Ziyang Zeng, Tingyang Chen, Jia Deng, Qiyao Wang, Shuo Tang, Yike Yang, Tao Ren, Wayne Xin Zhao 等 13 位作者
Scaling supervised fine-tuning for working agents requires not only broad coverage of professional tasks, but also expert demonstrations that faithfully carry out professional workflows to completion. However, missing or inconsistent task information can disrupt expert demonstrations, weakening the resulting supervision. We introduce WorkFactory, a scalable task-construction framework that turns professional intents into executable task environments through hybrid intent sourcing, semantic workflow planning, and execution-based repair. Starting from these intents, WorkFactory constructs a Semantic Workflow Graph that explicitly represents information requirements, dependencies, and locations within the task environment. The graph guides the instantiation of task environments, while an adversarial execution-refinement loop exposes missing or inconsistent information and steers targeted repairs that preserve the original intent. Using WorkFactory, we construct 112,686 tasks spanning 612 occupations and train the WorkPilot family at 4B, 27B, and 35B-A3B scales on the resulting teacher trajectories. All three models improve over their respective base models across four professional-work benchmarks, with WorkPilot-35B-A3B rising from 879 to 1565 Elo on GDPval-AA v2.1. Scaling experiments with up to 110k tasks further show that larger models benefit more from additional task data but exhibit diminishing returns at smaller data scales, while WorkPilot's long-horizon execution evolves from verbose trajectories toward more efficient task completion. Controlled component studies quantify the negative effects of task-environment defects on downstream performance and show how semantic workflow planning and execution-based repair yield downstream gains through improved task-environment quality. The subset of data are available at https://github.com/IQuestLab/WorkFactory.
Your LR Schedule May Not Be Principled: Effective Learning Rate Scheduling in Weight-Normalized Training
Qifan Yu, Xinyu Ma, Xingyan Bin, Di He
AI 速览：论文指出权重归一化训练的收益主要来自有效学习率轨迹而非归一化本身，并提出以有效学习率为调度对象的 AIR 调度。
Weight-normalized (WN) training is emerging as a compelling alternative to conventional weight decay (WD), with recent studies reporting substantial pretraining loss gains. We show that these gains are driven primarily by the induced effective learning rate (ELR) trajectory rather than by normalization alone. Matching ELR trajectories largely collapses the training loss dynamics of WN and WD. Schedule design should therefore target ELR rather than the nominal learning rate (LR) alone. Building on this principle, we propose AIR (Affine Inverse-square-Root decay), which combines a theoretically derived $1/\sqrt{t}$ decay with an affine adjustment to satisfy prescribed peak and terminal LRs over a finite training horizon. We further identify late-stage ELR scheduling as a key source of its gains, because a shared late schedule attenuates the effect of earlier LR choices. Experiments on 128 B-token Mixture-of-Experts pretraining demonstrate that AIR outperforms cosine and other candidate schedules in training loss, downstream performance, and hyperparameter robustness, reducing terminal training loss by approximately 0.01 relative to cosine at matched cumulative LR. Applying the same ELR trajectory to WD reproduces the corresponding training dynamics, showing that the benefits extend beyond fixed-norm training and establishing ELR as the principled scheduling object.
Formula-Driven Supervised Learning (FDSL) enables vision pre-training without real images, but existing methods cast pre-training as classification on discretized labels, discarding continuity in the generative process. We propose Formula-Driven Supervised Metric Learning (FDSML), which derives relational supervision from image-generation parameters beyond discrete class labels. FDSML combines FineFractalDB, our refined fractal dataset with higher intra-class variation and more stable structures, with a Parameter-Aware Triplet (PA-Triplet) Loss that assigns positives and negatives according to distances in the underlying Iterated Function System parameter space. This objective trains embeddings to respect parameter-derived relative similarity orderings. Experiments show that FDSML improves transferability to image retrieval while remaining competitive on downstream classification and improving robustness under several corruption settings. On ResNet-50, Recall@1 improves from 65.6% to 74.0% on Stanford Online Products and from 56.2% to 78.4% on Stanford Cars over FDSL pre-training. These results suggest that parameter-derived relational supervision provides a stronger synthetic pre-training signal than label-based classification for transfer beyond recognition.
Human cognition flexibly draws on a shared perceptual foundation to understand environments, attend to referenced entities, reason about spatial relationships, and act. Endowing aerial robots with an analogous functional organization remains challenging, as existing systems are often specialized for individual capabilities, dependent on offboard computation, or unable to sustain closed-loop operation on resource-constrained platforms. We present OpenFlyV2, a unified onboard vision-language-action architecture that uses a shared vision-language representation to support semantic understanding, language-grounded perception, spatial reasoning, and continuous aerial control within a single model. A compact action expert converts the shared multimodal context into short-horizon flight waypoints, while motion-aware temporal memory and action priors reduce the computational cost of closed-loop execution. We construct \dataset{}, a simulation-based dataset comprising approximately 70,000 navigation trajectories and 60,000 vision-language instruction--response pairs. OpenFlyV2 generates action chunks, each comprising 12 waypoints, entirely onboard at 10 Hz using an NVIDIA Jetson Orin NX. It transfers zero-shot from simulation to physical quadrotor flight, achieving a 48\% navigation success rate across four real-world environments without real-world aerial training data. Extensive results demonstrate that broad vision-language competence and continuous closed-loop flight can coexist within a single model on a resource-constrained aerial platform. Code will be made publicly available at https://github.com/Ideal-ljl/AeroBrain.
Lu Wang, Zhuoran Jin, Longxiang Wang, Junhong Chen, Yupu Hao, Yao Xu, Daojian Zeng, Yubo Chen 等 11 位作者
Recent advances in multimodal large language models have increased interest in interaction paradigms for video understanding. The evolution from offline to streaming and proactive interaction has gradually shifted models from passive responders toward more proactive agents that can decide when to answer. However, existing proactive paradigms typically assume that the information required to answer a query is complete, leaving models to only wait or answer. In the real world, however, information is often incomplete due to uncertainty from either the video or the user. In proactive paradigms, such uncertainty can cause the model to wait forever or generate highly hallucinatory responses. We therefore argue that proactive interaction should be redefined with a third action: ask, allowing models to actively ask for missing information. To study when and what models should ask before answering, we introduce AskStream, the first multi-turn benchmark spanning video and user uncertainty for streaming video understanding, with four task categories and eight subtasks. It contains 438 videos with 2,714 manually annotated frame-level time windows across three types: wait, ask, and answer. Across both multiple-choice and open-ended settings, even the strongest model attains only 52.18% on our proposed Ask-F1 metric, highlighting substantial limitations in proactive information seeking under uncertainty.
Financial agents must form forecasts and investment judgments from decision-time evidence, with predictions evaluated against subsequent outcomes. Existing benchmarks assess financial knowledge, research workflows, or forecasting, but evaluating the full research process requires a strictly frozen information set and professional evaluation criteria. We introduce Investor's Last Exam (ILE), an equity-research benchmark led by experienced buy-side investors and fund managers. Its three components are a versioned financial database spanning 3.43 million document versions and nine source categories, supplemented by structured market data, with validated timestamps governing retrieval and computation; a task taxonomy organized around professional analyst deliverables; and an expert-led framework linking scoring criteria to sources, calculations, and assumptions through financial dependency graphs. Evaluation spans 86 assignments and jointly assesses research quality, forecast accuracy, and simulated investment performance. GPT-6 Astra and Claude Opus 5.5 lead overall at 68.1 and 66.9 out of 100, yet even Astra scores only 52.5 on Forecast & Model. Our results show that even state-of-the-art models fail to translate overall research capability into predictive performance. ILE shifts the evaluation unit from isolated capabilities to the evidence--assumption--judgment chain, allowing researchers to locate where analytical advantages fail to translate into forecasts and investment judgments, and providing a traceable, diagnostic foundation for training and evaluating financial agents.
This paper introduces Quaternion Approximate Networks (QUAN), a novel deep learning framework that leverages quaternion algebra for rotation equivariant image classification and object detection. Unlike conventional quaternion neural networks attempting to operate entirely in the quaternion domain, QUAN approximates quaternion convolution through Hamilton product decomposition using real-valued operations. This approach preserves geometric properties while enabling efficient implementation with custom CUDA kernels. We introduce Independent Quaternion Batch Normalization (IQBN) for training stability and extend quaternion operations to spatial attention mechanisms. QUAN is evaluated on image classification (CIFAR-10/100, ImageNet), object detection (COCO, DOTA), and robotic perception tasks. In classification tasks, QUAN achieves higher accuracy with fewer parameters and faster convergence compared to existing convolution and quaternion-based models. For objection detection, QUAN demonstrates improved parameter efficiency and rotation handling over standard Convolutional Neural Networks (CNNs) while establishing the SOTA for quaternion CNNs in this downstream task. These results highlight its potential for deployment in resource-constrained robotic systems requiring rotation-aware perception and application in other domains. Code: https://cwru-aism.github.io/ QUANpaper/
Anouar Imel, Imad Charradi, Abderrahman Sghir, Khalil El Khammar
When an LLM group's discussion is compressed into a short memory for a final decision maker, what should the memory keep? Two natural heuristics keep the rarest claims or the claims relevant to the decision criteria. On hidden-profile tasks both look attractive, because how widely a fact is held is confounded with whether it helps. We introduce CrossProfile, a diagnostic benchmark of 60 templates with automatically verified answer keys, in which the ownership of supporting, decoy, and irrelevant facts is crossed while their text is held fixed (480 tasks). Across Qwen2.5-7B, Llama-3.1-8B, and two DeepSeek models, both heuristics fail in predictable places. Selecting by relevance readmits decoys, because decoys are on topic: even a perfect relevance filter loses 23 to 31 points to rarest-first selection in the classic hidden-profile configuration. Rarest-first selection is no better than random selection on average and wins only where the rare facts happen to be the useful ones. In both cases accuracy follows the balance of evidence the memory keeps, not the policy that kept it. The decision model adds a third failure, a form of social proof: it gives extra weight to facts attributed to many agents. Removing the attribution tags, or reassigning them at random with line order fixed, raises accuracy in the classic configuration by 13 to 24 points. In natural discussions the evidence is lost before any memory acts: the complete transcript yields lower accuracy than no notes at all.
Adrien Petralia, Alexandre Gavaudan, Artjom Joosen, Etienne Goffinet, Richard Kurle, Yassine El Ouahidi, Yassir Bendou, Abdelaziz Bounhar 等 356 位作者
We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We achieve this with three ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts; a covariate framework that natively supports known-future covariates within a causal architecture; and a two-stage pipeline that pretrains exclusively on a diverse, large-scale synthetic corpus before post-training on a mix of real and synthetic data with extended context. This report describes the model design and its benchmark results. Weights for all four models are available on Hugging Face and inference code on GitHub.
Henrique Schechter Vera, Sahil Dua, Jovan Andonov, Sara Smoot, Sebastian Russo, Ivan Llanos, Max Matveev, Jae Hyeon Yoo 等 73 位作者
We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements such as simplified per-layer embeddings and a multi-stage curriculum with modality and objective progression which ensure strong quality across unimodal, cross-modal, and mixed-modal evaluations. Notably, in comprehensive embedding benchmarks across all modalities, EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger. Furthermore, our design offers flexibility for resource-constrained environments, supporting selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights. We present ablations analyzing key decisions and release EmbeddingGemma 2 to promote further research.
Jiafu Tang, Yifan Xu, Eric Tang, Junjie Zhao, Shipeng Luo, Te Sun, Haoxuan Hu, Yiyan Ji 等 14 位作者
Video captioning aims to preserve visual information through compact descriptions, predominantly using natural language. However, precisely expressing fine-grained spatial structures and temporal dynamics in natural language often leads to lengthy and potentially ambiguous descriptions. In this work, we explore Code as Video Caption, a paradigm where executable programs serve as a complementary representation of visual content through graphical primitives, layout rules, and functions of time. To examine whether code captions effectively preserve video information, we compare them with natural language captions on downstream question answering and video reconstruction tasks. Our results show that code captions achieve superior reasoning accuracy and reconstruction fidelity. Despite these advantages, realizing the potential of code captioning depends on whether current multimodal large language models (MLLMs) can generate faithful code captions. To systematically assess this capability, we introduce CodeCap-Bench, which comprises 323 videos spanning five categories of structured 2D content, along with fine-grained annotations and a multi-level evaluation framework. Extensive experiments reveal that while leading proprietary MLLMs demonstrate promising code captioning capabilities, they still struggle to faithfully capture complex temporal dynamics, with open-source models lagging substantially behind.
Yonghao Yu, Lang Huang, Runyi Li, Toshihiko Yamasaki
The vision-language-action (VLA) and world-action model (WAM) communities debate whether control policies should inherit pretrained knowledge from vision-language models or from visual generators. The debate, however, bundles three separable choices: backbone initialization, the objective applied to future observations, and whether the machinery for that future loss is kept at inference. A gap between a VLA and a WAM therefore cannot simply be attributed to any one of them. We show that decoupling these choices yields a common framework for both families, which we term Just Action Model (JAM): a pretrained visual backbone with an action expert, trained with an action loss and a foresight loss. The foresight loss trains current-observation features to predict frozen encodings of the recorded future through a prediction head discarded after training, so no visual generation is involved in either training or inference, and the backbone can be chosen freely. We first compare objectives within fixed policy architectures, then evaluate the foresight recipe across backbone families. The resulting JAM policies outperform their matched FastWAM and ImageWAM counterparts on LIBERO and LIBERO-Plus, and achieve similar or slightly higher success rates on RoboTwin. Compared with FastWAM, JAM improves the success rates from 97.0%/55.3%/91.2% to 98.2%/64.5%/91.4%, respectively. With cached foresight targets, JAM reduces estimated online training FLOPs per sample per update by 73% relative to FastWAM. Our results show that future supervision improves control without visual generation or additional inference machinery, and strong foresight performance is attainable with either generator-initialized or vision-language backbones. Project Page: https://y0uroy.github.io/JAM
Donghyeon Cho, Joowon Kim, Seungho Shin, Eunho Yang
Diffusion vision-language models (DVLMs) process their entire input at every denoising step. Since visual tokens make up 89-97% of this input on the benchmarks we study, most of the decoding computation is spent repeatedly processing the same image. Visual token pruning methods partially relieve this burden by removing part of the visual tokens, but they lose the removed information and still process the remaining image at every step. To measure how often the image is actually needed, we compare LLaDA-V's predictions with and without the image at every decoding step and find that removing the image changes a token being unmasked at only 22.8% of steps on average and at as few as 7.9%. We therefore propose LOWN, a decoding method that decides at every step whether to process the image and how many tokens to unmask. LOWN predicts the masked tokens without the image and processes the image only when this prediction disagrees with the last image-conditioned prediction. Agreement certifies the earlier prediction for reuse, and the number of certified tokens sets how many are unmasked in that step. We further observe that the need for the image differs across benchmarks, so LOWN's adaptive decision is needed rather than a fixed schedule. Without training or a tuned threshold, LOWN retains 99.6% of LLaDA-V's score on ten benchmarks at 35.3% of the decoding FLOPs on average. LOWN also generalizes across LLaDA-based DVLMs, composes orthogonally with visual token pruning, and runs under key-value caching for further efficiency gains. Code and project page are available at https://hyeon9698.github.io/LOWN/
Multimodal intelligence can be viewed as human-guided co-evolution between worlds and agents through interaction. However, scaling this process faces two bottlenecks: constructing heterogeneous, verifiable environments is costly, and manual Harness iteration cannot keep pace with growing task diversity and complexity. A key insight is that tasks with different modalities and surface forms can share the same underlying capability requirements. We therefore introduce The Trinity, a recursive self-improvement framework coupling world generation with Harness evolution through three roles: Creator, Explorer, and Evolver. Drawing on verified execution trajectories, Evolver shifts the basis of task organization from surface categories to shared requirements, reorganizing the original task tree into a hierarchical capability tree. We introduce Capability-Driven Real2Sim2Real: for heterogeneous tasks at the same node, Creator abstracts their common requirements into simplified, verifiable simulated worlds for capability learning, while evaluation on original tasks measures transfer. Along this capability tree, Explorer progressively evolves reusable capability modules through selective inheritance, forming a Harness tree whose validated updates accumulate across iterations within a fixed topology. The resulting Harnesses specialize in different capabilities, motivating a Capability-Aware Harness Router that uses public task information to select compatible module configurations from the tree. Our evolved Harnesses improve mean performance by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks.
Yao Huang, Zhiyuan Zeng, Yu Wang, Zhitong Zhu, Dingcheng Duan, Hao Li, Shangzhi Xu, Yifan Wang 等 17 位作者
As frontier language models grow increasingly capable and autonomous, independent third-party evaluations are critical for characterizing their systemic safety risks. While closed-source releases are routinely accompanied by comprehensive system cards, open-weight models receive less scrutiny regarding frontier risks. To address this gap, we present an independent safety evaluation of DeepSeek V4.1 Flash-1017, the latest version of V4.1 Flash that departs from V4 with a new architecture, scaled-up training, and strong agentic capabilities. Drawing on the EU General-Purpose AI Code of Practice, we evaluate the model across 34 benchmarks, including 5 in-house suites, spanning evaluation integrity, loss of control, and harmful manipulation. Conforming to the AEF-1 standard, we compare V4.1 Flash-1017 against V4.1 Flash, earlier DeepSeek V4 models, and leading closed-source systems (GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash). Overall, V4.1 Flash-1017 shows substantial safety improvements over the V4 generation and approaches or matches leading closed-source baselines on several dimensions, while still exhibiting notable behaviors that warrant continued attention. Building on our findings, we highlight three key observations and recommendations. First, because safety can shift and even regress across minor updates, developers should integrate safety evaluation into incremental releases. Second, as final-turn outputs and end-state checks alone cannot reliably detect loss-of-control behaviors, deployments require process-level observability across both reasoning traces and action logs. Third, because public benchmarks elicit strong evaluation awareness and face data contamination risks, developing private benchmarks is essential for trustworthy frontier evaluation.
Jeonghoon Kim, Woojin Chung, Woomin Song, Cheonbok Park, Nako Sung, Jinwoo Shin
How much of zero-shot benchmark accuracy is captured by word-level alignment between a pre-training corpus and the benchmark target? We study this question through word-level unigram cross-entropy (UCE), a tokenizer-agnostic, model-free measure of lexical alignment between a reference corpus and a target benchmark. Likelihood training shapes the model's predictive distribution from the pre-training corpus, and zero-shot scores under likelihood-based evaluation depend in part on the lexical alignment between that corpus and the benchmark target. Cleanly tracing how the lexical imprint of pre-training data scales with corpus--benchmark alignment requires controlled from-scratch pre-training that varies corpus identity while holding the training recipe fixed. Across this controlled sweep, spanning $10$ zero-shot benchmarks, $4$ pre-training corpora, and $3$ model scales, we find that corpus--benchmark lexical alignment, UCE, consistently tracks zero-shot accuracy on every benchmark, with corpora more aligned with the benchmark (lower UCE) achieving higher accuracy. The same signal extends from measurement to control at finer scopes. At the document level, strengthening the imprint by selecting training documents to maximize lexical alignment with the benchmark produces the accuracy gains predicted by the mechanism. At the prompt level, matching the imprint per prompt by routing each prompt to the drafter whose domain corpus is most lexically aligned with it produces measurable throughput gains over a generalist-drafter baseline. Taken together, pre-training corpora leave a measurable lexical imprint on base-model zero-shot behavior, and UCE provides a model-free probe of lexical alignment between reference and target data.
Outcome RL raises pass@1 but not pass@k at large k (Yue et al., 2025). But every pass@1 is scored under a token budget. Every recipe also trains under one. On physics and math problems at fixed token budgets, we compare the pass@1 of 20 released RL models (1.5B to 32B) with that of the thinking models they started from. We ask whether each pass@1 gain is changed behaviour or new knowledge. We divide each change in pass@1 into three parts. Finishing measures answering before the budget runs out. Reliability measures being right once finished. Reach measures problems RL solves that its base answered wrongly in all eight tries, minus problems the base solves that RL answered wrongly in all eight. We find four things. (1) Where the base often hits the token budget before answering, most of the change, gain or loss, is finishing (82% at 4k and 8k). Given more tokens, the base closes most of the finishing gap itself. (2) Where the base has enough tokens, the gain is reliability: the RL model is graded right more often, even on problems both finish. (3) By our definition, RL rarely solves what its base cannot. Reach adds at most 2.1 percentage points on any of our four physics and math sets. We checked 144 physics problems that only the RL model solved. Of these, 3.5% hold up when resampled or read. In a random sample, nine of ten new RL steps a human confirmed also appear in the base's samples. (4) In a small-scale RL run on an 8B model, the gain on new problems is finishing, with reliability flat. Supplying each problem's physical setup was not shown to improve on plain RL.
Long-horizon agents continuously encounter reusable procedural knowledge, yet existing systems typically rely either on model adaptation or external memory, trading persistent learning for rapid reuse. This limitation is especially important for Web agents that repeatedly execute related multi-step workflows. We propose HaMo, a Harness-Model collaborative framework that combines fast non-parametric experience accumulation with slow parametric adaptation. HaMo distills trajectories into strategy-level experience cards for immediate retrieval, while an experience-conditioned continual learning objective trains the model to interpret, ground, and execute retrieved strategies across changing contexts. A retrieval-coverage-based selection mechanism further identifies compact and complementary trajectories for parameter updates. Experiments show that HaMo outperforms harness-only and model-only adaptation under within-domain and cross-domain shifts, improving both retention and adaptation. These results highlight the importance of jointly preserving procedural experience and internalizing the competence to transfer and execute it across tasks.
Language agents increasingly act through external tools and interactive environments, making the interface between the model and its executable capabilities part of the decision process. Model Context Protocol (MCP) and command-line interfaces (CLI) can expose the same semantic operations through markedly different model-facing representations. We study whether these representational differences change agent behavior when executable capabilities are matched. We introduce Matched Interface Comparison and Analysis (MICA), which renders MCP and CLI from a shared semantic action contract and routes decoded actions through the same runtime. Across six models and three interactive environments, matched interfaces produce substantial differences in task performance and interaction cost. These effects extend beyond surface validity to semantic action selection, failure pathways, and recovery. Matched-state probes localize sensitivity to capability presentation, realization, and feedback. Paired natural trajectories reveal structured divergence in 94.3% of 2,004 episodes and show that early differences propagate through subsequent interaction. This analysis identifies interface realization as a direct intervention boundary. We therefore introduce Decoupled Action Realization and Targeted Repair (DART), which preserves validated semantic commitments while repairing interface-specific realization. DART improves benchmark score in all six evaluated settings by up to 9.0 points and reduces MCP token cost in all three environments by up to 27.1k tokens per task. Together, these results establish interface representation as part of the closed-loop decision environment shaping agent behavior.
Yihao Hu, Yitong Lin, Pan Wang, Zhihao Wen, Hang Wang, Yiwen Zhao, Deyuan Liu, Yutong Li 等 11 位作者
Training a spatial VLM to answer which side an object lies on does not require it to preserve the metric displacement that determines that answer, while supervision for a distance estimate may leave categorical relations largely unconstrained. When post-training emphasizes one readout, improvements can therefore concentrate on that interface while transfer to others remains weak or even degrades; we call this observed specialization interface lock-in. ChronoReplay addresses it by replaying heterogeneous spatial supervision as chronologically ordered episodes, retaining source-native targets, and balancing supervision across families. Under the same observation access, a single ChronoReplay checkpoint improves OVO-S-Bench, OST-Bench, and UCS-Bench together, including a 13.41-point gain on OST-Bench, and achieves the highest three-benchmark mean among the supervision settings compared here. Same-data controls further separate the ingredients of this transfer: answer serialization redistributes gains across OST and UCS, family-balanced learning improves over ordinary joint SFT and natural turn weighting, and chronological replay exceeds a shifted-order control by about 2.41 OST points across two paired seeds. Adaptive frame selection then changes evidence access while keeping the trained model fixed, providing an additional inference-time gain at lower visual cost. Streaming spatial performance depends on what distinctions supervision preserves, how heterogeneous histories are organized during learning, and what causal evidence inference makes available at test time.
Xuan He, Sining Zhoubian, Ruozhen Yang, Yuhao Wen, Zhineng Wang, Yuyang Hu, Ke Yang, Yuyu Zhang 等 11 位作者
Memory is increasingly used in long-horizon agents, yet the net value of building memory over direct retrieval from raw history remains largely unclear. Through a large-scale evaluation across multiple benchmarks and model backbones, we find many dedicated memory systems underperform naive RAG in many cases, indicating a negative net value. Taking RAG as a reference, we dissect the net value of memory into the gains and losses over direct retrieval, termed memory benefit and memory tax. These gains and losses arise because memory process changes how the raw information is organized and accessed. We view such changes as the double-edged deviations from the raw history. They are originally designed and introduced to benefit the task, for example by reorganizing facts or distilling preferences, but some may also remove or obscure evidence, giving rise to memory tax. Since memory benefit reflects the expected gains of its design, we then focus more on the overlooked memory tax, study how it arises from the system and evaluation side, and generalize the concept to more application scenarios of memory. We hope the notion of memory tax can provide a new lens and shift more attention from standalone performance to the net value of memory, informing more reliable memory designs and evaluations. We will open-source the code and results to the community. Project Page: https://hexuan21.github.io/MemoryTax/
Yunfei Xie, Howard H. Qian, Shiyi Lan, Kaiyu Hang, Chen Wei
Robot foundation models are typically evaluated through costly closed-loop rollouts, motivating low-cost offline proxies for model capability. The Platonic Representation Hypothesis (PRH) suggests that as representations improve, modalities such as vision and language converge toward a shared statistical model of reality. We ask whether motion trajectories form another such modality: as embodied models become more capable, should their sensory representations converge toward the structure of physical motion? To test this, we propose Sensorimotor Alignment ($S^2$), a training-free offline metric that measures how well a model's sensory representation aligns with the geometry of physical motion. We find that $S^2$ correlates positively with downstream robot task success across model families, making it a low-cost offline proxy for prioritizing models before closed-loop evaluation. $S^2$ can also identify promising vision-language model (VLM) backbones before costly fine-tuning into vision-language-action models (VLAs). Beyond this correlation, $S^2$ offers representation-level explanations for phenomena observed in VLAs and world action models (WAMs). With $S^2$, we extend PRH to robotics and make sensorimotor alignment measurable, offering a low-cost, representation-level view of robot foundation models that closed-loop success rates alone cannot reveal.
Md Khairul Islam, Zeyu Xia, Adam S. Bolton, Judy Fox
Standard sequence models encode position by token rank rather than physical coordinate, typically forcing practitioners to destructively resample continuous physical signals such as spectra observed across heterogeneous instruments. We identify this coordinate mismatch as a structural limitation in cross-instrument representation learning, and propose OmniSpectra, a wavelength-native foundation model for spectral representation learning: a continuous-coordinate masked autoencoder that discards the integer pixel indexing of sequence modeling. Mapping physical coordinates to a logarithmic basis turns the multiplicative scaling induced by cosmological redshift into a uniform additive shift. In continuous rotary position embeddings, this shared shift cancels from relative token distances rather than producing grid-dependent positional distortion. We further counter the spectral bias inherent to implicit neural decoders with a physics-grounded, resolution-gated Fourier decoder that provides anti-aliasing. Pretrained on eight surveys and benchmarked against seven baselines, our model reaches state-of-the-art performance with only 12.2M parameters while outperforming substantially larger models, indicating that encoding physical structure can substitute for parameter scale.
Large language models have shown outstanding performance on text-first tasks such as coding, math, and writing. Recent efforts are shifting models' training to include scientific problems that often live in highly specialized domains beyond text, such as protein structure and biological images. However, in these scenarios, their general-purpose accuracy compared with specialized algorithms remains in question. One such domain is cell tracking, which requires spatiotemporal awareness of cell dynamics. To evaluate this, we created a new benchmark, CellTrack-Bench, which assesses Vision-Language Models' (VLMs) cell-tracking-related performance against specialized trackers. The full benchmark draws from 15 standard cell-tracking datasets across four imaging modalities and comprises 1,537 problems corresponding to 9,041 cell-linking decisions. We selected a hard subset requiring greater reasoning effort by identifying failures shared across distinct algorithmic trackers. With this proposed dataset, we: (i) evaluated different environment harnesses for VLMs; (ii) benchmarked 9 frontier closed-source models and 13 open-weight models; and (iii) provided a detailed analysis of their failure modes. Code, the stored model outputs, and a pipeline that regenerates the benchmark from the source datasets are available at https://github.com/royerlab/celltrack-bench.
We introduce OvisOCR3, a 0.8B model for a broad range of OCR-related tasks. OvisOCR3 unifies document parsing, general OCR, key information extraction (KIE), document figure captioning, and cross-page table reconstruction in a single model. We train the model on real and synthetic data through continued pretraining and use domain-balanced mid-training to balance supervision across tasks with different data volumes and response lengths. Multi-teacher on-policy distillation transfers expertise from five domain-specific 4B teachers into the shared model, followed by mixed-task reinforcement learning with domain-specific rewards. We compare OvisOCR3 with general-purpose vision-language models and specialized OCR models on public and in-house benchmarks covering all five task families. In these comparisons, it achieves an overall score of 96.74 on OmniDocBench v1.6 and the highest overall scores on PureDocBench, Real5-OmniDocBench, and MDPBench. On OCRBench v2, it ranks first, scoring 65.9 and 61.3 on the English and Chinese splits, respectively. It also outperforms the baselines on OCRBench and all four KIE benchmarks. On DocFigCapBench, it achieves the highest localization F1 and a caption score close to that of Qwen3.8-Max. For cross-page tables, it leads in Relation F1 on CrossTabBench and truncated-table TEDS on MPDocBench-Parse. These results demonstrate that a compact 0.8B model can achieve strong performance across diverse OCR-related tasks. OvisOCR3 is available at https://huggingface.co/ATH-MaaS/OvisOCR3.
Synthesizing a monocular video from a new camera trajectory requires preserving the observed scene geometry and motion while completing regions that become newly visible. Feed-forward 4D reconstructors recover the structure of the observed scene, whereas video diffusion models provide a generative prior for completing unseen content. We introduce Holo4D, a video reshooting framework that uses reconstructed 3D trajectories as a holistic source-view 4D scaffold. We project these trajectories into the requested views and encode their geometry and source appearance to guide a pretrained video diffusion model. Together with the source video and target cameras, the scaffold grounds generation in reconstructed observations while leaving unsupported regions to the generative prior. Experiments demonstrate the value of trajectory-based conditioning for camera control and novel-view synthesis. The framework accommodates different 4D reconstructors without retraining the video generator, allowing it to benefit from future advances in source reconstruction. More details can be found in our project page.