minimizing redundancy and noise. Extensive experiments demonstrate that our method consistently identifies near-optimal training subsets and achieves superior performance on both medical and general-domain benchmarks.

cs.LG↗
arXiv2025

Transductive Visual Programming: Evolving Tool Libraries from Experience for Spatial Reasoning

Spatial reasoning in 3D scenes requires precise geometric calculations that challenge vision-language models. Visual programming addresses this by decomposing problems into steps calling specialized tools, yet existing methods rely on either fixed toolsets or speculative tool induction before solving problems, resulting in suboptimal programs and poor utilization of induced tools. We present Transductive Visual Programming (TVP), a novel framework that builds new tools from its own experience rather than speculation. TVP first solves problems using basic tools while accumulating experiential solutions into an Example Library, then abstracts recurring patterns from these programs into reusable higher-level tools for an evolving Tool Library. This allows TVP to tackle new problems with increasingly powerful tools learned from experience. On Omni3D-Bench, TVP achieves state-of-the-art performance, outperforming GPT-4o by 22% and the previous best visual programming system by 11%. Our transductively learned tools are used 5x more frequently as core program dependency than inductively created ones, demonstrating more effective tool discovery and reuse. The evolved tools also show strong generalization to unseen spatial tasks, achieving superior performance on benchmarks from SpatialScore-Hard collection without any testset-specific modification. Our work establishes experience-driven transductive tool creation as a powerful paradigm for building self-evolving visual programming agents that effectively tackle challenging spatial reasoning tasks. We release our code at https://transductive-visualprogram.github.io/.

cs.CV↗
arXiv2025

Classification of atmospheric traveling waves at cloud level

We classify within the quasi-geostrophic framework all types of traveling waves in zonal bands of the planetary atmosphere at cloud level according to their wave speeds. This classification pertains to waves of all amplitudes, going beyond the small-amplitude perturbative regime. It provides a structurally robust criterion for determining which traveling-wave profiles are dynamically possible and we show that each wave classification type was observed on Jupiter or Saturn. Building on this classification, we also investigate the related rigidity issue for large-amplitude traveling waves and waves propagating near shear flows. Our study offers a unified quantitative characterization of the intrinsic constraints for traveling waves in the quasi-geostrophic regime of planetary atmospheric flow.

math.AP↗