arXiv subjects

Wenjun Wang

Publications and source records attributed to Wenjun Wang.

At least 19 recordsLinked to original sources

arXiv2026

NPCBench: A Clinical Apprenticeship Benchmark for Multimodal Care-Pathway Reasoning in Nasopharyngeal Carcinoma

Current medical benchmarks typically break down clinical competence into isolated factual questions, single-image tasks, or static case vignettes. These tests score individual answers, but reveal less about whether a model can integrate subspecialty knowledge with multimodal evidence and track a patient's condition throughout a course of care. To close this gap, we introduce NPCBench, a clinical apprenticeship benchmark for nasopharyngeal carcinoma (NPC). NPCBench mirrors specialist training, from guideline and foundational knowledge to case planning, longitudinal review, CarePath integration, and open-ended consultation. It contains 4,777 component and consultation items and a 100-patient CarePath cohort with 417 scored diagnostic and care tasks. The episode-level evaluation tests whether diagnostic and care answers are jointly correct within a care pathway. By linking stepwise tasks to patient-level care trajectories, NPCBench evaluates whether models ground decisions in visual evidence, interpret changes in patient state, and meet clinical requirements in consultation. Across 30 evaluated models, we observe a composition gap: strong performance on individual tasks does not ensure jointly correct diagnostic and care answers for the same patient. Among 21 multimodal models, strict CarePath success reaches at most 23%. Paired controls and physician calibration further examine evidence grounding, longitudinal revision, and consultation judging. Overall, NPCBench broadens the evaluation of medical artificial intelligence beyond static question-answering paradigms to an apprenticeship-oriented assessment of clinical workflows, thereby offering a testbed for evidence-based and temporally coherent reasoning in oncology care.

cs.AI↗

arXiv2026

Experience Memory Graph: Extracting Reusable Correction Knowledge from Agent Trajectories

Large Language Model (LLM) agents have shown remarkable capabilities in autonomous decision-making by generating sequential trajectories of states, actions, and observations. However, in complex, long-horizon tasks, these agents frequently suffer from compounding errors and struggle to recover from failures. Despite advances in experience memory, extracting reliable and transferable correction knowledge from historical trajectories remains challenging. Existing methods, whether based on test-time reflection or offline memory construction, largely rely on LLMs to diagnose failures and formulate corrective strategies, which can produce brittle or task-specific insights that are difficult to generalize. To address this challenge, we propose Experience Memory Graph (EMG), a framework that reformulates agent failure recovery as a graph matching problem. During offline construction on the training set, we convert both failed exploration trajectories and successful expert trajectories into directed action decision graphs. By matching these graphs, we extract common subgraphs and graph edit paths that explicitly indicate how to correct failures, and store them in a memory graph with intra-task nodes and cross-task edges. At test time, EMG retrieves relevant insights and guides the agent in a single test-time attempt without additional reflection-and-replay rounds. Experiments on ALFWorld and ScienceWorld demonstrate consistent improvements over the evaluated memory construction and competitive baselines across seen and unseen splits.

cs.AI↗

arXiv2026

Same Target, Different Origins: Aligning Latent Rollout for Consistent PDE Emulation

Neural PDE solvers commonly generate multi-step predictions through autoregressive rollout, recursively applying a learned evolution operator and feeding each prediction back as input for the next step. Approximation errors can therefore propagate through the rollout, causing predictions at a shared target time to depend on their temporal origins and rollout depths. For a well-posed deterministic PDE flow, however, states at different times along the same physical trajectory should, under exact evolution, reach the same target state. To this end, we propose Align Latent Rollout (ALR), a regularization approach that promotes consistency across rollout origins through the neural PDE solver's internal representations. A second-order Taylor analysis of latent discrepancy propagation motivates two objectives: zeroth-order latent state consistency, which aligns representations at a shared target time, and first-order latent dynamics consistency, which aligns local evolution responses to shared perturbations. ALR can be integrated into different neural PDE solvers by regularizing suitable internal representations, without modifying their backbone architectures. Experiments across multiple PDE benchmarks and neural-operator backbones show that ALR improves prediction accuracy while reducing shared-target discrepancies across rollout origins.

cs.LG↗

arXiv2026

ReAT: Residual Action Tokenization for Autoregressive Robot Policies

Autoregressive robot policies rely on action tokenizers to convert continuous action chunks into compact sequences of discrete prediction targets. Existing tokenizers can reconstruct action chunks from discrete codes, but often leave implicit what each individual token contributes to the reconstructed action. We introduce Residual Action Tokenization (ReAT), which gives each token an explicit role in refining the current action reconstruction. It uses the decoded reconstruction as feedback to construct the next token from the remaining action error. This closes the loop between encoding and decoding, making each new token responsive to the errors left by the actual decoded prefix. Across 19 manipulation tasks spanning three simulation benchmarks and real-world settings, 8-token ReAT policies outperform the evaluated baselines in task success; on LIBERO, even 4-token ReAT surpasses the evaluated 8-token action-tokenization baselines at half the token budget. These results suggest a path toward action tokenization that supports self-corrective policy learning, helping autoregressive policies learn to progressively refine their own action plans.

cs.RO↗

arXiv2026

From Parameters to Behaviors: A Survey of Model Fusion for Large Language Models

Model fusion integrates the capabilities from source models into a single target model. As of June 2026, Hugging Face hosts more than 2M models. This growing pool provides a rich base for model reuse and capability integration. Yet existing surveys often cover only separate parts of this space, and they do not provide a unified definition or a systematic taxonomy. This survey defines model fusion and organizes prior work into three levels: parameter-level, representation-level, and behavior-level fusion. We also review related metrics, benchmarks, and applications, summarize current challenges, and identify future directions. Our goal is to provide a clear map of this area and support future work on model fusion. A comprehensive list of papers about model fusion is available at https://github.com/Baicaihaochi/Awesome-Model-Fusion-Survey.

cs.CL↗

arXiv2026

FeatCal: Feature Calibration for Post-Merging Models

Model merging combines task experts into one model and avoids joint training, retraining, or deploying many expert models, but the merged model often still underperforms task experts. We study this performance gap through feature drift, the difference between features produced by the merged model and by the expert on the same input. Our theory decomposes this drift into upstream propagation and local mismatch, tracks how it propagates and combines through later layers in forward order, and links final feature drift to output drift. This view motivates FeatCal, which uses a small calibration set to calibrate the merged model weights layer by layer in forward order, reducing feature drift while staying close to merged weights and preserving the benefits of model merging. FeatCal uses an efficient closed-form solution to update model weights, with no gradient descent, iterative optimization, or extra modules. On the main CLIP and GLUE benchmarks, FeatCal beats Surgery and ProbSurgery, the closest post-merging calibration baselines: 85.5% vs. 77.0%/78.8% on CLIP-ViT-B/32 Task Arithmetic (TA) and 85.2% vs. 83.7%/82.2% on FLAN-T5-base GLUE. On CLIP-ViT-B/32, 8 examples per task reach 82.9%, and 256 examples per task take 53 seconds, about 4x faster than both baselines, showing better sample efficiency and lower calibration cost.

cs.LG↗

arXiv2026

Beyond Captions: Context-Grounded Reconstruction for Biomedical Multimodal Continued Pretraining

Biomedical figures are explained not by captions alone but by body-text passages that discuss them. Yet current multimodal corpora typically reduce figures to isolated image-caption pairs, discarding this crucial context. Existing pipelines either omit this context or append it without enforcing the figure references that support each attachment, which can create unsupported image-text attachments and incoherent discourse. We introduce context-grounded reconstruction, a source-grounded framework that converts PubMed Central Open Access (PMC-OA) records into referentially coherent interleaved sequences. It recovers captions and source text, attaches context only through article-native figure references, repairs non-contiguous context, and prunes unsupported images. Starting from these reconstructed sequences, PMC-InterCPT first filters records for text quality and medical relevance, then applies evidence-aware allocation to form a 9.63B-token corpus for continued pretraining (CPT) of generative medical MLLMs. With fixed supervised fine-tuning (SFT), PMC-InterCPT improves Qwen3.5-4B-Base by 1.46 medical-average points and 3.11 general/scientific-average points over a token-matched raw source control, and surpasses a 42% larger raw-data run. Gains transfer to Qwen3.5-2B-Base and LLaVA-OneVision-1.5-4B-Base. Controlled ablations show that context-grounded reconstruction, rather than simply appending article context, is central to useful biomedical multimodal CPT.

cs.CL↗

arXiv2026

OnlineQAT: On-Policy Distillation for Ultra-Low-Bit Large Language Models

Quantization-aware training (QAT) can recover much of the accuracy lost when large language models are compressed below four bits. Existing re- covery stages, however, are commonly optimized on fixed completions or teacher-generated answers, whereas the deployed quantized model condi- tions on prefixes generated by itself. Quantization errors can therefore move the model into states that are absent from offline recovery data. We introduce OnlineQAT, a two-stage framework that first obtains a usable low-bit initialization through block-wise QAT and then performs on-policy distillation (OPD) on student-generated responses. At each visited pre- fix, a frozen full-precision teacher provides a sampled reverse-KL training signal. On Qwen3-1.7B, OnlineQAT obtains the best average among the compared quantized methods: 57.28 at W3A16 and 32.52 at W2A16, im- proving over ReasoningQAT by 2.90 and 0.44 points, respectively. The results suggest that student-visited states provide a useful recovery signal beyond fixed-completion training, particularly at three bits.

cs.CL↗

arXiv2026

Cluster-Graph Edit Distance: Optimal Explicit Embeddings, Metric Proxies, and Complexity

The cluster graphs on $n$ vertices, the disjoint unions of complete graphs, have the integer partitions of $n$ as their isomorphism classes, and the quotient edit distance $q^*(λ,μ)=\min_{σ\in S_n}|E(G_λ)\triangleσE(G_μ)|$ makes that set a metric space. Its geometry and its complexity both issue from one identity: $q^*$ is an affine function of the maximum of $\lVert X\rVert_F^2$ over the contingency tables with margins $λ$ and $μ$. Our main result is an explicit optimal embedding. The weighted dyadic sums of the Ferrers staircase, taken at the critical exponent $\frac14$, give a map $F_n$ into $\ell_2^{\,<4n}$ that acts on a single partition and is computable in $O(n)$ time, and its distortion is $Θ(n^{1/4})$. That order is optimal, since $c_2(\mathcal K_n)=Θ(n^{1/4})$: the lower half follows from a $Θ(\sqrt n)$-dimensional Hamming cube of partitions and Enflo's theorem, so the determination needs no other external input. The analytic core is a scale-free inverse inequality for every integer sequence with $v(1)=v(N+1)=0$ and $v(s)-v(s+1)\in s\mathbb Z$: its critical dyadic energy is at least $\lVert v\rVert_1^2/(63504\sqrt{\mathrm{TV}(v)})$. Combinatorially the same identity yields two explicit $\ell_1$ models, the vertex-mass metric on sorted degree sequences with $\frac12δ_1\le q^*<\frac32δ_1$ and the block-energy metric with $q^*\le B\le2q^*-1$, both constants optimal; hence $c_1(\mathcal K_n)\le2$, and an $O(n\log n)$-time algorithm returns an alignment of cost below $2q^*$ carrying the certificate $q^*\in[\lceil(B+1)/2\rceil,B]$. Computationally, deciding $q^*(λ,μ)\le Q$ is strongly NP-complete and admits no FPTAS, while the farthest alignment is polynomial-time solvable. The best constant in the inverse inequality remains open; an exactly solvable chirp family caps it at $\frac23$.

cs.DS↗

arXiv2026

Breaking the mutual exclusivity between metallicity and ferroelectricity in a non-polar covalent semiconductor via orbital selective doping

The mutual exclusion of ferroelectricity and metallic conductivity is a long-standing tenet because itinerant electrons screen long-range Coulomb forces that stabilize the bulk polar order. Here, we break this paradigm by heavily doping a non-polar covalent semiconductor of cubic silicon carbide (3C-SiC) with nitrogen. This introduces heavy electron doping, inducing metallicity and driving a structural transition from the non-polar F-43m to the polar R3m symmetry via the pseudo-Jahn-Teller effect. Remarkably, we provide direct, atomic-scale visualization of about 180° polarization reversal under an external voltage bias in a ferroelectric metal. The strongly directional character of antibonding orbitals occupied by conduction electrons prevents them from screening the local Si-C polarization, resulting in the coexistence of metallicity and ferroelectricity. Ferroelectric tunnel junctions demonstrate nonvolatile memory properties with a well-defined high-resistance state (HRS) and low-resistance state (LRS), an ultrahigh response speed (~50 ns), an ultralow operating voltage (1 V), an endurance exceeding 85927 cycles, and a projected retention time of 100 years. Our results provide a novel strategy for pioneering ferroelectricity in a metal, a new ferroelectric metal platform for exploring exotic properties, and a ferroelectric device with high performance that meets the requirements for low consumption and high-speed non-volatile devices.

cond-mat.mtrl-sci↗

arXiv2026

Unsigned Frenet Data of Closed Space Curves: Exact Fibres, Generic Rigidity, and Conditional Stability

A closed positively curved space curve is determined by its curvature and signed torsion up to an orientation-preserving rigid motion; that sign is the only place the ambient orientation enters. We ask what survives its loss, for closed embedded curves in $\mathbb R^3$ with $κ>0$ compared pointwise in a common arclength label. The answer is governed by the branch invariant $c(τ)$, the number of components left by the infinite-order zero set of $τ$: the smooth signed lifts of $|τ|$ number exactly $2^{c(τ)}$, and reduce to $\{τ,-τ\}$ precisely when $c(τ)\le1$. Hence a given unsigned datum is carried by at most $2^{c(τ)}$ classes modulo $SE(3)$, and by a single $E(3)$-orbit when $c(τ)\le1$. Both extremes occur: for arbitrary knot types $K_1,\dots,K_m$ there is a datum with fibre exactly $2^m$ classes modulo $SE(3)$, realising all connected sums of the $K_i$ and their mirrors; under a chirality hypothesis these are $2^m$ knot types. Conversely, curves with only simple torsion zeros are open and dense, hence residual, among parametrised $C^r$ embeddings ($r\ge4$), and each is determined up to $E(3)$ by its datum. No uniform quantitative form of this rigidity exists; but on each stratum $Δ=\inf_s\sqrt{τ^2+(τ')^2}\geδ>0$ with uniform $C^5$ and curvature bounds the orbit distance obeys a log-Lipschitz bound, whose optimal constants diverge as $δ\downarrow0$ on the strata containing a fixed exact ambiguous pair. The engine is a one-dimensional inverse estimate for the signed square root, logarithmically optimal at that level.

math.DG↗

arXiv2026

A Comprehensive FP8 Training Recipe for Reasoning-Enhanced Language Models

The immense computational cost of training Large Language Models (LLMs) presents a major barrier to innovation. While FP8 training offers a promising solution with significant theoretical efficiency gains, its widespread adoption has been hindered by the lack of a comprehensive, open-source training recipe. To bridge this gap, we introduce an end-to-end FP8 training recipe that seamlessly integrates continual pre-training and supervised fine-tuning. Our methodology employs a fine-grained, hybrid-granularity quantization strategy to maintain numerical fidelity while maximizing computational efficiency. Through extensive experiments, including the continue pre-training of models on a 160B-token corpus, we demonstrate that our recipe is not only remarkably stable but also essentially lossless, achieving performance on par with the BF16 baseline across a suite of reasoning benchmarks. Crucially, this is achieved with substantial efficiency improvements, including up to a 22% reduction in training time, a 14% decrease in peak memory usage, and a 19% increase in throughput. Our results establish FP8 as a practical and robust alternative to BF16, and we will release the accompanying code to further democratize large-scale model training.

cs.CL↗

arXiv2026

When Can You Correct Distribution Drift in Temporal Graph Generation? A Sharpening--Drift Tension and an Impossibility for Observation-Based Correction

Generative models of temporal graphs are trained on one stretch of an evolving network and deployed on the next, and they degrade badly in the gap. We show this degradation is derivable, general, and not fixable from observations. The masked flow-matching loss decomposes exactly, with no independence assumption, into an irreducible entropy plus a divergence whose derivative along the training path is positive precisely for structures rare during training and common at deployment, diverging as their training probability goes to zero. Empirically the trade-off is a power law with exponent $-0.605$ ($R^2=0.9977$), and drift raises the sampler's error floor without changing how many steps reach it: across seven well-powered conditions the drift-period marginal error varies by at most $6\%$ over a $50\times$ range of sampling budgets, while the floor sits $2.2\times$ to $34.3\times$ above the in-period floor. Because the deployment period is observed, correction looks like a matter of measurement. It is not. We prove that any corrector measurable with respect to past observations leaves at least the conditional variance of the statistic it tracks, and that trend extrapolation beats trusting the last observation only when $μ^2>v(1-2ρ)$. Both premises are measurable and both go the wrong way: the drift is trendless and mean-reverting, with a one-step innovation as large as the drift itself. An oracle removes $60\%$ of the error, the best observation-based corrector recovers $5.7\%$ of that, and extrapolation is strictly worse than doing nothing clever.

cs.LG↗

arXiv2026

DiffAttn: Diffusion-Based Drivers' Visual Attention Prediction with LLM-Enhanced Semantic Reasoning

Drivers' visual attention provides critical cues for anticipating latent hazards and directly shapes decision-making and control maneuvers, where its absence can compromise traffic safety. To emulate drivers' perception patterns and advance visual attention prediction for intelligent vehicles, we propose DiffAttn, a diffusion-based framework that formulates this task as a conditional diffusion-denoising process, enabling more accurate modeling of drivers' attention. To capture both local and global scene features, we adopt Swin Transformer as encoder and design a decoder that combines a Feature Fusion Pyramid for cross-layer interaction with dense, multi-scale conditional diffusion to jointly enhance denoising learning and model fine-grained local and global scene contexts. Additionally, a large language model (LLM) layer is incorporated to enhance top-down semantic reasoning and improve sensitivity to safety-critical cues. Extensive experiments on four public datasets demonstrate that DiffAttn achieves state-of-the-art (SoTA) performance, surpassing most video-based, top-down-feature-driven, and LLM-enhanced baselines. Our framework further supports interpretable driver-centric scene understanding and has the potential to improve in-cabin human-machine interaction, risk perception, and drivers' state measurement in intelligent vehicles.

cs.CV↗

arXiv2026

E-PMQ: Expert-Guided Post-Merge Quantization with Merged-Weight Anchoring

Low-resource deployment constraints have made model quantization essential for deploying neural networks while preserving performance. Meanwhile, model merging has become an increasingly practical low-resource strategy for integrating multiple task- or domain-specialized experts into a single model without joint training or multi-model serving. Together, quantization and model merging enable an efficient low-resource deployment pipeline by integrating multiple experts into one low-bit model. We formulate this setting as Post-Merge Quantization (PMQ). We show that directly applying post-training quantization (PTQ) to a merged model is unreliable because two distinct deviations are coupled: the quantization deviation introduced by low-bit reconstruction and the expert-relative merging deviation inherited from model merging. To mitigate these deviations, we propose E-PMQ, an expert-guided PMQ framework that uses source expert weights to provide expert- guided output targets during layer-wise calibration, together with merged-weight anchoring to stabilize the calibration and preserve the integrated behavior of the merged model. On CLIP-ViT-B/32 eight-task merging, E-PMQ improves 4-bit GPTQ from 65.0% to 73.6% under Task Arithmetic and from 69.1% to 74.8% under TIES-Merging. On harder settings, E-PMQ improves GPTQ from 34.8% to 76.7% on 20-task CLIP-ViT-L/14 and from 78.26% to 83.34% on FLAN-T5- base GLUE. These results demonstrate that E-PMQ enables effective post-merge quantization and low-bit deployment.

cs.CL↗

arXiv2026

Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training

Continual post-training aims to extend large language models (LLMs) with new knowledge, skills, and behaviors, yet it remains unclear when sequential updates enable capability transfer and when they cause catastrophic forgetting. Existing methods mitigate forgetting through sequential fine-tuning, replay, regularization, or model merging, but offer limited criteria for determining when incorporating new updates is beneficial or harmful. In this work, we study LLM continual post-training through three questions: What drives forgetting? When do sequentially acquired capabilities transfer or interfere? How can compatibility be used to control update integration? We address these questions through task geometry: we represent each post-training task by its parameter update and study the covariance geometry induced by the update. Our central finding is that: forgetting can be considered as a state-relative update-integration failure, it arises when the covariance geometries induced by tasks misalign with the geometry of the evolving model state. Sequential updates transfer when they remain compatible with the model state shaped by previous updates, and interfere when state-relative geometry conflict becomes high. Motivated by this finding, we propose Geometry-Conflict Wasserstein Merging (GCWM), a data-free update-integration method that constructs a shared Wasserstein metric via Gaussian Wasserstein barycenters and uses geometry conflict to gate geometry-aware correction. Across Qwen3 0.6B--14B on domain-continual and capability-continual settings, GCWM consistently outperforms data-free baselines, improving retention and final performance without replay data. These results identify geometry conflict as both an explanatory signal for forgetting and a practical control signal for LLM continual post-training.

cs.LG↗

arXiv2026

Cross-Vehicle 3D Geometric Consistency for Self-Supervised Surround Depth Estimation on Articulated Vehicles

Surround depth estimation provides a cost-effective alternative to LiDAR for 3D perception in autonomous driving. While recent self-supervised methods explore multi-camera settings to improve scale awareness and scene coverage, they are primarily designed for passenger vehicles and rarely consider articulated vehicles or robotics platforms. The articulated structure introduces complex cross-segment geometry and motion coupling, making consistent depth reasoning across views more challenging. In this work, we propose \textbf{ArticuSurDepth}, a self-supervised framework for surround-view depth estimation on articulated vehicles that enhances depth learning through cross-view and cross-vehicle geometric consistency guided by structural priors from vision foundation model. Specifically, we introduce multi-view spatial context enrichment strategy and a cross-view surface normal constraint to improve structural coherence across spatial and temporal contexts. We further incorporate camera height regularization with ground plane-awareness to encourage metric depth estimation, together with cross-vehicle pose consistency that bridges motion estimation between articulated segments. To validate our proposed method, an articulated vehicle experiment platform was established with a dataset collected over it. Experiment results demonstrate state-of-the-art (SoTA) performance of depth estimation on our self-collected dataset as well as on DDAD, nuScenes, and KITTI benchmarks.

cs.CV↗

arXiv2026

Language Models as Messengers: Enhancing Message Passing in Heterophilic Graph Learning

Graph neural networks (GNNs) have become a standard paradigm for graph representation learning, yet their message passing mechanism implicitly assumes that messages can be represented by source node embeddings, an assumption that fails in heterophilic graphs. While existing methods attempt to address heterophily through graph structure refinement or adaptation of neighbor aggregation, they often overlook the semantic potential of node text, relying on suboptimal message representation for propagation and compromise performance on homophilic graphs. To address these limitations, we propose LEMP4HG, a novel language model (LM)-enhanced message passing approach for heterophilic graph learning. Specifically, for text-attributed graphs (TAG), we leverage a LM to explicitly model inter-node semantic relationships from paired node texts, synthesizing semantically informed messages for propagation. To ensure practical efficiency, we further introduce an active learning-inspired strategy guided by a tailored heuristic, MVRD, which selectively enhances messages for node pairs most affected by message passing. Extensive experiments demonstrate that LEMP4HG consistently outperforms state-of-the-art methods on heterophilic graphs while maintaining robust performance on homophilic graphs under a practical computational budget.

cs.AI↗