Bingran You, Wenjun Ke, Ria Rosenauer, Bo Liu, Wenbo Chen, Min Chen, Shuqi Xu, Yue Ma 等 72 位作者
AI agents are playing an increasingly important role in physics research. However, existing physics benchmarks only evaluate fixed questions, isolated sub-tasks, or paper reproduction with given solutions, and rarely check research quality alongside correctness. We introduce FrontierPhysics, an end-to-end benchmark with 56 tasks made by 42 physicists based on their own research, covering nine domains across theoretical, computational, and experimental physics. Each task requires agents to review the literature, develop a research plan, conduct research and experiments, and produce structured outputs and scientific manuscripts. Combining deterministic verifiers with expert-authored rubrics, FrontierPhysics evaluates whether agents' research is correct and publishable. Across 22 models in 26 configurations, no configuration passes more than 46 of its 168 rollouts. Manual review of agent trajectories shows that frontier agents' main weaknesses are insufficiently rigorous reasoning and limited expert intuition, rather than gaps in physics knowledge or execution errors.
Vision-Language-Action Models (VLAs) enable robots to perform manipulation tasks by leveraging Vision-Language Models (VLMs) to process visual observations and language instructions. However, the action prediction objective itself lacks fine-grained task-specific perceptual supervision, making it prone to relying on task-irrelevant visual shortcuts, leading attention away from target objects and key interaction areas, thereby limiting its operational performance and generalization ability. To address this, we propose Automatic Perception-Guided Attention Alignment for Robot Manipulation ($A^3$-VLA), a framework that derives task-specific visual guidance from the manipulation process itself. Specifically, $A^3$-VLA first generates heuristic candidate regions for target objects and the robot from interaction-induced visual changes, and refines them with a visual segmentation model offline. These regions serve as explicit perceptual guidance through two complementary objectives: an attention allocation loss that encourages sufficient focus on task-critical regions when language or robot state tokens query the visual input, and a contrastive learning loss that aligns the representations of candidate target regions with language instructions while separating them from task-irrelevant regions. Extensive experiments on simulated and real-world manipulation tasks demonstrate that $A^3$-VLA consistently improves multiple VLA backbones, with an average success rate gain of 6.5% - 25.2% across benchmarks.
This work presents UGD-FM, a novel unaligned image guided denoising framework based on cross-modal conditional flow matching. Unlike previous approaches that assume spatially aligned inputs or only handle small-baseline rectified pairs, we address a more challenging setting where severe target noise and large-range two-dimensional cross-modal misalignment coexist. Specifically, we formulate guided denoising as a progressive flow matching process, allowing image restoration and aligned guidance aggregation to mutually reinforce each other. To support efficient few-step inference, we introduce a stage-focused and inference-consistent training strategy that focuses on task-critical denoising stages, and mitigates the training-inference mismatch. We further propose adaptive sparse guidance propagation for efficient and reliable large-range matching, regularized by perceptual and geometric auxiliary losses. Experiments demonstrate that UGD-FM achieves state-of-the-art denoising performance with only one or two inference steps, and is compatible with modern single-image restoration backbones.
Current world models for explicit world simulation typically follow two distinct pathways: (1) video generative models (e.g., Sora), which simulate realistic appearances but fail to preserve 3D consistency; and (2) 3D generative models (e.g., Marble), which generate geometrically consistent scenes but often yield static, less realistic observations. We present AEON, a unified framework that bridges video-based and 3D-based world models and combines the benefits of both. The key insight of AEON lies in a generative model that operates within the text- or image-conditioned video latent space while producing renderable 4D graphics primitives---specifically, 4D Gaussian Splatting (4DGS)---as the explicit world representation. We train a 4D reconstruction model to predict both 4DGS and camera poses from monocular videos, serving as our initialized 4D decoder. Furthermore, we design a domain adapter that translates generative video latents into reconstructive 4D latents through a distillation process on the video VAE. AEON naturally supports video-based world modeling by rendering 4DGS at predicted camera trajectories, and constructs a persistent 4D environment with scene dynamics for realistic world simulation. Consequently, AEON empowers a diverse range of world modeling applications and achieves state-of-the-art performance on standard benchmarks. Project Page: https://weiqi-zhang.github.io/AEON
Large language model (LLM) shopping assistants now recommend products in multi-turn conversations. However, reliable evaluation faces Data scarcity and Difficulty Standardizing Criteria. To address these obstacles, we introduce CARTBench for CART (Complex and Ambiguous Requests with Timelines). For Data scarcity, we curate cases from real commercial shopping needs across 17 retail domains, pairing timestamped dialogue with 1,603 product records. For Difficulty Standardizing Criteria, we develop case-specific rubrics guided by Ideal Response Principles, distilled from excellent replies manually selected from successfully completed orders. These principles also guide Principle-Guided CART Prompting, our structured recommendation strategy. The experimental benchmark contains 1,328 expert-annotated criteria across Core, Excellence, and Safety. To answer the CART question, we evaluate 7 flagship commercial API models, including GPT-6.1-Sol and Claude-Opus-5, and 35B+ parameters open-weight models. In the first full-suite evaluation under the frozen rubrics, 3 of 10 models exceed 90% on Core, while none exceeds 45% on Excellence. Therefore, we identify an understanding-to-execution gap between required behavior and high-quality purchase advice. To examine these limitations, we diagnose shared failures in product evidence and unresolved needs, linked to changing requirements and decisive unknowns. Code and data are available at https://github.com/Da1yuqin/CARTBench.
Large language models (LLMs) can memorize and reproduce training sequences verbatim, raising privacy and copyright concerns. Existing mitigation methods apply interventions uniformly, often degrading predictions at the many positions unrelated to verbatim recall. We study a token-level log-likelihood ratio between a target model and a domain-competent reference with lower exposure to the evaluated sequences. Exposure-controlled experiments support its use for localizing positions associated with sequence-level verbatim recall. Across evaluated settings, a minority of positions carries a disproportionate share of the positive signal, motivating selective activation correction. From a first-order surrogate of the memorization-reduction objective, we derive a probe-steer framework, which decomposes intervention into a probe that flags high-signal activations and a steer that applies a correction only when the probe exceeds a threshold. We propose Gated Subspace Steering (GSS) as its practical instantiation: under a background normalization constraint, the probe-steer pairs that are optimal for this surrogate follow in closed form from the leading singular vectors of a whitened gradient-weighted activation matrix. Across four benchmarks, GSS achieves memorization suppression comparable to strong baselines, with favorable utility trade-offs in several settings, while requiring 10-100$\times$ less compute than optimization-based alternatives. Code is available at https://github.com/ubc-tea/Mitigating-Memorization-Where-It-Happens.
Fast rollout generation for long-context reinforcement learning with verifiable rewards (RLVR) requires reducing KV-cache traffic without changing the response groups used for policy optimization. In our profiled RLVR workloads, attention accounts for 72--76\% of autoregressive decoding kernel time. Sparse attention reduces these reads, but sampling directly from a sparse policy changes the response groups used for learning. For group-relative optimization, we show that even exact response-wise reweighting does not generally recover the expected dense-policy update because the advantage of each response depends on the rewards of other group members. We propose Self-Sparse Speculative Rollouts (S$^3$R). It uses sparse attention to propose tokens, then generates responses through full-attention verification and exact speculative sampling. Drafting and verification share the same policy parameter, eliminating the need to train or update a separate drafter. The resulting response groups follow the dense-policy distribution. Under S$^3$R, we characterize speedup conditions and the costs of drafting and verification. We found that dense verification still requires the complete historical KV cache. A KV-aware scheduler therefore limits KV-cache repeated offloading and restoration while maintaining rollout parallelism. % OR thrashing In a matched ablation with an aggressive sparse drafting budget, S$^3$R maintains stable training while direct sparse sampling with training-side correction becomes unstable. On long-context mathematical reasoning, S$^3$R achieves $1.39\times~1.35\times$ the end-to-end training throughput of autoregressive rollouts on Qwen3 model, with comparable task performance.
Large language models are deployed as long-horizon agents that compose diverse tools across MCP services into executable workflows. Existing MCP environment-scaling methods are data-centric: they first synthesize a database and scenario, then derive tools and tasks from them. As a result, tool interfaces may not match real ones, tool compositions are not grounded in execution, and tasks rarely span more than one environment. We propose ToolSpace, a tool-centric framework that reverses this order. ToolSpace mines data, state, and retry dependencies from adjacent calls in execution traces of real MCP services and open agent corpora, using deterministic rules rather than LLM judgment, and clusters the resulting weighted tool-composition graph into environments. Databases are synthesized only for state-mutating environments, while query and stateless tools replay recorded returns. ToolSpace then performs state-aware random walks with rollback and decompiles each chain into a natural-language task with assertion-based checks, keeping only samples that replay successfully from the initial state. Applied to AutomationBench and tau3-Bench Banking, the pipeline produces harder Pro versions on which 7 frontier models drop by up to 38.2 points. Across seven long-horizon tool-use benchmarks, ToolSpace raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score (43.2 with RL) and GLM-5.2 from 44.6 to 56.6, on par with Claude Opus 5. Ablations show that execution-witnessed edges clearly outperform LLM-judged ones, and performance keeps improving with data and environment scale.
Zhiheng Li, Zongyang Ma, Jiaxian Chen, Cheng Cui, Yi Liu, Ziqi Zhang, Chunfeng Yuan, Bing Li 等 10 位作者
Optical Character Recognition (OCR) is among the oldest and most widely deployed applications of artificial intelligence. This survey reviews its development, focusing on the Foundation Model Era. We define OCR through faithful transcription: within a specified visual scope, OCR should completely reconstruct the text, notation, and document structure observable in the source image, and every textual, notational, or structural element in the output must be verifiable from visual evidence in that image. We trace the development from specialist OCR models toward unified models, examining joint text detection and recognition, structured sequence generation, and the use of pretrained vision-language models (VLMs). We organize contemporary OCR systems into five coexisting families: modular toolkits and task-specific models, specialized document VLMs, hybrid workflows, general-purpose VLMs, and commercial products and managed services. We compare their architectures, data, and training, examining hallucination and omission, insufficient spatial grounding, and inference cost in generative OCR. We review evaluation from character and scene text recognition to parsing individual structural elements and full-page transcription, highlighting gaps in consistent output definitions, attribution of performance to individual components, and coverage of real capture conditions and visible content and structure. Finally, we summarize the challenges facing OCR and discuss directions for future research. We continuously track related work at https://github.com/zhihengli-casia/Awesome-OCR-in-the-Foundation-Model-Era.
Industrial recommender systems allocate substantial computation to pointwise candidate scoring, yet separate reranking and mix-ranking stages still assemble the final exposure slate using hand-crafted rules and specialized low-capacity models. Mix-ranking combines product cards, short videos, livestreams, and advertising cards supplied by distinct upstream business channels. Although these rules and models encode valuable domain knowledge, they are inherently limited in leveraging additional data and computation. They constrain the combinatorial space of candidate slates and determine objective trade-offs through manual tuning. To enable scalable and effective slate generation, we reformulate reranking and mix-ranking as a unified learning problem whose performance can improve with computation. Specifically, we instantiate this formulation across the reranking and mix-ranking stages and propose OneTrans-GSR, a single decoder-only policy that extends our deployed OneTrans backbone to generate the exposure slate end-to-end. During training, a reward model evaluates complete slates and guides reinforcement learning toward slate-level and long-term objectives. During decoding, business constraints are explicitly enforced. Offline evaluation shows that slate quality consistently improves as reward-model and policy compute increase. Deployed in a large-scale industrial recommendation system, OneTrans-GSR mixes heterogeneous content types within a unified slate. In online A/B tests, it improves GMV/u by 2.63%, advertiser value by 2.06%, and repurchase rate by 2.03%, while reducing end-to-end latency across the reranking and mix-ranking stages by 13%. These results mark a paradigm shift in slate generation: moving beyond human-engineered heuristics toward a general learned policy creates a scalable path for further algorithmic optimization and compute-driven improvement.
Jie Liu, Xusheng Tang, Chao Wang, Zhengxiao Yan, Yuchen Dai, Qi Zhang, Conghe Song
High-resolution land-cover maps support ecological monitoring, but consistent historical time series remain scarce because aerial imagery varies across years and sensors. We developed a dual-track workflow to produce temporally coherent 1-m land-cover maps from National Agricultural Imagery Program (NAIP) imagery without historical training samples. For higher-quality imagery from 2009-2017, we used a pretrained U-Net/ResNet-34 model and refined boundaries with the Segment Anything Model (SAM). For lower-quality imagery from 2004-2008, we reconstructed land cover through an NLCD-based spatiotemporal bridging and label back-casting pipeline. Evaluation across six counties in North Carolina and Pennsylvania yielded overall accuracies of 0.887 (2014), 0.886 (2017), and 0.733 (2004). Structure, Water, Wetland, and Cropland consistently achieved high F1-scores. The workflow integrates GeoAI and multiple data sources to address temporal heterogeneity in historical aerial imagery, supporting high-resolution land-change analysis and landscape planning across the conterminous United States.
Large language models (LLMs) increasingly shape recommendations across chatbots, e-commerce, and user-generated content (UGC) platforms, creating demand for Generative Engine Optimization (GEO) to improve item visibility in generated responses. Existing GEO studies are mainly content-centric, optimizing visibility under general-purpose queries. However, a shared item description usually serves users with different needs and decision goals, resulting in its visibility gains being dependent on the request. To investigate how changing intent affects visibility, we formulate IntentGEO to evaluate shared rewrites across need specificity and decision progression. We revisit existing GEO methods and find that their visibility gains vary across user intents. More critically, learning-based prompt optimization can underperform its handwritten seed on average. Tracing these updates reveals prompt drift from explicit recommendation instructions toward generic promotional tactics. To address this drift, we propose AMO (Anchored Meta-Optimization), which preserves recommendation instructions in a fixed core while learning appended strategies from citation feedback. Across four domains and ten intent states, AMO improves average citation-rank gain by 35% over the strongest prior learning-based baseline and leads on nine states. Ablations show that learned additions improve on the fixed core alone and that AMO outperforms full-prompt optimization initialized from that core. It also achieves the highest gains in an agentic recommendation pipeline and, after separate adaptation, on Doubao, a commercially deployed, closed-source online assistant. These findings highlight user intent as a key dimension of GEO evaluation and support anchored adaptation across recommendation needs and decision stages.
Rapid advances in cowork agents are shifting human-AI collaboration toward real-world automation, delivering software, designs, and other artifacts. However, such productivity work demands both high-quality delivery on complex productivity tasks and continued monitoring and decision-making over long horizons, a combination that remains an open challenge. We introduce CoworkOS, a general-purpose harness for long-horizon work that separates task strategy from execution management: the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. Its three-layer architecture enables direct execution or delegation to Subagents, synchronously or asynchronously, under stable responsibilities. Through a shared application layer, humans and agents inspect and revise the same evolving documents, presentations, CAD models, and Blender scenes, so human-in-the-loop participation can enter both through conversation and through direct modification of the work itself. To sustain continuity across agents and sessions, CoworkOS further introduces hierarchical workspaces and persistent semantic memory, so working files, key decisions, and long-running Job logs remain shareable, reusable, and resumable after interruption. Across 12 benchmarks and 6 categories of real-world applications, CoworkOS proves highly competitive with frontier agent applications such as Codex and DeepSeek Harness. As a sustained-work case study, CoworkOS uses GLM-5.3 as its backbone to conduct RSI-driven automated post-training on Qwen3.6-35B-A3B. Over five rounds, the model's average score across eight benchmarks rose from 33.19 to 42.95 (+9.76). Looking ahead, we believe CoworkOS opens a meaningful path toward general cowork-agent intelligence.
Generalization remains a central challenge in robotics, calling for policies that address domain variations from multiple dimensions such as diverse embodiments, camera poses, and/or scene layouts. In order to approach this problem, we introduce PointPilot, a task-focused and 3D aligned manipulation framework built on a 3D space point-flow prediction. Rather than modeling the global scene, PointPilot only represents task-focused objects (robot grippers treated as objects), avoiding the prediction of potentially redundant scene parts unrelated to the manipulation tasks. Moreover, PointPilot aligns 3D spatial representation in one unified coordinate and supports closed-loop and fully automatic action execution across different manipulation modes from object point flow without a separately trained action expert, while existing works typically need an extra action head or are (semi-)automatic and single-arm mode specific. Given sensor observations and language instructions, our model jointly learns subtask planning, task-relevant object grounding, and 3D point flow prediction in an end-to-end way. By processing heterogeneous data in a shared 3D point flow space, PointPilot is able to absorb demonstrations from different embodiments. Extensive simulative and real world experiments validate our method. Notably, we demonstrate zero-shot embodiment generalization with PointPilot. A single trained policy can be deployed on previously unseen embodiments and successfully complete manipulation tasks without any additional fine-tuning.
Long-lived agents rely on persistent memory of users, tasks and decisions, and this memory must stay consistent with a world that keeps changing. When an upstream fact changes, however, memory systems can revise the contradicted memory itself while the beliefs, plans and actions derived from it remain accepted, even though they have lost their support. We therefore formulate dependency-aware state revision, which asks not only what changed but also which dependent states should no longer be trusted. To address it, StateGraph maintains an evolving state dependency graph whose edges are recovered from interaction histories. Its core design is counterfactual verification: each candidate edge is tested by asking whether the dependent state would survive the loss of its premise, which separates support dependence from semantic relatedness, and revision then invalidates a dependent state only when a strict dependency fails and no independent support path survives. To test this design, we introduce StateChangeBench, a controlled benchmark of 50 author-constructed cases annotated with cascade depth, hard negatives and action consequences. On it, StateGraph's Impact-F1 margin over a CUPMem reimplementation is 6.9 percentage points at depth 1 and 26.0 at depth 3+. On public long-term memory benchmarks, its point estimates are comparable to or higher than those of the evaluated baselines on most datasets. These results provide initial evidence that explicit dependency tracking can improve selective memory revision under changing conditions.
Recent advances in video generation have enabled generative video compression to reconstruct details lost during coding using pretrained spatiotemporal priors. However, a substantial gap remains between codec-induced degradation and the perturbation trajectories learned by flow-matching models. In this paper, we identify a compression-dependent directional mismatch that degradation-strength-based timestep calibration alone cannot resolve. Compressed latents remain strongly aligned across codec operating points, while their reconstruction residuals vary substantially in magnitude and direction. To bridge this gap, we propose RIFT, a Residual-Informed Flow Transfer framework for one-step generative video compression. At its core, RIFT introduces a Residual Geometry Inducer (RGI) that converts the pretrained flow prediction into a content-adaptive restoration update using decoder-visible codec cues. RGI integrates flow-timestep conditioning with a learned full-band residual-strength estimate and content-conditioned direction and gain modulation, aligning the orientation and magnitude of the update with the compression-dependent residual geometry. Extensive experiments on three public video compression benchmarks show that RIFT outperforms competing generative video compression methods. The code is publicly available on RIFT.
Latent Path Exploration via Potential-Guided Subtrajectory Balance
Xiaoling Zeng, Qi Zhang
AI 速览：提出 SubLAT，用相对答案潜力与 Subtrajectory Balance 引导语言模型在连续潜在空间中的随机推理路径探索。
Continuous latent reasoning enables language models to perform intermediate computation without generating explicit reasoning traces. Advances in stochastic latent reasoning allow multiple paths to be sampled in continuous space, but stochasticity alone does not ensure effective exploration. A key challenge is guiding stochastic transitions between intermediate latent states toward useful reasoning paths. We introduce SubLAT, which couples a relative answer potential with Subtrajectory Balance to learn stochastic latent exploration. The relative potential evaluates the answer support of sampled states against replacement states, with a correctness signal at the terminal state. Subtrajectory Balance relates potential differences to forward and backward transition densities across subtrajectories of different lengths. The model is jointly trained with answer supervision, CoT supervision, and representation distillation. Experiments on arithmetic and deductive reasoning show increasing solution coverage with larger sampling budgets across in-domain and out-of-domain benchmarks, as well as improvements on several structural OOD benchmarks. Trajectory analyses reveal progressive divergence among sampled latent paths and higher mean relative potentials for correct trajectories across arithmetic benchmarks.
Shiqi Zhang, Guochen Zhou, Zesheng Wei, Hai Zhao, Zhisong Zhang
Web agents are typically evaluated with goals fully specified before execution and fixed thereafter, whereas real users often revise their requirements during interaction. To study such interactions, we introduce DynaRev, a benchmark of 181 tasks to evaluate web agents under dynamic goal revision. DynaRev uses semantic milestones to deliver revisions at task-defined stages, covers additive revisions, unconditional overrides, and runtime-conditioned overrides, and evaluates correctness, responsiveness, and efficiency through trajectory-grounded evaluation. Five-model evaluations show that agents often satisfy only part of the task requirements, with performance varying across task sources, revision types, and models. Further analyses reveal varied milestone timing, mixed effects of dynamic delivery, and contrasting revision-handling outcomes, with gains from instruction rewriting while preserving browser state and losses from browser resets. DynaRev provides a testbed for studying how web agents adapt to evolving user goals. The code and data are available at https://github.com/zsq259/DynaRev.
Recent advances in autoregressive video diffusion have enabled sequential and streaming video generation. However, long-horizon generation requires increasingly large KV caches, making efficient compression without sacrificing quality challenging. Existing methods mostly select historical frames based on attention scores, but their context decisions remain coarse. When multiple frames are generated in the same chunk, these methods often apply a shared history selection to the whole chunk, score historical frames solely by attention, and assign head-wise budgets either uniformly or by attention-pattern heuristics rather than explicit head-importance estimation. We show that frames within the same generated chunk can depend on distinct historical frames, that the same historical frame can receive different attention scores as its relative temporal distance to the current frames changes, and that masking different heads induces unequal generation degradation. Motivated by these findings, we propose Focused Forcing, a training-free KV selection method that focuses cached history along both generated-frame and head dimensions. For each generated frame, Focused Forcing preserves the most relevant and distinctive historical frames by combining attention scores with diversity scores of historical frames, while assigning larger budgets to heads with higher estimated importance. Across multiple autoregressive generation paradigms, Focused Forcing achieves up to 1.48x end-to-end acceleration without training, while improving visual quality and text alignment. Our code is available at https://github.com/caipeiliang2004/Focused-Forcing.
On-device deployment of large language models (LLMs) is increasingly adopted due to latency and privacy requirements, but exposes proprietary models to reverse engineering and intellectual property theft. Trusted Execution Environments (TEEs) provide hardware-enforced isolation for on-device inference, yet most TEE-based parameter obfuscation relies on unilateral matrix transformations that obfuscate only the output side. We identify a vulnerability in this design, termed the Solution Determinacy Property (SDP): the obfuscation secret is uniquely determined, exposing the structure of the true key and enabling attackers to break the protection. Under our attacks, representative methods GroupCover and ArrowCloak are reduced to negligible protection, with attack accuracy reaching over 2.4 times the full-shielding black-box baseline and recovering near-complete model functionality. To address this vulnerability, we propose BiShield-TEE, a bilateral framework that induces Ambiguous Solutions through joint obfuscation of input and output dimensions, hiding the true key among many equally valid candidates. Experiments across diverse models and datasets show that BiShield-TEE confines attack accuracy to the full-shielding black-box baseline, while achieving higher inference efficiency than prior methods on real hardware. Our code is available at https://github.com/laduiw/BiShield-TEE.
Yi Pan, Miao Pan, Qi Lu, Jiaming Huang, Man Zhang, Siteng Huang, Xin Li, Jie Zhang 等 11 位作者
Vision-Language-Action (VLA) foundation models have recently achieved strong progress in embodied intelligence. To reduce policy-call frequency while preserving temporal coherence, most generative policies adopt an action chunk mechanism, executing multiple future actions in an open-loop manner under a fixed action horizon. However, this "predict-then-blindly-execute" paradigm sacrifices closed-loop reactivity: in contact-rich physical interactions, even small local perturbations can rapidly amplify within the open-loop blind spot, leading to compounding errors and ultimately task failure. To address this limitation, we propose VLA-Corrector, a lightweight corrective inference framework for action-chunked VLA policies. Without modifying the backbone policy weights, VLA-Corrector introduces a lightweight Latent-space Vision Monitor (LVM) that continuously compares predicted and actual visual feature evolution, enabling online detection of visual dynamics deviations. Once persistent deviation is detected, the system triggers a truncation event, discards the remaining stale actions, and invokes corrective replanning via Online Gradient Guidance (OGG). The detect-and-correct mechanism of VLA-Corrector naturally induces an event-triggered adaptive action horizon: it preserves long-horizon execution when the current chunk remains reliable, and invokes short-horizon corrective replanning when execution begins to drift. In doing so, VLA-Corrector mitigates the trade-off imposed by static horizons between execution robustness and policy-call frequency. It can be integrated into different VLA models without further retraining the VLA backbone, interrupting compounding errors while preserving much of the efficiency benefit of action chunking and substantially improving robustness in long-horizon, contact-rich robotic manipulation tasks.
Guoqiang Zhang, Kexin Tan, Ming Zhang, Li Ju, Wenqing Jing, Zhonghan Yue, Jiayi Chen, Shiqiang Wu 等 15 位作者
Large language models (LLMs) are increasingly used in peer review at major AI conferences, yet novelty remains a persistent weak point. Existing benchmarks assess novelty as a single holistic score, making it difficult to diagnose which dimension a model misjudges or whether its evidence is faithful. We present NovGauge, a human-anchored benchmark for fine-grained novelty assessment diagnosis. The benchmark contains 619 paper pairs and 50 multi-paper sets, drawn from two expert sources: ICLR reviewer overlap claims and survey co-citations. Instances are independently labeled along three dimensions: task, problem, and method, capturing application goals, technical challenges, and solution approaches. We propose a cascading diagnostic pipeline that verifies per-dimension correctness, evidence grounding, and logical support. Evaluation of 18 LLMs shows hallucination rates ranging from 0% to 39% across dimensions, and among non-hallucinated correct-positive judgments, over 70% cite evidence fails to logically support the stated reason. The best-performing model, GPT-5.5, achieves 43-72% Verified F1 across dimensions, while most models retain less than half of their raw F1 after faithfulness verification. These results suggest that current LLMs remain far from reliable scientific novelty assessment, particularly when correctness is conditioned on faithful evidence grounding.
Junzhang Liu, Hao Yuan, Yiming Wu, Qi Zhang, Long Chen, Zhecan Wang
Effective video agents must learn not only how to answer, but also where to look next. Transferring this capability to smaller models is challenging because each observation serves two purposes: supporting an answer and guiding the search for further evidence. Textual descriptions may capture what is currently visible while omitting cues needed for subsequent acquisition. We introduce Latent Agentic Reasoning for Video Analysis (LARVA), a framework that distills multi-round video agents through jointly learned latent reasoning and visual observations. Rather than communicating through captions, an observer passes compact continuous representations to a reasoner that selects the next acquisition action or produces the final answer. Crucially, action and answer supervision propagate through this interface, training the observer to communicate information useful for downstream decisions. Observation reconstruction and reasoning-distribution alignment provide auxiliary supervision from textual teacher trajectories. Across five video-understanding benchmarks, LARVA achieves strong performance with 4B and 7B backbones. On VideoHolmes, Video-MME, and LVBench, the 7B model outperforms a textual CoT student trained on the same teacher trajectories by 5.1, 2.5, and 3.7 percentage points, respectively. These results highlight the value of jointly learning how agents represent evidence and act on it when distilling video-understanding capabilities.
Xuan Wang, Tianqi Zhang, Gabrielle De Micheli, Yujin Nam, Keming Fan, Augusto Vega, Minxuan Zhou, Tajana Rosing
Fully Homomorphic Encryption (FHE) enables computation on encrypted data and is promising for privacy-preserving computing. However, FHE deployment remains constrained due to massive computational, memory, and data-movement overheads. Existing acceleration strategies often overlook the storage I/O bottleneck caused by the enormous size of FHE data in large-scale applications, as most prior works focus on small benchmarks that fit in main memory. Our studies show that state-of-the-art FHE systems spend 71\% to 85\% of their execution time moving data between storage and compute units rather than performing useful computation. To address the I/O bottleneck, we present Rosemary, a heterogeneous FHE acceleration system that incorporates a novel in-storage processing (ISP) architecture with compiler-level and algorithmic optimizations. Our key innovations include 1) an adaptable hardware-aware FHE engine inside a NAND flash-based solid-state drive (SSD) for efficient and cost-effective in-storage FHE processing, 2) a Throughput-Data-Aware Program Allocator (TDAP) that partitions FHE workloads into heterogeneous programs, and 3) a Compiler-Level Ciphertext-Aware Optimizer (CLCAO) that exploits ISP-friendly FHE kernels for low-cost ISP execution. Across a diverse set of FHE workloads, including encrypted deep learning inference and cloud database queries, Rosemary achieves geometric-mean speedups of $185.7\times$ over state-of-the-art CPU/GPU implementations and $3.7\times$ over prior FHE accelerators, while improving energy efficiency by up to $3.9\times$ and Energy-Delay-Area-Product (EDAP) by up to $16.5\times$.
Pascal Jutras-Dubé, Patrick Pynadath, Jeremy Lu, Yuan Gao, Ruqi Zhang
We introduce sticky jump diffusions (SJDs), continuous-time Markov processes on $\mathbb{R}^d$ whose discrete anchors are token embeddings. In forward time, anchors release their mass at a hazard rate and the released mass diffuses in the continuous ambient space; time reversal couples a score-driven SDE with a sticky jump kernel whose rate and destination are fixed by flux balance with the forward law. We estimate the score and the per-anchor reverse hazards from a single denoising classifier via denoising hazard matching, the hazard analogue of denoising score matching, with simulation-free cross-entropy training. SJD recovers masked diffusion, continuous diffusion, and hybrid diffusion as limits. Its reversal explains features that each family treats as given: the mask of masked diffusion carries no evidence about the source token because the unsticking kernel of every anchor collapses to the same absorbing point; the terminal projection of continuous diffusion is required due to the absence of atoms in its forward marginal, without which flux balance yields no reverse jumps; and the update rules of hybrid diffusion (commit rate, destination, and drift) all follow from flux balance rather than from separate design. Beyond these limits, the unsticking kernel becomes a design space: a cross-position blending corrupts each position toward a blend of its neighbors' clean values or embeddings, turning dependency structure such as spatial locality or a constraint graph into an inductive bias of the corruption itself, and improves over the identity-kernel hybrid on CIFAR-10, Text8, and Sudoku.
Recent work has explored improving agents by jointly evolving their harnesses and models, but often takes a ''potpourri'' approach that bundles together new tools, new decision-making procedures, and model adaptation to the evolved harness under a single notion of agent improvement. In this paper, we instead investigate how agents can improve their decision-making procedures. In particular, we propose EvoIn, an agent fine-tuning framework that bridges evolution and internalization. EvoIn first analyzes agent execution traces to evolve and validate new decision-making procedures by temporarily instantiating them in the harness. The validated procedures guide the agent to generate improved reasoning traces. These traces are then rewritten into self-contained reasoning traces, removing explicit references to harness instructions while expressing the induced decision logic as the model's own reasoning. Finally, EvoIn fine-tunes the model on the rewritten traces, internalizing these procedures so that the improved decision-making persists without the evolved harness at inference time. We evaluate EvoIn on diverse benchmarks and find that it consistently enables agents to learn stronger decision-making procedures, raising the pass rate by 10.9 points in-domain and by 9.2 points out-of-domain. Results further show that the internalized decision procedures generalize to unseen tasks. Case studies show that agents can learn to decide how to solve a task before solving it, for example by checking a document's length to choose between reading it in full and searching it. EvoIn is also broadly applicable, showing consistent improvements on another model family.
Hybrid large language models (LLMs) have emerged as a promising way to combine the modeling capability of the Attention mechanism with the efficiency of state space models (SSMs). However, most existing hybrid LLMs still inherit a static design philosophy: once the Attention-to-SSM layout is chosen, every request must follow the same hybrid execution. This one-layout-fits-all design is inefficient for real-world serving workloads, where different requests rely on contextual information in different ways. This paper presents Morph, a serving system that co-designs the model architecture and runtime for request-adaptive hybrid LLM inference. Its model architecture pairs Attention and SSM layers within each block and uses prefill-oriented gating to select a request-specific layout that remains fixed throughout decoding. Its serving runtime preserves batching efficiency under divergent layer choices through selective batching and enables correct prefix reuse through hybrid-aware context caching. Experiments on six datasets show that Morph achieves a better quality-efficiency trade-off than existing static hybrid baselines. System-level evaluation further shows that Morph improves the serving throughput by up to 3.87x over baselines while substantially reducing inference latency.
Long-context multimodal document question answering (QA) has become an important yet challenging task for modern vision-language models (VLMs), as it requires reasoning over evidence dispersed across pages and modalities within limited input budgets. Current pipelines adopt retrieval-augmented generation (RAG) to select multimodal evidence for downstream VLMs. However, existing systems mainly rely on static selection rules over fixed-granularity retrieval units such as pages or blocks, resulting in information loss during retrieval and suboptimal performance in downstream reasoning. To alleviate these issues, we formulate the core task as budget-aware evidence selection over structured document evidence and propose HyperDoc, a document RAG framework that constructs compact yet sufficient evidence while preserving document relations. Specifically, HyperDoc constructs a document hypergraph to explicitly capture structural, logical, and cross-modal relationships among multimodal document elements. Over this structured evidence space, HyperDoc performs adaptive cost-aware hyperedge selection based on query intent and input budget. Extensive experiments on three representative benchmarks demonstrate that HyperDoc outperforms various types of state-of-the-art baselines.
Interactive world models hold immense potential for high-stakes applications in robotics, game engines, and immersive digital twins, yet existing benchmarks primarily focus on perceptual similarity rather than the physical-causal coherence essential for real-world utility. We introduce WMEval, a comprehensive benchmark that evaluates interactive world models against Unity-based physical simulations. By leveraging Unity as a high-fidelity ground-truth oracle, we shift evaluation beyond perceptual visual assessment toward the objective state-based verification required for downstream applications. Spanning 9 diverse scenes, WMEval assesses action controllability, scene and character consistency, prompt adherence, kinematics, fluid mechanics, physical causality, and lighting. By executing matched action sequences in both the simulator and the world model, WMEval enables evaluation at both the perceptual level and the physical-causal level, using simulator-derived states as reference. Experiments on state-of-the-art models reveal significant room for improvement in dimensions that remain hidden by standard perceptual metrics but are vital for real-world deployments. Our code is available at: https://github.com/hao-ai-lab/WMEval.
Generative recommendation (GR) predicts the next item through its semantic ID (SID). Existing methods commonly fit SID vocabularies to pretrained language model (PLM) representations of item descriptions, incurring additional training costs and making vocabulary coverage dependent on the fitting domains. We observe that pretrained sparse autoencoders expose PLM features with human-interpretable semantic annotations, allowing feature indices to serve as a reusable SID vocabulary. However, reusing this vocabulary requires preserving item distinctions under compact token budgets and learning to favor target items over candidates with shared dominant features. It also requires extending feature interpretability from item content to retrieval decisions. To address these challenges, we propose API, a framework that Adapts the Pretrained Interpretable vocabulary for effective and explainable GR without additional vocabulary learning. Within API, Pretrained Interpretable IDs (PIDs) retain selected feature indices and relative activation strengths for item discrimination. With PIDs as supervision, the Semantic Prediction Adapter (SPA) learns user preferences from interaction histories through joint feature-index and strength prediction. SPA combines both predictions to score candidate items, with item-level supervision encouraging the observed next item to outrank its alternatives. API further explains candidate score differences through feature contributions paired with their pretrained semantic annotations. Extensive experiments show that API achieves competitive performance across backbones and decoding schemes, while preserving item distinctions under compact token budgets and enabling explainable retrieval.