跳到正文
研踪PaperTrail

版本更新

查看已发布论文的最新修订,在详情页对照实际观测到的版本。

部分采集异常

全站更新概况

+1084新增论文
+883其中 AI 相关
+130已有论文修订
+94新增 AI 导读

arXiv 最近完成扫描:2026-11-18 04:28 · 全站最近入库:2026-11-18 19:45各信源进度 →

统计口径与数据新鲜度

统计 2026-11-18 03:53 — 2026-11-19 03:53(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-18 04:28最近发现新增版本:2026-11-18 19:45库内论文最新更新时间:2026-11-18 08:00

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

26,870 已收录论文22,374 AI 相关8,156 有修订112 HF 社区发现

论文修订列表

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 8,156 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-18 04:27(北京时间)
版本更新 · v3arXiv:2608.22788

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

Tianqi Xu, Lu Lv, Haoyang Huang, Wenjie Huang, Zhanming Shen, Yuhao Shen, Baolin Zhang, Xinyi Hu 等 18 位作者

原文摘要

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

cs.AI · cs.LG

版本更新 · v3arXiv:2611.00800

A Discrete-Time Variational Objective for Degenerate Diffusion and Equivalence to Flow Matching

Rares Dolga, Harshad Dahake, David Barber

原文摘要

Diffusion models and more recently Flow models have widespread application. In this work we consider degenerate Diffusion models in which the per-step generator is deterministic. We construct a modified variational framework and show a formal equivalence between Flow models and Degenerate Diffusion models. For our discrete-time variational objective we show that the optimal degenerate generator recovers the data distribution in the limit of an infinite number of generator steps. Our work therefore provides a unified perspective, showing that discrete-time implementations of flow matching and the discrete-time Degenerate Diffusion model can be considered within the same simple variational framework, removing the need for the continuous-time machinery normally used to justify flow matching models.

cs.AI

版本更新 · v4arXiv:2501.13229

The Yoneda embedding in simplicial type theory

Daniel Gratzer, Jonathan Weinberger, Ulrik Buchholtz

原文摘要

Riehl and Shulman introduced simplicial type theory (STT), a variant of homotopy type theory which aimed to study not just homotopy theory, but its fusion with category theory: $(\infty,1)$-category theory. While notoriously technical, manipulating $\infty$-categories in simplicial type theory is often easier than working with ordinary categories, with the type theory handling infinite stacks of coherences in the background. We capitalize on recent work by Gratzer et al. defining the $(\infty,1)$-category of $\infty$-groupoids in STT to define presheaf categories within STT and systematically develop their theory. In particular, we construct the Yoneda embedding, prove the universal property of presheaf categories, refine the theory of adjunctions in STT, introduce the theory of Kan extensions, and prove Quillen's Theorem A.

cs.LO · math.AT · math.CT

版本更新 · v2arXiv:2611.06984

AdaFloor: Adaptive Floor Planning and Rank-Time Reuse for Long-Tail MoE RL Rollouts

Jia Wei, Ziyi Qiu, Yi Xu, Xuehai Qian

原文摘要

Synchronous MoE reinforcement-learning post-training suffers from response-length imbalance. Existing methods reduce request tails, yet completed ranks remain bound to Expert Parallelism (EP) collectives and expert ownership. In a representative 16-accelerator Qwen3-30B-A3B rollout, this dummy run traps 32.83% of rank-time. AdaFloor converts this EP-coupled capacity into bounded leases for auxiliary work to improve system resource efficiency through rank-time reuse. Historical prompt-tail prediction and a calibrated guard construct the KV-safe layout. AdaFloor then selects the floor-specific safe quorum schedule with the largest predicted release area. Its runtime shrinks the EP group and leases detached ranks until full-world restoration. Across three updated-policy seeds on DeepScaleR mathematical reasoning rollouts, its capacity-oriented Natural floor2 configuration with TailGuard reduces Vanilla's 29393.9 rank-seconds of dummy run by 96.15%. Rollout shortening eliminates 18693.6 rank-seconds, while coordinated exits convert another 9568.6 rank-seconds into reusable capacity, leaving 1131.7 rank-seconds. For sidecar reuse, its Planned residency floor4 configuration without TailGuard serves 3134 Qwen3-8B GSM8K queries per epoch at 99.46% accuracy and incurs only a 2 to 3% primary throughput reduction against the same configuration with sidecar execution disabled. Its latency-oriented Planned residency floor4 configuration with TailGuard reduces rollout time by 24.09% relative to Vanilla and by 2.31% relative to LengthSort with TailGuard. On public Eurus-2-RL code rollouts with the same model and 16-rank topology, AdaFloor reduces rollout time by 20.56% and dummy run by 89.13% while generated work and mean reward remain stable.

cs.AI · cs.DC

版本更新 · v3arXiv:2609.08572

AgentGrad: Intervention-guided Prompt Optimization for Multi-Agent Systems

Jaewon Chu, Jinwoo Seo, Jaewon Cho, Jeehye Na, Yunyang Xiong, Youngdae Kim, Hyunwoo J. Kim

原文摘要

Large language model (LLM)-based multi-agent systems (MAS) achieve strong performance by employing specialized multiple agents, yet their performance depends on the prompt design of each agent. For MAS prompt optimization, textual gradient methods that guide prompt updates using natural-language feedback have emerged as a leading paradigm. In this paper, we identify limitations in two stages of existing textual gradient approaches: gradient extraction and gradient aggregation. In gradient extraction, previous works select a prompt to update without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision over the corresponding agent's intermediate output. In gradient aggregation, individual gradients are randomly grouped and concatenated, often mixing unrelated failure modes and producing prompts that fail to generalize. To address these limitations, we propose AgentGrad, a prompt optimization framework for multi-agent systems based on sequential intervention and semantic textual gradient abstraction. For each failure, sequential intervention injects a hint into one agent at a time to identify the target agent whose modification resolves the failure. The modified output of the target agent then serves as agent-level supervision for extracting a fine-grained gradient. Semantic textual gradient abstraction clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient that captures the shared corrective pattern. Experimental results show that AgentGrad achieves state-of-the-art performance across five MAS benchmarks while reducing wall-clock optimization time by 2.5 times and optimization cost by 21.8% on average compared to the next-best baselines. Our code is available at https://github.com/mlvlab/AgentGrad.

cs.AI

版本更新 · v2arXiv:2611.07084

DMuon-2: Scaling Matrix Optimizers in Data Parallelism with Topology-Aware Matrix Ownership and Communication

Starrick Liu, Ruizhi Chengze, Lufang Chen, Michael Zhao, Ran Yu, Xiaofan Li, Dance Kuzi, Yue Zhang 等 13 位作者

原文摘要

Matrix optimizers such as Muon update whole weight matrices, which conflicts with the element- or row-wise sharding of data-parallel runtimes. Once parameters are sharded as well, matrix optimizers incur extra communication or repeated computation. We present DMuon, a matrix-granular optimizer runtime for PyTorch FSDP (ZeRO-2/3) and HSDP. DMuon assigns each matrix a unique owner, which performs the Newton--Schulz (NS) iteration and optimizer update on the full matrix, and balances the load across owners by computational cost. On HSDP, ownership is global across replicas, so the cluster keeps a single copy of gradients and optimizer state, and updated parameters are published across nodes as soon as they are ready, overlapping subsequent computation. This layout makes every collective sparse and variable in length, so DMuon schedules these collectives as their data becomes ready and routes them along the cluster's two-level network. On up to 256 H100 GPUs, across three dense LLMs and two VLA models, DMuon matches or exceeds the training throughput of PyTorch FSDP2 with AdamW, reaching 1.05$\times$ that of FSDP2 with AdamW at the median on 256 GPUs. The optimized parameter gathering and gradient reduction absorb the extra communication and computation a matrix optimizer requires. Against a gather-then-replicate Muon baseline, DMuon's throughput is 2.1 times as high at the median.

cs.DC

版本更新 · v2arXiv:2507.15079

Isotonic Quantile Regression Averaging for uncertainty quantification of electricity price forecasts

Arkadiusz Lipiecki, Bartosz Uniejewski

原文摘要

Quantifying the uncertainty of forecasting models is essential to assess and mitigate the risks associated with data-driven decisions, especially in volatile domains such as electricity markets. In this paper, we propose a computationally efficient isotonic variant of Quantile Regression Averaging (QRA), called Isotonic Quantile Regression Averaging (iQRA), for generating probabilistic forecasts from ensembles of point forecasts. Building on QRA, we introduce stochastic order constraints as a hyperparameter-free regularization mechanism that reduces computational complexity while maintaining competitive forecast accuracy. We further demonstrate that the same isotonic regularization principle can be applied to the smoothed estimator of QRA. In an extensive forecasting study of two European day-ahead electricity markets, we show that iQRA and its smoothed variant consistently rank among the top state-of-the-art postprocessing methods in terms of reliability and sharpness. It produces well-calibrated prediction intervals, with smoothed iQRA providing superior reliability even when compared against coverage-oriented conformal prediction. We further introduce a battery arbitrage evaluation, showing that the leading probabilistic methods achieve the highest profit per trade and that probabilistic risk control improves the economic value of trading decisions. Isotonic regularization offers a scalable hyperparameter-free variable selection for linear quantile regression, making iQRA attractive for practical decision-support systems in energy markets.

cs.LG · q-fin.ST · stat.AP

版本更新 · v2arXiv:2605.18190

Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network

Grigory Bartosh, David Ruhe, Emiel Hoogeboom, Jonathan Heek, Thomas Mensink, Tim Salimans

原文摘要

Diffusion models achieve state-of-the-art generative performance but suffer from high computational costs during inference due to the repeated evaluation of a heavy neural network. In this work, we propose Dual-Rate Diffusion, a method to accelerate sampling by interleaving the execution of a heavy high-capacity context encoder and a light efficient denoising model. The context encoder is evaluated sparsely to extract high-dimensional features, which are effectively reused by the light denoising model at every step to refine the sample efficiently. This approach significantly accelerates inference without compromising sample quality. On ImageNet benchmarks, Dual-Rate Diffusion matches the performance of standard baselines while reducing computational cost by a factor of $2$-$4$. Furthermore, we demonstrate that our method is compatible with distillation techniques, such as Moment Matching Distillation, enabling further efficiency gains in few-step generation.

cs.LG · cs.CV

版本更新 · v6arXiv:2604.13899

Do We Still Need Humans in the Loop? Human vs. LLM Annotation in Active Learning for TikTok Hate Speech Detection

Ahmad Dawar Hakimi, Lea Hirlimann, Isabelle Augenstein, Hinrich Schütze

原文摘要

Annotating data remains a costly bottleneck for supervised NLP. Active learning (AL) reduces the number of human labels needed by selecting only the most informative instances, while instruction-tuned LLMs attack the same bottleneck from the other side, making labels cheap enough to annotate entire corpora. This raises two questions: can LLM labels replace human labels within the AL loop, and does AL remain necessary when entire corpora can be cheaply labeled? We investigate both by training supervised hate speech classifiers on a new dataset of 278K German political TikTok comments, comparing human and LLM annotation under matched conditions. LLM annotation at scale outperforms human-supervised classifiers at roughly one-tenth the cost, for both a closed-source (GPT-5.2) and an open-weight (Qwen3.5-122B-A10B) LLM, and the advantage is robust under soft-label evaluation. It hinges on the annotation interface: only a two-question decomposition mirroring the human annotation task unlocks it. AL provides no reliable advantage over random sampling in our prefiltered pool. Error structure depends on the LLM: only GPT-5.2 matches the human FP/FN balance, while other variants over-flag border-control and economic-competition discourse. Humans remain essential as evaluators; for training labels, the question shifts to which LLM, which interface, and what shape of pool.

cs.CL · cs.AI

版本更新 · v2arXiv:2611.04241

AceAction-5M: Hierarchical Goals for Generalist Action Learning Across 3D Game Worlds

Kailing Li, Ruoxuan Zhang, Zheyuan Zhou, Boyang Chu, Yingshan Lei, Hongxia Xie, Liang Du, Zixun Sun

原文摘要

Video games provide a practical setting for collecting vision-language-action (VLA) data, whose acquisition in physical environments is costly. Existing datasets, however, remain limited by restricted game coverage, sparse language supervision, or temporal hierarchies that do not necessarily reflect changes in task objectives. We introduce AceAction, a dataset of human gameplay across multiple 3D games for learning skill composition and subgoal transitions. AceAction organizes demonstrations around complete tasks, linking overall instructions, ordered subgoals, and local skill instructions to temporally aligned visual observations and keyboard-and-mouse actions. To construct these annotations, we develop a vision-language model pipeline that identifies subgoals from changes in objectives or interaction targets, anchors navigation goals to visually recognizable waypoints, and composes adjacent subgoals into task instructions. The resulting demonstrations cover navigation, reconnaissance, combat, interaction, and recovery, with explicit stage boundaries that connect task progression to fine-grained control. We assess annotation quality through model-based verification and human inspection. An audit of 500 trajectories yields a 95.8% pass rate across independent reviewer assessments. AceAction provides a data foundation for studying how agents execute, combine, and switch skills to follow complex instructions in dynamic game environments.

cs.AI

版本更新 · v3arXiv:2605.06505

PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization

Murat Bilgehan Ertan, Xiaochen Zhu, Phuong Ha Nguyen, Marten van Dijk, Srinivas Devadas

原文摘要

Differentially private (DP) zeroth-order (ZO) fine-tuning of large language models (LLMs) inherits DP's harsh utility-privacy trade-off. PAC Privacy, an instance-based notion, offers an alternative. It bounds the mutual information (MI) between the secret input and the output, and calibrates the noise to the stability of the output under the input distribution. We introduce PACZero, a family of ZO fine-tuning mechanisms under PAC Privacy. For stability, PACZero quantizes each update to a single bit, the sign of the ZO update aggregated over the secret subset of the training data. On many steps all candidate subsets agree on this sign, so the bit can be released at no cost in MI. On the remaining steps the bit passes through a noisy channel calibrated to a per-step MI budget. We study two channels, the standard additive Gaussian channel of PAC Privacy (PACZero-AGC) and a binary asymmetric channel with optimal bit flipping (PACZero-BAC), along with their zero-budget versions at $\mathrm{MI} \to 0$, PACZero-AGC$_0$ and PACZero-BAC$_0$. At zero MI, nothing released during training, including the fine-tuned model, carries any information about which secret subset was sampled, so no adversary inferring that subset can do better than its prior. In particular, the success rate of any membership inference attack (MIA) is bounded by 50% under a balanced prior. To our knowledge, this is the first such guarantee for LLM fine-tuning, and DP matches the MIA bound only at $\varepsilon = δ= 0$ and infinite noise. We evaluate on SST-2 and SQuAD with models from 1.3B to 8B parameters, under LoRA and full fine-tuning. Both channels retain high utility under tight MI budgets, and even PACZero-BAC$_0$, which guarantees zero MI, stays within 2 points of non-private ZO fine-tuning with the same number of steps. To our knowledge no DP method achieves usable utility matching this MIA-resistance level.

cs.LG · cs.AI · cs.CR

版本更新 · v2arXiv:2611.02984

Classification Aggregation with Multiple Memberships

Federico Fioravanti, Leo Kurata, Vasilis Stamatis

原文摘要

People face the problem of classifying objects into categories while aggregating their opinions. We axiomatically study such a classification aggregation problem with multiple memberships, meaning that each individual assigns one category to every object, whereas the resulting classification may assign more than one category. Under the natural settings and axioms (i.e., Independence and Supportiveness), each of our new axioms -- Responsibility, Split Stability and Rejection Stability -- characterises what we call the junta rules. The rules assign a category to an object exactly when some member of a fixed coalition does so. The coalitions that can secure an assignment form a grill, a topological structure implied by an ultrafilter. Hence, the study parallels the literature on Arrow's theorem characterised by the decisive coalition structures. An infinite population setting is also investigated to clarify the differences among the axioms.

econ.TH · cs.GT · cs.MA

版本更新 · v2arXiv:2610.02946

When Predicting Nothing Beats SAM 3: Revisiting Evaluation in Video Object Segmentation

Jihwan Hong, Woohyeon Park, Jaeik Kim, Jaeyoung Do

原文摘要

Video Object Segmentation (VOS) in complex and long videos is increasingly important for real-world applications, where target objects often appear only intermittently within long temporal horizons. However, existing benchmarks largely focus on temporally salient objects that remain visible for most of the video. To address this gap, we introduce FaVOS (A Benchmark for Video Object Segmentation with Fractional Temporal Visibility), a benchmark designed to evaluate VOS methods under low temporal visibility. We show that, in this regime, the standard J&F metric can collapse VOS evaluation into absence classification, because empty predictions receive high rewards on target-absent frames. Consequently, even a trivial empty-mask predictor can outperform strong models such as SAM 3, revealing a fundamental mismatch between current metrics and practical VOS performance. To mitigate this issue, we propose Volumetric J&F, which evaluates mask sequences as spatio-temporal volumes and reduces the dominance of target-absence rewards while preserving sensitivity to segmentation quality and temporal structure. Project page: https://aidaslab.github.io/FaVOS.

cs.CV

版本更新 · v2arXiv:2611.07059

A hierarchy of physics-embedded neural operators for structural dynamic surrogate modeling

Amir Nasrollahi, Mehmet Sefa Orak

原文摘要

Structural-response surrogates must capture moving resonances, phase changes, and unequal-amplitude transfer channels during optimization or health-state changes. Physics-embedded learning seeks FEA-level accuracy at lower cost by embedding physics in inputs, forward passes, losses, and backpropagation. We propose Modal Rational (MR), mapping structural features to modal quantities for analytical frequency-response reconstruction. The Physics-Guided Modal Operator Network (PGMON) adds bounded frequency, damping, and residual-flexibility corrections. Modal Operator with Structural GELU and Anchored-SIREN Correction (MOSAIC) denotes the bounded spectral-discrepancy framework used here; MOSAIC-MR and MOSAIC-PGMON add such discrepancies. They support finite-dimensional maps and physical decoders for future neural operators; evidence is fixed-topology, not a completed operator comparison. Two 20,000-design populations are studied: an 80 m steel truss tower with variable member areas and attached masses, and a tapered shell beam with ten variable station masses. Archived tests cover ten designs and seven acceleration histories per structure; they are benchmarks, not seismic-design assessments. With 5% reference modal damping, MR, PGMON, and MOSAIC give mean displacement-history errors of 0.73-0.76% for the tower and 0.029-0.116% for the beam, against declared band-limited modal references. Supplied modal frequencies are distinguished from design-only use; unequal parameter counts and training settings are explicit. Error analysis separates learned modal discrepancies, truncation, and excitation weighting, explaining acceleration and difficult-tower limits and risks of blind use. Results support physically structured prediction but limit claims about generalization, computational savings, and added corrections. Next is developing and validating neural operators around these architectures.

cs.CE · math-ph · physics.comp-ph

版本更新 · v2arXiv:2605.07847

Measuring and Mitigating the Distributional Gap Between Real and Simulated User Behaviors

Shuhaib Mehri, Philippe Laban, Sumuk Shashidhar, Marwa Abdulhai, Sergey Levine, Michel Galley, Dilek Hakkani-Tür

原文摘要

As user simulators are increasingly used for interactive training and evaluation of AI assistants, it is essential that they represent the diverse behaviors of real users. While existing works train user simulators to generate human-like responses, whether they capture the broad and heterogeneous distribution of real user behaviors remains an open question. In this work, we introduce a method to measure the distributional gap between real and simulated user behaviors, validated through a human study and ablations. Given a dataset of real and simulated conversations, our method extracts representations of user behavior from each conversation, quantizes them into discrete distributions via clustering, then computes divergence metrics. We provide the first systematic evaluation of 24 LLM-based user simulators on coding and writing tasks, and reveal a large distributional gap from real users that varies across model families, scales, and behavioral facets. Pairwise comparisons show that most simulators behave similarly, while a few stand apart. Combining behaviorally complementary simulators brings the resulting distribution closer to real users compared to either simulator on its own. Finally, a TF-IDF analysis of the clusters surfaces interpretable patterns of behaviors that simulators capture, miss, and hallucinate.

cs.CL

版本更新 · v2arXiv:2610.30430

Tight Bounds for Paging with Delay and Joint Replenishment with Varying Opening Costs

Joseph, Naor, Tomer Tsachor

原文摘要

We study weighted paging with delay, a generalization of classical paging in which requests may be served after their arrival, incurring delay costs while awaiting service. This model allows the algorithm to delay service and batch service requests over time, but also introduces significant challenges due to overlapping requests. While the problem is well understood in the unweighted setting, where optimal $O(\log k)$-competitive algorithms are known, the weighted case remains more elusive. The best previously known randomized algorithm, due to Gupta, Kumar, and Panigrahi, achieves an $O(\log k \log n)$-competitive ratio, leaving a gap between weighted and unweighted settings. We obtain a randomized $O(\log k)$-competitive algorithm, matching the lower bound inherited from classical paging. The core of our approach is a structural decomposition that separates the evolution of the cache over time from scheduling additional page loads, so as to make the solution feasible. We develop this approach first for paging with time windows, the special case in which each request must be served within a specified time interval. Given a sequence of cache configurations, we reduce the problem of augmenting this sequence to a feasible solution to a variant of the Joint Replenishment Problem with deadlines (JRP-D) with time-dependent opening costs. We design a constant-competitive online algorithm for this variant under a prediction model whose required information is available online in our paging reduction. Combining this algorithm with an existing online construction of cache configurations yields an $O(\log k)$-competitive algorithm for paging with time windows. We then extend the framework to obtain the same guarantee for weighted paging with a general delay function.

cs.DS

版本更新 · v2arXiv:2606.06423

RiskFlow: Fast and Faithful Safety-Critical Traffic Scenario Generation

Jie Li, Qi Lan, Yining Tang, Yu Shen, Yi Zhou, Yuhao Wei, Guofa Li, Hao Sun

原文摘要

Safety-critical traffic scenario generation is essential for evaluating autonomous driving systems under rare but high-risk interactions. Existing diffusion-based methods offer strong controllability in closed-loop generation, but their iterative denoising process is computationally expensive and may accumulate sampling and guidance errors over long rollouts, causing unrealistic motion artifacts such as abnormal acceleration and off-road behavior. To address these issues, we propose RiskFlow, a closed-loop safety-critical traffic generation framework that reformulates future behavior generation as one-step flow transport in the action space during training. At test time, RiskFlow generates safety-critical interactions through a selective guidance procedure in which time-to-collision-based critical-agent ranking selects high-risk agents and an adversarial interaction objective directly refines their output actions. Vehicle-dynamics rollout and map-aware regularization then translate guided actions into feasible trajectories while discouraging off-road behavior. By intervening on selected agents in the output space, RiskFlow avoids repeated corrections of intermediate denoising states, thereby mitigating accumulated guidance errors and accelerating closed-loop scenario generation. Experiments on the nuScenes dataset under closed-loop evaluation show that RiskFlow achieves a strong adversariality-realism trade-off across multi-agent and long-horizon settings. Compared with representative baselines, RiskFlow consistently improves realism while maintaining competitive safety-critical generation capability, and substantially reduces inference time for evaluation.

cs.RO · cs.AI

版本更新 · v3arXiv:2607.00466

ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving

Sangjin Choi, Sukmin Cho, Yifan Xiong, Ziyue Yang, Youngjin Kwon, Peng Cheng

原文摘要

Equal load does not mean equal decode latency in prefill-decode (PD) disaggregated mixture-of-experts (MoE) serving. In bandwidth-bound decode, more distinct experts mean more weight loads and higher latency. We present ELDR, a lossless, expert-locality-aware decode router. Prefill signatures predict decode expert usage. Balanced $K$-means fits per-decoder centroids with background refitting. Locality-band routing selects the least-loaded worker among close matches. A block-granular cache keeps signatures exact under prefix caching. We evaluate ELDR in vLLM on up to 40 GPUs. Across three MoE models on 12 public task benchmarks and multilingual WildChat, it reduces median time-per-output-token (TPOT) by 5.4-14.5% on average across request rates and up to 17.8% over the best of four load-balancing baselines, preserving model outputs and token-to-expert assignments.

cs.DC

版本更新 · v2arXiv:2605.18094

Learning to Solve Compositional Geometry Routing Problems

Mingfeng Fan, Jianan Zhou, Jiaqi Cheng, Yifeng Zhang, Jie Zhang, Guillaume Adrien Sartoretti

原文摘要

We study the Compositional Geometry Routing Problem (CGRP), a unified superclass of traditional routing problems that covers point-only, line-only, area-only, and arbitrary hybrid task geometries, providing a broad abstraction for real-world routing scenarios. Beyond standard point-based routing, CGRP with non-point tasks can be inherently asymmetric, tightly coupled travel routes with the intrinsic path, and enlarges the action space with numerous feasible yet often irrelevant options, thereby posing significant challenges for both representation learning and decision-making. To address these challenges, we propose DiCon, a differential attention-assisted solver with contrastive learning, as a plug-and-play framework that tackles the problem from two complementary angles. First, we introduce a differential attention mechanism that actively suppresses the probability mass on less competitive candidate actions. Second, we design a double-level contrastive learning objective to promote robust global instance representations and regularize geometry-aware task representations. Extensive experiments demonstrate that DiCon achieves strong performance, broad versatility, and superior generalization across diverse CGRP instances with different compositions.

cs.AI

版本更新 · v2arXiv:2607.15873

A Physically-Realisable Quantum Programming Language with Quantum-Controlled Quantum Channels

Kengo Hirata, Takeshi Tsukada

原文摘要

In contrast to a classical bit, which can only take the value $0$ or $1$, its quantum counterpart---a qubit---can exist in a superposition of $0$ and $1$. This is a superposition of data values, and it is then tempting to superpose programs as well. Indeed, a particular superposition of programs, known as the quantum SWITCH, has attracted considerable attention. A natural way to superpose programs is conditional branching on quantum data, and several studies have pursued this idea. However, there is a well-known semantic difficulty with this approach, and the fundamental question of whether the languages proposed in prior work can be physically realised remains open. This paper proposes a physically realisable quantum programming language with quantum branching over quantum channels. We first analyse the semantic difficulty from an implementation perspective and show that the difficulty is caused by arbitrariness in the implementation. We then design a language that is free from this arbitrariness. As expected from its implementation-oriented design, our language is physically realisable. We argue for this from two directions: first, we give a method for compiling programs into quantum circuits; and second, we show that our programs stay within the physically admissible range, namely, that they neither yield probabilities greater than one nor violate causal inequalities. The absence of such violations is established through the inclusion of our language in QC-QC, a notion of realisability studied in physics.

cs.PL · cs.LO · quant-ph

版本更新 · v2arXiv:2610.21613

To Think or Not to Think: Pre-Decisional Reasoning Budgets for Referring Audio-Visual Segmentation

Haojie Hu, Senda Chen, Ying Shen, Lin Zhang

原文摘要

Multimodal reasoning systems increasingly assume that longer chain-of-thought (CoT) uniformly improves downstream grounding. We show that this assumption fails in referring audio-visual segmentation (Ref-AVS): simple queries can be harmed by excessive reasoning, creating an "overthinking trap" associated with phrase and localization drift, while genuinely ambiguous queries still benefit from multi-step reasoning. To systematically study this, we construct counterfactual reasoning-budget labels under a five-fold out-of-fold evaluation protocol. Our results recast adaptive Ref-AVS reasoning as a pre-decisional representation problem: before the first generated token, the generation-onset state contains linearly decodable information predictive of whether additional reasoning will improve segmentation. A linear probe predicts binary reasoning need with 70.1% accuracy, substantially outperforming text-only features and uncertainty-based proxy signals. As a router, the onset-state controller improves over downstream confidence routing and closely matches Short-then-Decide while using far fewer tokens. Proxy comparisons show that the onset representation contains predictive information not fully recovered by the evaluated hand-crafted features. Reusing this onset state for lightweight, single forward-pass routing reduces autoregressive token consumption by 60% while retaining about 96% of always-long performance. It further provides a cost-quality trade-off on $R^2$-AVSBench without target-label tuning, rather than an accuracy improvement over always-long. Error analysis links harmful long reasoning to phrase and localization drift, illustrating the benefit of bypassing unnecessary reasoning on simple queries. Ultimately, our results support efficient adaptive reasoning in Ref-AVS through pre-decisional internal representations.

cs.MM

版本更新 · v2arXiv:2610.22457

Erase the Mark, Preserve the Motion: Temporally Consistent Removal of Video Watermarks

Yifan Liao, Zhen Sun, Zhicong Huang, Jiaming He, Zongmin Zhang, Xi Ai, Cheng Hong, Xinhu Zheng 等 9 位作者

原文摘要

Recent advances in video generation have increased the demand for reliable watermarking for copyright protection, yet the robustness of video watermarks against removal attacks remains underexplored. Most watermark-removal methods originate from the image domain: signal optimization injects imperceptible pixel-domain perturbations, whereas generative resynthesis produces a visually similar instance through diffusion synthesis. When extended to videos, the former introduces temporally varying noise that causes visible flickering, while the latter regenerates fine-grained structures inconsistently across frames, leading to structural motion inconsistency. To evaluate video watermark robustness, we propose PoseRegen, a pose-consistent reconstruction framework that removes watermarks while preserving the original content and motion. PoseRegen first applies a temporally shared viewpoint transformation to disrupt watermark-bearing spatial patterns and then jointly reconstructs the transformed sequence with a pose-conditioned video model. A frozen video DiT processes the complete sequence with its reference-video latent, and full-sequence Flow Matching supervision keeps identity and background consistent across frames. To further preserve motion structure, we introduce a lightweight Pose-Conditioned Motion Consistency Adapter that adapts the frozen backbone to temporally ordered pose guidance and constrains body-part trajectories throughout reconstruction. Experiments on eleven video watermarking schemes show that PoseRegen reduces the average bit accuracy to 0.505, near random guessing, while outperforming the strongest baseline by 9.34% on motion-related VBench metrics. These results reveal a practical vulnerability of current video watermarks and highlight the need to evaluate robustness jointly with temporal and motion quality.

cs.CR · cs.CV

版本更新 · v2arXiv:2610.33798

Whispers from the Neighbors: Rethinking Caption Consensus as Inference-Time Evidence for Vision-Language Retrieval

Khang Nhat Hoang Vo, Belal Shoer, Abhirama Subramanyam Penamakuri

原文摘要

An image is rarely described only once. Retrieval benchmarks and many collections attach several captions to each picture, yet dual encoders still judge every image-caption pair in isolation. We ask what these neighboring captions contribute when a frozen encoder consults them at inference time. Fusing image similarity with similarity to an image's own captions improves text-to-image R@1 by 3.4 to 16.8 points across nine frozen encoders when the captions come from the same annotators as the queries. The gain falls to 1.9 to 7.2 points when an off-the-shelf captioner writes them, and to at most one point when they come from a different annotation source; a paraphrase-trained text encoder strengthens the text side further. A natural refinement, scoring the part of the image that the captions leave unexplained, appears to add much more under a leave-one-out evaluation. We show that this advantage arises because the correct candidate's supports are built differently from those of the negative candidates, with set size and query removal each contributing about half, and that under a fixed support index the residual reduces to a reweighting of what the captions already provide. On compositional benchmarks, the captions help by matching candidates to human descriptions of the image, without improving image-text grounding. The code is available at: https://github.com/nhatkhangcs/clip_consensus

cs.CV · cs.CL

版本更新 · v4arXiv:2609.24997

VideoGen-Agent: Reinforcing Video Generation Agents

Binxu Li, Haoyi Duan, Yuhui Zhang, Yaohui Zhang, Zihao Lin, Kaituo Feng, Hui Yuan, Suozhi Huang 等 13 位作者

原文摘要

Recent advances in video generative models have enabled high-fidelity, temporally coherent video generation. However, these models often struggle to satisfy prompts requiring specialized knowledge, specific identities, physical consistency, or ordered events. In this paper, we present VideoGen-Agent, a multimodal agent trained through multitask agentic reinforcement learning to use external tools for video generation. The agent coordinates augmentation, generation, and verification tools through multi-turn interactions, using the prompt and intermediate observations to guide its decisions. We train a shared policy on a category-balanced dataset spanning six tasks. Supervised fine-tuning on teacher-generated trajectories establishes tool-use behavior, which is then refined through reinforcement learning. A category-aware hybrid reward evaluates tool-call validity, task-appropriate tool use, and generated video quality. We further introduce VABench, a held-out benchmark of 600 prompts covering procedural knowledge, single- and multi-entity identity preservation, physical consistency, scene composition, and multi-shot temporal structure. On VABench, VideoGen-Agent improves over its base text-to-video generator by 19.1 points, from 56.5 to 75.6. Upgrading the generation tools further raises the score to 86.1 without additional agent training. Human raters prefer the upgraded configuration over the strongest standalone baseline in 84.3% of comparisons. These results support learning tool use across video-generation tasks and show that the trained agent can benefit from subsequent advances in generation tools. Project page: https://andyca111.github.io/VideoGen_Agent/

cs.CV

版本更新 · v2arXiv:2606.05516

Dominant-Layer ZO: A Single Layer Dominates Zeroth-Order Fine-Tuning of LLMs

Wanhao Yu, Ziyan Wang, Zheng Wang, Abeer Matar Almalky, Yihang Zuo, Shuteng Niu, Sen Lin, Adnan Siraj Rakin 等 10 位作者

原文摘要

Zeroth-order (ZO) optimization enables memory-efficient fine-tuning of large language models (LLMs) using only forward passes, but it remains unclear how useful adaptation is distributed across layers. In this work, we reveal a surprising phenomenon: ZO fine-tuning is sharply dominated by a single decoding layer. Across multiple LLM families and downstream tasks, fine-tuning this dominant layer alone consistently matches or even exceeds full-model ZO fine-tuning. We further show that the dominant layer is task-agnostic but model-specific, and can be identified before training through a simple inference-only analysis of activation outliers. Specifically, the dominant layer consistently aligns with the first activation-outlier layer in the pre-trained model. To explain this phenomenon, we analyze how perturbation effects propagate under ZO optimization. We find that the dominant layer combines two key properties: high perturbation sensitivity and early placement in the residual stream, allowing perturbation-induced effects to propagate and accumulate through remaining subsequent decoding layers. As a result, this layer produces disproportionately strong and stable optimization signals under forward-only updates. Extensive experiments on LLaMA2-7B and Qwen3-8B across nine benchmarks show that dominant-layer ZO fine-tuning improves average performance over full-model MeZO and LoRA-based ZO fine-tuning while achieving up to 4.52$\times$ training speedup.

cs.LG

版本更新 · v5arXiv:2605.11651

Hide to See: Reasoning-prefix Masking for Visual-anchored Thinking in VLM Distillation

Seonghoon Yu, Dongjun Nam, Byung-Kwan Lee, Jeany Son

原文摘要

Recent think-answer approaches in VLMs, such as Qwen3-VL-Thinking, boost reasoning performance by leveraging intermediate thinking steps before the final answer, but their computational cost becomes substantial, especially for larger VLMs. To distill such capabilities into compact think-answer VLMs, a primary objective is to improve the student's ability to utilize visual evidence throughout its reasoning trace, as long think-answer traces suffer from visual forgetting issues. To this end, we introduce a novel think-answer distillation framework that encourages the student to anchor its thinking on visual information by masking the student's salient reasoning prefixes. To compensate for such masked textual cues, the student is encouraged to rely more on visual evidence as an alternative source of information during distillation. Our masking strategies include: 1) token-wise salient reasoning-prefix masking, which masks high-influence reasoning prefixes selectively for each next-token prediction, and 2) self-paced masking budget scheduling, which gradually increases the masking scale according to distillation difficulty, measured by the discrepancy between teacher--student distributions. In the distillation phase, the student is guided by our salient reasoning-prefix mask, which blocks both future tokens and salient reasoning cues, in place of the standard causal mask used for auto-regressive language modeling. Experimental results show that our approach outperforms recent open-source VLMs, VLM distillation, and self-distillation methods on multimodal reasoning benchmarks, while further analyzes confirm enhanced visual utilization along the student thinking process.

cs.CV · cs.AI · cs.CL

版本更新 · v5arXiv:2511.21638

Iterative Policy Optimization: Batch-Online Multi-Turn RL via Single-Turn RL

Daniel R. Jiang, Ankur Samanta, Yukai Yang, Jalaj Bhandari, Rémi Munos, Tyler Lu

原文摘要

Practical LLM agents often operate over multi-turn conversations where success is determined only after the full interaction ends. Most multi-turn RL methods train via on-policy rollouts, but unlike in single-turn RL, the policy cannot produce a trajectory alone, since an external environment must respond after each agent turn. For conversational agents, this environment is a user, but real users are generally unavailable inside the training loop and simulated users are difficult to build faithfully. Separately, real-world deployment is rarely fully online or fully offline. The common production pattern is called batch online, where the current policy is deployed to collect a batch of interaction data, then retrained on that batch and redeployed. We show that this batch-online setting creates an opportunity for applying the classical approximate policy iteration algorithm. Our central observation is that running standard token-level, single-turn GRPO with a learned turn-level $Q$-function as the reward model is a policy improvement step for the multi-turn problem. Building on this, we present Iterative Policy Optimization (IterPO), which alternates between fitting $Q^π$ from logged Monte Carlo returns via standard reward modeling (policy evaluation) and running single-turn GRPO against $Q^π$ (policy improvement). During policy improvement, the learned $Q^π$ scores candidate responses by their expected downstream return, avoiding the need for simulator rollouts of the remaining conversation. Overall, this allows us to do principled multi-turn RL using only single-turn RL methods without an interactive environment or user simulator inside the training loop. We demonstrate the effectiveness of IterPO on six multi-turn negotiation environments.

cs.LG

版本更新 · v3arXiv:2605.00020

AirFM-DDA: Air-Interface Foundation Model in the Delay-Doppler-Angle Domain for AI-Native 6G

Kejia Bian, Meixia Tao, Jianhua Mo, Zhiyong Chen, Leyan Chen

原文摘要

The success of large foundation models is catalyzing a new paradigm for AI-native 6G network design: wireless foundation models for physical-layer design. However, existing models often operate on channel state information (CSI) in the spatial-temporal-frequency (STF) domain, where multipath components are superimposed and structurally entangled. This hinders the learning of a universal channel representation. Their reliance on global attention also incurs prohibitive overhead. In this paper, we propose AirFM-DDA, an Air-interface Foundation Model in the Delay-Doppler-Angle (DDA) domain. AirFM-DDA reparameterizes CSI into the DDA domain to resolve multipath components along physically meaningful axes and employs window-based attention with frame-structure-aware positional encoding. Extensive experiments demonstrate transferability across scenarios, tasks, datasets, and antenna configurations. For channel prediction and estimation, AirFM-DDA generalizes zero-shot to unseen cities, achieving average normalized mean-square error (NMSE) gains of 4.9-8.5 dB over the strongest baselines. With only 10% labeled data, it achieves average gains of 12.0 percentage points in Top-1 accuracy for beam prediction and 3.4 percentage points in F1 score for line-of-sight (LoS) identification. It further transfers across simulated datasets and adapts to measured data and different antenna arrays. Compared with global attention, window-based attention reduces training and inference costs by nearly an order of magnitude.

cs.LG · cs.AI · cs.IT · eess.SP

版本更新 · v2arXiv:2609.30967

MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens

Subhojyoti Mukherjee, Md Mehrab Tanjim

原文摘要

Most work on improving large language models treats accuracy as the sole objective. We argue that the harness, the Python code surrounding the model that constructs prompts, routes calls, and parses outputs, is a first-class design surface whose quality is inherently multi-objective: an accurate harness that refuses no unsafe request, or that consumes an order of magnitude more tokens, is not a good harness. We present Meta-Harness, a system that casts harness design as search over three per-domain objectives (accuracy, behavioural safety, and token cost) solved by an agentic proposer (Claude Code) with full filesystem access to prior harness source, execution traces, and scoring artifacts. Our central finding is that a singlephase joint-reward proposer (MoMHa) outperforms every alternative, including a two-phase "accuracy then tokens" ablation, scalar-only feedback, and an accuracy-only baseline. We evaluate on seventeen domains: seven synthetic capability suites, seven real-world public benchmarks (HumanEval, MBPP, Spider, FEVER, MMLU-Pro, LawBench, NuminaMath), and three U-SafeBench-derived user-specific safety domains, using a 12-model fleet spanning four families. On the synthetic track MoMHa achieves a joint mean of 0.482 versus 0.198-0.422 for ten baselines, winning $7 / 10$ per-domain columns; on the real-world track it scores 0.461 versus 0.377 for the strongest baseline (DSPy), winning 5/7 columns, demonstrating that harness strategies transfer to unseen benchmarks without retraining on 8 of 12 target models. MoMHa attains the highest measured behavioral safety composite (U-SafeBench, 0.781) and uses 95 fewer tokens per example than the two-phase alternative. We will release all harness code, evaluation infrastructure, and crossmodel logs.

cs.AI

版本更新 · v2arXiv:2611.03285

Mesh-Agnostic Refinement Models for Latent Space Encoding of 3D Human Body Shapes

Chaima Zouaoui, Jonathan Cerqueira, Sylvain Arguillere, Martin Bauer, Nicolas Charon, Mohamed Daoudi, Emmanuel Hartman

原文摘要

Parametric human body models such as the Skinned Multi-Person Linear (SMPL) model provide compact representations of human shape and pose, but their reconstructions remain affected by both errors in estimated latent parameters and limitations in the decoder's expressive capacity. We propose a framework for refining these reconstructions by exploiting the geometric discrepancy between an initial decoded surface and a target mesh, without requiring vertex correspondences between them. The framework comprises three complementary approaches. Latent-VaRefine uses varifold distance gradients to correct the estimated latent code while preserving the original parametric representation. Vertex-VaRefine processes multiscale varifold-gradient features with a DiffusionNet backbone to predict vertex displacements in a canonical pose, yielding an approximately pose-equivariant refinement pipeline. ElaOpt directly optimizes surface deformations under elastic regularization and normal direction constraints. On AMASS, Vertex-VaRefine reduces the mean vertex-to-vertex error of OmniFit reconstructions from 1.181 to 0.342 cm, a 71.0 percent reduction, while combining Latent-VaRefine with ElaOpt reduces the asymmetric Chamfer distance of ArtEq reconstructions from 1.681 to 0.717 cm, a 57.3 percent reduction. Evaluations with SMPL-H and SMPL-X demonstrate the framework's adaptability to different body models, while additional experiments on remeshed targets and raw scans assess its robustness to changes in discretization and geometry beyond the parametric model. These results highlight the complementary strengths of latent space correction, learned vertex refinement, and elastic optimization, offering different trade-offs between preserving the model parameterization and improving geometric fidelity.

cs.CV