Search arXiv⌕ Search

arXiv subjects

Xin Liu

Publications and source records attributed to Xin Liu.

At least 37 records · Page 2Linked to original sources
arXiv2026

Harnessing VLM in Embodied RL: From Reward Regressor to Success Detector via Exploration-Augmented Key-Observation Gating

VLMs provide promising solutions to embodied reward design. Most current methods that train VLMs as reward regressors not only require extensive annotations but are also primarily designed for finetuning pretrained policies, and perform less well in the more challenging from-scratch RL setting. This issue stems from a fundamental mismatch: regressing precise rewards is misaligned with what VLMs naturally excel at. We posit that it is more aligned with VLMs' inherent multi-modal VQA capabilities for the models to function as success detectors, and propose VLM success reward via Exploration-Augmented Key-Observation Gating (VLMeg) to well harness VLM for embodied reward design. As the core of VLMeg, the proposed exploration-augmented key-observation gating first determines whether the current RL episode is judgment-worthy based on its similarity to a generated success trajectory. Then, the key observation with the lowest confusion from the judgment-worthy episode will be selected for VLM evaluation as episodic reward. This gating mechanism not only avoids redundant calls but also enables higher detection accuracy of VLM. Considering the sparsity of judgment-worthy episodes, the advanced exploration incentives based on the generated trajectory are introduced, enabling more efficient exploration toward judgment-worthy trajectories for more VLM reward triggers and thus gaining higher RL efficiency. Extensive experiments in MetaWorld demonstrate VLMeg achieves state-of-the-art from-scratch RL, outperforming advanced VLM-based baselines on 12 challenging reward-free tasks. Furthermore, VLMeg also enables effective RL without any ground-truth reward in 3 challenging real-world experiments. Code will be open after acceptance at https://github.com/Raina915/VLMeg.

cs.LG↗
arXiv2026

CowCare: A Unified Visual Framework for Viewpoint-Robust Cattle Tracking and Spatiotemporal Heat-Stress Analysis

Reliable individual monitoring is essential for dairy cattle welfare and early heat stress detection in Precision Livestock Farming (PLF). However, translating raw multi-camera videos into automated diagnostics remains challenging due to disconnected tracking and recognition pipelines and complex physical environments. Dense mutual occlusions and cross-camera viewpoint variations fragment identities, while environmental noise, including camera vibrations, and limited agricultural data obscure subtle heat stress cues such as rapid breathing and flank undulations. To bridge low-level visual features and high-level behavioral diagnostics, we propose CowCare, a unified framework combining viewpoint-robust tracking with spatiotemporal heat stress recognition. Its tracking module, CowTrack, maintains cross-camera identity consistency through a View-Decoupled Aggregation Network (VDA-Net). Instead of relying on static global representations, VDA-Net aggregates localized textures under heavy occlusion and uses a Viewpoint Decoupling Transformer (VDT) to suppress transient viewpoint variations. CowTrack further incorporates a causal online self-supervision mechanism for continuous adaptation to dynamic farm scenes. The resulting tracklets are subsequently processed by TSDF-Net, a fine-grained video understanding network. Its Temporal Stabilization and Aggregation Module (TSAM) suppresses high-frequency jitter, while a Diffusion Semantic Prior (DSP) uses frozen pre-trained diffusion semantics to enhance subtle behavioral signals and mitigate overfitting. Extensive evaluations on two real-world farm benchmarks demonstrate the framework's effectiveness: CowTrack achieves 83.48% HOTA and 87.72% IDF1 for global multi-object tracking on MVCow, and TSDF-Net achieves 87.72% Top-1 accuracy on HSCow. These results establish CowCare as a robust end-to-end framework for automated livestock health monitoring.

cs.CV↗
arXiv2026

A Unified Benchmark for Dynamic Medical Treatment Reinforcement Learning

Medical treatment recommendation poses several challenges to reinforcement learning (RL): patient physiology evolves in continuous time, measurements and interventions are performed at irregular intervals, and treatment effects vary substantially across individuals. Existing RL formulations and simulated environments, however, are based on discrete-time MDPs with fixed decision intervals. Thus, it remains difficult to evaluate whether RL methods can handle time-interval-dependent disease progression, personalized treatment response, and safety between consecutive measurement points. To address this gap, we introduce MedGym, a benchmark environment for dynamic treatment recommendation. MedGym models longitudinal patient evolution in a continuous-time framework and constructs a configurable medical RL benchmark from clinical data by using Physics-Informed Neural Networks. The resulting benchmark enables direct comparison between discrete-time and continuous-time methods under irregular treatment timing and patient-specific dynamics. Furthermore, MedGym supports evaluation from clinically important perspectives, such as personalization and trajectory-level safety. By providing a standardized and configurable benchmark for continuous-time dynamic treatment, MedGym enables more realistic and informative evaluation of medical RL methods.

cs.LG↗
arXiv2026

STAPLE: Static Sparse Upcycling of Pretrained Checkpoints via Embedding Modules

Sparse computation can scale the parametric capacity of LLMs without proportionally increasing active computation, but pretraining sparse architectures remains expensive and data-hungry. Existing sparse upcycling methods mainly target Mixture-of-Experts models, where expert homogeneity, sample inefficiency, and system overheads can limit practical scalability. Static sparse memory offers predictable access and token-specific updates that can help address these challenges, but naively integrating it into pretrained computation yields poor returns. In this work, we introduce STAPLE, a recipe built upon STEM for integrating static sparse memory into pretrained models. Whereas STEM replaces FFN up-projections with token-indexed embeddings, STAPLE retains them and adds the embeddings as a residual pathway, preserving pretrained computation while learning complementary token-specific information in a scalable manner. On OLMo-2 1B, STAPLE improves over continued dense training by 4.27 points on HumanEval, 2.80 on MBPP, 2.27 on GSM8K, and 1.65 on DROP. Average performance increases with static capacity, and improvements persist on the larger OLMo-2 7B backbone. STAPLE also outperforms static sparse baselines such as STEM and an Engram-style model, and achieves a higher average score than an MoE upcycling baseline while decoding 2x faster with less than half its GPU model memory and measured GPU energy per token in our throughput-oriented decoding evaluation. Importantly, despite introducing billions of parameters late in training, our host-offloaded implementation keeps generation throughput within 1.5% and training throughput within 3% of the dense baseline, while caching only the ~5% most frequent embeddings on GPU. These results show that static sparse memory can provide a sample-efficient, scalable, and sustainable path for expanding pretrained models without sparse pretraining from scratch.

cs.LG↗
arXiv2026

DeepDISC on LSST Data Preview 1: Scene-Level Detection, Deblending, and Star/Galaxy Classification in the ECDFS and EDFS Fields

We present the first application of DeepDISC, a scene-level deep-learning framework, to real LSST data, using the Data Preview 1 (DP1) coadded images of the Extended Chandra Deep Field-South and the Euclid Deep Field-South fields. In a single pass, DeepDISC jointly detects, segments, and classifies sources as stars or galaxies, rather than treating these tasks as separate steps as in traditional pipelines. We warm-start from a model pretrained on LSST simulations and fine-tune it on DP1 six-band images, using the DP1 catalog's extendedness flag (point-like versus extended) as provisional class labels. The classification part of the network is then fine-tuned alone on a curated sample of stars and galaxies with spectroscopic, Gaia, and photometric identifications, so the final output is a star/galaxy classification. Fine-tuning on real DP1 observations substantially improves source recovery relative to the un-fine-tuned model, raising galaxy and star completeness from 40.2% and 13% to 84.2% and 49.1% under strict class-aware matching; under class-agnostic matching the fine-tuned model reaches 84.9% and 74.3%. On a common sample of sources that neither classifier saw in training, DeepDISC classifies to within about one percentage point of a published photometry-based Random Forest for the same field (97.4% vs. 98.5%); at fainter magnitudes DeepDISC labels more sources as stars than the Random Forest, and independent labels favor the Random Forest in most of these disagreements. Cross-matching to deeper HST/CANDELS imaging on a held-out footprint gives an unrecognized-blend fraction for DeepDISC of 14.9% (95% CI 12.5-17.8%), consistent with or modestly below the 18.0% we measure for the production DP1 catalog. This work establishes a baseline for DeepDISC on real LSST data and a foundation for self-supervised pretraining and for LSST Data Preview 2.

astro-ph.IM↗
arXiv2026

TraceQuery: learning what to recompute from full-denoising trajectory in diffusion language models

Diffusion large language models (dLLMs) enable flexible generation through iterative masked denoising, but incur substantial inference cost because token representations are repeatedly recomputed across denoising steps. Existing cache-based methods reduce this cost by reusing intermediate results across denoising steps and recomputing only a subset of positions at each step, chosen either on a fixed schedule or by heuristic signals of change. Our trajectory experiments instead point to decoding trajectory as the more relevant quantity. Cached generation remains accurate when it stays close to the full-denoising trajectory, with errors concentrated around unmask-order deviations and reduced as longer reference prefixes are recovered. This suggests a more direct principle: limiting representation error helps, but keeping generation on the full-denoising trajectory is a more direct route to accurate decoding. We therefore formulate efficient dLLM inference as trajectory-guided querying. Rather than broadly repairing representation error, we use full-denoising trajectories to supervise which unresolved positions should receive limited fresh computation. Based on this formulation, we introduce TRACEQUERY, a lightweight sparse decoding framework that predicts a small set of near-future trajectory-relevant positions, queries only those candidates, and lets the backbone select which one to commit from their fresh predictions. TRACEQUERY delivers the strongest quality-computation trade-off among cache-based approaches, sustaining high accuracy with substantially fewer fresh token computations, and generalizes across tasks, model variants, and families. Code: https://github.com/fish1928/tracequery.git

cs.CL↗
arXiv2026

TimelyDAgger: Timing-Aware Expert Querying for VLA Policy Improvement

DAgger improves robot policies by aggregating expert supervision from states visited during policy execution. Robot-gated DAgger automates expert queries, allowing the robot to decide when to request expert takeover. While existing gates emphasize detecting the need for assistance, takeover timing also shapes the content of these demonstrations and their value for policy learning. We propose TimelyDAgger, combining Bridge-PCA monitoring of internal vision-language-action (VLA) features with Feedback-guided Threshold Adaptation based on expert behavior to improve takeover timing. We introduce an evaluation framework linking failure detection, takeover timing, and policy improvement, including Target-Aligned Supervision Ratio (TASR) for assessing supervision quality without retraining. Experiments show that takeover timing affects policy learning, with TimelyDAgger achieving competitive failure detection and higher post-training success in most evaluated settings under matched expert-action budgets. Project website: https://seen-e.github.io/TimelyDagger/.

cs.RO↗
arXiv2026

ForestQuery: Boundary-Aware and Spatially Anchored Query Learning for Unified Forest Point Cloud Segmentation

Forest point cloud segmentation is fundamental for fine-grained 3D forest scene understanding, yet remains challenging due to irregular tree structures, severe occlusions, density variations, and ambiguous instance boundaries. Recent query-based forest segmentation methods have shown promise for unified semantic and instance prediction, but they still insufficiently exploit forest-specific spatial structure and account for boundary uncertainty. In this paper, we propose ForestQuery, a boundary-aware and spatially anchored query learning framework for unified forest point cloud segmentation. ForestQuery enhances instance and semantic query learning through two complementary designs. Specifically, boundary uncertainty is explicitly modeled to guide reliable instance query construction and modulate query optimization through adaptive loss reweighting. Meanwhile, spatially anchored semantic query enhancement (SA-SQE) introduces learnable 3D anchors encoding forest vertical stratification priors to enrich semantic queries with explicit spatial references. We evaluate ForestQuery on multiple public forest point cloud benchmarks and a self-collected annotated real-world dataset. Extensive experiments demonstrate consistent improvements in both individual-tree segmentation and semantic segmentation across diverse forest scenes. Code and data are publicly available at https://zhan994.github.io/ForestQuery

cs.CV↗
arXiv2026

Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents

Large language models (LLMs) rely on long-horizon tool invocation sequences for complex tasks, where each invocation can alter the task state and condition subsequent decisions. In long-horizon tool use, final-outcome rewards provide weak credit assignment over long interaction traces. Step-level rewards can offer more targeted feedback, but obtaining reliable step supervision often requires human or LLM judgment, or additional rollouts to estimate the downstream effect of an intermediate decision. In this paper, we argue that effective tool-use agents should estimate the long-horizon value of a possible next tool invocation before executing it. This objective requires comparative supervision over alternative invocations under the same context, while logged trajectories only contain the invocation that was actually taken. Therefore, we propose Comparative Inference for Tool-use Agents (CITA). CITA trains a Comparative Inference Model (CIM) from paired signals that combine observed tool behavior, scalable supervision from a Bayesian tool-graph simulator, and semantic judgments from LLM-based comparison. The resulting CIM learns to estimate how likely a possible next tool invocation is to support final task success under the current context. Across three tool-use benchmarks and multiple backbone LLMs, CITA consistently improves Tool F1 and task success. Additional analysis shows that CIM learns accurate step-level value estimates for comparative tool choices.

cs.AI↗
arXiv2026

Training LLM Judges from Language Feedback via Position-Selective Self-Distillation

We study training LLM judges from natural language feedback, especially for subjective tasks where the verdict depends strongly on which evaluation criteria the judge invokes and how it weighs them. The dominant approach, outcome-supervised RL (e.g., GRPO), credits every token in the rollout with a single scalar determined only by the accuracy of the final verdict, providing no separate credit at the criterion-choice tokens and ignoring the rich language feedback (e.g., preference rationales) that naturally accompanies preference labels. Self-Distillation (SD) is one natural way to use this language feedback: the same model, conditioned on this feedback, acts as a teacher providing dense, position-level supervision. However, not all positions carry equally useful signal. Using the per-position entropy shift between teacher and student, we identify two regimes: context sharpening, where the teacher concentrates probability on a particular feedback-aligned criterion expression, and context spreading, where the teacher distributes probability across multiple feedback-aligned alternatives. We interpret these patterns as follows: sharpening encourages memorization of a particular criterion expression, whereas spreading promotes semantic understanding by preserving these alternatives. Motivated by this asymmetry, we introduce position masking based on the entropy shift that retains the lower tail of the entropy-shift distribution. Experiments show that masking higher-entropy-shift positions improves out-of-distribution generalization over naive SD. The resulting self-distilled judges outperform judges trained with outcome-supervised RL by 2-9 percentage points on the evaluated subjective subcategories, while remaining competitive on objective ones.

cs.CL↗
arXiv2026

Beyond State-as-Action: Exploiting Command-State Discrepancy for Robot Imitation Learning

Constructing action targets from measured robot motion is an established approach in imitation learning. Under interaction constraints, however, command-state discrepancy may reflect control demands that motion alone does not capture. We investigate when this information matters and how to exploit it. Across three real-robot tasks, task and phase analyses reveal larger supervision gaps under constrained interaction, while selective command retention provides evidence of locally useful command information. Building on these findings, we propose Command-State Discrepancy Weighting (CSDW), which accounts for robot response times and combines subsequent progress, persistent unmet demand, and demand changes into continuous weights for command supervision. The method requires no task-phase annotations or changes to policy architecture or inference. CSDW improves over uniform command supervision on constrained tasks, while methods perform similarly in the less constrained task. Project page: https://seen-e.github.io/CSDW/.

cs.RO↗
arXiv2026

Switching Anomalous Hall and Nernst Responses by Nonmagnetic N Occupation at Fixed Noncoplanar Mn Antiferromagnetic Order

Nonmagnetic atomic occupation can control anomalous transverse transport by modifying magnetic symmetry without changing the underlying magnetic order. We demonstrate this effect using controlled $γ$-Mn, Mn$_4$N, and MnN reference states with the same lattice constant and identical noncoplanar all-in-all-out Mn magnetic configurations, while varying only the occupation of the N sublattice. The anomalous Hall and anomalous Nernst responses exhibit a pronounced zero--finite--zero evolution across the series despite the unchanged Mn spin order. In $γ$-Mn and MnN, the high magnetic symmetry enforces complete cancellation of the Brillouin-zone-integrated Berry curvature. In Mn$_4$N, N occupation lowers the magnetic symmetry while preserving inversion and breaking the relevant twofold rotational symmetries, thereby lifting the cancellation constraint and permitting an uncompensated Berry-curvature contribution along the [111] direction. The resulting finite anomalous Hall conductivity reaches $-126$~S/cm near the Fermi level. These results establish nonmagnetic sublattice occupation as a symmetry-control parameter for Berry-curvature-driven transport in compensated antiferromagnets, independent of changes in the magnetic order.

cond-mat.mtrl-sci↗
arXiv2026

Randomized Branch Methods with Inexact Subproblems for Bouligand Stationarity in Linear and Quadratic Programs with Complementarity Constraints

We develop randomized branch methods for finding Bouligand-stationary (B-stationary) points of linear and quadratic programs with complementarity constraints (LPCCs and QPCCs) using only linear programming subproblems. The methods exploit the finite-union geometry of the feasible set and search for first-order descent on randomly selected compatible branches. For LPCCs, an exact method optimizes over sampled branches, while an inexact simplex method can accept an improving branch-feasible vertex before solving the sampled penalty problem to optimality. For QPCCs, including problems with indefinite quadratic objectives, branch quadratic programs are replaced by linearized trust-region subproblems; a ratio test ensures actual decrease, and thresholded sampling detects branches that emerge only at accumulation points. Under the stated assumptions, the LPCC methods stabilize after finitely many changes at B-stationary points almost surely. For the QPCC method, finite termination yields a B-stationary point, and every accumulation point of an infinite run is B-stationary almost surely. Experiments on bilevel-induced instances, instances arising from inverse quadratic programming, and sparse affine generalized Nash equilibrium instances, together with 129 MacMPEC embedding tests, show that the methods return points with competitive objective quality and runtimes on large-scale complementarity systems.

math.OC↗
arXiv2026

Do Emotion Concepts Generalize Across Sources, Modalities, and Architectures in Vision-Language Models?

Recent studies suggest that large language models encode emotion concepts as structured internal representations, but most existing work focuses on text and a single architecture. Therefore, we ask, do emotion concepts generalize across sources, modalities, and architectures in vision--language models (VLMs)? To address this, we construct CMES (Cross-Modal Emotion Stimuli), a multi-source collection of emotion-conditioned stories, real facial expressions, synthetic portraits, and synthetic emotion-evoking scenes. For each stimulus source, we extract a separate set of six Ekman emotion vectors from each of three VLMs. We report four main findings as follows: 1) Image-derived emotion vectors form a low-dimensional geometry similar to that of text-derived vectors. Valence is relatively stable across sources, while arousal varies more. 2) Text- and image-derived emotion vectors have modest cosine similarity but still show held-out cross-modal correspondence. Text-derived vectors can also steer image interpretation. 3) Cross-architecture correspondence remains even when native cosine is near zero. Transformations estimated from generic ImageNet activations recover both correspondence and causal transfer without using the six emotion vectors or their labels. 4) After aligning representations across architectures, we construct a shared emotion subspace that preserves affective geometry and selective steering effects. The corresponding consensus emotion vectors also generalize to a held-out fourth architecture at two model sizes. These results suggest that emotion representations can share relational structure and causal effects across sources, modalities, and architectures, even when individual vector directions differ.

cs.CV↗
arXiv2026

Convergent representations of elastic-wave structure emerge across deliberately distinct seismic training routes

Elastic-wave propagation connects laboratory fracture with earthquakes, raising the prospect of seismic models that reuse waveform structure across physical scales. Here we show that the phase picker PNSN and the multi-task model SeismicXM organize unseen laboratory acoustic emissions around common waveform relations. Their representations agree on 512 held-out records relative to matched random networks. Four earthquake-model families also track displaced laboratory arrivals without fitting target neural weights. Receiver-function classification is accessible to frozen pretrained features, although standardized random features remain competitive. Noise correlations test the extension to continuous propagation outputs: a nonlinear SeismicXM ensemble improves on random topology, while a period-wise median remains more accurate. Waveform interventions and synthetic controls reveal that the observed agreement depends on feature calibration and the waveforms being compared. These results establish convergent waveform organization across distinct seismic training routes and show that cross-scale arrival transfer is already accessible to compact models.

physics.geo-ph↗
arXiv2026

Does High Signal-to-Noise Ratio Identify Better Training Data for Seismic Deep Learning?

Hard signal-to-noise ratio (SNR) screening removes weak seismic records, but does retaining clearer waveforms improve learning enough to compensate for the discarded data? We test this question in phase picking and ambient-noise dispersion, first matching training counts and then restoring the full eligible dispersion pool at fixed compute. Phase-picking screening gave no consistent unfiltered-test benefit under fine-tuning or random initialization, including a control matched on distance and phase composition. Dispersion screening increased error at the original five-epoch endpoint but improved on an equally sized unscreened set after longer training. Retaining all 29,788 eligible fit paths instead of 9,929 randomly sampled paths reduced unfiltered-test error by 3.0% across three seeds at the same update budget. Full-available and strict-SNR training then had similar mean errors, 0.0405 and 0.0406 km s$^{-1}$, with mixed seed-paired ordering; strict training retained a 9.9% advantage on the SNR-selected test. In two-day monitoring, SNR regulated pick streams but event recovery depended on its measurement definition. High SNR did not reliably identify better training data: gains depended on the task, training schedule, and test domain, and fixed-count comparisons omitted the benefit of retaining additional valid examples.

physics.geo-ph↗
arXiv2026

Bond-selective modulation of scalar spin chirality in strained Mn4N

Scalar spin chirality (SSC) underlies a variety of Berry-phase-driven transport phenomena in noncoplanar magnets. Mn4N is a unique ferrimagnetic system in which both collinear and noncoplanar magnetic configurations have been reported at different lattice parameters, featuring vanishing and finite SSC, respectively. However, the microscopic mechanism governing the competition between these magnetic states and the associated modulation of SSC remains unclear. This issue is particularly important for Mn4N because its high magnetic ordering temperature (TN = 740 K) provides an attractive platform for exploring chiral magnetic states and their associated topological functionalities at elevated temperatures. Here, using first-principles calculations, we investigate the strain-driven evolution of the magnetic ground state and SSC in Mn4N and uncover the microscopic origin of the collinear-to-noncoplanar magnetic transition. We demonstrate that tensile strain continuously stabilizes the noncoplanar configuration and enhances SSC, as quantified by the magnitude of the chirality-order vector. Orbital-resolved crystal orbital Hamilton population and charge-density-difference analyses reveal that strain selectively weakens the Mn3c-N hybridization while preserving direct Mn3c-Mn3c interactions, leading simultaneously to the activation of the Mn3c in-plane magnetic moment and the suppression of N-mediated ferromagnetic superexchange between nearest-neighbor Mn3c atoms. This bond-selective electronic response governs the magnetic ground state and consequently controls the emergence and enhancement of SSC in Mn4N. Our work establishes bond selectivity as an effective strategy for engineering SSC in high-temperature ferrimagnetic materials.

cond-mat.mtrl-sci↗
arXiv2026

Rufus-Air: An Open LLM Post-Training Recipe

Rufus-Air is an open and reproducible post-training recipe on GLM-4.5-Air-Base (106B-A12B), organized as a serial pipeline of eight stages: SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. We document the data, reward design, infrastructure, stage order, and stagewise results needed to reproduce the recipe. Stages progress from basic to advanced capabilities and from hard, verifiable rewards to softer judge-based signals. Training builds on open-source components and public data, much of it used as released, without new human annotation or an in-house distillation teacher. Our main findings are that (i) diverse, high-quality SFT establishes a strong capability floor; (ii) difficulty filtering keeps RL prompts within a productive learning range; (iii) reward reliability provides a practical principle for ordering stages; and (iv) infrastructure and engineering choices are part of the recipe, not just an implementation detail. Rufus-Air improves over the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models.

cs.CL↗