arXiv subjects

Zheng Li

Publications and source records attributed to Zheng Li.

At least 73 records · Page 4Linked to original sources

arXiv2026

A Systematic Analysis of Hybrid Linear Attention

Transformers face quadratic complexity and memory issues with long sequences, prompting the adoption of linear attention mechanisms using fixed-size hidden states. However, linear models often suffer from limited recall performance, leading to hybrid architectures that combine linear and full attention layers. Despite extensive hybrid architecture research, the choice of linear attention component has not been deeply explored. We systematically evaluate various linear attention models across generations - vector recurrences to advanced gating mechanisms - both standalone and hybridized. To enable this comprehensive analysis, we trained and open-sourced 72 models: 36 at 340M parameters (20B tokens) and 36 at 1.3B parameters (100B tokens), covering six linear attention variants across five hybridization ratios. Benchmarking on standard language modeling and recall tasks reveals that superior standalone linear models do not necessarily excel in hybrids. While language modeling remains stable across linear-to-full attention ratios, recall significantly improves with increased full attention layers, particularly below a 3:1 ratio. Our study highlights selective gating, hierarchical recurrence, and controlled forgetting as critical for effective hybrid models. We recommend architectures such as HGRN-2 or GatedDeltaNet with a linear-to-full ratio between 3:1 and 6:1 to achieve Transformer-level recall efficiently. Our models are open-sourced at https://huggingface.co/collections/m-a-p/hybrid-linear-attention-research-686c488a63d609d2f20e2b1e.

cs.CL↗

arXiv2026

Latency-Aware Service Placement using Neural Combinatorial Optimisers for Edge--Cloud Systems

The growth of Internet of Things (IoT) applications and latency-sensitive services has increased the demand for efficient service placement across compute continuum platforms, such as edge--cloud systems. Modern applications are decomposed into interdependent microservices deployed over heterogeneous infrastructures, making placement under resource and network constraints an intractable NP-hard combinatorial optimisation problem. This study proposes a latency-aware Edge Placement Neural Combinatorial Optimiser (EP-NCO), a learning-based framework for service placement in compute continuum platforms. EP-NCO employs a dual-graph model to capture resource relationships and service dependencies within both computing infrastructure and application structure. Graph neural networks (GNNs) learn structural embeddings of infrastructure nodes and service components, whereas reinforcement learning policies construct feasible placements that account for execution latency, communication link delays, and bandwidth-sharing effects. Extensive simulations across multiple system scales demonstrate that EP-NCO consistently achieves high-quality placement decisions, reducing the total service response time by 46%--50% compared with metaheuristics (genetic algorithm and particle swarm optimisation) and by 25%--35% compared with controlled RL ablation baselines. Once trained, EP-NCO enables fast online inference, making it a practical solution for dynamic large-scale edge--cloud environments with hundreds of computing nodes, hosting thousands of applications, which is significantly beyond the capability of current scheduling systems.

cs.DC↗

arXiv2026

ProMed: Shapley Information Gain Guided Reinforcement Learning for Proactive Medical LLMs

Interactive medical questioning is essential in clinical consultations, where physicians must actively gather necessary patient information. Yet existing medical Large Language Models (LLMs) predominantly follow a reactive paradigm, risking diagnostic errors by answering before seeking sufficient details. To bridge this gap, we propose ProMed, a reinforcement learning framework that transitions LLMs toward a proactive paradigm, enabling them to ask clinically valuable questions before decision-making. Central to ProMed is the Shapley Information Gain (SIG) reward, which quantifies a question's clinical utility as the amount of newly acquired information, while considering its contextual importance via Shapley values. We integrate SIG into a two-stage training pipeline: (1) SIG-Guided Model Initialization uses Monte Carlo Tree Search to construct high-reward interaction trajectories for supervision, and (2) SIG-Augmented Policy Optimization, with a novel SIG-guided Reward Distribution Mechanism that prioritizes informative questions for fine-grained optimization. Experiments on partial-information medical benchmarks show that ProMed significantly outperforms state-of-the-art methods by 6.29% on average and delivers a 54.45% gain over the reactive paradigm, and generalizes robustly to out-of-domain cases. Our codes are available at https://github.com/hxxding/ProMed.

cs.CL↗

arXiv2026

EvoRubrics: Dynamic Rubrics as Rewards via Adversarial Co-Evolution for LLM Reinforcement Learning

Rubric-based rewards offer interpretable and fine-grained optimization signals for reinforcement learning in open-ended tasks where verifiable answers are unavailable. However, pre-constructed rubrics remain static throughout training, creating a fundamental mismatch with the evolving policy: fixed criteria gradually lose discriminative power as the model improves, leading to reward saturation and potential hacking. Recent dynamic rubric methods partially address this but rely on external frontier models or ground-truth answers, and update rubrics only at coarse granularity. We propose EvoRubrics, a co-evolutionary RL framework where a Policy LLM and a Rubric Generator jointly improve through adversarial interaction within each training step. As the policy improves under the rubric generator's guidance, the rubric generator adapts its criteria to remain discriminative and informative, enabling evaluation to track the policy in real time and naturally inducing an automatic curriculum. Experiments show that EvoRubrics consistently outperforms static and dynamic rubric baselines across benchmarks. The learned Rubric Generator further generalizes as a transferable reward model. Notably, even a fully self-supervised variant without any external supervision achieves meaningful gains, suggesting that co-evolution between generation and evaluation alone can provide sufficiently rich learning signals. Our code is publicly available at https://anonymous.4open.science/r/EvoRubrics-2155/.

cs.LG↗

arXiv2026

FILIC: Dual-Loop Force-Guided Imitation Learning with Impedance Torque Control for Contact-Rich Manipulation Tasks

Many contact-rich manipulation tasks require precise force regulation. However, most imitation learning (IL) policies remain position-centric and lack explicit force awareness, and adding force/torque sensors to collaborative robot arms is often costly and requires additional hardware design. To overcome these issues, we propose FILIC, a Force-guided Imitation Learning framework with impedance torque control. FILIC integrates a Transformer-based IL policy with an impedance controller in a dual-loop structure, enabling compliant force-informed, force-executed manipulation. For robots without force/torque sensors, we introduce a cost-effective end-effector force estimator using joint torque measurements through analytical Jacobian-based inversion while compensating with model-predicted torques from a digital twin. Experiments show that FILIC significantly outperforms vision-only and joint-torque-based methods, achieving safer, more compliant, and adaptable contact-rich manipulation. The source code is available at https://github.com/OpenGHz/mujoco-wrench-estimator.git.

cs.RO↗

arXiv2026

The Moving Target of Urban Equity: Spatiotemporal Demand and Double Disadvantage in Hefei, China

Equitable access to essential urban services is a pillar of modern planning, yet most accessibility models rely strictly on static residential locations, ignoring how demand shifts throughout the daily loop. This study introduces a population-based, temporally differentiated framework to examine the resulting "moving target" of urban equity, focusing on medical facilities and green spaces in Hefei, China. Utilising large-scale mobile phone GPS data, we construct dynamic residential and workplace population exposure surfaces to capture shifting hourly demand. We then evaluate accessibility via network-based travel times paired with a novel per-capita provision metric that accounts for real-time demand competition. We define \textit{double disadvantage} as the co-occurrence of poor spatial accessibility and insufficient per-capita service availability. Counterintuitively, the results reveal that double-disadvantaged areas cluster primarily along the inner suburban belt rather than the remote periphery, where per-capita service provision remains relatively sufficient. Furthermore, temporal shifts drastically alter equity landscapes: daytime workplace concentrations intensely exacerbate demand competition in urban job centres. These findings demonstrate that urban inequality depends heavily on spatiotemporal population flows rather than just the fixed location of services. Ultimately, achieving true urban equity requires dynamic planning interventions that address time-varying demand rather than focusing solely on static, home-based metrics.

physics.soc-ph↗

arXiv2026

Quantization Meets Spikes: Nearly Lossless Conversion at the First Timestep via Polarity Multi-Spike Mapping

Spiking neural networks (SNNs) offer advantages in computational efficiency via event-driven computing, compared to traditional artificial neural networks (ANNs). While direct training methods tackle the challenge of non-differentiable activation mechanisms in SNNs, they often suffer from high computational and energy costs during training. As a result, ANN-to-SNN conversion approach remains a valuable and practical alternative. These conversion-based methods aim to leverage the discrete output produced by the quantization layer to obtain SNNs with low latency. Although the theoretical minimum latency is one timestep, existing conversion methods have struggled to realize such ultra-low latency without accuracy loss. Moreover, current quantization approaches often discard negative-value information following batch normalization and are highly sensitive to the hyperparameter configuration, leading to degraded performance. In this work, we, for the first time, analyze the information loss introduced by quantization layers through the lens of information entropy. Building on our analysis, we introduce polarity multi-spike mapping (PMSM) framework and a hyperparameter initialization strategy tailored for the quantization layer. Our method achieves nearly lossless ANN-to-SNN conversion at the extremity, i.e., the first timestep, while also leveraging the temporal dynamics of SNNs across multiple timesteps to maintain stable performance on complex tasks. Extensive experiments on six image and neuromorphic datasets consistently demonstrate that PMSM achieves nearly lossless accuracy at the first timestep. Remarkably, despite operating under ultra-low-latency constraints, PMSM surpasses state-of-the-art direct training methods on multiple benchmarks.

cs.NE↗

arXiv2026

The Benchmark Illusion: Pruned LLMs Can Pass Multiple Choice but Fail to Answer

Compressing large language models reduces memory use and inference cost, but it can also create failures that standard benchmarks miss. A pruned model may still perform well on multiple-choice evaluations, yet fail to answer the same question in open generation. We ask what pruning changes: does it erase the correct answer, or does it make the answer harder to produce as the top output? We study this question with multilingual question answering, tracking the same questions before and after pruning. We find a benchmark illusion. Under high-sparsity pruning, especially Wanda, models often fail in greedy open generation while still selecting the correct answer under multiple-choice scoring. In these recognition-only errors, the answer is usually not gone, but demoted: it often reappears with beam search, sampling, or one in-context example. Overall, multiple-choice benchmarks can overstate the usability of compressed LLMs, creating an evaluation blind spot. Compressed models should be tested on what they can produce, not only on what they can recognize.

cs.CL↗

arXiv2026

Instant Personalized Large Language Model Adaptation via Hypernetwork

Personalized large language models (LLMs) tailor content to individual preferences using user profiles or histories. However, existing parameter-efficient fine-tuning (PEFT) methods, such as the ``One-PEFT-Per-User'' (OPPU) paradigm, require training a separate adapter for each user, making them computationally expensive and impractical for real-time updates. We introduce Profile-to-PEFT, a scalable framework that employs a hypernetwork, trained end-to-end, to map a user's encoded profile directly to a full set of adapter parameters (e.g., LoRA), eliminating per-user training at deployment. This design enables instant adaptation, generalization to unseen users, and privacy-preserving local deployment. Experimental results demonstrate that our method outperforms both prompt-based personalization and OPPU while using substantially fewer computational resources at deployment. The framework exhibits strong generalization to out-of-distribution users and maintains robustness across varying user activity levels and different embedding backbones. The proposed Profile-to-PEFT framework enables efficient, scalable, and adaptive LLM personalization suitable for large-scale applications.

cs.CL↗

arXiv2026

UniLab: A Heterogeneous Architecture for Robot RL Beyond GPU-Dominant Paradigms

Simulation-based RL for contemporary robot control is increasingly organized around GPU-resident simulation: physics, rollout collection, and learning are placed on a single GPU-centric execution path. This paradigm has greatly improved training speed, but it has also encouraged a default assumption that efficient training requires physics to reside on the GPU. We revisit this assumption. Our view is that, in simulation-dominated robot control, the essential question is not which processor runs physics, but whether simulation throughput, policy learning, and runtime synchronization form an efficient end-to-end loop. We present UniLab, a heterogeneous CPU-simulation / GPU-learning architecture that decouples CPU-parallel simulation from GPU policy updates through a unified runtime for data movement, buffering, and synchronization. UniLab is implemented as a complete and extensible training system using MuJoCoUni and MotrixSim CPU-batched physics backends, supporting PPO, FastSAC, FlashSAC, and APPO. On representative simulation-based robot control tasks, UniLab improves end-to-end training efficiency by 3--10$\times$ under the same hardware configuration, while reducing dependence on the NVIDIA CUDA-based software stack and supporting cross-platform execution on the Apple macOS platform and the AMD ROCm and Intel XPU accelerator backends. These results show that GPU simulation is an effective path to efficient training, but not a necessary one, broadening the practical system choices available for robot RL training. Project page: https://unilabsim.github.io.

cs.RO↗

arXiv2026

DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding

Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. Predicting an entire block at once requires a sufficiently capable draft model and effective utilization of the target model's internal knowledge. However, the state-of-the-art method DFlash constrains all draft layers to share a single fused representation derived from only a few target layers, limiting per-layer expressiveness and hindering further scaling of draft capacity. In this paper, we present \modelname, which flares out the narrow conditioning bottleneck of DFlash through a lightweight layer-wise fusion mechanism: each draft layer attends to its own learnable combination of a broad set of target layers at negligible overhead, simultaneously injecting richer target knowledge and providing every draft layer with a distinct input. This enhanced per-layer expressiveness enables scaling the draft model to deeper architectures with consistent gains. We further scale training data from 800K to 2.4M samples to fully exploit the enlarged capacity. On six benchmarks spanning mathematical reasoning, code generation, and conversation, \modelname attains average wall-clock speedups of 5.52x on Qwen3-4B, 5.46x on Qwen3-8B, and 3.91x on GPT-OSS-20B, improving over DFlash by roughly 11\%, 8\%, and 5\% respectively. Our code is available at https://github.com/Tencent/AngelSlim.

cs.CL↗

arXiv2026

ProductWebGen: Benchmarking Multimodal Product Webpage Generation

Crafting a product display webpage from a source product image, along with layout and visual content instructions, holds significant practical value for domains such as marketing, advertising, and E-commerce. Intuitively, this task demands strict visual consistency across product displays and high-fidelity instruction following to jointly generate renderable HTML code. These requirements on controllability and instruction-following are closely aligned with the core features of advanced multimodal generative models, such as image editing models and unified models. To this end, this paper introduces ProductWebGen to systematically benchmark the product webpage generation capacities of these models. We organize ProductWebGen with 500 test samples covering 13 product categories; each sample consists of a source image, a visual content instruction, and a webpage instruction. The task is to generate a product showcase webpage including multiple consistent images in accordance with the source image and instructions. Given the mixed-modality input-output nature of the task, we design and systematically compare two workflows for evaluation -- one uses large language models and image editing models to separately generate HTML code and images (editing-based), while the other relies on a single UM to generate both, with image generation conditioned on the preceding multimodal context (UM-based). Empirical results show that editing-based approaches achieve leading results in webpage instruction following and content appeal, while UM-based ones may display more advantages in fulfilling visual content instructions. We also construct a supervised fine-tuning dataset, ProductWebGen-1k, with 1,000 groups of real product images and LLM-generated HTML code. We verify its effectiveness on the open-source UM BAGEL. The data and code are available at https://github.com/SJTU-DENG-Lab/ProductWebGen.

cs.CV↗

arXiv2026

Can BEV Perception Gracefully Degrade under Sensor Failures?

Despite the remarkable success of multi-modal bird's-eye view (BEV) perception in autonomous driving, current systems exhibit a critical vulnerability: existing fusion mechanisms are highly brittle to sensor corruptions, often causing catastrophic performance degradation. This vulnerability largely stems from the fact that standard fusion frameworks typically integrate multi-modal representations in a static manner, leading to a precipitous performance collapse under missing or corrupted modalities. In contrast, we show that graceful degradation is achievable through active modality reliability assessment. To this end, we present Grace-BEV, a lightweight and plug-and-play framework that enforces active reliability awareness during multi-modal fusion. Instead of relying on computationally expensive cross-modal interactions, Grace-BEV leverages the aligned BEV space to explicitly assess modality trustworthiness via a TrustGate Router and dynamically recalibrate feature integration using the FailSafe Fusion Block. Furthermore, we devise a Three-Phase Training strategy with Modality Dropout to prevent modality dominance and encourage balanced cross-modal learning under unreliable inputs. Extensive experiments on nuScenes-R and nuScenes-C show that Grace-BEV maintains robust performance across diverse corruption settings. Notably, under catastrophic LiDAR failures where standard baselines collapse to 0.0% mean Average Precision (mAP), Grace-BEV restores performance to as high as 34.7% mAP. Moreover, it improves clean accuracy by up to 1.4%, achieving a strong trade-off between robustness and efficiency.

cs.CV↗

arXiv2026

Topological Lifshitz transition-induced bipolarity of anomalous Nernst effect in kagome magnet YCo3

The kagome lattice, renowned for hosting topological band structures and rich magnetic behaviors, offers an exceptional setting to investigate unconventional transport in magnetic topological systems. Controlling the polarity of the anomalous Nernst effect (ANE) is crucial for designing flexible thermoelectric devices, such as thermopiles, where the ability to switch the thermoelectric voltage sign can dramatically enhance energy conversion efficiency and output. Here, we demonstrate such a bipolar ANE in the kagome magnet YCo3, driven by a temperature-induced topological Lifshitz transition. With a Curie temperature TC~225 K, sizable anomalous Hall and Nernst effects emerge below TC. Supported by the first-principles calculations, the AHE and ANE are suggested to be dominated by the intrinsic mechanism. Furthermore, the intrinsic anomalous Hall conductivity exhibits a piecewise-linear dependence on magnetization, with an abrupt slope change near 100 K, consistent with the Karplus-Luttinger mechanism. Concurrently, the anomalous Nernst coefficient SAyx reverses its sign around the same temperature, realizing the crucial bipolarity. These anomalies could be interpreted as a topological Lifshitz transition, enabled by the evolution of Co moments that could shift the Fermi level relative to Weyl nodes. Our work reveals YCo3 as a prototypical kagome magnet where temperature and magnetism directly govern both Weyl node topology and the bipolar ANE, opening a pathway to magnetically control thermoelectric output in topological quantum materials.

cond-mat.mtrl-sci↗

arXiv2026

IFMTBench: A Comprehensive Benchmark for Multilingual Translation Instruction Following

Modern translation workflows demand more than semantic equivalence. Users routinely require models to preserve JSON or HTML schemas, honor curated glossaries, disambiguate with provided context, and match prescribed registers, often several at once. Conventional metrics such as BLEU and xCOMET capture semantic fidelity but provide little signal on constraint adherence, while general instruction following benchmarks ignore the cross-lingual nature of translation. We introduce \bench, a benchmark for multilingual translation instruction following covering seven languages, with 4,506 single-constraint and 2,838 multi-constraint items spanning six constraint dimensions and five compositional patterns with instructions issued in all seven languages. Constraints are split into a gating subset verified by deterministic checkers and a continuous subset scored by a rubric-based LLM judge, combined under a multiplicative rule that resists reward hacking. Evaluating 15 models reveals systematic gaps that prior protocols miss: Instruction following scales with size more sharply than translation quality, glossary and structured-format constraints dominate the difficulty gradient, and general instruction following rankings correlate only weakly with translation behavior. Our benchmark are available at https://github.com/Tencent-Hunyuan/Hy-MT2/tree/main/IFMTBench.

cs.CL↗

arXiv2026

HardMTBench: Stress-Testing Chinese-English Translation on Knowledge-Intensive Domains

General-purpose machine translation benchmarks such as FLORES-200 have reached a saturation regime on Chinese-English pairs, where modern large language models cluster within a narrow band of high scores. Across 22 systems, FLORES-200 zh-en GEMBA scores fall in a 7.87-point range with a standard deviation of 2.29, which compresses the separation between systems on knowledge-intensive domains such as finance, healthcare, law, and science and technology. We introduce HardMTBench, a difficulty-aware diagnostic benchmark for bidirectional Chinese-English domain translation. HardMTBench covers 12 domains and contains 10,000 hand-curated source sentences with reference translations, packaged as 20,000 directional test items. A three-stage construction pipeline builds a domain-balanced candidate pool of 84{,}566 pairs, applies an LLM-based multi-signal judge over knowledge density, translation difficulty, terminology load and reference correctness, and assembles the final test set under a hardness fusion rule with per-domain quotas. Across 22 systems spanning general LLMs, commercial engines and specialised MT models, HardMTBench widens the cross-system GEMBA range by roughly a factor of two over FLORES-200, induces visible rank reorderings, and exposes domain-specific terminology and knowledge weaknesses that quality-only metrics tend to flatten. All data and code are open-sourced at https://github.com/jasonNLP/HardMTBench.

cs.CL↗

arXiv2026

SEAL: Self-Evolving Agentic Learning for Conversational Question Answering over Knowledge Graphs

Knowledge-based conversational question answering (KBCQA) confronts persistent challenges in resolving coreference, modeling contextual dependencies, and executing complex logical reasoning. Existing approaches often suffer from inaccuracies and prohibitive computational costs, particularly when processing intricate queries over large knowledge graphs. Specifically, large language models (LLMs) tend to generate syntactically invalid or semantically misaligned logical forms for complex multi-hop or aggregation queries, while conventional entity-relation linking methods face an exponentially growing candidate space. To address these limitations, we introduce SEAL, a novel two-stage semantic parsing framework grounded in self-evolving agentic learning. In the first stage, an LLM extracts a minimal S-expression core capturing the essential semantics, which is then refined by an agentic calibration module to correct syntactic inconsistencies and align entities and relations with the knowledge graph. The second stage employs template-based completion guided by question-type prediction to construct a fully executable S-expression. Crucially, SEAL incorporates a self-evolving mechanism integrating local and global memory with a reflection module, enabling continuous adaptation from dialog history and execution feedback without explicit retraining. Extensive experiments on the SPICE benchmark demonstrate that SEAL achieves state-of-the-art performance in multi-hop reasoning, comparison, and aggregation tasks, validating notable gains in both structural accuracy and computational efficiency.

cs.CL↗

arXiv2026

Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild

Hy-MT2 is a family of fast-thinking multilingual translation models designed for complex real-world scenarios. It includes three model sizes: 1.8B, 7B, and 30B-A3B (MoE), all of which support translation among 33 languages and effectively follow translation instructions in multiple languages. Multi-dimensional evaluations show that Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. The 7B and 30B models outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B model also surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall. Moreover, when paired with AngelSlim's 1.25-bit extreme quantization for on-device deployment, the lightweight 1.8B model requires only 440 MB of storage and achieves a 1.5x inference speedup.

cs.CL↗