Searcharxiv⌕ Search

arXiv subjects

Zihan Wang

Publications and source records attributed to Zihan Wang.

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

Humanity's Last Exam

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

cs.LG↗
arXiv2026

Evidence for Magnon-Assisted Exciton Radiation in a Layered Antiferromagnetic Semiconductor

Layered van der Waals (vdW) magnetic semiconductors open a new avenue for exploring intertwined excitonic and magnetic phenomena. Here, we investigate this interplay in the vdW antiferromagnet MnPS3, uncovering an exceptionally long exciton lifetime (~100 μs) below the Néel temperature (T_N). We demonstrate that the exciton lifetime is governed by phonon-assisted nonradiative recombination and thus exhibits a strong temperature dependence. In contrast, the exciton radiative recombination shows a distinct temperature dependence that is sensitive to magnetic order. Below T_N, the temperature dependence of the radiative recombination rate is consistent with a magnon-assisted emission pathway, while above T_N it reflects the combined effects of short-range spin correlations and phonons. These findings not only establish MnPS3 as a compelling candidate for excitonic devices due to its long-lifetime and correlation with magnetic orders but also provide crucial insights into the interplay between excitons, spins, and lattice in vdW magnetic semiconductors.

cond-mat.mtrl-sci↗
arXiv2026

LogiDroid: Individual Functional Test Generation via Business Logic Extraction and Adaptation

Functional testing is essential for verifying that the business logic of mobile applications aligns with user requirements. Despite its importance, functional testing remains heavily dependent on manual effort due to two core challenges. First, acquiring and reusing business logic from unstructured requirements remains difficult, which hinders the understanding of specific functionalities. Second, a significant semantic gap exists when adapting business logic to the diverse GUI environments, which hinders the generation of test cases for specific mobile applications. To address the preceding challenges, we propose LogiDroid, a two-stage approach that generates individual functional test cases by extracting business logic and adapting it to target applications. First, in the Knowledge Retrieval and Fusion stage, two LLM-based agents are employed to construct a functional test dataset, retrieve relevant test cases, and extract structured business logic for the target functionality. Second, in the Context-Aware Test Generation stage, two other LLM-based agents jointly analyze the extracted business logic and the real time GUI environment to incrementally generate context adaptive functional test cases. This design allows LogiDroid to accurately understand application semantics and use domain expertise to generate complete test cases with verification assertions. We assess the effectiveness of LogiDroid using two widely-used datasets that cover 28 real-world applications and 190 functional requirements. Experimental results show that LogiDroid successfully tested 40% of functional requirements on the FrUITeR dataset (an improvement of over 25% compared to the state-of-the-art approaches) and 65% on the Lin dataset (an improvement of over 55% compared to the state-of-the-art approaches). These results demonstrate the significant effectiveness of LogiDroid in functional test generation.

cs.SE↗
arXiv2026

MASTE: A Multi-Agent Pipeline for Zero-Shot Aspect Sentiment Triplet Extraction

Aspect Sentiment Triplet Extraction (ASTE) requires jointly identifying (aspect, opinion, sentiment) triples from a given review sentence. While large language models (LLMs) achieve strong zero-shot performance on many NLP benchmarks, their effectiveness on ASTE remains limited, as single-pass generation forces the model to determine span boundaries, opinion grouping, and sentiment polarity in a single decoding step. Common remedies, such as few-shot in-context learning and chain-of-thought prompting, offer only marginal improvements and rely heavily on either in-domain demonstrations sampled from labeled training data or carefully engineered reasoning prompts, neither of which is broadly available in zero-shot deployment. Inspired by the classical agent paradigm, we propose MASTE, a multi-agent pipeline for zero-shot Aspect Sentiment Triplet Extraction. MASTE decomposes ASTE into four sequential stages, where specialized agents handle different compositional subtasks with explicit conditioning on prior outputs. This design enables entirely training-free zero-shot ASTE and generalizes across different backbones and datasets. Extensive experiments on four ASTE benchmarks show that MASTE substantially outperforms zero-shot and chain-of-thought LLM baselines under the same backbone, narrowing the gap to fully supervised methods without using any labeled triplets. Code is available at https://github.com/Hankerlove/MASTE.

cs.CL↗
arXiv2026

Allo{SR}$^2$: Rectifying One-Step Super-Resolution to Stay Real via Allomorphic Generative Flows

Real-world image super-resolution (Real-SR) has been revolutionized by leveraging the powerful generative priors from Diffusion Models (DMs) and Flow Matching (FM). However, existing one-step methods typically replace Gaussian noise with degraded low-resolution (LR) latents at initialization, introducing a substantial distribution shift that further leads to trajectory deviation and prior collapse under extreme acceleration. To overcome these limitations, we propose Allo{SR}$^2$, a novel FM-based framework that rectifies one-step SR flows via allomorphic generative flows to maintain high-fidelity generative realism. Specifically, we utilize SNR-Guided Trajectory Initialization to identify a statistically aligned intermediate state along the pre-trained path to integrate LR representations into the generative flow. To ensure a stable, low-curvature path for one-step inference, we propose Flow-Anchored Trajectory Consistency (FATC), which explicitly regularizes the velocity field of the underlying probability flow. Furthermore, we develop Allomorphic Trajectory Matching (ATM), a self-adversarial distillation strategy that jointly models the SR flow and the generative flow within a unified velocity field, enabling one-step Real-SR while preserving the generative prior. Extensive experiments on both synthetic and real-world benchmarks demonstrate that Allo{SR}$^2$ achieves state-of-the-art performance in one-step Real-SR, offering a superior balance between fidelity and realism while maintaining extreme efficiency.

cs.CV↗
arXiv2026

Seeing and Reflecting: Multimodal Memory-Enhanced Agent Collaboration for Recommendation

Large language model (LLM)-based agentic recommender systems show promise in modeling user preferences through natural-language reasoning, yet they remain limited by text-centric inputs and coarse-grained memory updates, making agents prone to missing visual evidence, semantic noise, and preference drift. To address these limitations, we propose MMEACR, a Multimodal Memory-Enhanced Agent Collaboration framework for recommendation. MMEACR introduces a dual-track memory architecture that separates interpretable agent reasoning from fine-grained multimodal matching. In the reasoning track, collaborative User and Item Memory Agents maintain persistent multimodal memories and update them through an attribute-guided reinforcement-and-reflection mechanism. In the matching track, a decoupled multi-modal embedding memory is built from raw interaction narratives and item images to preserve detailed cross-modal signals beyond structured memory updates. The two tracks are integrated through weighted Reciprocal Rank Fusion to produce robust and interpretable rankings. Experiments on three real-world domains show that MMEACR achieves strong overall performance against competitive LLM-based and agent-based baselines, with notable gains in visually grounded recommendation scenarios.

cs.IR↗
arXiv2026

Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator

Embodied navigation aims to build agents that interpret multimodal goals, reason in 3D space, and reach target destinations reliably in the real world. However, progress remains constrained by the lack of scalable, high-fidelity, and physically grounded interactive environments. Although real-world scanned datasets offer visual realism, they are limited by scale. In contrast, synthetic simulators scale more easily but often exhibit large sim-to-real gaps. We introduce Image2Sim, a real-time neural simulation framework that constructs high-quality interactive environments from posed RGB-D image sequences. The central idea is to decouple 3D spatial anchoring from photorealistic observation synthesis. For scene construction, Image2Sim uses a feed-forward feature Gaussian model that lifts posed RGB-D observations into a 3D feature-Gaussian representation in a single pass. For rendering, we propose a Geometry-Aware One-Step Pixel Flow model that transforms sparse and noisy Gaussian projections into high-quality panoramic RGB-D observations. Image2Sim also serves as a fully automated embodied data engine that generates high-fidelity observations, executable actions, and diverse navigation instructions at scale. It converts large collections of videos and images into nearly 20K interactive scenes and synthesizes more than 10 million navigation training samples. Navigation models trained entirely in these neural environments achieve strong improvements on major benchmarks and transfer effectively to real-world zero-shot settings. These results suggest that scalable neural simulation can serve as a practical training substrate for embodied navigation at scale.

cs.CV↗
arXiv2026

Planning with the Views

Can VLMs predict how each camera move changes the view, and plan many such moves ahead? We call this capability view planning, requiring (1)understanding how a single action transforms the view, and (2)composing many such transformations across multi-turn plans to identify a target view. We probe both abilities in our proposed ViewSuite, a 3D point-cloud environment on real ScanNet scenes. Across 13 frontier VLMs, a critical planning gap emerges: they possess basic view-action knowledge but fail to compose it across multi-turn plans, with the gap widening as viewpoint distance grows. To close this gap, we propose an iterative framework that alternates self-exploration with view graph distillation. The key insight is that all exploration trajectories, regardless of their outcome, collectively form a view graph that compactly captures how viewpoints connect across a scene. Distilling this graph into diverse supervised tasks reshapes the policy distribution and overcomes the sparse rewards that stall pure RL. This improves Qwen2.5-VL-7B from 2.5% to 47.8% on interactive view planning, surpassing GPT-5.4 Pro (18.5%) and Gemini 3.1 Pro (21.4%). Self-exploration emerges as a promising path toward VLMs that can actively reason and plan in 3D space. Code and Data are at https://viewsuite.github.io.

cs.AI↗
arXiv2026

Generalised Probabilistic Theories

We give an introduction to research associated with the generalised probabilistic theories framework, also known as the convex framework. States are real vectors representing lists of probabilities of measurement outcomes. Convex combinations of the vectors represent probabilistic combinations of different state preparations. Transformations are real matrices. Measurement outcomes are represented by functionals of the states, inner products of the state with a real vector, whose values are the probability of the measurement outcome in question. The framework generalises quantum theory. We describe the operational meaning of the framework, and how the concepts can be defined in terms of cones of states and measurement outcome vectors. We describe how the classical and quantum probability theories are represented in the framework. We describe Bell non-locality and the theory with super-quantum non-locality known as box world. We discuss generalised Hamiltonian mechanics in the discrete case and in continuous phase space, including the role of negativity of the phase space density in contextuality and tunnelling.

quant-ph↗
arXiv2026

Stage-Transition Dense Reward Modeling for Reinforcement Learning

Reinforcement learning for long-horizon robotic manipulation is often limited by sparse and delayed rewards, while manually designing dense shaping signals is costly and brittle to changes in environments and object configurations. This work proposes Stage-Transition Dense Reward (STDR), a visual reward-learning framework that converts unstructured expert videos into logically grounded dense rewards for training RL agents from scratch. STDR leverages semantic understanding to infer a task's stage structure from demonstrations, and delivers two complementary learning signals during online training: (i) stage-transition feedback that provides goal-directed reward, and (ii) within-stage progress feedback that supplies fine-grained guidance toward completing each stage. Furthermore, an out-of-distribution (OOD) detection mechanism and a grasping regulation module are integrated to enhance robustness and prevent reward hacking. Experiments on 14 manipulation tasks across MetaWorld, ManiSkill, and Franka Kitchen show that STDR consistently improves sample efficiency and success rates over multiple baselines, and matches or surpasses handcrafted dense rewards on several challenging tasks. Real-robot evaluations further indicate that STDR assigns stable, progress-aligned rewards on successful executions while producing appropriately low rewards for failures, suggesting robustness to visual noise and better-calibrated reward assignment across settings.

cs.RO↗
arXiv2026

DAIN: Dynamic Agent-Based Interaction Network for Efficient and Collaborative Multimodal Reasoning

Current multimodal fusion approaches, particularly those based on static Mixture-of-Experts (MoE) architectures, often struggle to provide the adaptive and efficient collaborative reasoning required by complex real-world applications. We introduce the Dynamic Agent-based Interaction Network (DAIN), which reconceptualizes multimodal fusion as a dynamic, multi-agent collaborative process. DAIN employs a context-aware Meta-Controller that dynamically schedules sparse activation of specialized interaction agents and orchestrates compressed inter-agent communication for consensus-building. The framework is guided by a multi-objective loss function that jointly optimizes task accuracy, agent specialization, and operational efficiency through sparse activation and communication regularization. Comprehensive evaluations across five diverse benchmarks -- ADNI, MIMIC-IV, MM-IMDB, CMU-MOSI, and ENRICO -- establish DAIN as a new state-of-the-art, delivering significant performance improvements including a 2.6\% accuracy gain on ADNI. Ablation studies verify the critical roles of both dynamic scheduling and agent communication. Furthermore, DAIN offers enhanced interpretability by exposing context-dependent agent roles and collaboration patterns while maintaining computational efficiency through sample-wise sparse agent activation. Our work demonstrates the promise of dynamic, agent-based paradigms for multimodal reasoning.

cs.CL↗
arXiv2026

Interpretable Information-Decomposed Brain Graph Learning for fMRI-based Disease Diagnosis

Resting-state functional magnetic resonance imaging (rs-fMRI) has enabled non-invasive mapping of functional brain interactions for computer-aided diagnosis, yet most existing approaches reduce inter-regional relationships to correlation-based edge weights. Such representations capture co-fluctuation strength but obscure how information is shared across brain regions. Because brain disorders may disrupt not only connectivity strength but also the organization of redundancy, uniqueness and synergy, traditional functional connectivity may miss disease-relevant information structures. Here we introduce IID-GCN, an interpretable graph learning framework that decomposes rs-fMRI interactions into redundancy, uniqueness and synergy graphs using partial entropy decomposition. These information-specific graphs separately characterize shared, region-specific and jointly emergent components of brain activity. A multi-channel graph convolutional network then integrates the decomposed graphs through edge recalibration, cross-information interaction, ROI-attention readout and channel-attentive fusion. Across three datasets, IID-GCN consistently captures complementary diagnostic information beyond traditional functional connectivity. The learned information profiles reveal disorder-specific patterns of altered redundancy, uniqueness and synergy, suggesting that brain diseases reshape functional information organization rather than merely changing connection strength. These results establish information-decomposed brain graphs as an interpretable representation for rs-fMRI-based diagnosis. Our code is available at https://github.com/Zdy12/IID-GCN.

q-bio.NC↗
arXiv2026

A hidden reionization prior biases cosmological inference

Precision cosmology assumes that cosmic reionization was a single smooth transition. We show that this assumption is in tension with current observations: the Planck optical depth, patchy kinetic Sunyaev--Zel'dovich (PkSZ)limits from SPT and ACT, and the Ly$α$ forest endpoint cannot be simultaneously reproduced by any viable monotonic ionization history. Non-parametric reconstructions and Planck EE polarization provide independent support for an additional ionization component at $z \gtrsim 12$. Incorporating this early phase relaxes the upper bound on the summed neutrino mass to $\sum m_ν< 0.39$~eV ($95\%$ CL) and shifts $σ_8$ toward values preferred by weak-lensing surveys, both consequences of the standard $A_s$--$τ_e$ degeneracy. These shifts arise from relaxing a hidden prior on reionization shape rather than from new physics, and identify reionization shape as an implicit prior in cosmological inference. Next generation of CMB and 21~cm experiments will be able to test this directly,which will convert what has been an unflagged systematic on $σ_8$ and $\sum m_ν$ into a quantified statistical uncertainty.

astro-ph.CO↗
arXiv2026

ReportLogic: Evaluating Logical Quality in Deep Research Reports

Users increasingly rely on Large Language Models (LLMs) for Deep Research, using them to synthesize diverse sources into structured reports that support understanding and action. In this context, the practical reliability of such reports hinges on logical quality: whether the report's claims and arguments are explicitly supported and can be trusted as a basis for downstream use, rather than merely appearing fluent or informative. However, current evaluation frameworks largely overlook this requirement. To bridge this gap, we introduce ReportLogic, a benchmark that quantifies report-level logical quality through a reader-centric lens of auditability. Specifically, ReportLogic adopts a hierarchical taxonomy that evaluates whether readers can (1) trace an on-topic report structure with a unified analytical arc (Macro-Logic), (2) understand the progression with necessary context (Expositional-Logic), and (3) verify conclusions via explicit claim--support (Structural-Logic). Based on this taxonomy, we construct a human-annotated rubric-guided dataset and train an open-source LogicJudge for scalable evaluation. We further evaluate judge robustness via adversarial attacks, showing that off-the-shelf LLM judges are frequently influenced by superficial cues (e.g., verbosity), and reasoning modes can mask broken support relations. Overall, our results provide actionable guidance for building more robust logic evaluators and improving the logical reliability of LLM-generated reports.

cs.CL↗
arXiv2026

Unifying Search and Recommendation in LLMs via Gradient Multi-Subspace Tuning

Search and recommendation (S&R) are core to online platforms, addressing explicit intent through queries and modeling implicit intent from behaviors, respectively. Their complementary roles motivate a unified modeling paradigm. Early studies to unify S&R adopt shared encoders with task-specific heads, while recent efforts reframe item ranking in both S&R as conditional generation. The latter holds particular promise, enabling end-to-end optimization and leveraging the semantic understanding of LLMs. However, existing methods rely on full fine-tuning, which is computationally expensive and limits scalability. Parameter-efficient fine-tuning (PEFT) offers a more practical alternative but faces two critical challenges in unifying S&R: (1) gradient conflicts across tasks due to divergent optimization objectives, and (2) shifts in user intent understanding caused by overfitting to fine-tuning data, which distort general-domain knowledge and weaken LLM reasoning. To address the above issues, we propose Gradient Multi-Subspace Tuning (GEMS), a novel framework that unifies S&R with LLMs while alleviating gradient conflicts and preserving general-domain knowledge. GEMS introduces (1) \textbf{Multi-Subspace Decomposition}, which disentangles shared and task-specific optimization signals into complementary low-rank subspaces, thereby reducing destructive gradient interference, and (2) \textbf{Null-Space Projection}, which constrains parameter updates to a subspace orthogonal to the general-domain knowledge space, mitigating shifts in user intent understanding. Extensive experiments on benchmark datasets show that GEMS consistently outperforms the state-of-the-art baselines across both search and recommendation tasks, achieving superior effectiveness.

cs.IR↗
arXiv2026

A symmetric relaxation method for entire two-dimensional cellular networks and its implications

To simulate the relaxation of an entire 2D cellular network, this study proposes a symmetric relaxation method for both inner and marginal vertices. The relaxations of these two types of vertices are determined by the central angle symmetry of associated cells and the angle symmetry at each vertex, but with different major considerations. Trimmed Voronoi networks with varying irregularity are used as initial networks for the relaxation simulation. In particular, we propose a regular hexagon disordering method to generate Voronoi networks and find that the inner cells of networks with an irregularity value of one exhibit a conserved edge number distribution, as found in other 2D cellular networks. Simulation results agree with the von Neumann-Mullins law for both inner and marginal cells, and a modified equation including a geometric correction term significantly improves prediction quality. The Aboav-Weaire law and Lewis law are also reproduced, with the latter showing that relaxed cells tend to approach the ellipses' maximum inscribed polygons. Analysis of edge length, interior angle, and shape index reveals that symmetric relaxation inhibits T1 (neighbour exchange) topological transitions by reducing short edges while increasing area disparity among neighbouring cells. The findings suggest that T1 events may be triggered when force disequilibrium overcomes the stabilising effect of symmetric relaxation, providing a possible mechanistic explanation for T1 in 2D foams.

physics.bio-ph↗
arXiv2026

On the Adversarial Robustness of Multimodal LLM Judges

Multimodal Large Language Models (MLLMs) are increasingly used as automated judges, e.g., for image quality and safety assessment. However, their adversarial robustness remains largely unexplored, threatening the fairness and reliability of automated judging. To bridge this gap, we introduce RobustMLLMJudge, the first general framework for evaluating the adversarial robustness of general-purpose MLLMs when functioning as judges. It covers diverse attacks against popular judge approaches across quality and safety evaluation scenarios. Using RobustMLLMJudge, we reveal that i) different MLLM judges are highly vulnerable to score-inflating adversarial attacks; and ii) although effective, these attack methods face a critical challenge due to unique constraints in the evaluation protocols of MLLM judges. We further propose MGSIA, namely Manifold-Guided Semantic Induction Attack, a novel method that bypasses these constraints to enable more effective and transferable attacks on MLLM judges. The core idea of MGSIA is to combine affirmative semantic induction with high-score manifold alignment: it maximizes the probability that judges yield affirmative responses (e.g., "Yes") to binary semantic queries, while regularizing adversarial representations toward high-score centers estimated from proxy protocols. Together, these objectives yield transferable score-inflating perturbations. Extensive experiments demonstrate the superiority and generalizability of MGSIA in deceiving advanced MLLM judges under different evaluation scenarios, highlighting the need for robust MLLM judges. Code and data will be made available at https://github.com/mala-lab/RobustMLLMJudge.

cs.CV↗
arXiv2026

Factorized Latent Reasoning for LLM-based Recommendation

Large language models (LLMs) have recently been adopted for recommendation by framing user preference modeling as a language generation problem. However, existing latent reasoning approaches typically represent user intent with a single latent vector, which struggles to capture the inherently multi-faceted nature of user preferences. We propose Factorized Latent Reasoning (FLR), a novel framework for LLM-based sequential recommendation that decomposes latent reasoning into multiple disentangled preference factors. FLR introduces a lightweight multi-factor attention module that iteratively refines a latent thought representation, where each factor attends to distinct aspects of the user's interaction history. To encourage diversity and specialization, we design orthogonality, attention diversity, and sparsity regularization objectives, and dynamically aggregate factor contributions for the final prediction. We further integrate FLR with an efficient reinforcement learning strategy based on group-relative policy optimization, enabling stable alignment directly in the latent reasoning space. Experiments on multiple benchmarks show that FLR consistently outperforms strong baselines while improving robustness and interpretability.

cs.IR↗