Searcharxiv⌕ Search

arXiv subjects

Hailin Zhang

Publications and source records attributed to Hailin Zhang.

At least 19 recordsLinked to original sources
arXiv2026

KnitTrans: Coordinated Trajectory Transfer for Effective and Efficient LLM Reinforcement Learning

Trajectory transfer from rollout to training affects both learning quality and system efficiency in large language model reinforcement learning (RL). Enforcing batch-wide training requirements can delay the admission of completed trajectories, resulting in higher architectural staleness: the additional parameter-version gap from disaggregated pipelining relative to colocated execution. Beyond admission, misalignment between data access and training consumption causes redundant movement and unnecessary waiting when transferring large trajectory payloads. We present KnitTrans, a system that coordinates the rollout-to-training path through three designs. First, constraint-preserving mini-batch admission allows flexibility in early mini-batch composition, while checking that enough capacity remains to satisfy the remaining quotas. Second, on-demand trajectory data access matches storage units to consumer assignments and overlaps micro-batch data preparation with computation. Third, multimodal encoding deduplication supports reuse across a mini-batch while preserving gradient contributions from all backbone micro-batches. Experiments show up to 93% lower architectural staleness than with train-batch admission, narrowing the learning-quality gap between disaggregated execution and colocated execution. On multimodal workloads, KnitTrans reduces transfer latency by two orders of magnitude and achieves a speedup of up to 5.8 times in transfer plus training over the baseline.

cs.DC↗
arXiv2026

Prorsa: Unleashing the Potential of Prefill-Decode Disaggregation in Agentic Reinforcement Learning

Agentic reinforcement learning (RL) is commonly used to enhance the capabilities of large language models (LLMs) in solving complex multi-turn tasks. During rollout, agentic RL exhibits a workload pattern distinct from that of traditional RL: multi-turn interactions repeatedly introduce prefill requests, causing prefill and decode workloads to coexist throughout rollout. This workload makes prefill-decode (PD) disaggregation, a design widely used for online LLM serving, promising for speeding up agentic RL rollout. This paper shows that PD disaggregation can improve rollout efficiency by (a) mitigating interference among data-parallel attention modules, (b) enabling stage-specific parallelism, and (c) reducing end-to-end trajectory completion time. However, a fixed prefill-decode allocation may fail to sustain these benefits as workload composition varies across tasks and over time. We therefore present Prorsa, an agentic RL framework for prediction-guided prefill-decode resource allocation. Prorsa observes the lifecycle of every trajectory to forecast short-horizon prefill and decode demand. These forecasts then guide worker allocation decisions and trigger role switching of rollout engines, aligning prefill and decode capacity with anticipated workload demand. Across synchronous and asynchronous agentic RL settings, Prorsa achieves end-to-end rollout speedups of 1.11x-4.05x over prefill-decode colocation.

cs.DC↗
arXiv2026

MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

Reinforcement learning (RL) is the central training paradigm for advancing large foundation models towards self-improvement. This report introduces the MiMo-V2.6 series, an omni-modal family that pushes the frontier of model intelligence by scaling RL compute. Prior to RL, we conduct mid-training on a broad multimodal corpus to provide ample exploration space, and build a solid infrastructure on the pretrained hybrid-SWA architecture to support subsequent scale-up. We scale RL compute along three dimensions: (1) larger batches and higher throughput, with an asynchronous training that consumes 1,568 samples and 2.7-3.7B tokens per step at context lengths of up to 1M; (2) more diverse and complex environments, spanning code, general, visual, and cyber domains under a mixture of agent harnesses; and (3) more grader compute, via groupwise agentic grading that yields more accurate reward signals for long-horizon tasks and steers the model towards shorter, more token-efficient solutions. To keep training stable at scale, we freeze the MoE router and establish a multi-layer defense against reward hacking. We further build infrastructure for mixed-task agentic RL, including a unified trajectory representation, high-concurrency multi-framework rollout, decoupled control and data planes, and training-inference consistency. We open-source the training dynamics, RL environments, and RL framework to facilitate reproduction and further research on scaled RL and model self-improvement.

cs.CL↗
arXiv2026

Coupling Perception and Reasoning in Federated Multimodal Graph Foundation Models

Federated multimodal graph foundation models (GFMs) aim to adapt pretrained multimodal models to decentralized graph data, where each client owns a private multimodal graph and cannot share raw information. These models typically combine a multimodal Encoder that extracts semantic evidence from heterogeneous modalities and a graph neural network (GNN) that performs relational reasoning over graph structures. However, existing federated GFM adaptation methods mainly update graph-side modules while keeping the multimodal Encoder frozen, limiting adaptation to \emph{how information is propagated} while fixing \emph{what information is extracted}. Through empirical studies, we reveal that Encoder and GNN adaptations are not independent: Encoder adaptation is affected by graph relations, while cross-client module swapping reveals substantial pairing sensitivity between separately parameterized Encoder and GNN updates. Motivated by this observation, we propose \textbf{FedCORE}, a federated adaptation framework that represents Encoder and GNN updates through a shared low-dimensional latent state. FedCORE jointly optimizes this core from multimodal and structural signals and performs federated evolution directly in the shared state space, preserving compatibility between perception and reasoning adaptations. Extensive experiments demonstrate that FedCORE reduces the Encoder--GNN pairing gap from $30.6$ to $5.9$, corresponding to an $80.7\%$ reduction over independent joint adaptation.

cs.LG↗
arXiv2026

HySparse2: Hybrid Sparse Attention with Two-Level KV Sharing

Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. To meet these demands, we introduce HySparse2, a hybrid sparse attention architecture with two-level KV sharing. At the outer level, KV Bridging adopts a YOCO-style self-decoder and cross-decoder structure, but bridges only full-attention layers. The self-decoder uses hybrid sliding-window attention (SWA), while the cross-decoder uses hybrid sparse attention. The KV caches for full-attention layers in the cross-decoder are generated from the hidden states of full-attention layers in the self-decoder. At the inner level, HySparse2 retains HySparse's core KV Reuse design with two refinements. First, it replaces block-level sparsity with token-level sparsity for finer long-context retrieval. Second, it removes the separate SWA branch from sparse layers and instead forces a sliding window of recent tokens into the sparse selection. This two-level KV sharing allows all cross-decoder KV caches to be constructed from self-decoder hidden states. Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers. On an 80B-A3B MoE model, HySparse2 outperforms HySparse and Hybrid SWA on long-context retrieval and multi-turn agentic tasks, while substantially reducing prefill computation and KV-cache storage.

cs.CL↗
arXiv2026

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

Training capable coding agents via reinforcement learning (RL) requires diverse tasks with reliable verifiers. Open-source codebases offer a rich source of such tasks, while existing methods typically rely on development artifacts such as issues and commits, limiting the range of tasks that can be extracted. To better scale RL environments, we present CodeMidas, an agentic pipeline that turns implemented functionality in existing codebases into executable RL environments using source code as its only task-specific input. CodeMidas allocates agentic compute to every stage of environment construction: agents explore implemented functionality to formulate behavioral specifications, construct tests grounded in execution of the original code, and validate and filter candidate tasks through execution checks and repeated solution rollouts. The resulting dataset has 5,545 training tasks from 3,185 open-source codebases spanning 23 programming languages and 15 technical domains. Training MiMo-V2.5 on these tasks with GRPO improves performance on all five diverse benchmarks, covering issue repair (DeepSWE + 11.7%), whole-program construction (ProgramBench +17%), and terminal work (Terminal-Bench v2.1 +8.5%). Ablations show that increasing the number of high-quality training tasks improves performance. Trajectory analysis shows the RL-trained agent demonstrates better behaviors like increasing codebase exploration and more diverse self-verification. These results establish source code as a scalable foundation for constructing RL environments that improve coding agents across diverse software tasks.

cs.AI↗
arXiv2026

Graph-Aware Group Testing with Locally Clustered Infections

Group testing has been widely used to identify infected individuals with a limited number of tests, typically under the assumption of independent infections. Recent studies have exploited correlations among individuals, but often require additional information beyond the contact graph, such as community structures, interaction strengths, or detailed infection dynamics. Such information may be unavailable, incomplete or unreliable in practice. In this work, we assume that only the contact graph is known and develop a graph-aware group testing framework that exploits localized infection clustering in pooling design, fundamental limits of recovery, and decoding. Specifically, we propose an optimal transport-based pooling design that incorporates graph proximity and pooling constraints into a unified optimization framework. We prove that, under mild conditions, the proposed design eliminates uninfected individuals with higher probability than the Bernoulli pooling design, reducing the feasible search space for decoding. Then, we characterize the family of possible infected sets induced by localized infection clustering and derive necessary conditions on the number of tests required for exact recovery, revealing a lower testing requirement than that under the combinatorial prior. For decoding, we model the infection states of the population as a piecewise-constant graph signal and propose a graph total variation regularized decoder. We establish sufficient conditions for exact recovery under the Bernoulli pooling design in both noiseless and noisy settings, and show that O(Klog(n/K)) tests are sufficient under mild conditions in the noiseless case. Extensive simulations on synthetic and real-world networks demonstrate the effectiveness of the proposed framework and the benefit of exploiting graph-induced correlations in group testing.

eess.SP↗
arXiv2026

Robust Recovery of Sparse Support in Constrained Group Testing

In the early stage of a pandemic, rapidly identifying a small number of infected individuals through large-scale screening is critical for pandemic control. Group testing has been widely used to improve testing efficiency and numerous studies have investigated the problem under noisy measurements, typically modeled as bit-flipping of test outcomes. However, these methods do not consider the constraints imposed by dilution, pool size, and the limit of detection (LOD), which can lead to false negatives when the viral load in a pool falls below the LOD. In addition, liquid dispensing errors, common in laboratory settings, affects diluted viral loads in a nonlinear manner. In this work, we introduce a novel measurement model that characterizes the process of sample pooling and dilution, incorporating LOD-induced binary quantization as well as liquid dispensing errors. For the case with known sparsity level, we propose a low-complexity decoding algorithm and provide theoretical guarantees for exact support recovery under both noiseless and noisy settings. For the case with unknown sparsity level, we develop a blind support recovery algorithm, along with a heuristic variant to enhance robustness, which can achieve exact support recovery with only O(klogn) measurements. Extensive simulations show that the proposed algorithms outperform existing combinatorial group testing algorithms, validating the effectiveness, efficiency and robustness in large-scale screening.

eess.SP↗
arXiv2026

An Ultra-Low Profile Metasurface for Enhanced Non-Diffractive OAM Beam Generation Based on Sparse Feed Array

This paper proposes a novel ultra-low profile transmitted metasurface to generate enhanced non-diffractive orbital angular momentum (OAM) beams, employing a sparse feed array (SFA) to create a quasi-plane wave excitation for the first time. Our simulation indicates that with uniform amplitude excitation, the non-diffraction performance of Bessel beam produced by metasurface, surpasses that of conventional single-feed excitation. Based on this principle, a 5 * 5 sparse feed array is introduced and positioned less than one wavelength from the metasurface, ensuring a quasi-uniform amplitude excitation across all units. Further, this metasurface leverages its flexible phase control capability and integrates the spatial phase, OAM phase, and axicon phase to generate an enhanced high-order Bessel beam. The simulated results confirm a successful non-diffractive Bessel beam generation carrying OAM with mode l = +2, exhibiting reduced beam divergence and higher gain. This design also offers benefits of ultra-low profile, high aperture efficiency, low structural complexity.

eess.SP↗
arXiv2026

A Wideband Multibeam Planar Quasi-Yagi Array Antenna Based on Parasitic Pixel Strips

This letter presents a low-complexity parasitic pixel strip (PPS) structure for array antenna to generate six radiations with flexible beam control. To realize multibeam operation, three groups of symmetrical strip-shaped pixels are integrated with the feed antenna, the connection state of which can be controlled by PIN diode. Leveraging the reconfigurable parasitic configuration, the coupling path of the antenna can be strategically manipulated, achieving the desired multibeam property. Thus, by utilizing the PPS structure in conjunction with a planar quasi-Yagi antenna, this single antenna can operate at four states with beam direction towards 40°, -40°, 0°, along with a dual-beam at +/-45°, within the frequency range from 5.0 to 5.6 GHz. Moreover, a prototype of a 1 * 2 array with the PPS structure is also fabricated to achieve six beams with a scan range of 40°, 20°, 0°, -20°, -40° and a dual-beam at +/-30°. Notably, only four PIN diodes are utilized to implement six radiations, verifying that this approach effectively minimizes the excessive use of switches and extends the variety of beams, which provides a low complexity method for multibeam array antenna without complicated beamforming network. These antennas satisfy the benefits of wideband, planner structure, low structural complexity, low cost, and flexible beam control.

eess.SP↗
arXiv2026

Full-Pipeline Inference Optimization for MiMo-V2.5 Series: Pushing Hybrid SWA Efficiency to the Limit

We present a full-pipeline inference optimization for the MiMo-V2.5 model family, which combines Hybrid Sliding Window Attention (Hybrid SWA), sparse Mixture-of-Experts (MoE), and multimodal encoders. While Hybrid SWA can ideally reduce both attention compute and KVCache storage significantly compared to Full Attention, realizing these gains in production requires substantial engineering effort. We systematically optimize the KVCache system with layerwise prefetch, SWA-aware prefix cache trees, and specialized placement strategies, achieving strict $O(W)$ SWA storage and high cache hit rates. We further build GCache, a high-performance distributed cache infrastructure with RDMA-optimized networking, and develop a KVCache-affinity router to reduce computation while preserving load balancing. We also optimize for multimodal inputs, including GPU image preprocessing, parallel video decoding, and multimodal cache sharing. Together, these optimizations constitute the first large-scale LLM serving system in production that efficiently covers the Hybrid SWA + MoE + multimodal composite architecture.

cs.AR↗
arXiv2026

RaBitQCache: Rotated Binary Quantization for KVCache in Long Context LLM Inference

Long-context Large Language Model inference is severely bottlenecked by the massive Key-Value (KV) cache, yet existing sparse attention methods often suffer from static fixed-budget (Top-k) retrieval or rely on proxy scores that are computationally expensive and biased. To address these limitations, we propose RaBitQCache, a novel sparse attention framework that utilizes randomized rotated binary quantization and high-throughput binary-INT4 arithmetic to efficiently estimate attention weights. Our proxy score serves as an unbiased estimator with a proven error bound, enabling adaptive Top-p retrieval that dynamically adjusts the token budget based on actual attention sparsity. We further implement a hardware-aware system with asynchronous pipelining and lazy updates to mask overhead. Evaluations demonstrate that RaBitQCache significantly accelerates inference and reduces memory I/O while preserving generation quality compared to state-of-the-art baselines. Code is available at https://github.com/Sakuraaa0/RaBitQCache.git.

cs.LG↗
arXiv2026

MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training

Modern large language models (LLMs) rely on reinforcement learning during post-training to push specific capabilities, yet integrating multiple capabilities into one model remains hard. Existing methods, such as Off-Policy Finetune and Mix-RL, are either inefficient or lose performance. In this work, we propose Multi-teacher On-Policy Distillation (MOPD), a post-training paradigm for combining the capabilities of multiple domain RL teachers: we first run per-domain specialised RL to obtain a set of domain teachers, then distill these teachers into the student on its own rollouts. This eliminates exposure bias and provides a dense optimization signal. On Qwen3-30B-A3B, MOPD outperforms Mix-RL, Cascade RL, Off-Policy Finetune, and Param-Merge baselines, inheriting nearly all of each teacher's capability. MOPD also enables parallel, independent development of domain teachers, removing the cross-domain coupling typical of multi-domain post-training. MOPD has been deployed in the post-training of MiMo-V2-Flash, an industrial-scale frontier model, demonstrating its practical value for capability integration in frontier-scale LLMs.

cs.CL↗
arXiv2026

ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement Learning

Agentic reinforcement learning (RL) has emerged as a transformative workload in cloud clusters, enabling large language models (LLMs) to solve complex problems through interactions with real world. However, unlike traditional RL, agentic RL demands substantial external cloud resources, e.g., CPUs for code execution and GPUs for reward models, that exist outside the primary training cluster. Existing agentic RL framework typically rely on static over-provisioning, i.e., resources are often tied to long-lived trajectories or isolated by tasks, which leads to severe resource inefficiency. We propose the action-level orchestration, and incorporate it into ARL-Tangram, a unified resource management system that enables fine-grained external resource sharing and elasticity. ARL-Tangram utilizes a unified action-level formulation and an elastic scheduling algorithm to minimize action completion time (ACT) while satisfying heterogeneous resource constraints. Further, heterogeneous resource managers are tailored to efficiently support the action-level execution on resources with heterogeneous characteristics and topologies. Evaluation on real-world agentic RL tasks demonstrates that ARL-Tangram improves average ACT by up to 4.3$\times$, speeds up the step duration of RL training by up to 1.5$\times$, and saves the external resources by up to 71.2$\%$. This system has been deployed to support the training of the MiMo series models.

cs.DC↗
arXiv2026

EEG-Based Brain-LLM Interface for Human Preference Aligned Generation

Large language models (LLMs) are becoming an increasingly important component of human--computer interaction, enabling users to coordinate a wide range of intelligent agents through natural language. While language-based interfaces are powerful and flexible, they implicitly assume that users can reliably produce explicit linguistic input, an assumption that may not hold for users with speech or motor impairments, e.g., Amyotrophic Lateral Sclerosis (ALS). In this work, we investigate whether neural signals can be used as an alternative input to LLMs, particularly to support those socially marginalized or underserved users. We build a simple brain-LLM interface, which uses EEG signals to guide image generation models at test time. Specifically, we first train a classifier to estimate user satisfaction from EEG signals. Its predictions are then incorporated into a test-time scaling (TTS) framework that dynamically adapts model inference using neural feedback collected during user evaluation. The experiments show that EEG can predict user satisfaction, suggesting that neural activity carries information on real-time preference inference. These findings provide a first step toward integrating neural feedback into adaptive language-model inference, and hopefully open up new possibilities for future research on adaptive LLM interaction.

eess.SP↗
arXiv2026

HySparse: A Hybrid Sparse Attention Architecture with Oracle Token Selection and KV Cache Sharing

This work introduces Hybrid Sparse Attention (HySparse), a new architecture that interleaves each full attention layer with several sparse attention layers. While conceptually simple, HySparse strategically derives each sparse layer's token selection and KV caches directly from the preceding full attention layer. This architecture resolves two fundamental limitations of prior sparse attention methods. First, conventional approaches typically rely on additional proxies to predict token importance, introducing extra complexity and potentially suboptimal performance. In contrast, HySparse uses the full attention layer as a precise oracle to identify important tokens. Second, existing sparse attention designs often reduce computation without saving KV cache. HySparse enables sparse attention layers to reuse the full attention KV cache, thereby reducing both computation and memory. We evaluate HySparse on both 7B dense and 80B MoE models. Across all settings, HySparse consistently outperforms both full attention and hybrid SWA baselines. Notably, in the 80B MoE model with 49 total layers, only 5 layers employ full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.

cs.CL↗
arXiv2026

Reinforcement Learning for Chain of Thought Compression with One-Domain-to-All Generalization

Chain-of-thought reasoning in large language models can trigger an "overthinking trap": longer rollouts raise cost and latency yet often yield unreliable accuracy gains. Existing methods use global, static controls that may suppress needed reasoning. We propose mastery-gated, sample-level, soft reinforcement learning compression that penalizes long rollouts only when the model already solves the problem and has produced a shorter rollout. Across benchmarks, it cuts response length by 20-40% with comparable or higher accuracy and generalizes across domains: a model trained on math spontaneously shortens unseen tasks (code, instruction following, general-knowledge QA) without hurting accuracy. We further show two-way transfer between non-agent CoT and tool-use agents: non-agent training reduces SWE-Bench Verified rounds by 13%, while compressing a thinking agent cuts SWE trajectories by 67% tokens and 52% rounds and shortens non-agent outputs by up to 44%. Compression is thus not cosmetic brevity, but an inherent computation policy -- what to keep, and what to forget.

cs.CL↗
arXiv2026

JudgeRLVR: Judge First, Generate Second for Efficient Reasoning

Reinforcement Learning with Verifiable Rewards (RLVR) has become a standard paradigm for reasoning in Large Language Models. However, optimizing solely for final-answer correctness often drives models into aimless, verbose exploration, where they rely on exhaustive trial-and-error tactics rather than structured planning to reach solutions. While heuristic constraints like length penalties can reduce verbosity, they often truncate essential reasoning steps, creating a difficult trade-off between efficiency and verification. In this paper, we argue that discriminative capability is a prerequisite for efficient generation: by learning to distinguish valid solutions, a model can internalize a guidance signal that prunes the search space. We propose JudgeRLVR, a two-stage judge-then-generate paradigm. In the first stage, we train the model to judge solution responses with verifiable answers. In the second stage, we fine-tune the same model with vanilla generating RLVR initialized from the judge. Compared to Vanilla RLVR using the same math-domain training data, JudgeRLVR achieves a better quality--efficiency trade-off for Qwen3-30B-A3B: on in-domain math, it delivers about +3.7 points average accuracy gain with -42\% average generation length; on out-of-domain benchmarks, it delivers about +4.5 points average accuracy improvement, demonstrating enhanced generalization.

cs.CL↗