arXiv subjects

Han Zhang

Publications and source records attributed to Han Zhang.

At least 73 records · Page 4Linked to original sources

arXiv2026

FitText: Evolving Agent Tool Ecologies via Memetic Retrieval

Efficient reasoning is not only a matter of shortening an answer trace; for tool-using agents, it also depends on whether the agent is reasoning over the right action space. As API ecosystems scale to tens of thousands of endpoints, the semantic gap between user requests and tool documentation makes this problem concrete: static retrieval from the initial query can fail before planning begins, and stronger planning alone cannot recover a missing tool. We study this problem as budgeted test-time retrieval and introduce FitText, a training-free framework that makes the tool interface revisable during execution by generating, refining, and evolving natural-language pseudo-tool descriptions as retrieval probes. FitText supports serial refinement, parallel exploration, and Memetic Retrieval, which adds evolutionary selection, local refinement, and tool memory to avoid redundant search. On StableToolBench (16,464 APIs), Memetic FitText reaches an 84.3% pooled pass rate, improving +26.7 points over static retrieval, +22.2 over Single-Pass, +23.2 over Re-Invoke, and +27.5 over Xu-style root refinement. It leads on every evaluated current model, with gains growing alongside model capability, and produces the largest improvements on ambiguous multi-tool tasks where dynamic re-retrieval restores correct candidates after early mistakes. At 40-way concurrency, parallel population execution keeps batched wall-clock at 1.01x Single-Pass despite the added search work.

cs.AI↗

arXiv2026

Similarity Is Not Logic: Factored Inference for Dual-Encoder Vision-Language Models

Dual-encoder vision-language models (VLMs) expose a similarity interface that enables zero-shot retrieval but fails compositional constraints: queries like "umbrella and no person" retrieve images containing both, even when concept detection is reliable. We trace this to an interface-level Bag-of-Concepts effect, where similarity scores approximate mean pooling of concept evidence regardless of operators. Although operator-dependent signals exist in text embeddings, they are too weak or misaligned to affect rankings. Fine-tuning does not reliably resolve this failure because the dominant bottleneck is how similarity aggregates evidence rather than what encoders represent. We propose factored inference, which separates evidence extraction from constraint execution, and introduce LCSE (Logic-Constrained Score Editing), a training-free method that executes constraints externally using concept scores from frozen encoders. We also introduce FACTOR-Bench, where LCSE achieves 85.5% accuracy versus 73.2% for the best fine-tuned baseline, 90.7% when applied to SigLIP 2, and improves NegBench COCO MCQ accuracy from 27.2% to 65.2% while preserving retrieval performance.

cs.CV↗

arXiv2026

Sources of Inequity and Fairness Risks in Wellbeing Sensing

Passive sensing for wellbeing uses smartphones and wearables to continuously collect human behavioral data and applies ML/AI models to infer psychological states and behaviors (e.g., depression, cognitive load). These systems are increasingly adopted in high-stakes settings (e.g., hospitals, universities), yet fairness research remains limited---primarily to post-hoc, identity-based comparisons of model performance. However, passive sensing combines heterogeneous sensing infrastructures, indirect behavioral inference, and longitudinal deployment---characteristics that, while not exclusive to the domain, are jointly pronounced here and raise two underexplored questions: (1) what additional sources of inequity arise from these characteristics, and (2) how do such inequities propagate beyond algorithmic audits across the system lifecycle? To address this gap, we conducted semi-structured interviews with 14 researchers and practitioners across five countries, examining how fairness risks emerge and are negotiated across the full passive sensing lifecycle. Our findings empirically characterize five situated sources of inequity (e.g., comfort with monitoring, behavioral regularity) that systematically shape fairness risks beyond identity-based attributes. We further synthesize 15 fairness risks and corresponding mitigation strategies across the lifecycle, from study design to deployment. Finally, we identify structural barriers that constrain fair practice in reality, and argue that enabling fair passive sensing requires both individual researcher efforts and ecosystem-level governance support from funders, publication venues, and deploying institutions.

cs.HC↗

arXiv2026

Neo-Classic: A Benchmark for Evaluating Linguistic-Aesthetic Reasoning in Classical Chinese Poetry

While Large Language Models (LLMs) achieve high accuracy on established Classical Chinese Poetry benchmarks, it remains challenging to distinguish transferable Linguistic-Aesthetic Reasoning from reliance on familiar pre-training patterns. To address this issue, we introduce Neo-Classic, an evaluation benchmark that combines a constructionist Out-of-Sample (OOS) dataset with a suite of reverse understanding probes. Unlike traditional benchmarks that rely on verification or generation over historical corpora, Neo-Classic comprises strictly metrical poetry authored by contemporary experts, reducing the possibility of direct retrieval. We evaluate state-of-the-art models, including Qwen3-Max, Gemini-3-Pro, and DeepSeek-V3.2, across five behavioral probes designed to test hierarchical constraint satisfaction. Our results reveal two primary limitations. First, a performance gap of 20 to 50 percent emerges when models transition from historical to contemporary texts. Second, models exhibit substantial difficulties in discourse-level ordering tasks, with standard accuracy remaining low (0 to 13 percent). Although expert-level guidance improves the performance of reasoning-enhanced models to 36 percent, a notable gap with human experts persists. These findings suggest that while current LLMs capture local formal patterns, they struggle with global hierarchical planning required for robust Linguistic-Aesthetic Reasoning.

cs.CL↗

arXiv2026

Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models

While Mixture-of-Experts (MoE) scales capacity via conditional computation, Transformers lack a native primitive for knowledge lookup, forcing them to inefficiently simulate retrieval through computation. To address this, we introduce conditional memory as a complementary sparsity axis, instantiated via Engram, a module that modernizes classic $N$-gram embedding for O(1) lookup. By formulating the Sparsity Allocation problem, we uncover a U-shaped scaling law that optimizes the trade-off between neural computation (MoE) and static memory (Engram). Guided by this law, we scale Engram to 27B parameters, achieving superior performance over a strictly iso-parameter and iso-FLOPs MoE baseline. Most notably, while the memory module is expected to aid knowledge retrieval (e.g., MMLU +3.4; CMMLU +4.0), we observe even larger gains in general reasoning (e.g., BBH +5.0; ARC-Challenge +3.7) and code/math domains~(HumanEval +3.0; MATH +2.4). Mechanistic analyses reveal that Engram relieves the backbone's early layers from static reconstruction, effectively deepening the network for complex reasoning. Furthermore, by delegating local dependencies to lookups, it frees up attention capacity for global context, substantially boosting long-context retrieval (e.g., Multi-Query NIAH: 84.2 to 97.0). Finally, Engram establishes infrastructure-aware efficiency: its deterministic addressing enables runtime prefetching from host memory, incurring negligible overhead. We envision conditional memory as an indispensable modeling primitive for next-generation sparse models.

cs.CL↗

arXiv2026

MAVEN: A Multi-stage Agentic Annotation Pipeline for Video Reasoning Tasks

Training Vision Language Models (VLMs) for video event reasoning requires high-quality structured annotations capturing not only what happened, but when, where, why, and with what consequence, at a scale manual labelling cannot support. We present MAVEN (Multi-stage Agentic Video Event aNnotation), a multi-stage agentic pipeline that turns raw videos into multi-task training data with Chain-of-Thought (CoT) reasoning traces, organized around a designated Event of Focus. At its core, MAVEN synthesizes a Multi-Scale Spatio-Temporal Event Description (MSTED) from three complementary caption levels; this explicit intermediate serves as the sole input to downstream Q&A generation across multiple task formats. Crucially, MAVEN supports agent-driven domain adaptation: given a new video dataset and target question examples, the agent redesigns all prompts top-down without manual re-engineering. A hierarchical refinement loop further classifies annotation errors against a taxonomy, traces root causes to the originating pipeline stage, and applies targeted edits that rewrite prompts or modify the pipeline structure itself, iteratively improving data quality. We apply MAVEN to label over 5,300 traffic videos and fine-tune Cosmos-Reason2-8B on the resulting data. On a private CCTV evaluation set, fine-tuning surpasses both Gemini 2.5 Pro and 3.1 Flash, including a $+38.8$-point gain in MCQ accuracy over zero-shot. On AccidentBench, CCTV-only training lifts Cosmos-Reason2 by $+10.7$ MCQ points and matches Gemini 2.5 Pro despite seeing no dashcam videos; adding agent-adapted dashcam annotations narrows the gap to Gemini 3.1 Flash, and RL post-training pushes overall performance past both Gemini baselines. Qualitative results on warehouse surveillance and public safety videos further show the agentic workflow readily adapts the pipeline to new domains.

cs.CV↗

arXiv2026

From Foundation to Application: Improving VLA Models in Practice

Despite recent progress of VLA foundation models, the disparity between laboratory conditions and real-world applications continues to impede their practical implementation. To bridge this gap, we present LingBot-VLA 2.0, which advances LingBot-VLA through improvements in three functional domains. (1) Generalization across tasks and embodiments. Compared to the previous version, we revamp the data processing pipeline and curate around 60,000 hours of data for pretraining, including 50,000 hours of robot trajectories spanning 20 robot configurations and 10,000 hours of egocentric human videos. (2) Expanded action space in addition to dual-arm hardware platforms. In particular, our system accommodates degrees of freedom for the heads, waists, mobile bases, and dexterous hands, thereby empowering the robots to tackle more complex tasks in practical scenarios. (3) Predictive dynamics modeling for improved temporal reasoning. Specifically, we formulate future prediction as a proxy task, facilitated by a video representation model for semantic priors and a depth estimation model for geometric cues. Evaluations on the GM-100 benchmark, conducted in a generalist setting, validate the beneficial impact of these proposed modifications. Furthermore, benefiting from the expanded pretraining data that covers whole-body degrees of freedom, LingBot-VLA-2.0 demonstrates strong cross-embodiment long-horizon mobile manipulation capability across the two robotic platforms.

cs.RO↗

arXiv2026

ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning

Agentic reinforcement learning (ARL) has rapidly gained attention as a promising paradigm for training agents to solve complex, multi-step interactive tasks. Despite encouraging early results, ARL remains highly unstable, often leading to training collapse. This instability limits scalability to larger environments and longer interaction horizons, and constrains systematic exploration of algorithmic design choices. In this paper, we first propose ARLArena, a stable training recipe and systematic analysis framework that examines training stability in a controlled and reproducible setting. ARLArena first constructs a clean and standardized testbed. Then, we decompose policy gradient into four core design dimensions and assess the performance and stability of each dimension. Through this fine-grained analysis, we distill a unified perspective on ARL and propose SAMPO, a stable agentic policy optimization method designed to mitigate the dominant sources of instability in ARL. Empirically, SAMPO achieves consistently stable training and strong performance across diverse agentic tasks. Overall, this study provides a unifying policy gradient perspective for ARL and offers practical guidance for building stable and reproducible LLM-based agent training pipelines.

cs.AI↗

arXiv2026

Downlink Performance Analysis of Pinching Antenna Systems: WDMA or NOMA?

This paper presents an analytical framework for downlink pinching antenna systems (PASS) employing waveguide division multiple access (WDMA) and non-orthogonal multiple access (NOMA). A unified channel model is developed to capture antenna deployment, user spatial distribution, and path loss. Closed-form and single-integral expressions for the outage probability and average achievable rate are derived and validated via Monte Carlo simulations. The results show that NOMA achieves higher spectral efficiency at high transmit signal-to-noise ratio (SNR) due to successive interference cancellation (SIC), whereas WDMA offers more reliable performance at low to moderate SNR but suffers from an outage floor and rate saturation at high SNR. Moreover, WDMA performance is more sensitive to the user spatial distribution due to the spatially dependent inter-waveguide interference. These findings provide design insights for access-scheme selection and antenna placement in PASS.

eess.SP↗

arXiv2026

Registering the 4D Millimeter Wave Radar Point Clouds Via Generalized Method of Moments

4D millimeter wave radars (4D radars) are new emerging sensors that provide point clouds of objects with both position and radial velocity measurements. Compared to LiDARs, they are more affordable and reliable sensors for robots' perception under extreme weather conditions. On the other hand, point cloud registration is an essential perception module that provides robot's pose feedback information in applications such as Simultaneous Localization and Mapping (SLAM). Nevertheless, the 4D radar point clouds are sparse and noisy compared to those of LiDAR, and hence we shall confront great challenges in registering the radar point clouds. To address this issue, we propose a point cloud registration framework for 4D radars based on Generalized Method of Moments. The method does not require explicit point-to-point correspondences between the source and target point clouds, which is difficult to compute for sparse 4D radar point clouds. Moreover, we show the consistency of the proposed method. Experiments on both synthetic and real-world datasets show that our approach achieves higher accuracy and robustness than benchmarks, and the accuracy is even comparable to LiDAR-based frameworks.

cs.RO↗

arXiv2026

Exascale Hybrid Numerical-AI Ensembles for Operational Flood-Season Forecasting in East Asia: 15-km Decadal Hindcasts and 1-km High-Resolution Capability

Seasonal forecasting of summer rainfall in East Asia remains a grand challenge, as predictability at 3 to 6 month lead times is constrained by the spring predictability barrier, weak large-scale signals, and localized nonlinear convective extremes. We address this challenge with CAPES, which integrates a kilometer-resolution coupled regional model with atmosphere, land, and ocean components and a data-driven AI seasonal forecasting system. At 15 km resolution, the fused workflow combines 174 numerical members from varying start times, physics schemes, and parameter perturbations with 1,600 AI members generated from initial and physical perturbations. Using the full LineShine system, CAPES completes ten annual 1,774-member hindcasts for 2016 to 2025 within 14.6 hours, improving the mean prediction score from ECMWF's 71.8 to 75.9 and delivering a major gain in operational forecasting capability. The 1-km configuration further enables fine-scale typhoon simulation and establishes the feasibility of kilometer-scale fused ensemble forecasting on a one-week timescale.

cs.CE↗

arXiv2026

Circular Quasiconformal Deturbulence: Geometry-Based Restoration from Multiple Turbulent Frames

Imaging through inhomogeneous media often results in severe distortions, posing significant challenges to downstream image-processing tasks. The lack of clean paired images makes supervised learning impractical, motivating unsupervised restoration approaches. In this work, we propose the Circular Quasi-Conformal Deturbulence (CQCD) framework, an unsupervised approach that reconstructs distortion-free images from multiple frames using a circular architecture. The framework minimizes reconstruction errors by jointly estimating forward and backward transformations between distorted observations and the restored image. A key advancement of CQCD is the integration of computational quasi-conformal geometry, which encourages bijective non-rigid deformations and improves the well-posedness of both forward and inverse mappings for cycle consistency. The deformation field is further regularized to preserve structural coherence and reduce non-physical artifacts such as folding or tearing. Additionally, tight-frame blocks are employed to effectively encode distortion-sensitive features, enhancing the precision of the restoration process. To assess the effectiveness of the proposed framework, extensive evaluations are conducted on synthetic and real-world image datasets. Experimental findings indicate that CQCD not only surpasses existing state-of-the-art deturbulence techniques in restoration quality but also achieves highly accurate deformation field estimation.

cs.CV↗

arXiv2026

SelPE: Progressive Selection for Private Structured Text Synthesis

Many data-driven applications rely on structured textual records, such as clinical triage notes and financial transaction logs, for downstream learning and decision-making. In privacy-sensitive domains, access to such records is strictly regulated, often resulting in only a small number of available private examples for model development and analysis. Yet existing differential privacy data synthesis methods fall short: tabular techniques cannot faithfully model free-form text, while text-based approaches often break structural constraints. We propose SelPE, a selection-guided progressive evolution framework for small-sample private structured text synthesis. Rather than relying on noisy aggregation or private model training, SelPE concentrates privacy budget on a sequence of multi-batch top-1 selections, enabling efficient guidance under tight privacy constraints. To support faithful and valid synthesis, SelPE decouples semantic abstraction from schema realization via a two-stage generation pipeline, and evaluates candidates using a multi-channel distance kernel that jointly models textual, categorical, and numeric fields in their native representations. A non-private contrastive expansion mechanism further promotes diversity without incurring additional privacy cost. Extensive Experiments demonstrate that SelPE consistently improves structural validity, fidelity, and downstream utility under strict differential privacy budgets, particularly in low-data regimes.

cs.CR↗

arXiv2026

On Zeckendorf-Niven numbers and arithmetic progressions

A positive integer is Zeckendorf-Niven (respectively, Lucas-Niven) if it is divisible by the number of summands in its Zeckendorf decomposition (respectively, Lucas decomposition). We show that there exist infinitely many Zeckendorf-Niven numbers and Lucas-Niven numbers in every arithmetic progression. Furthermore, we provide bounds on the maximum number of consecutive Zeckendorf-Niven terms in certain arithmetic progressions.

math.NT↗

arXiv2026

Robust probabilistic measurement of structural-functional module consistency in infant brain development

Brain network is commonly divided into modules for analyzing their functionally segregated roles for group-level analysis in neuroimaging studies. Here, we introduce stochastic modules within brain networks for a robust probabilistic measurement of structural-functional module consistency (SFMC) in a group of subjects. Specifically, a stochastic module can be regarded as the chance of a brain region across subjects potentially being assigned to a group-level sub-network, characterized as an assignment probability for this brain region. This novel method has two advantages for evaluating inhomogeneous modules in brain networks. The first is that it can robustly evaluate the consistency between brain structural and functional modules whose population sizes are not necessary the same, and the second is that it is able to take into account the inter-individual variability of the modules for the groups. Moreover, compared with the conventional structural-functional coupling approach, our stochastic module-based method reveals a more pronounced decline in the coupling between structure and function, indicating stronger developmental reorganization. Our results using the dataset from Baby Connectome Project (BCP) show that the SFMC decreases from 0 to 5 years old, and is greater in primary brain regions, such as visual areas, while lower in more advanced cognitive regions, including those related to attention, control, and default mode network.

q-bio.NC↗

arXiv2026

Reinforcing Dual-Path Reasoning in Spatial Vision Language Models

Spatial VLMs have made substantial progress in geometric perception, yet complex spatial reasoning requiring multi-step inference over depth, distance, and scene relations remains challenging. Moreover, different spatial queries call for fundamentally different strategies: some are best addressed through purely linguistic, step-by-step deduction, while others require explicit 3D grounding before quantitative inference. We present Dual-Path Spatial Reasoning via Reinforcement Learning for Spatial VLMs (SR-REAL), a unified framework that equips a spatial VLM with two complementary reasoning paths: Language-Only Reasoning (LOR), which performs step-by-step linguistic deduction, and Detect-Then-Reason (DTR), which detects 3D geometric cues (e.g., centers or bounding boxes) via region tokens before explicit geometric inference. SR-REAL begins with a cold-start supervised fine-tuning stage that constructs LOR and DTR chain-of-thought supervision and exposes a region-to-3D interface, followed by RL that optimizes the policy model with accuracy and format rewards; for DTR, a discrete center-based detection reward further refines geometric alignment. Across diverse spatial benchmarks, SR-REAL significantly outperforms spatial VLM baselines: (i) a single RL-trained model supports both reasoning paths, with DTR excelling in region-aware tasks through precise 3D localization and LOR enhancing general spatial reasoning; (ii) jointly training both paths fosters mutual reinforcement; (iii) high-quality, blended cold-start data is crucial for stable RL optimization; and (iv) the model generalizes across datasets and domains without per-task tuning, demonstrating positive transfer between LOR and DTR.

cs.CV↗

arXiv2026

Training LLMs with Reinforcement Learning over Digital Twin Representations for Reasoning-Intensive Surgical VideoQA

Surgical video question answering requires multi-step reasoning across semantic, spatial, and temporal dimensions. Existing methods architecturally compress videos into discrete token representations and couple visual perception with reasoning. This approach fragments continuous spatial-temporal relationships and has been shown to restrict multi-step reasoning capabilities. We introduce a reinforcement learning (RL) framework that trains large language models (LLMs) to decouple perception from reasoning by operating over digital twin representations constructed from surgical foundation models. Additionally, we introduce hierarchical representations across frame, temporal window, and procedure levels with probabilistic uncertainty estimates. Finally, we propose a novel reward that combines format validation with accuracy assessment through clinical plausibility evaluation and uncertainty-aware calibration for training. To demonstrate the capabilities of this approach, we introduce REAL-Colon-Reason, a colonoscopic benchmark with 2000 question-answer pairs across three complexity levels. We achieve state-of-the-art performance on REAL-Colon-Reason and two existing surgical VideoQA benchmarks REAL-Colon-VQA and EndoVis18-VQA.

cs.CV↗

arXiv2026

Dual-Transformer Aided Hierarchical Deep Reinforcement Learning for Robust RIS-Assisted Near-Field Communications

The deployment of extremely large aperture arrays (ELAAs) in sixth-generation (6G) networks is expected to shift communications into the near-field regime, where spherical-wave propagation enables distance-aware beamfocusing but remains highly vulnerable to physical blockages that cause non-line-of-sight (NLoS) conditions. To resolve this inherent vulnerability, reconfigurable intelligent surfaces (RIS) can be utilized to circumvent these blockages and effectively establish reliable NLoS communication links. In envisioned deployment scenarios, accurately acquiring instantaneous CSI and predicting sudden blockages is profoundly challenging due to the prohibitive pilot overhead associated with massive passive arrays and the unpredictable mobility of environmental scatterers. To address this, we propose the Dual-Transformer Hierarchical Deep Reinforcement Learning (DT-HDRL) framework, which integrates two specialized transformer models with a two-timescale hierarchical control agent. The first transformer integrates a ray-tracing digital twin prior with distance-aware geometric correction features to yield rapid and precise CSI estimates, while a complementary vision transformer (ViT) processes sequential camera frames to forecast impending blockages prior to link degradation. These predictive outputs are then fed directly into the hierarchical control agent. Within this architecture, a high-level controller processes the slow-timescale blockage predictions to jointly dictate the user transmission path (line-of-sight (LoS) or RIS-assisted NLoS) and RIS active/sleep scheduling, whereas a low-level controller employs the fast-timescale CSI estimates to perform joint base station (BS) beamfocusing and RIS phase-shift optimization.

eess.SP↗