arXiv subjects

Weihang Chen

Publications and source records attributed to Weihang Chen.

14 recordsLinked to original sources

arXiv2026

ExFold: Unified Expert Folding for Training-Free MoE Prefill-Decode Acceleration

Mixture-of-Experts (MoE) models scale capacity for strong quality while keeping per-token compute bounded through sparse expert activation. Yet low-latency MoE serving is increasingly challenging, because it spans two inference phases with fundamentally different bottlenecks: prefill is dominated by token-wise expert computation, whereas decode is constrained by memory traffic from the batch-wise activated expert set. However, existing training-free acceleration methods optimize only a single resource proxy, either the experts each token executes or the experts a batch activates, and either discard the excluded experts' contribution or leave it only implicitly approximated. In this paper, we propose ExFold, a unified training-free expert-folding framework for jointly accelerating MoE prefill and decode. ExFold casts both prefill and decode as one budgeted output-approximation problem: execute only a phase-specific constrained expert set while projecting the contribution of budget-excluded experts onto retained experts using calibrated scalar projectors. Motivated by the observation that many expert outputs are directionally aligned but differ in magnitude, ExFold calibrates a pairwise scalar-projector matrix on unlabeled data and uses it at inference time to fold excluded expert contributions into retained experts. Under this view, prefill acceleration becomes token-level Top-K folding, and decode acceleration becomes batch-level expert-pool folding. The two phases differ only in how retained experts are selected, while excluded contributions are recovered by one shared folding mechanism. We implement ExFold as a plug-and-play plugin in vLLM, with a lightweight expert-folding CUDA kernel, delivering up to 1.41x TTFT and 2.45x TPOT speedups while retaining about 99% of the original average quality.

cs.LG↗

arXiv2026

Screenshots or Tools? Eliciting Tool Use and Managing Multimodal Context in Hybrid GUI-MCP Computer-Use Agents

Hybrid computer-use agents can act through screenshots or call text tools. We find that having a tool available does not settle which way the effect goes. Under one identical GUI-MCP harness on the OSWorld-MCP benchmark (309 tasks), the same MCP tools improve a reasoning model by +4.0pp and degrade a non-reasoning model by -5.9pp (5 runs each, both beyond 2 SE). What separates the two is tool-decision behavior. The non-reasoning policy ignores, misnames, or falsely terminates around tools. The reasoning model avoids these failures, yet still calls a tool on only 55/309 tasks, 23.9% of the tool-reachable ones. We call this shortfall the adoption gap. Both levels of the problem share one cause: the model already has a cheaper route and is never trained to take it. Multi-turn RL probes that cause. At the action level, a dense tool bonus raises spreadsheet adoption 0.03 -> 0.33 and carries into greedy decoding, but held-out accuracy does not follow. Behavior is steerable; competence is not. The bottleneck lies in tool-call semantics. At the context level, a successful tool call often makes the next screenshot redundant. Dropping it and halving image history cuts input tokens by about a third, at a small accuracy cost. Retraining under the same observation rule removes that cost. The compressed agent then reaches 37.8% against 33.0% for the uncompressed operating point, at 53% of the input cost, and closes the rich-lean gap on a pre-registered degraded subset to zero. Tools help when the model chooses and integrates them, and current hybrid agents leave many such choices unused. Code and checkpoints: https://github.com/redai-infra/hybrid-routing-agent

cs.AI↗

arXiv2026

HYPIC: Accelerating Hybrid-Attention LLM Serving with Position-Independent Caching

In retrieval-augmented generation and agentic LLM serving, prompts are assembled from independent segments into long contexts, making the prefill stage dominate per-request cost. Two directions have emerged to reduce this cost: position-independent caching (PIC) admits KV reuse for non-contiguous segments shared across requests, while hybrid-attention models cut computation by replacing most full-attention layers with linear attention. However, they cannot coexist: applying existing PIC methods to hybrid-attention models breaks down because per-token KV-cache reuse primitives do not transfer to the per-request recurrent state. We present Hypic, the first system to accelerate hybrid-attention LLM serving with position-independent caching. For linear-attention layers, we identify the segment-cumulative transition operator as the missing algebraic primitive and cache it alongside each segment's zero-start end-state, enabling near-exact and constant-time composition of independently cached segments. For the remaining full-attention layers, existing PIC methods also fail because linear layers do not expose the per-token hidden states needed for selective recomputation. We show that the largest deviations concentrate at segment beginnings and construct a small seam window that propagates hidden states through the hybrid-attention stack to repair cross-segment attention. Finally, Hypic introduces segment parallelism, which exploits PIC's segment-level self-containment to parallelize cache-miss prefill across instances, turning long cold requests into an accelerable workload. Evaluated across four hybrid-attention models and five workloads, Hypic reduces time-to-first-token by $3.25\times$ on average and improves QPS by $1.66\times$ over Prefix Cache, while preserving task quality with a 1.71-point gap from Full Recompute.

cs.DC↗

arXiv2026

Pair-In, Pair-Out: Latent Multi-Token Prediction for Efficient LLMs

Long chain-of-thought reasoning has made autoregressive decoding the dominant inference cost of modern large language models. Existing methods target either the input side (latent compression) or the output side (speculative decoding and multi-token prediction, MTP), but the two lines of work have been pursued independently. Moreover, output-side methods must incur an expensive verifier pass to validate the unreliable draft tokens predicted by MTP. To address these issues, we propose \textbf{Pair-In, Pair-Out (PIPO)}, which unifies both sides by viewing a latent compressor and an MTP head as mirror-image operations: the compressor folds two input tokens into one latent representation, while the MTP head unfolds one hidden state into one additional output token. To remove the verifier cost without sacrificing reliability, PIPO trains a lightweight confidence head that decides whether draft tokens should be accepted. We observe that On-Policy Distillation (OPD) naturally matches the rejection-sampling criterion of speculative decoding, so the confidence head can be trained alongside OPD with negligible extra cost. Experiments on AIME 2025, GPQA-Diamond, LiveCodeBench v6, and LongBench v2 with Qwen3.5-4B and 9B backbones show that PIPO improves pass@4 over regular decoding by up to $+7.15$ points, while delivering up to $2.64\times$ first-token-latency and $2.07\times$ per-token-latency speedups. Project Page: GitHub.com/RedAI-Infra/PIPO.

cs.CL↗

arXiv2026

Hint Tuning: Less Data Makes Better Reasoners

Large reasoning models achieve high accuracy through extended chain-of-thought but generate 5--8 more tokens than necessary, applying verbose reasoning uniformly regardless of problem difficulty. We propose Hint Tuning, a data-efficient approach that teaches models to calibrate reasoning depth. Our key insight: the corresponding instruct model serves as an ideal difficulty probe. By testing what the instruct model can solve with varying guidance, we automatically construct training data across three states: No-Hint (direct answer), Sparse-Hint (minimal prefix), and Full-Hint (complete reasoning). This converts the abstract challenge of difficulty labeling into a measurable consistency check between the instruct and reasoning models. With only 1K self-annotated samples, Hint Tuning achieves 24--66% token reduction (31.5% average) across mainstream reasoning models (Qwen3-Thinking, DeepSeek-R1-Distill) at multiple scales (4B--32B) while maintaining competitive accuracy on five benchmarks. Unlike methods requiring massive distillation datasets or expensive RL, we achieve superior efficiency through simple alignment with the instruct model's capabilities. Code and data are available at https://github.com/redai-infra/hint-tuning.

cs.CL↗

arXiv2026

Relax: An Asynchronous Reinforcement Learning Engine for Omni-Modal Post-Training at Scale

Reinforcement learning (RL) post-training has proven effective at unlocking reasoning, self-reflection, and tool-use capabilities in large language models. As models extend to omni-modal inputs and agentic multi-turn workflows, RL training systems face three interdependent challenges: heterogeneous data flows, operational robustness at scale, and the staleness -- throughput tradeoff. We present \textbf{Relax} (Reinforcement Engine Leveraging Agentic X-modality), an open-source RL training engine that addresses these challenges through three co-designed architectural layers. First, an \emph{omni-native architecture} builds multimodal support into the full stack -- from data preprocessing and modality-aware parallelism to inference generation -- rather than retrofitting it onto a text-centric pipeline. Second, each RL role runs as an independent, fault-isolated service that can be scaled, recovered, and upgraded without global coordination. Third, service-level decoupling enables asynchronous training via the TransferQueue data bus, where a single staleness parameter smoothly interpolates among on-policy, near-on-policy, and fully asynchronous execution. Relax achieves a 1.20$\times$ end-to-end speedup over veRL on Qwen3-4B on-policy training. Its fully async mode delivers a 1.76$\times$ speedup over colocate on Qwen3-4B and a 2.00$\times$ speedup on Qwen3-Omni-30B, while all modes converge to the same reward level. Relax supports R3 (Rollout Routing Replay)~\cite{ma2025r3} for MoE models with only 1.9\% overhead, compared to 32\% degradation in veRL under the same configuration. It further demonstrates stable omni-modal RL convergence on Qwen3-Omni across image, text, and audio, sustaining over 2{,}000 steps on video without degradation. Relax is available at https://github.com/rednote-ai/Relax.

cs.CL↗

arXiv2025

Towards Large-scale Generative Ranking

Generative recommendation has recently emerged as a promising paradigm in information retrieval. However, generative ranking systems are still understudied, particularly with respect to their effectiveness and feasibility in large-scale industrial settings. This paper investigates this topic at the ranking stage of Xiaohongshu's Explore Feed, a recommender system that serves hundreds of millions of users. Specifically, we first examine how generative ranking outperforms current industrial recommenders. Through theoretical and empirical analyses, we find that the primary improvement in effectiveness stems from the generative architecture, rather than the training paradigm. To facilitate efficient deployment of generative ranking, we introduce GenRank, a novel generative architecture for ranking. We validate the effectiveness and efficiency of our solution through online A/B experiments. The results show that GenRank achieves significant improvements in user satisfaction with nearly equivalent computational resources compared to the existing production system.

cs.IR↗

arXiv2025

ThinTact:Thin Vision-Based Tactile Sensor by Lensless Imaging

Vision-based tactile sensors have drawn increasing interest in the robotics community. However, traditional lens-based designs impose minimum thickness constraints on these sensors, limiting their applicability in space-restricted settings. In this paper, we propose ThinTact, a novel lensless vision-based tactile sensor with a sensing field of over 200 mm2 and a thickness of less than 10 mm.ThinTact utilizes the mask-based lensless imaging technique to map the contact information to CMOS signals. To ensure real-time tactile sensing, we propose a real-time lensless reconstruction algorithm that leverages a frequency-spatial-domain joint filter based on discrete cosine transform (DCT). This algorithm achieves computation significantly faster than existing optimization-based methods. Additionally, to improve the sensing quality, we develop a mask optimization method based on the generic algorithm and the corresponding system matrix calibration algorithm.We evaluate the performance of our proposed lensless reconstruction and tactile sensing through qualitative and quantitative experiments. Furthermore, we demonstrate ThinTact's practical applicability in diverse applications, including texture recognition and contact-rich object manipulation. The paper will appear in the IEEE Transactions on Robotics: https://ieeexplore.ieee.org/document/10842357. Video: https://youtu.be/YrOO9BDMAHo

cs.RO↗

arXiv2024

Optimizing Personalized Federated Learning through Adaptive Layer-Wise Learning

Real-life deployment of federated Learning (FL) often faces non-IID data, which leads to poor accuracy and slow convergence. Personalized FL (pFL) tackles these issues by tailoring local models to individual data sources and using weighted aggregation methods for client-specific learning. However, existing pFL methods often fail to provide each local model with global knowledge on demand while maintaining low computational overhead. Additionally, local models tend to over-personalize their data during the training process, potentially dropping previously acquired global information. We propose FLAYER, a novel layer-wise learning method for pFL that optimizes local model personalization performance. FLAYER considers the different roles and learning abilities of neural network layers of individual local models. It incorporates global information for each local model as needed to initialize the local model cost-effectively. It then dynamically adjusts learning rates for each layer during local training, optimizing the personalized learning process for each local model while preserving global knowledge. Additionally, to enhance global representation in pFL, FLAYER selectively uploads parameters for global aggregation in a layer-wise manner. We evaluate FLAYER on four representative datasets in computer vision and natural language processing domains. Compared to six state-of-the-art pFL methods, FLAYER improves the inference accuracy, on average, by 5.40\% (up to 14.29\%).

cs.LG↗

arXiv2024

DexSim2Real$^{2}$: Building Explicit World Model for Precise Articulated Object Dexterous Manipulation

Articulated objects are ubiquitous in daily life. In this paper, we present DexSim2Real$^{2}$, a novel framework for goal-conditioned articulated object manipulation. The core of our framework is constructing an explicit world model of unseen articulated objects through active interactions, which enables sampling-based model predictive control to plan trajectories achieving different goals without requiring demonstrations or RL. It first predicts an interaction using an affordance network trained on self-supervised interaction data or videos of human manipulation. After executing the interactions on the real robot to move the object parts, we propose a novel modeling pipeline based on 3D AIGC to build a digital twin of the object in simulation from multiple frames of observations. For dexterous hands, we utilize eigengrasp to reduce the action dimension, enabling more efficient trajectory searching. Experiments validate the framework's effectiveness for precise manipulation using a suction gripper, a two-finger gripper and two dexterous hand. The generalizability of the explicit world model also enables advanced manipulation strategies like manipulating with tools.

cs.RO↗

arXiv2023

Enhancing Generalizable 6D Pose Tracking of an In-Hand Object with Tactile Sensing

When manipulating an object to accomplish complex tasks, humans rely on both vision and touch to keep track of the object's 6D pose. However, most existing object pose tracking systems in robotics rely exclusively on visual signals, which hinder a robot's ability to manipulate objects effectively. To address this limitation, we introduce TEG-Track, a tactile-enhanced 6D pose tracking system that can track previously unseen objects held in hand. From consecutive tactile signals, TEG-Track optimizes object velocities from marker flows when slippage does not occur, or regresses velocities using a slippage estimation network when slippage is detected. The estimated object velocities are integrated into a geometric-kinematic optimization scheme to enhance existing visual pose trackers. To evaluate our method and to facilitate future research, we construct a real-world dataset for visual-tactile in-hand object pose tracking. Experimental results demonstrate that TEG-Track consistently enhances state-of-the-art generalizable 6D pose trackers in synthetic and real-world scenarios. Our code and dataset are available at https://github.com/leolyliu/TEG-Track.

cs.CV↗

arXiv2023

TransTouch: Learning Transparent Objects Depth Sensing Through Sparse Touches

Transparent objects are common in daily life. However, depth sensing for transparent objects remains a challenging problem. While learning-based methods can leverage shape priors to improve the sensing quality, the labor-intensive data collection in the real world and the sim-to-real domain gap restrict these methods' scalability. In this paper, we propose a method to finetune a stereo network with sparse depth labels automatically collected using a probing system with tactile feedback. We present a novel utility function to evaluate the benefit of touches. By approximating and optimizing the utility function, we can optimize the probing locations given a fixed touching budget to better improve the network's performance on real objects. We further combine tactile depth supervision with a confidence-based regularization to prevent over-fitting during finetuning. To evaluate the effectiveness of our method, we construct a real-world dataset including both diffuse and transparent objects. Experimental results on this dataset show that our method can significantly improve real-world depth sensing accuracy, especially for transparent objects.

cs.RO↗

arXiv2023

Sim2Real$^2$: Actively Building Explicit Physics Model for Precise Articulated Object Manipulation

Accurately manipulating articulated objects is a challenging yet important task for real robot applications. In this paper, we present a novel framework called Sim2Real$^2$ to enable the robot to manipulate an unseen articulated object to the desired state precisely in the real world with no human demonstrations. We leverage recent advances in physics simulation and learning-based perception to build the interactive explicit physics model of the object and use it to plan a long-horizon manipulation trajectory to accomplish the task. However, the interactive model cannot be correctly estimated from a static observation. Therefore, we learn to predict the object affordance from a single-frame point cloud, control the robot to actively interact with the object with a one-step action, and capture another point cloud. Further, the physics model is constructed from the two point clouds. Experimental results show that our framework achieves about 70% manipulations with <30% relative error for common articulated objects, and 30% manipulations for difficult objects. Our proposed framework also enables advanced manipulation strategies, such as manipulating with different tools. Code and videos are available on our project webpage: https://ttimelord.github.io/Sim2Real2-site/

cs.RO↗

arXiv2018

Unsupervised Cross-dataset Person Re-identification by Transfer Learning of Spatial-Temporal Patterns

Most of the proposed person re-identification algorithms conduct supervised training and testing on single labeled datasets with small size, so directly deploying these trained models to a large-scale real-world camera network may lead to poor performance due to underfitting. It is challenging to incrementally optimize the models by using the abundant unlabeled data collected from the target domain. To address this challenge, we propose an unsupervised incremental learning algorithm, TFusion, which is aided by the transfer learning of the pedestrians' spatio-temporal patterns in the target domain. Specifically, the algorithm firstly transfers the visual classifier trained from small labeled source dataset to the unlabeled target dataset so as to learn the pedestrians' spatial-temporal patterns. Secondly, a Bayesian fusion model is proposed to combine the learned spatio-temporal patterns with visual features to achieve a significantly improved classifier. Finally, we propose a learning-to-rank based mutual promotion procedure to incrementally optimize the classifiers based on the unlabeled data in the target domain. Comprehensive experiments based on multiple real surveillance datasets are conducted, and the results show that our algorithm gains significant improvement compared with the state-of-art cross-dataset unsupervised person re-identification algorithms.

cs.CV↗