arXiv subjects

Minghao Li

Publications and source records attributed to Minghao Li.

At least 19 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

Element-Aware Group Learning for E-Commerce Image Generation

Recent advances in image generation and editing have made prompt quality a key bottleneck for e-commerce creatives. Vision-language models (VLMs) can generate image-editing prompts from product images and metadata, but further improving their prompt-writing capabilities requires post-training with feedback from the generated images. Group Relative Policy Optimization (GRPO) is a natural framework for such outcome-level reward optimization. However, it assigns credit only at the full-prompt level, even though image quality often depends on specific design elements such as composition, background, and the presentation of selling points. Existing fine-grained credit assignment methods typically require step-level supervision or learned critics. To address this, we propose EAGLE-GRPO (Element-Aware Group Learning for E-Commerce Image Generation), which decomposes the group-centered reward over predefined elements. We cast element-level credit assignment as a kernel ridge regression problem and derive a closed-form solution, without additional rollouts or separate credit-assignment models. This yields interpretable per-element advantages and more precise policy updates. Experiments show that EAGLE-GRPO sustains performance gains over more training steps before plateauing and generates prompts that produce higher-quality e-commerce images than competitive VLM prompt-writing baselines.

cs.CV↗

arXiv2026

Before Agents Speak: Pre-hoc Failure Risk Inference in Multi-Agent Systems

LLM-based multi-agent systems (MAS) have exhibited remarkable capabilities in collaborative reasoning and decision-making, yet their interconnected communications introduce new systemic risk: localized hallucinations can propagate along agent communication chain, amplify through interactions, and ultimately trigger cascading failures. Existing countermeasures predominantly follow a post-hoc paradigm, identifying failures only after unsafe behaviors emerge, by which time harmful effects may have already spread throughout the agent network. To tackle this problem, we investigate a complementary pre-hoc approach and propose HalluProp, a Propagation-aware Hallucination inference framework that estimates individual agent failures and emergent system-level hallucination risks before inter-agent interaction. First, we model intrinsic hallucination risks by identifying fine-grained semantic misalignment between agent roles and task queries. We then characterize inter-agent risk propagation by modeling both semantic influence and communication topology. Finally, we integrate these two risks via a differentiable Noisy-OR inference mechanism to derive a systemic diagnosis. Extensive experiments show that HalluProp accurately localizes faulty agents, achieving an average AUROC of 84.6%, while enabling sub-second diagnosis with over $65\times$ speedup over post-hoc methods. By facilitating early intervention through upstream screening, HalluProp effectively complements post-hoc methods, highlighting the potential of pre-hoc risk inference for building more reliable multi-agent systems.

cs.CR↗

arXiv2026

The Model in the Middle: Toward AI-Native Real-Time Communication

Full-duplex omni models are transforming human--AI interaction from turn-based exchanges into continuous multimodal conversations in which speaking, listening, and reasoning unfold concurrently. Rather than viewing the model as a replacement for a human endpoint, we argue for a new perspective: the model is a stateful computational middlebox inside a human-centered feedback loop, with network transport, model serving, and user playback jointly shaping how the interaction evolves. This perspective breaks the traditional boundaries among stages designed around local objectives. Rather than optimizing them in isolation, an AI-native real-time stack should allow the state of each stage to shape the actions of the others. We explore three cross-stage coordination opportunities: network-aware inference scheduling, execution-aware transport prioritization, and playback control that accounts for both network and model variability. We are building Conflux to explore these ideas, and preliminary results show substantial improvements in response latency and playback deadline adherence under network degradation. More broadly, we call for an AI-native real-time communication stack that resolve the joint control problem spanning communication, computation, and playback.

cs.NI↗

arXiv2026

Towards a Systems Foundation for Agentic Cloud Management

Agentic cloud management is emerging as a practice to automate laborious operations, minimize toil, and improve responsiveness. Despite the rapid development of autonomous management agents, we argue that the fundamental missing piece is a systems foundation to enable safe, effective operations across agents and between agents and human operators. In this paper, we advocate for the need of such a systems foundation and share our efforts on developing CloudWeaver, an agentic management substrate that works across existing cloud-user interfaces and future agent-native interfaces. Specifically, we discuss how CloudWeaver (1) scopes the context of individual agent sessions with local views of cloud resources and (2) coordinates concurrent management operations on shared cloud resources. CloudWeaver offers strong safety guarantees and attributable feedback in the presence of conflicting intents, while preserving concurrency between independent operations. We validate CloudWeaver using a representative Azure API workload.

cs.MA↗

arXiv2026

HNSW with Accuracy Guarantees Using Graph Spanners

Hierarchical Navigable Small World (HNSW) graphs serve as the industry standard due to their logarithmic complexity and strong empirical performance. However, HNSW relies on greedy graph traversal, a heuristic that provides no theoretical guarantees of correctness. In this paper, we propose a novel "Certify-then-Rectify" framework that bridges the gap between the speed of heuristic search and the rigor of exact retrieval. Rather than discarding HNSW, our approach first employs a distribution-free statistical certifier to dynamically evaluate the quality of a standard HNSW search with minimal overhead. If certification indicates that the retrieved neighbors are of low quality, the framework safely escalates to a rigorous exact recovery algorithm. To make this exact recovery computationally feasible, we reinterpret the HNSW graph as a geometric spanner and utilize Extreme Value Theory to stochastically estimate its maximum empirical stretch factor. This allows us to mathematically bound the maximum distance of true nearest neighbors. Extensive evaluations on benchmark datasets demonstrate that our tiered framework delivers the average-case speed of HNSW while ensuring the worst-case correctness of exact search and outperforming other applicable approaches.

cs.DB↗

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

Beyond Absolute Scores: Relative Edit-induced Difference for Generalizable Image Aesthetic Assessment

Traditional Image Aesthetic Assessment (IAA) methods mainly rely on regressing absolute Mean Opinion Scores (MOS). However, such a paradigm overlooks the inherently dynamic nature of human aesthetic perception, which relies on subconscious comparison against implicit visual references. Consequently, the lack of causal reasoning regarding aesthetic differences prevents models from learning generalizable aesthetic principles, thus limiting their generalization across diverse scenarios. In this work, we rethink the IAA task and propose Relative Edit-induced Difference Aesthetic learning (RED-Aes), a novel framework that leverages controllable image editing models to simulate the human aesthetic reasoning process. Instead of fitting absolute score distributions, RED-Aes explicitly learns the visual factors that drive aesthetic changes. To support this paradigm, we construct the RED-20k dataset, which comprises editing-based image pairs, quantitative aesthetic differences, and Chain-of-Thought (CoT) reasoning. Furthermore, we introduce a three-stage training strategy guided by a relative ranking consistency reward, optimizing the model solely via relative supervision. Extensive experiments demonstrate that RED-Aes achieves state-of-the-art performance on multiple public benchmarks, exhibiting superior generalization capabilities.

cs.CV↗

arXiv2026

TransportBench: A Comprehensive Benchmark for Non-Equilibrium Flow Transport

Scientific machine learning models, as versatile tools for numerical simulation and analysis, are increasingly transforming the landscape of fluid mechanics research. However, existing datasets and benchmarks are primarily limited to continuum fluids and provide limited support for non-equilibrium transport phenomena. To address this gap, we present TransportBench, a high-fidelity dataset and standardized benchmark for non-equilibrium flow transport, designed to reveal the strengths and limitations of neural network models across diverse flow regimes. Specifically, the dataset encompasses a broad physical spectrum, covering continuum and rarefied regimes, low-speed and hypersonic flows, inert and chemically reactive gases, and both translational and internal-energy non-equilibrium effects. Built upon this dataset, we systematically benchmark representative neural architectures using unified evaluation protocols to probe key challenges in learning non-equilibrium flows, including robustness to shock-dominated discontinuities and multi-scale effects, as well as generalization across geometry and physical parameters. Numerical results demonstrate that model performance exhibits a pronounced dependence upon the specific flow characteristics. No single architecture consistently performs best for all the tasks. Instead, different architectural inductive biases provide distinct advantages in capturing smooth flow fields, shock-induced discontinuities, and high-order non-equilibrium statistics. By jointly providing the non-equilibrium flow dataset and model benchmark, TransportBench offers a new testbed for the development, evaluation, and diagnosis of scientific machine learning methods for fluid transport beyond the Navier-Stokes hydrodynamics. The benchmark datasets and implementation codes are available under the MIT license.

physics.comp-ph↗

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

ScaleAcross Explorer: Exploring Communication Optimization for Scale-Across AI Model Training

The rapid scaling of large language model training requires distributing GPU resources across multiple data center buildings and regions. We refer to such paradigm as "scale-across" training. As infrastructure expands, the system design space becomes increasingly intricate, encompassing new model architectures, hardware heterogeneity, and evolving communication patterns. Drawing from Meta's production experience, we highlight the complexities of deploying training jobs across a few data centers housing hundreds of thousands of GPUs. To accelerate exploration of the large design space and to enable efficient training for frontier model development, we conduct in-depth characterization of three key design dimensions: parallelism placement, parallelism scheduling, and network layer technologies. We then propose ScaleAcross Explorer, an optimizer that considers the interplay of design dimensions and holistically optimizes scale-across training. Testbed experiments and simulations demonstrate up to 64.62% training speedups over production configuration and up to 37.59% training speedups over the state-of-the-art baseline across a wide range of design points.

cs.DC↗

arXiv2026

SVR-MAD: A Bayesian-Inspired Framework for Posterior-Guided Multi-Agent Debate

Multi-Agent Debate (MAD) improves LLM-agent accuracy but suffers from rapid context growth, limiting scalability in larger multi-agent settings. Existing methods prune low-utility communications using prior signals, such as token-level log-likelihoods or LLM self-reported confidence. However, these signals become unreliable under hallucination, degrading the accuracy of MAD methods that rely on them. We propose SVR-MAD, a Bayesian-inspired MAD framework that treats pre-debate signals as priors and debate outcomes as posterior-style evidence for estimating agent correctness. SVR-MAD uses this evidence to incrementally construct the communication graph, prioritizing agents whose answers survive peer challenges. Experiments across multiple LLMs and benchmarks show that SVR-MAD reduces token cost by up to 61% while matching or improving accuracy relative to the most accurate competing MAD baseline.

cs.MA↗

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

Metric-Normalized Posterior Leakage (mPL): Attacker-Aligned Privacy for Joint Consumption

Metric differential privacy (mDP) strengthens local differential privacy (LDP) by scaling noise to semantic distance, but many machine learning (ML) systems are consumed under joint observation, where model-agnostic, per-record guarantees can miss leakage from evidence aggregation. We introduce metric-normalized posterior leakage (mPL), an attacker-aligned, distance-calibrated measure of posterior-odds shift induced by releases, and show that for single or independent releases, uniformly bounding mPL is equivalent to mDP. Under joint observation, however, satisfying mDP may still leave mPL high because learned aggregators compound evidence across correlated items. To make control practical, we formalize probabilistically bounded mPL (PBmPL), which limits how often mPL may exceed a target budget, and we operationalize it via Adaptive mPL (AmPL), a trust-and-verify framework that perturbs, audits with a learned attacker, and adapts parameters (with optional Bayesian remapping) to balance privacy and utility. In a word-embedding case study, neural adversaries violate mPL under joint consumption despite per-record mDP perturbations, whereas AmPL substantially lowers the frequency of such violations with low utility loss, indicating PBmPL as a practical, certifiable protection for joint-consumption settings.

cs.LG↗

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↗

arXiv2026

Towards Visual Query Segmentation in the Wild

In this paper, we introduce visual query segmentation (VQS), a new paradigm of visual query localization (VQL) that aims to segment all pixel-level occurrences of an object of interest in an untrimmed video, given an external visual query. Compared to existing VQL locating only the last appearance of a target using bounding boxes, VQS enables more comprehensive (i.e., all object occurrences) and precise (i.e., pixel-level masks) localization, making it more practical for real-world scenarios. To foster research on this task, we present VQS-4K, a large-scale benchmark dedicated to VQS. Specifically, VQS-4K contains 4,111 videos with more than 1.3 million frames and covers a diverse set of 222 object categories. Each video is paired with a visual query defined by a frame outside the search video and its target mask, and annotated with spatial-temporal masklets corresponding to the queried target. To ensure high quality, all videos in VQS-4K are manually labeled with meticulous inspection and iterative refinement. To the best of our knowledge, VQS-4K is the first benchmark specifically designed for VQS. Furthermore, to stimulate future research, we present a simple yet effective method, named VQ-SAM, which extends SAM 2 by leveraging target-specific and background distractor cues from the video to progressively evolve the memory through a novel multi-stage framework with an adaptive memory generation (AMG) module for VQS, significantly improving the performance. In our extensive experiments on VQS-4K, VQ-SAM achieves promising results and surpasses all existing approaches, demonstrating its effectiveness. With the proposed VQS-4K and VQ-SAM, we expect to go beyond the current VQL paradigm and inspire more future research and practical applications on VQS. Our benchmark, code, and results will be made publicly available.

cs.CV↗

arXiv2026

BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing?

Current code-agent benchmarks primarily evaluate localized issue resolution within a single target repository, leaving under-tested many software engineering tasks that require external knowledge or broader repository-level changes. We introduce BeyondSWE, a 500-instance benchmark drawn from 246 real-world GitHub repositories to evaluate code agents beyond single-repository bug fixing. BeyondSWE covers four representative settings: cross-repository issue resolution, domain-specific issue resolution, dependency-driven migration, and document-to-repository generation, spanning both broader knowledge scope and broader resolution scope. Our evaluation shows that BeyondSWE remains far from saturated: the best OpenHands-based agent reaches 46.12 average score, while the strongest Codex harness with GPT-5.4 (xhigh) reaches 56.65 under a search-aware prompt. To study whether external information access closes this gap, we use SearchSWE as a controlled diagnostic baseline for search-augmented coding. Search access improves most models and substantially helps some tasks, but the gains remain limited and uneven, showing that current agents still struggle to convert retrieved information into precise, version-compatible, and locally actionable code changes. These results suggest that deep search for coding remains an open problem: progress requires agents that can reliably combine external evidence with repository-local reasoning and execution-based verification.

cs.CL↗