Searcharxiv⌕ Search

arXiv subjects

Yu Wang

Publications and source records attributed to Yu Wang.

At least 145 records · Page 8Linked to original sources
arXiv2026

The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and The Channel, Not the Content, Decides What Works

Dense per-step supervision is the standard remedy for sparse-reward long-horizon LLM agents: reward the policy for predicting its next observation, which looks provably safe under potential-based shaping. Published prediction-reward and auxiliary-loss variants report both successes and instabilities; we supply the controlled account: 74 preregistered arms dissect one fixed prediction signal under GRPO across ALFWorld, WebShop, a synthetic POMDP, and Qwen3-1.7B/4B/8B, varying only the delivery mechanism. (1) Every run sustaining this difference-form reward under untouched std normalization (no filtering, dynamic-sampling, or decoupling mitigations) collapses: eleven runs across scales, coefficients, group sizes, and groupings (the floor-bound synthetic environment stalls instead); ALFWorld runs end in an absorbing state (prediction accuracy -> 1.0, success -> 0): the optimizer builds the "dark room". The algebra is one line: in all-fail groups z-scoring cancels the shaping coefficient; removing only std normalization restores baseline parity. (2) A signal's danger is set by its within-group variance trajectory, plus hackability as a second axis; it retrodicts every reward-channel collapse and survives preregistered prospective tests. (3) The same signal as a teacher-forced auxiliary loss is harmless on ALFWorld at 4B, but the gain is not the signal's: content-free placebos as a class match or beat gold at both matched seeds (s0: 78.8 vs 68.6; s42: 67.9 vs 57.9); the auxiliary update is the regularizer. (4) At 8B the recipe turns bistable: gold full-weight locks two of three seeds; every content-free or reduced-weight arm stays healthy. No ALFWorld or WebShop reward-channel variant measurably beats its matched-normalization baseline and no gold signal measurably outperforms its content-free placebo: the delivery channel, not the content, decides; which channel is safe is regime-dependent.

cs.LG↗
arXiv2026

A termolecular reaction theory for gas-phase nucleation based on long-range intermolecular forces

The birth of a new phase is usually described thermodynamically, but in the gas phase it begins as chemistry. Here, we show that gas-phase nucleation can be described as a termolecular reaction network controlled primarily by long-range intermolecular forces. In single-component mixtures, dimer formation emerges as a direct termolecular process, whereas in binary mixtures a chaperon mechanism can enhance nucleation, with the second component acting as a catalyst for cluster formation. By incorporating cluster evaporation, the same framework can be extended beyond the collision limit, providing a molecular route to nucleation in regimes where larger critical clusters become relevant. We test the theory against unary and binary nucleation of water, toluene, and butane, obtaining agreement in absolute rates within one order of magnitude across the explored temperature and density ranges. These results identify long-range intermolecular forces as molecular drivers of gas-phase nucleation and establish elementary termolecular chemistry as a predictive route to cluster formation.

physics.atom-ph↗
arXiv2026

Tevatron Meets Megatron: Expert-Parallel LLM Reranker Training on an Academic Budget

Modern reranking recipes---billion-scale cross-encoders, mixture-of-experts (MoE) backbones, and distillation against strong teachers---have outpaced the training infrastructure available to most academic groups. Existing Tevatron reranker training relies on the Hugging Face Trainer with DeepSpeed or PyTorch FSDP1, but these backends lack efficient support for large-scale MoE training. We present Tevatron 3.0, which integrates a Megatron-Core training backend into Tevatron while preserving its data pipeline, evaluation workflow, and Hugging Face-compatible checkpoints. We benchmark existing distributed training configurations against the new backend, showing that Megatron matches FSDP reranker quality and training efficiency under comparable data-parallel settings, is up to 22% faster in the recommended single-node configuration, and supports both LoRA and full-parameter fine-tuning. Crucially, expert parallelism enables training a 30B-parameter Qwen3-30B-A3B MoE reranker, which is infeasible with PyTorch FSDP1. Using this framework, we conduct a controlled comparison of MoE versus dense models, LoRA versus full-parameter tuning, and distillation versus contrastive training on BEIR-15 with three first-stage retrievers, and report serving throughput for Hugging Face and vLLM. We find that the MoE reranker matches dense 8B quality while activating less than half as many parameters and achieving substantially higher inference throughput. We will release the framework and trained checkpoints.

cs.IR↗
arXiv2026

A Geometric Inverse Source Problem for Stochastic Parabolic Equations

This paper addresses the geometric inverse problem of simultaneously recovering two unknown deterministic source supports in a stochastic parabolic equation, where one source appears in the drift term and the other in the diffusion term. We establish that partial boundary flux measurements alone uniquely determine both supports. Moreover, we prove the existence of minimizers for a perimeter-regularized objective functional. For smooth interfaces, we derive a Hadamard-type boundary representation of the shape derivative. To the best of our knowledge, this is the first such formula for the simultaneous recovery of a drift-source support and a diffusion-source support in an SPDE setting. The derivative exhibits a genuinely stochastic two-channel structure: the adjoint state governs the sensitivity of the drift-source interface, while the martingale component controls the sensitivity of the diffusion-source interface. Based on this formula, we develop a shape-gradient reconstruction method. Numerical experiments demonstrate its effectiveness and its capacity to distinguish between the two source channels under both full and partial boundary observations.

math.OC↗
arXiv2026

Pay for The Second-Best Service: A Game-Theoretic Approach Against Dishonest LLM Providers

The widespread adoption of Large Language Models (LLMs) through Application Programming Interfaces (APIs) induces a critical vulnerability: the potential for dishonest manipulation by service providers. This manipulation can manifest in various forms, such as secretly substituting a proclaimed high-performance model with a low-cost alternative, or inflating responses with meaningless tokens to increase billing. This work tackles the issue through the lens of algorithmic game theory and mechanism design. We are the first to propose a formal economic model for a realistic user-provider ecosystem, where a user can iteratively delegate $T$ queries to multiple model providers, and providers can engage in a range of strategic behaviors. As our central contribution, we prove that for a continuous strategy space and any $ε\in(0,\frac12)$, there exists an approximate incentive-compatible mechanism with an additive approximation ratio of $O(T^{1-ε}\log T)$, and a guaranteed quasi-linear second-best user utility. We also prove an impossibility result, stating that no mechanism can guarantee an expected user utility that is asymptotically better than our mechanism. Furthermore, we demonstrate the effectiveness of our mechanism in simulation experiments with real-world API settings.

cs.GT↗
arXiv2026

Beyond Routing Weights: Faithful Response-Level Interpretation of Mixture-of-Experts Reward Models via Contribution Contrast

Reward models are central to learning from human preferences, yet identifying what drives their predictions remains challenging. Recent sparse Mixture-of-Experts (MoE) reward models seek to improve interpretability by routing prompts to specialized experts and characterizing experts through examples with high routing weights. However, routing weights only reveal which prompts an expert $\textit{receives}$, not how it $\textit{judges}$ responses, providing only a partial account of expert behavior. We therefore propose $\textbf{Co}$ntribution-$\textbf{Co}$ntrast ($\textbf{CoCo}$) response-level interpretation, which faithfully characterizes experts' roles using chosen-rejected response pairs with the largest contribution contrasts, jointly capturing routing and preference behavior. Across automatic and human evaluations, CoCo yields more coherent, faithful, and specialized interpretations than router-based, score-based, and sparse autoencoder-based alternatives while maintaining competitive reward modeling accuracy. To the best of our knowledge, this is the first systematic study of interpretation methods for MoE reward models.

cs.AI↗
arXiv2026

BCNet: Bronchus Classification via Structure Guided Representation Learning

CT-based bronchial tree analysis is essential for diagnosing lung and airway diseases, yet automatic bronchus classification remains challenging because bronchial topology varies substantially across individuals. We propose the Bronchus Classification Network (BCNet), a structure-guided framework that uses segment-level topological information from point clouds to improve voxel-level representation learning. BCNet contains two jointly trained branches: a Point-Voxel Graph Neural Network (PV-GNN) for segment classification and a Convolutional Neural Network (CNN) for voxel-wise labeling. The branches share a common convolutional backbone, allowing topology-aware supervision from the PV-GNN to enhance voxel-level features. During inference, only the CNN branch is required, so BCNet retains the computational efficiency of its CNN baseline. Experiments on BronAtlas demonstrate that BCNet outperforms state-of-the-art methods by more than 8.0% in F1-score for bronchus classification. We also introduce BronAtlas, an open-access benchmark for bronchial imaging analysis that contains high-quality voxel-wise annotations of anatomical and abnormal bronchial segments. BronAtlas provides a valuable resource for developing and evaluating advanced methods for bronchial tree analysis, disease diagnosis, and surgical planning.

eess.IV↗
arXiv2026

Do Unified Multimodal Models Think in One Space? A Lens Through Cross-Branch Steering

Unified multimodal models (UMMs) aim to integrate understanding and generation within a single architecture, yet it remains unclear whether these capabilities share a unified and transferable semantic space. This question is fundamentally challenging, as the two branches operate over heterogeneous representations (text tokens vs.\ visual latents) and distinct training objectives, making direct comparison difficult. To address this, we introduce \emph{cross-branch semantic steering}, an intervention-based framework that extracts semantic directions from one branch and applies them to the other. We show that steering vectors learned from the understanding branch can transfer to generation, enabling controllable image synthesis and improved semantic faithfulness. In contrast, the reverse direction consistently shows limited effectiveness. Our analysis suggests that this asymmetry may be related to a practical representational mismatch: understanding-derived vectors capture transferable, object-centric semantics, while generation-derived vectors primarily encode low-level appearance features. Our results reveal that architectural unification does not guarantee semantic alignment, and establish cross-branch steering as a practical tool for probing multimodal representations.

cs.CV↗
arXiv2026

Calibrated Pressure-Observable Born and Hessian Actions for Quantum-Assisted Waveform Inversion

We construct a pressure-consistent operator-and-readout interface for Born, adjoint, and Gauss--Newton actions in constant-density acoustic full-waveform inversion (FWI) using Schrödingerised propagation. The energy variables $π=c^{-1}\partial_t u$ and $q=\nabla u$ yield an auxiliary-space Hamiltonian, while physical pressure $p=cπ$ depends explicitly on wavespeed. Its derivative $D(cπ)[c_0](δc)=c_0δπ+δc\,π_0$ combines propagated wavefield sensitivity with a direct receiver-calibration term. Duhamel and receiver-row differentiation retain both contributions in the Born map, its adjoint, and the Gauss--Newton normal action. We prove a conditional consistency estimate with a periodic second-order finite-difference specialization and give a resource model for state preparation, normalization, quadrature, and selected-output measurement. A compiled nine-qubit instance realizes structured preparation, product-formula propagation, a derivative-LCU block, and calibrated pressure-overlap measurements. Bernoulli samples from ideal-circuit probabilities drive a four-parameter hybrid inversion. A two-qubit VQLS circuit represents the normalized update direction, while normal-system assembly, line search, and model refresh remain classical. Finite differences, tangent and reverse-adjoint recurrences, autodiff JVP/VJP evaluations, and explicit Jacobians verify the discrete Born, adjoint, and normal actions. Smooth periodic refinement confirms second-order convergence, whereas omitting receiver calibration leaves an order-one Born error and substantially changes the regularized Gauss--Newton direction. All ten predeclared finite-shot runs reduce the initial model error. These results specify the physical-pressure derivative and selected-output measurements needed to connect Schrödingerised propagation to a local FWI update.

quant-ph↗
arXiv2026

CAST: Game Solvers as Turn-Level Teachers for LLM Agents

Training large language models (LLMs) to act in long-horizon games is a promising step toward generalist decision-making, yet reinforcement learning with verifiable rewards (RLVR) relies on sparse final rewards that reveal little about which decisions determine success. Denser process signals could supply this missing turn-level credit, but existing sources are hard to keep both cheap and accurate. We observe that changes in a game solver's state value reveal whether an action advances the state toward success. Building on this insight, we propose CAST (Credit Assignment from Solver Teachers), which converts these value changes into solver advantages and injects them into RLVR as turn-level signals. We further show that, under a soft-optimal solver assumption, maximizing the solver advantage is equivalent to on-policy distillation from the solver, requiring only scalar values rather than teacher logits. Across Sokoban, Minesweeper, and Rush Hour, CAST outperforms all trained baselines on every game under both in-domain and unseen-difficulty evaluation and achieves the highest average zero-shot performance on ALFWorld and WebShop. Our code is available at https://github.com/Wloner0809/CAST.

cs.CL↗
arXiv2026

CoRenew: A large language model agent-based policy simulation platform for multifamily residential redevelopment

The difficulty of collective action remains a central challenge in the design of policies for multifamily residential redevelopment. Stakeholders continually adjust their decisions in response to evolving negotiation contexts and the reactions of others, meaning that when a policy intervenes and which stakeholders it targets can substantially reshape collective outcomes. Assessing these adaptive responses ex ante remains difficult because existing simulation models often rely on predefined behavioral rules. Here, we present CoRenew, an open-source platform that uses LLM-based agents to simulate negotiations among multiple stakeholders and evaluate the effects of alternative policy combinations. Integrating open source geographic and demographic data, the platform can generate synthetic residents, simulate negotiation dynamics under alternative policy settings and compares policy performance across competing objectives. It supports both numerical and semantic policy inputs and includes built-in tools for visualization and result export. We validate its behavioral realism against survey responses from 324 residents and a nine-month observed negotiation process from a real redevelopment case. With its modular and adaptable architecture, CoRenew can be used to assess policies across different institutional and cultural contexts.

cs.MA↗
arXiv2026

The afterglow of gamma-ray burst - supernova connections

The X-ray afterglow of several long gamma-ray bursts (LGRBs) associated with broad line type Ib/c supernovae (SNe) exhibits a standard non-thermal afterglow, commonly attributed to synchrotron emission from a relativistic jet, and an evolving thermal component whose physical origin is still debated. We investigate whether these thermal and non-thermal components can be described within a common analytical framework and whether their temporal evolution can provide insight into the physical connection between relativistic jets and SN ejecta. We combine a phenomenological description of the synchrotron emission of the non-thermal energy associated with the relativistic jet with a diffusion model describing the thermal evolution of the jet-affected portion of the SN ejecta. The resulting coupled system is solved analytically, leading to closed-form expressions for both non-thermal and thermal luminosities. We apply the model to the following systems: GRB 060218A/SN 2006aj and GRB 171205A/SN 2017iuk. The proposed formalism reproduces the main features of both thermal and non-thermal light curves. In both systems, the thermal luminosity follows a temporal evolution similar to that of the non-thermal component up to the plateau phase. The inferred thermal energy stored in the jet-affected ejecta is found to be a fraction of the energy coupled to the observed non-thermal emission. The characteristic timescales obtained from the fits suggest a direct link between the evolution of the jet and the thermal response of the expanding ejecta. Although simplified, the proposed framework provides a unified analytical description of LGRB afterglows and their thermal counterparts and offers a useful tool for investigating the physical connection between relativistic jets and SNe Ib/c.

astro-ph.HE↗
arXiv2026

Application-Driven Architecture Exploration for Cross-Layer Heterogeneous Systems

AI and HPC infrastructure increasingly serves workload portfolios that combine dense tensor computation, sparse kernels, large memory footprints, and communication-intensive collectives. Supporting these portfolios requires coordinated choices across accelerators, memory tiers, scale-up fabrics, and cluster networks. The resulting Cross-layer Heterogeneous System (XHS) design space is difficult to explore: hardware choices change legal task mappings, while rack power, switch radix, cabling, and cost constraints invalidate many candidates. We present CHASE, an application-driven framework that searches physically feasible XHS architectures through the workloads they must execute. CHASE represents candidates as hierarchical typed graphs and rejects designs that violate deployment constraints. It avoids intractable joint hardware-mapping search with a decoupled two-level loop: an inner mapper translates hardware-independent workload DAGs into topology-aware event traces, a calibrated event-driven simulator evaluates each mapping, and an outer telemetry-guided optimizer evolves the hardware graph. We evaluate CHASE on sparse-computing and LLM workloads. Its mapper remains within 6.06% of exhaustive optima while reducing mapping time by 60.5% on average relative to PEFT. Compute-model errors average 4.4-7.5%, and communication validation reproduces key trends across physical platforms. The outer search reaches near-global optima within 64 iterations. End-to-end case studies show that sparse workloads favor criticality-aware heterogeneous pods, whereas LLM inference favors scale-up islands; the resulting designs deliver 6.20$\times$ and 2.12$\times$ geomean speedups, respectively, while reducing cost and power relative to the baselines.

cs.DC↗
arXiv2026

The Curse of Precision: A Data Scaling Law for High-Precision Robotic Manipulation

While scaling laws for imitation learning have primarily focused on generalization in open-world settings, the relationship between data and precision in closed-world tasks like robotic assembly remains largely unexplored. This paper systematically investigates this relationship and introduces a novel scaling law. We find that to achieve a fixed success rate, the required number of demonstrations $N$ grows super-exponentially as the target precision $P$ approaches a limit $c$. This relationship is accurately captured by the model $\log(N) \propto 1/(P-c)$. Crucially, we reveal that the limit precision $c$ is not a static physical constant of the task but an emergent property of the entire agent system, including its sensors and expert policy. Through experiments on canonical manipulation tasks, we validate this law and demonstrate that improving system components, such as adding a wrist camera or using a more effective expert, measurably lowers $c$, thus expanding the system's achievable precision. Our work provides a new theoretical framework for precision in robotics and a quantitative metric to evaluate system capabilities. Furthermore, these findings provide a practical methodology for guiding the development and debugging of high-precision manipulation systems.

cs.RO↗
arXiv2026

DAIS: Dependency-Aware Intermediate QA Supervision for Complex Reasoning

Chain-of-thought (CoT) supervision exposes intermediate rationales, but flat rationale targets usually optimize a single reasoning sequence and provide limited supervision on how local conclusions should support later decisions. We introduce Dependency-Aware Intermediate QA Supervision (DAIS), a training-time framework that converts filtered teacher rationales into stage-level QA records. Each intermediate record predicts a local answer conditioned on the previous states needed for that decision, while the final-answer record keeps the original task format; evaluation therefore uses only the original input and optional context. Across GDPR, AIACT, MedQA, and FOLIO with multiple Qwen backbones, DAIS improves average final-answer accuracy over answer-only, flat chain-of-thought, and independent-QA baselines. On policy-compliance benchmarks, it achieves a largest gain of 5.6% and an average gain of 4.2% over the strongest non-DAIS baseline. Controlled ablations show that valid previous-state conditioning contributes beyond longer targets or additional intermediate text, supporting dependency-conditioned intermediate QA as a lightweight auxiliary supervision signal for standard final-answer inference.

cs.CL↗
arXiv2026

Convolution for Large Language Models

Large language models (LLMs) largely rely on Transformers, where self-attention provides global token interaction but does not explicitly encode the locality of natural language. We study whether lightweight depthwise convolutions can supply this local inductive bias without materially increasing model size. Our macro-level ablation compares convolution at 17 locations in a Qwen3 Transformer block and finds the best results when convolution is applied to the projected queries, keys, and values before attention. A subsequent micro-level study favors a residual depthwise convolution with kernel size $k=3$, without additional normalization or activation. Across Qwen3 models and several pre-training data budgets, this design improves the average accuracy on seven downstream benchmarks while adding less than $0.01\%$ parameters. A representation-level case study further suggests that the convolution makes repeated token IDs more sensitive to their immediate context. These results support depthwise convolution as a lightweight complement to self-attention for modeling short-range token interactions.

cs.CL↗
arXiv2026

PDD: Unleashing Economical and Flexible Heterogeneous LLM Inference via Cross-Datacenter Prefill-Decode Disaggregation

Interconnecting geographically dispersed clusters over wide-area Ethernet provides a scalable and cost-effective alternative to dedicated intra-datacenter heterogeneous clusters for Large Language Model (LLM) inference. However, this cross-datacenter disaggregation imposes heavy KV-cache transfers between clusters, causing substantial Time-to-First-Token (TTFT) latency, which is especially detrimental for agentic workloads characterized by long contexts, high cache hit rates, and short outputs. We propose PDD, a three-tier disaggregation architecture built upon prefill-decode (PD) disaggregation, consisting of Prefill, RelayDecode (RLD), and MainDecode (MD) instances. On Cluster A, Prefill instances perform the prefill computation, while RLD instances immediately receive the KV cache via high-speed RDMA and begin decoding, thereby overlapping the KV transfer to Cluster B over TCP-based Ethernet. MD instances on Cluster B receive the KV cache along with the tokens produced by RLD, and decoding is then seamlessly handed off from RLD to MD for completion. To maximize overall efficiency, PDD employs three core mechanisms: Decode-side RadixCache to alleviate bandwidth bottlenecks, an Extend-Decode Handoff mechanism for smooth control migration between RLD and MD, and multi-stage pipeline orchestration to manage complex inter-tier dependencies under high concurrency and long-term serving. We further design a low-cost, fine-grained heterogeneous deployment scheme that maximizes latency-masking efficiency at marginal cost. Compared to the intra-DC homogeneous PD baseline, PDD's cross-datacenter mapping of compute-intensive H100s and memory-bandwidth-optimized H200s achieves a Benefit-Cost Ratio (BCR) up to 37.5% higher in SLA-compliant goodput.

cs.AR↗
arXiv2026

Modality Discrepancy Transformer for Ambivalence and Hesitancy Recognition

Ambivalence and hesitancy (A/H) are affective states in which individuals express contradictory signals across facial, vocal, and linguistic channels. Automatically recognising A/H in clinical videos requires detecting cross-modal disagreement -- the signal that standard fusion methods suppress. Based on the conflict-aware multimodal fusion framework of Bekhouche et al., we present the Modality Discrepancy Transformer (MDT). MDT enriches the original 6-token design to a 9-token representation comprising three modality embeddings, three absolute-difference features, and three Hadamard-product discrepancy features learned through linear projections. These nine tokens undergo Transformer self-attention, with FiLM-based text-conditioned modulation and LoRA fine-tuning as core architectural components. A text-guided late fusion branch blends a text-only auxiliary head with the full multimodal output at inference. On the BAH dataset from the 3rd ABAW Challenge, MDT achieves 0.7408 Macro F1 on the labelled test split and 0.7368 on the private leaderboard, outperforming the strongest published baseline by over 10 points while training in under 20 minutes on a single GPU.

cs.CL↗