Searcharxiv⌕ Search

arXiv subjects

Wenjie Huang

Publications and source records attributed to Wenjie Huang.

At least 19 recordsLinked to original sources
arXiv2026

TailSieve: Partial-Rollout-Guided Tail Routing for LLM Rollouts

Large-scale rollouts have become a core component of modern LLM systems, spanning reinforcement learning (RL) post-training, on-policy distillation, and sampling-heavy evaluation. Unlike online serving, which typically targets request-level latency and throughput, synchronous rollouts must collect a complete batch before proceeding. A few long-tail generations can therefore dominate an entire rollout step, and uniform routing can slow these generations further by placing them in high-concurrency decoding batches. We present TailSieve, a partial-rollout-guided framework that jointly controls tail routing and replica allocation. Motivated by the observed persistence of prompt-level tail tendencies across policy updates, TailSieve uses partial-rollout completion status as a training-free signal for identifying candidate tail groups. Selected prompts are combined with newly sampled prompts, and all responses are generated from scratch under the current policy. A hierarchical controller uses response-work history and a serving model: a fast inner loop adjusts the number of isolated groups to balance completion times, while a slower outer loop adapts how many replicas serve these groups. The resulting low-concurrency execution also enables route-specialized speculative decoding with MTP or DFlash. Across the evaluated configurations, TailSieve achieves up to $1.67\times$ speedup with routing alone and $2.59\times$ with routing and speculation combined, both over uniform group routing.

cs.AI↗
arXiv2026

CARGO: Causal Attribution and Responsibility-Guided Preventive Safe Multi-Agent Reinforcement Learning for Cooperative Driving

This study presents Causal Attribution and Responsibility-Guided Optimization (CARGO), a multi-agent reinforcement learning framework for cooperative on-ramp merging in mixed traffic. A shared latent-context posterior and an action-conditioned response model estimate directed behavioral influence, with posterior uncertainty propagated into interaction-edge selection. Complementary attribution mechanisms couple signed influence with marginal team utility for cooperative reward shaping and assign responsibility from positive excess tail risk relative to a safe reference policy. Conditional diffusion generates future responses for action-dependent risk estimation. Responsibility-guided preventive optimization calibrates agent-specific risk-reduction requirements to local capacity within a policy trust region, using reserved safety margins, adaptive requirement scaling, and joint feasibility verification. Theoretical analysis establishes risk monotonicity and sufficient local conditions for non-decreasing environmental return under explicit assumptions. Experiments in highway-env across three traffic densities, together with controlled graph-recovery and attribution studies, support the effectiveness of the framework. Under high-density traffic, CARGO increases the success rate by 66.04\% and reduces the crash rate by 56.23\% relative to the best evaluated baseline for each metric.

cs.AI↗
arXiv2026

Raincheck: Deferred Multimodal Prefill for Efficient Multi-Turn MLLM Inference

Multi-turn conversations with multimodal large language models (MLLMs) reuse input context, but long multimodal inputs increase time to first token (TTFT). Query-aware hard pruning reduces prefill tokens but permanently excludes information that later questions may require. We propose Raincheck, a framework for deferred multimodal prefill that decouples initial token selection from later information availability. An existing pruning method selects tokens for initial prefill; Raincheck preserves and progressively prefills unselected tokens in source order using unused compute capacity during decode. Unfinished prefill continues after answer return during available scheduling gaps. Attention visibility control and affine state mapping enable context updates after generation begins in full-attention and hybrid recurrent models. Prefill-decode fusion reduces repeated weight reads. We evaluate Qwen3-VL-8B and Qwen3.5-9B with 30% initial multimodal-token retention. Across both models and four pruning methods, Raincheck recovers on average 77.7% and 60.4% of the quality lost to matched hard pruning on MultiVerse and ConvBench, respectively. On MultiVerse, the SGLang implementation reduces mean first-turn TTFT by up to 24.1% relative to full prefill, with less than 0.5% increase in mean time per output token (TPOT).

cs.CL↗
arXiv2026

SpecOPD: Branch-Parallel On-Policy Distillation for Speculative Decoding

Speculative decoding (SD) accelerates large language model inference by reducing the sequential cost of token-by-token generation. Advanced SD methods reduce proposal cost with parallel drafters that predict entire blocks in one forward pass, but overall speedup still depends on acceptance length. On-policy distillation (OPD) can push well-trained drafters beyond acceptance-length plateaus, yet conventional pipelines collect supervision through sequential draft-and-verify rollouts. Our key insight is that on-policy supervision need not inherit these dependencies: a draft block can be supervised using only its reference prefix and draft-sampled history. We propose SpecOPD, a branch-parallel OPD framework that parallelizes draft sampling and target supervision across independent branches drawn from the evolving drafter policy at reference anchors. We pack these branches into proposal trees to share target-prefix computation, using tree masks for full attention, branch-consistent indexer groups for sparse attention, and indexed state replay for linear attention to preserve each branch's history. We train the drafter with an E2E-TV surrogate that directly targets acceptance length. Our study spans 8B full-attention (GQA), 27B full-linear (GQA--GDN), and 125B sparse-linear (QSA--GDN) targets, three parallel draft architectures (DFlash, DSpark, and DFlash V2), and four task domains (dialogue, coding, mathematics, and agentic). SpecOPD achieves up to 41.45% higher mean acceptance length, $16.09\times$ supervised-token throughput over sequential-rollout OPD under native configurations, and up to $3.56\times$ generation speedup over autoregressive decoding.

cs.AI↗
arXiv2026

Hybrid-PCGC: Hybrid Point Cloud Geometry Compression Bridging Generalization-based Model and Implicit Neural Representation

Efficient point cloud geometry compression (PCGC) is essential for 3D sensing applications such as autonomous driving and digital twins. Hybrid methods combine the generalization of learned models with the adaptability of Implicit Neural Representations (INRs), but existing schemes often use complex architectures, rigid update strategies, and post-training weight compression, limiting adaptation efficiency and increasing bitstream overhead. This paper proposes Hybrid-PCGC, which formulates hybrid compression as a joint geometry-and-model bitrate allocation problem and optimizes instance-specific weight offsets as rate-constrained transmitted variables. It separates dataset-agnostic prior modeling from instance-specific adaptation through four components: a Frozen General Module (FGM), a Dual Trainable Adaptive Module (DTAM), an Adaptive Scope Controller (ASC), and a Supervised Offset Compression Module (SOCM). FGM captures global geometric regularities, while the parameter-efficient DTAM adapts to individual point clouds. ASC selects the update scope based on coding difficulty, and SOCM jointly optimizes and compresses weight offsets through differentiable entropy supervision during training. Experiments show that Hybrid-PCGC balances fidelity and bitrate, reducing bitrate by 13.96% and 16.33% over G-PCC on lossless and lossy benchmarks, respectively, and saving up to 12.33% bitrate over the state-of-the-art hybrid method on challenging dense point clouds.

cs.CV↗
arXiv2026

Codec-Compatible Learned Compression for 3D Gaussian Splatting

Video-based 3D Gaussian Splatting (3DGS) compression methods provide a practical route by representing Gaussian attributes as 2D maps and coding them with standard video codecs. Among these maps, high-dimensional non-DC spherical harmonics (SH-AC) maps contain the largest number of channels and account for a dominant portion of the bitstream. However, their statistics differ substantially from those of natural images, and the nonlinear dependencies among SH-AC coefficients remain insufficiently exploited. To address these issues, we propose CCT-GS, a Codec-Compatible learned Transform framework for 3DGS compression. CCT-GS first learns a generalized SH-AC transform that captures cross-scene regularities and converts SH-AC maps into compact codec-friendly bottleneck maps. It further introduces LoRA-based scene adaptation to recover scene-specific appearance details with lightweight side information. Experiments on MPEG GSC sequences demonstrate that CCT-GS consistently outperforms representative video-based baselines, achieving an average BD-rate reduction of 51.30% over the video-based anchor.

eess.IV↗
arXiv2026

Target-Aware Network Dismantling with Limited Budgets

During epidemics, limited resources dictate critical decisions about whom to isolate or monitor. Conventional public health strategies intuitively prioritize known infected individuals and their immediate contacts. Here, we reveal that this contact-centric approach is fundamentally suboptimal under realistic resource constraints. By formulating this challenge as target-aware network dismantling, we uncover a counter-intuitive, budget-dependent intervention transition. Generating-function analysis demonstrates that when intervention budgets are scarce, isolating structurally central individuals, rather than the immediate neighbours of the infected, protects a vastly larger population. The conventional neighbour-targeting strategy only becomes optimal as available resources increase. Because this critical transition point shifts dynamically with the intervention budget, network topology, and spatial distribution of infections, static strategies inevitably fail. To overcome this, we develop Risk-Aware Isolation Learning (RAIL), a deep reinforcement learning framework that adaptively selects optimal interventions based on the evolving residual network. Across diverse synthetic and real-world networks, RAIL consistently maximizes population protection and suppresses outbreak spread. Ultimately, our findings challenge prevailing epidemiological intuitions, establishing a generalizable framework for adaptive interventions against localized risks in public health and other complex systems.

cs.SI↗
arXiv2026

TTRSD: Test-Time Reinforcement Learning with Self-Distillation for Vision-Language Models

Test-time reinforcement learning enables vision-language models (VLMs) to adapt using unlabeled inputs. However, repeated sampling under fixed visual conditions can reinforce shared perceptual errors, while sequence-level rewards fail to isolate visual perception the foundational bottleneck that anchors multimodal reasoning risking the degradation of pre-trained reasoning capabilities. We propose TTRSD, a test-time reinforcement learning framework combining multi-view answer-level self-distillation with visual contrastive token selection. A shared policy aggregates teacher predictions across original, cropped, and downsampled views into an answer distribution. Student trajectories generated from the original image receive rewards based on the support for their final answers in this distribution. To allocate this feedback precisely toward perceptual bottlenecks, we compare the log-probabilities of the same sampled tokens under original and visually ablated inputs while holding their textual prefixes fixed, selecting visually sensitive positions for policy-gradient updates. TTRSD separates update direction, determined by group-relative advantages, from update position, determined by visual sensitivity, without requiring ground-truth labels, external verifiers, or a separate teacher. With only 20 unlabeled adaptation samples, TTRSD improves performance across seven benchmarks and three VLMs, raising InternVL3-2B's MMMU accuracy from 35.79% to 49.32%(+13.53%), demonstrating cross-dataset generalization while preserving inherent reasoning integrity.

cs.CV↗
arXiv2026

Pseudo-Dirac Inelastic Dark Matter in the Leptophobic $U(1)_B$ Model: Confronting the LUX-ZEPLIN High-Recoil Event with Collider Searches

Interpreting the $248$~keV nuclear-recoil candidate reported by LUX-ZEPLIN (LZ) as an endothermic transition needs two ingredients usually put in by hand: a stable dark state, and a coherent current that is purely off diagonal. Gauging baryon number supplies both. For the minimal dark charges $B(χ_L)=-3/2$ and $B(χ_R)=-9/2$, the scalars that generate the dark masses leave an exact residual parity $P=(-1)^{6B}$ stabilizing the lightest state, while the same Majorana masses make the vector current transition dominated as an identity rather than a choice. At $m_χ=1$~TeV and $δ=300$~keV the model is over-determined: the observed LZ interval, a leading-order relic estimate and the LHC dijet limit meet at the single perturbatively acceptable point $(M_{Z_B},g_B)\simeq(1.44~\mathrm{TeV},0.233)$. That point is allowed by currently existing searches but sits just below the trigger-level dijet boundary, so it is alive and not yet decided--and a re-run of that analysis on the Run-3 data at $\sqrt s=13.6$~TeV, already on tape, will exclude it or find it. Nor is the interpretation confined to that splitting: at $δ=200$~keV, the same construction predicts a heavier mediator at comparable coupling, $M_{Z_B}\simeq2.4$-$3.0$~TeV with $g_B\simeq0.15$-$0.34$, a range the HL-LHC dijet program will probe.

hep-ph↗
arXiv2026

Adaptive Finite-Budget Training for CVaR Risk-Aware Q-Learning

Risk-aware Q-learning (RaQL) provides a model-free, two-timescale estimator for dynamic risk objectives, but its finite-budget behavior remains fragile: fixed inner-loop hyperparameters can produce unstable value estimates, persistent Bellman residuals, and inefficient sample reuse. This paper proposes an adaptive training controller for Conditional Value-at-Risk (CVaR) RaQL and evaluates it on a daily Bitcoin trading task. The controller preserves the original CVaR estimator and Bellman fixed point; instead, it redesigns the training procedure through six coordinated mechanisms: per-cell inner-step sizing, outer-rate-matched decay synchronization, a short early correction for the VaR-like inner variable, a coverage-first-then-greedy sample allocation rule, progressive suffix aggregation of mature inner estimates, and data-driven calibration of key scales from online-observable quantities. Across 20 random seeds and 856,000 inner-transition samples, the controller reduces the mean empirical CVaR Bellman residual by approximately 85% relative to the fixed-parameter baseline (MeanBEQ: 1.2202 to 0.1854; MeanBEV: 1.1624 to 0.0535) and maintains stability across CVaR levels, discount factors, and training budgets. On the chronological out-of-sample test set, the learned policy attains a Sharpe ratio of 0.9281 with a maximum drawdown of 6.46% after transaction costs. Although buy-and-hold yields a higher cumulative return (35.43% vs. 23.61%), the adaptive policy achieves far lower volatility (9.57% vs. 47.93%), drawdown, and CVaR loss. These results demonstrate that adaptive finite-budget training design, applied solely to the training procedure without altering the risk objective, can materially improve the reliability and risk-adjusted performance of risk-aware Q-learning in financial applications.

cs.LG↗
arXiv2026

OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs

Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.

cs.AI↗
arXiv2026

Safety-Regulated Transfer Reinforcement Learning with Adaptive Teacher Guidance

We propose Safety-Regulated Adaptive Transfer Reinforcement Learning (SRATRL), a teacher--student framework that combines safety-triggered intervention, safety-adaptive value shaping, and policy-compatibility-based optimization for efficient target-domain adaptation. First, a safety-triggered closed-loop intervention strategy is developed that activates teacher guidance according to the instantaneous safety cost and adaptively adjusts the intervention threshold based on the student policy's recent safety performance, thereby providing timely safety supervision while progressively restoring student autonomy as its safety improves. Next, a safety-adaptive teacher-guided value-shaping scheme is introduced, in which a teacher-consistency signal is incorporated into the critic target, and its contribution is dynamically regulated by the safety-constraint multiplier, enabling stronger teacher guidance under elevated safety risks and gradually weakening such guidance as the safety constraint is better satisfied. In addition, a teacher-student policy-compatibility weighting approach is proposed to alleviate the adverse optimization effects caused by policy mismatch. It reweights teacher-intervened transitions according to the relative likelihood of the executed action under the teacher and student policies, thereby improving policy-update stability. Experimental results demonstrate that compared with a Proximal Policy Optimization with Lagrangian constraint baseline, the proposed method improves the average velocity by 6.90%, and reduces the crash ratio by 75.00%. These results demonstrate that the proposed method can reduce safety costs while maintaining competitive task efficiency.

cs.LG↗
arXiv2026

Shaping Sparse Rewards in Reinforcement Learning: A Semi-supervised Approach

In many real-world scenarios, reward signal for agents are exceedingly sparse, making it challenging to learn an effective reward function for reward shaping. To address this issue, the proposed approach in this paper performs reward shaping not only by utilizing non-zero-reward transitions but also by employing the \emph{Semi-Supervised Learning} (SSL) technique combined with a novel data augmentation to learn trajectory space representations from the majority of transitions, {i.e}., zero-reward transitions, thereby improving the efficacy of reward shaping. Experimental results in Atari and robotic manipulation demonstrate that our method outperforms supervised-based approaches in reward inference, leading to higher agent scores. Notably, in more sparse-reward environments, our method achieves up to twice the peak scores compared to supervised baselines. The proposed double entropy data augmentation enhances performance, showcasing a 15.8\% increase in best score over other augmentation methods

cs.LG↗
arXiv2026

Robust Data-Driven Quasiconcave Optimization

We investigate a data-driven quasiconcave maximization problem where information about the objective function is limited to a finite sample of data points. We begin by defining an ambiguity set for admissible objective functions based on available partial information about the objective. This ambiguity set consists of those quasiconcave functions that majorize a given data sample, and that satisfy additional functional properties (monotonicity, Lipschitz continuity, and permutation invariance). We then formulate a robust optimization (RO) problem which maximizes the worst-case objective function over this ambiguity set. Based on the quasiconcave structure in this problem, we explicitly construct the upper level sets of the worst-case objective at all levels. We can then solve the resulting RO problem efficiently by doing binary search over the upper level sets and solving a logarithmic number of convex feasibility problems. This numerical approach differs from traditional subgradient descent and support function based methods for this problem class. While these methods can be applied in our setting, the binary search method displays superb finite convergence to the global optimum, whereas the others do not. This is primarily because binary search fully exploits the specific structure of the worst-case quasiconcave objective, which leads to an explicit and general convergence rate in terms of the number of convex optimization problems to be solved. Our numerical experiments on a Cobb-Douglas production efficiency problem and a fair resource allocation problem demonstrate the tractability of our approach.

math.OC↗
arXiv2026

DRL-ORA: Distributional Reinforcement Learning with Online Risk Adaption

One of the main challenges in reinforcement learning (RL) is that the agent has to make decisions that would influence the future performance without having complete knowledge of the environment. Dynamically adjusting the level of epistemic risk during the learning process can help to achieve reliable policies in safety-critical settings with better efficiency. In this work, we propose a new framework, Distributional RL with Online Risk Adaptation (DRL-ORA). This framework quantifies both epistemic and implicit aleatory uncertainties in a unified manner and dynamically adjusts the epistemic risk levels by solving a total variation minimization problem online. The framework unifies the existing variants of risk adaption approaches and offers better explainability and flexibility. The selection of risk levels is performed efficiently via a grid search using a Follow-The-Leader-type algorithm, where the offline oracle also corresponds to a ''satisficing measure'' under a specially modified loss function. We show that DRL-ORA outperforms existing methods that rely on fixed risk levels or manually designed risk level adaptation in multiple classes of tasks.

cs.LG↗
arXiv2026

REPAIR: Robust Editing via Progressive Adaptive Intervention and Reintegration

Post-training for large language models (LLMs) is constrained by the high cost of acquiring new knowledge or correcting errors and by the unintended side effects that frequently arise from retraining. To address these issues, we introduce REPAIR (Robust Editing via Progressive Adaptive Intervention and Reintegration), a lifelong editing framework designed to support precise and low-cost model updates while preserving non-target knowledge. REPAIR mitigates the instability and conflicts of large-scale sequential edits through a closed-loop feedback mechanism coupled with dynamic memory management. Furthermore, by incorporating frequent knowledge fusion and enforcing strong locality guards, REPAIR effectively addresses the shortcomings of traditional distribution-agnostic approaches that often overlook unintended ripple effects. Our experiments demonstrate that REPAIR boosts editing accuracy by 10%-30% across multiple model families and significantly reduces knowledge forgetting. This work introduces a robust framework for developing reliable, scalable, and continually evolving LLMs.

cs.CL↗
arXiv2026

Generalizable Trajectory Prediction via Inverse Reinforcement Learning with Mamba-Graph Architecture

Accurate driving behavior modeling is fundamental to safe and efficient trajectory prediction, yet remains challenging in complex traffic scenarios. This paper presents a novel Inverse Reinforcement Learning (IRL) framework that captures human-like decision-making by inferring diverse reward functions, enabling robust cross-scenario adaptability. The learned reward function is utilized to maximize the likelihood of output by integrating Mamba blocks for efficient long-sequence dependency modeling with graph attention networks to encode spatial interactions among traffic agents. Comprehensive evaluations on urban intersections and roundabouts demonstrate that the proposed method not only outperforms various popular approaches in terms of prediction accuracy but also achieves 2.3 times higher generalization performance to unseen scenarios compared to other baselines, achieving adaptability in Out-of-Distribution settings that is competitive with fine-tuning.

cs.LG↗
arXiv2026

Light4GS: Lightweight Compact 4D Gaussian Splatting Generation via Context Model

3D Gaussian Splatting (3DGS) has emerged as an efficient and high-fidelity paradigm for novel view synthesis. To adapt 3DGS for dynamic content, deformable 3DGS incorporates temporally deformable primitives with learnable latent embeddings to capture complex motions. Despite its impressive performance, the high-dimensional embeddings and vast number of primitives lead to substantial storage requirements. In this paper, we introduce a \textbf{Light}weight \textbf{4}D\textbf{GS} framework, called Light4GS, that employs significance pruning with a deep context model to provide a lightweight storage-efficient dynamic 3DGS representation. The proposed Light4GS is based on 4DGS that is a typical representation of deformable 3DGS. Specifically, our framework is built upon two core components: (1) a spatio-temporal significance pruning strategy that eliminates over 64\% of the deformable primitives, followed by an entropy-constrained spherical harmonics compression applied to the remainder; and (2) a deep context model that integrates intra- and inter-prediction with hyperprior into a coarse-to-fine context structure to enable efficient multiscale latent embedding compression. Our approach achieves over 120x compression and increases rendering FPS up to 20\% compared to the baseline 4DGS, and also superior to frame-wise state-of-the-art 3DGS compression methods, revealing the effectiveness of our Light4GS in terms of both intra- and inter-prediction methods without sacrificing rendering quality.

cs.CV↗