arXiv subjects

Wei Hu

Publications and source records attributed to Wei Hu.

At least 19 recordsLinked to original sources

arXiv2026

SWE-Test: Benchmarking LLM Vulnerability Discovery via Input Prediction

Vulnerability discovery is becoming an important ability of large language model (LLM) agents: agents that silently miss real defects leave critical software exposed. Rigorously measuring this ability is therefore urgent, but existing benchmarks are gameable through data contamination, score recall against an unknowable vulnerability set, often rely on synthetic bugs, and report a single end-to-end verdict that cannot localize where an agent fails. Vulnerability discovery is a composite ability: an agent must comprehend source code, infer input constraints, construct inputs, execute them, and iteratively correct from feedback. We recast its measurement as an input-prediction task with a closed, deterministic ground truth: using coverage-guided fuzzing, we mine deep target branches in real-world C/C++ programs and ask an agent to predict an input that drives execution to a given branch. This decomposes discovery into three task modes over 22 real-world C/C++ programs spanning 15 domains. Open-loop and Feedback-enabled share 60 fixed-target task instances across 16 of these codebases (13 domains), testing input construction without and with a distance oracle to isolate code comprehension from feedback-driven correction. Online Arena instead removes the predefined target and scores path exploration by coverage gain on a separate, partially overlapping pool of 11 programs; agents collectively confirmed 13 distinct bugs across six programs. Evaluating 15 default-effort model-scaffold configurations, the best reaches only 55.0% pass rate in the Feedback-enabled mode, and the mean across seven paired Claude Code configurations is 36.4% with feedback versus 19.3% without. Decomposing failures, we find constraint inference, not navigation, is the dominant bottleneck. We release SWE-Test with a turnkey evaluation environment.

cs.SE↗

arXiv2026

Polished but Unresolved: Identifying Late-Stage Pressure States in Long-Horizon Tool-Use Agents

Long-horizon tool-use agents need not only to search and plan, but also to decide when to finalize. We study late-stage pressure states, in which an agent is biased toward submitting a final answer that appears complete and polished while key constraints remain unresolved. We first train a linear probe to show that this pressure state is identifiable from the agent's hidden states. Then, we use activation interventions along this pressure direction and find that shifting the hidden states changes both the pressure score and whether the agent continues tool use or submits early. Through controlled context manipulations, we further see that the pressure is mitigated by constraint clarity and action mapping. Based on these findings, we propose Probe-Sensed Pressure Relief (PSPR), a plugin that applies lightweight pressure relief direction under moderate pressure and moves to structured organization under high pressure risk. Experiments on multiple long-horizon benchmarks show that our method consistently strengthens existing agent methods.

cs.AI↗

arXiv2026

Dense Process Supervision for Search Agents via Fact Utility Estimation

Reinforcement learning (RL) for search agents typically relies on outcome rewards. However, it often fails to achieve effective credit assignment, due to the unclear value of intermediate steps. It is hard to separate their contributions from the final result. In this paper, we propose a dense process supervision method based on fact utility estimation, which models the reasoning process as the accumulation of discrete evidence facts. We first extract structured facts from raw observations and organize them into an explicit fact store. To support credit assignment, we then cluster semantically equivalent facts and infer the posterior utility of each fact cluster using Bayesian estimation over group rollouts. Finally, we convert the estimated fact utilities into dense step-level rewards to guide RL training. Experiments on seven single-hop and multi-hop QA benchmarks show that our method consistently outperforms existing baselines. Ablation studies validate clear relative improvements on multi-hop QA compared to outcome reward-only training.

cs.CL↗

arXiv2026

How Do Prompt Variations Affect Energy Consumption in On-Device LLMs?

Large language models (LLMs) are increasingly deployed on mobile devices, making energy efficiency a key deployment constraint, yet the energy impact of prompt design remains underexplored. This paper aims to understand how two prompt properties, cognitive load and phrasing pattern, shape the energy behavior of on-device LLM inference. We conduct a broad empirical study covering prompt properties, datasets, models, and devices, with phase-level profiling that separates prefill and decode energy. We find that cognitive load primarily affects the energy cost per token, while phrasing pattern affects energy largely through token usage. Our energy-quality analysis further shows that prompt design reshapes the attainable frontier differently across models, highlighting the need for model-aware prompt design in energy-efficient on-device LLM inference. Code, datasets, and scripts are available at https://amai-gsu.github.io/PromptProperty/.

cs.CL↗

arXiv2026

Toward Effective and Reliable LLM Agents via Dynamic Ontology

Large language model (LLM) agents rely heavily on knowledge encoded in model parameters or presented as unstructured context. In domain-specific tasks, this leaves important semantic connections implicit. This often results in incomplete evidence use and brittle multi-step decisions. Ontologies offer a way to externalize domain concepts and relations as machine-interpretable structures, but constructing task-usable ontologies traditionally requires substantial effort from domain experts and is difficult to scale. Automatic construction is also challenging: an ontology that appears semantically plausible may not contain the relational structures needed for actual decision making. We present OaK, an ontology-as-a-kernel framework that dynamically constructs and refines task-oriented ontologies for LLM agents. Given task requirements and training data, OaK constructs an ontology and its knowledge graph, generates task-adaptation functions for graph reasoning, and uses judge feedback to iteratively refine both. By making relevant concepts and relations explicit, the ontology grounds knowledge retrieval and multi-step decision making. We evaluate OaK on TravelPlanner, CRMArenaPro, and ToolQA. Results show that OaK improves standard LLM agents, strengthens evidence grounding, and boosts the reliability of multi-step reasoning.

cs.AI↗

arXiv2026

Rickard's question on standard derived equivalences

In 1991 Rickard asked whether every derived equivalence of algebras over a common field is standard. We construct a series of examples of non-standard derived equivalences, and thus answer the question negatively. Moreover, we give a necessary and sufficient condition for the question to be true.

math.RT↗

arXiv2026

ReGround: Restoring Visual Grounding in Multi-Step Reasoning through Self-Diagnosis and Visual Re-Examination

Vision-Language Models (VLMs) often lose visual grounding during multi-step reasoning: as reasoning chains grow longer, later inference steps rely increasingly on language priors rather than image evidence. We identify a consistent benchmark-level signature associated with this degradation: across 2,510 re-examined samples from four benchmarks, attention entropy over image tokens typically decreases during Round 1 and rises again after image re-injection. However, we find that effective visual re-examination requires two complementary ingredients: image re-injection and targeted self-diagnosis. Without targeted diagnosis, re-examination can even hurt performance, whereas accurate self-diagnosis yields substantial gains -- a swing of several points on key benchmarks, indicating that diagnostic quality is a key factor in whether re-examination helps or hurts in our setting. We present ReGround, a two-stage framework that teaches VLMs to self-diagnose grounding failures and selectively re-examine visual evidence, without architectural modifications or external tools. Through capability bootstrapping, a stronger variant from the same model family provides diagnostic scaffolding only during data construction, while the policy model learns to diagnose autonomously at inference time and retains most of the assisted gains. Experiments on eight benchmarks across two VLM backbones demonstrate consistent gains, especially on visually intensive multi-step reasoning tasks, while incurring only modest inference overhead relative to tool-augmented baselines. Project page: https://sespoir.github.io/reground-page/ . Code: https://github.com/sespoir/ReGround .

cs.CV↗

arXiv2026

Motion Beyond Morphology: Bootstrapping Cross-Category Motion Transfer from Abstract Motion Representations

Video motion transfer aims to animate a target object using dynamics from a reference video. Existing formulations largely rely on fixed structural correspondence, which becomes ill-defined when reference and target objects differ substantially in morphology, articulation, or deformation mechanisms. We introduce Motion Beyond Morphology, a perspective that seeks to transfer motion beyond fixed structural correspondence, by preserving dynamics that remain meaningful across different target morphologies. To realize this, we propose a two-stage framework. Stage~I learns complementary multi-granularity abstract motion views and uses them to bootstrap cross-category video pairs that preserve transferable dynamics across diverse morphologies. Stage~II internalizes this supervision into direct reference-video-conditioned generation, removing the need for explicit motion extraction at inference. We further introduce OpenVMT-Dataset and OpenVMT-Bench for training and evaluating image- and text-conditioned motion transfer across Same, Near, and Far category gaps. Extensive experiments demonstrate state-of-the-art motion fidelity and target preservation. Project page: https://miniz233.github.io/MotionBeyondMorphology/

cs.CV↗

arXiv2026

Higher cluster tilting objects in locally finite triangulated categories

We study higher cluster tilting objects through covering functors from derived categories of hereditary algebras. The covering formalism reduces the existence problem to the equivariant problem of finding \(G\)-stable \(d\)-cluster tilting objects in \(d\)-cluster categories. For triangulated categories with finitely many indecomposable objects this gives a complete ADE existence criterion and explicit counting formulas. We also prove that, in finite Frobenius models, the endomorphism algebras of all \(d\)-cluster tilting objects are derived equivalent. Applications are given to Cohen--Macaulay finite categories, finite noncommutative crepant resolutions, and rigidity dimensions of representation-finite self-injective algebras.

math.RT↗

arXiv2026

SonoCLIP: Mask-Guided Region-Aware Vision-Language Pretraining for Fetal Ultrasound Analysis

Vision-language foundation models have shown strong potential in medical image analysis. Although foundation models for ultrasound imaging have recently emerged, the domain remains particularly challenging due to severe speckle noise, acquisition variability, and subtle anatomical boundaries, leading to high inter-observer variability. Existing CLIP-based models rely primarily on global image-text alignment, limiting their sensitivity to clinically decisive local structures. We propose SonoCLIP, the first million-scale region-controllable fetal ultrasound vision-language foundation model that integrates segmentation masks as mask-channel visual prompts within the vision encoder, enabling joint global-local contrastive representation learning. To support scalable region-text alignment, we introduce a sigmoid-based pairwise contrastive loss that improves stability under large-scale supervision. We further curate a 1.44M-image multimodal fetal ultrasound dataset spanning 24 standard planes for large-scale pretraining. Extensive cross-center evaluations demonstrate that SonoCLIP achieves superior zero-shot transfer performance under both global and mask-guided inference, establishing a controllable and clinically oriented foundation model for fetal ultrasound analysis. Our code and data are available at https://github.com/Harrison-one/SonoCLIP.

cs.CV↗

arXiv2026

Erase-then-Delta Attention: Decoupling Erase and Write Addresses in Delta-Rule Linear Attention

Delta-rule linear attention improves recurrent memory updates by correcting what is already stored at the current write address before writing new content. However, the active correction is still anchored to that same write address. As a result, stale information stored at a different address cannot be actively removed before new content is written elsewhere. We propose Erase-then-Delta Attention (EDA), a memory update rule that decouples where to erase from where to write. The key insight is that recurrent memory models should not only correct the current write, but also selectively suppress outdated memory at an independently chosen address. Concretely, our method first applies a targeted erase step along a learned erase direction, and then performs the standard delta-style corrective write along the current write direction. This preserves the corrective behavior of delta-rule updates while expanding their memory-management capacity. Language-model pretraining experiments across dense 2.5B and MoE 25B-A2.8B model families show that EDA performs best in both settings. The gain persists after 80B-token long-context midtraining of the MoE models, where EDA also performs best in long-context evaluations from 4k to 128k contexts. A compact update analysis and memory-state probes suggest why: EDA keeps the delta-rule corrective write intact while allocating an additional cleanup path most strongly when passive decay is weak. These results suggest that recurrent memory models should decide not only what to write, but also what stale information to erase and where.

cs.CL↗

arXiv2026

EIBench: A Simulator-Based Benchmark and Turn-Credit RL for Emotion Management

Emotional intelligence (EI) in Large Language Models (LLMs) is often evaluated through static understanding tasks or single-response dialogue generation. However, emotion management is interactive: a good model should not only recognize a user's emotion, but also improve the user's emotional and relational state over several turns. We introduce EIBench, a simulator-based benchmark for interactive emotion management. EIBench contains 2,222 scenarios, with 2,009 for training and 213 for held-out testing. The scenarios are organized by a 2x2 taxonomy covering Support, Defense, Repair, and Charm, which together capture different forms of support, boundary maintenance, trust repair, and rapport building. In each scenario, an LLM simulator plays the user, updates an emotion-relation state after each turn, and maps the final state to an anchor-based score. This design makes EIBench both an evaluation benchmark and a training environment: the final state gives the outcome reward, while the per-turn state updates provide dense feedback for RL. We evaluate 15 open- and closed-source LLMs. Current models perform well on support and rapport-building scenes, but struggle with boundary maintenance under user pressure. To improve the EI ability of LLMs, we propose Centered Turn-Credit GRPO (CTC-GRPO), a GRPO extension that reuses the simulator's per-turn state updates as dense turn-level feedback while preserving the final outcome reward. CTC-GRPO improves Qwen3-8B from -22.4 to +22.4 on EIBench and also improves on out-of-distribution evaluations including SAGE (+12.4) and EQBench3 (+20.9%). Our results show that simulator-tracked user states can support both evaluation and training for multi-turn emotion management.

cs.CL↗

arXiv2026

Intelligent Multimodal Retrieval and Reasoning for Geospatial Knowledge Discovery on the I-GUIDE Platform

Geospatial knowledge discovery increasingly requires search across heterogeneous artifacts: datasets, maps, notebooks, software, publications, and the provenance links among them. Conventional geoportals support metadata and spatial filtering, but they rarely provide semantic retrieval, graph-aware provenance traversal, and conversational synthesis in one integrated system. This paper presents I-GUIDE Smart Search, a production multimodal geospatial retrieval-augmented generation (RAG) system embedded in the I-GUIDE Platform, and reports on its design, deployment, and evaluation. The system combines production-maintained OpenSearch keyword, vector, and spatial indexes with a Neo4j knowledge graph and an iterative RAG pipeline for memory-aware query augmentation, reasoning, retrieval-method routing, relevance grading, grounded generation, hallucination and relevance checking. In a single-A100 RAG deployment, I-GUIDE Smart Search supports interactive use up to about 100 concurrent simulated users, reaching 4.4 requests per second with p50 latency near 25 seconds despite 20-50 LLM calls per query. For answer quality, we evaluate a four-category benchmark of 170 unique human-filtered user-facing queries, together with ten intent-specific probe sets generated from the deployed indexes and graph. Smart Search improves retrieved evidence coverage and judged answer quality over non-retrieval and naive-RAG baselines, with the clearest gains on exact-identifier, spatially constrained, simple-recommendation, and domain-specific factual queries requiring current indexed evidence. We distill transferable deployment lessons for spatial RAG systems, covering spatial metadata quality, graph provenance, retrieval routing, interface contracts, refusal-aware evaluation, latency-cost tradeoffs, and the role of the user interface in deployed geospatial cyberinfrastructure.

cs.IR↗

arXiv2026

Decoding Semantic Categories from Picture-Naming EEG

Picture naming requires the transformation of visual object information into a spoken lexical response through perceptual, semantic, lexical, and articulatory processes. This study asked whether semantic-category information is recoverable from high-density EEG during overt picture naming. Sixteen native French-speaking participants performed a picture-naming task using line drawings. Picture labels were embedded with a multilingual text-embedding model and organized into nine interpretable semantic categories, providing a data-driven semantic target space for neural decoding. EEG activity was represented channel-wise using a pre-trained single-channel EEG encoder over an early post-stimulus window, a later naming-related window, and their combination. Nine-class decoding showed above-chance semantic-category discrimination in all temporal representations. Balanced accuracy increased from 0.562 in the early window to 0.610 in the naming-related window, and reached 0.781 when both windows were combined, with a maximum Macro-F1 of 0.784. Class-level F1 scores showed consistent gains across semantic categories, and sensor-level decoding maps indicated spatially distributed category information. These findings suggest that semantic-category structure is reflected in EEG activity during overt picture naming and that early and naming-related temporal windows provide complementary information. The results support the use of modern neural decoding methods as tools for investigating lexical-semantic processing in spoken language production.

q-bio.NC↗

arXiv2026

Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws

Classical scaling laws for language model pretraining balance model size against training dataset size under a fixed compute budget, assuming abundant data and a single pass over the corpus. As training compute grows faster than the supply of natural language data, pretraining is likely to enter a data-constrained, compute-rich regime where models train for multiple epochs over a finite dataset. We study data-constrained pretraining along two axes, regularization and scaling. For regularization, we study masked-input regularization (MIR), an auxiliary next-token prediction loss on randomly masked inputs. MIR tests whether the random masking central to diffusion language models can benefit autoregressive pretraining without architectural changes or inference overhead. Across 72M to 1.4B parameter models, we find that MIR added on top of strong weight decay improves validation loss over autoregressive strong-weight-decay-only models, with downstream gains at 1.4B. For scaling, we propose SoftQ, a scaling law that couples model size and data size to capture their interaction under repeated data. Classical alternatives such as the Chinchilla law use an additive form that decouples these terms, making them misspecified in the data-constrained regime. We find that SoftQ fits data-constrained experiments substantially better than these alternatives, and estimates MIR's gains as equivalent to roughly 1.3 times as much unique training data. We release our code at https://github.com/yixinw-lab/dc_pretrain.

cs.LG↗

arXiv2026

Harnessing Structural Context for Entity Alignment Foundation Models

Entity alignment (EA) aims to identify equivalent entities across heterogeneous knowledge graphs (KGs) and is a key component of knowledge fusion and cross-KG reasoning. The recent EA foundation model demonstrates that alignment knowledge, once pretrained, can be directly applied to diverse previously unseen KG pairs. However, it still underuses structural context in two places: cross-KG interaction is weak during encoding, and final candidate ranking still relies too heavily on coarse similarity. We address these limitations with ContextEA, an enhanced encoder-decoder framework for transferable EA. On the encoder side, we introduce a cross-KG interaction encoder that unifies the two KGs with anchor bridges and performs earlier relation-aware cross-graph propagation. On the decoder side, we introduce a structural calibration decoder that calibrates alignment scores with entity-level, neighborhood-level, relation-level, and anchor-aware structural evidence. This design strengthens both structural context construction and structural context exploitation while remaining lightweight. Experiments on 29 EA datasets in OpenEA, SRPRS, and DBP show consistent gains over strong transferable baselines. Notably, the pretrained ContextEA already surpasses the finetuned baselines on all three benchmark groups, demonstrating substantially stronger transfer to unseen KGs. These results suggest that explicitly harnessing structural context is an effective direction for improving EA foundation models.

cs.CL↗

arXiv2026

BUP-TR: Bayesian Underdetermined Projection Trust-Region Methods for Derivative-Free Optimization

Underdetermined quadratic interpolation is a central model-construction tool in model-based derivative-free trust-region methods: it limits sampling costs but leaves an affine family of interpolating quadratics. Classical solvers select one element of this family by prescribing a fixed norm or model-change measure, such as the least-Frobenius-change Hessian update in Powell-type methods. We introduce BUP-TR (Bayesian Underdetermined Projection Trust-Region), which instead completes the model by projecting a prior quadratic onto the affine interpolation set in the precision norm supplied by the prior. The same precision matrix defines a spectral geometry certificate, MAP-poisedness, and a repair mechanism for interpolation sets. Under standard smoothness assumptions, uniform precision bounds, MAP-poisedness, and a trust-region-scale prior-accuracy condition, the hard-MAP models are fully linear. Consequently, BUP-TR attains global first-order convergence and O(epsilon^{-2}) evaluation complexity, with geometry-repair evaluations included. A NEWUOA-style implementation, BUP-NEWUOA, improves fixed-budget performance on the reported benchmark suite at moderate and stringent accuracy targets while retaining the computational structure of a Powell-type trust-region method.

math.OC↗

arXiv2026

MATRO: Metric-Aware Trust-Region Optimization with Fully Quadratic Models

Model-based derivative-free trust-region methods build local interpolation models and restrict trial steps to regions where those models are reliable. This paper studies the shape of that region. When an objective is poorly scaled or locally anisotropic, a Euclidean ball can be governed by the steepest local direction and can restrict progress along directions of slow variation. We propose MATRO (Metric-Aware Trust-Region Optimization), a fully quadratic interpolation framework in which the trust region is the ellipsoid s^T M_k s <= Delta_k^2. For any positive definite metric M_k, the induced variable y = M_k^{1/2}s converts the ellipsoidal subproblem into a standard Euclidean trust-region subproblem, so model decrease, ratio tests, radius updates, poisedness, and fully quadratic error bounds can be stated in induced coordinates under a uniform metric contract. The metric is selected from the interpolation Hessian: positive definite quadratics yield a unique volume-normalized curvature metric that isotropizes the induced Hessian and gives a truncated Newton step, while indefinite fitted Hessians motivate an absolute-curvature metric that balances curvature magnitudes without changing curvature signs. Under the standard fully quadratic assumptions and the metric contract, MATRO retains the first-order evaluation-complexity order O(n^2 epsilon^{-2}). Experiments on More-Wild benchmarks, controlled anisotropy tests, and two-dimensional trajectories show that curvature-shaped regions are most effective when the interpolation Hessian captures stable local anisotropy, while dense linear algebra is most visible at loose accuracies or on inexpensive analytic tests.

math.OC↗