Hidir Yesiltepe, Julius Berner, Arash Vahdat, Pinar Yanardag
AI 速读 基于摘要
PMD 通过匹配教师后验均值来蒸馏扩散模型,在保持多样性的同时实现少步生成。
cs.CV
查看原文摘要
Diffusion distillation accelerates image and video generation by compressing a pretrained teacher's iterative sampling process into a few network evaluations. Distribution-based methods achieve this by matching the student's output distribution to the teacher's. This matching is often driven by reverse-KL objectives, whose mode-seeking bias favors dominant teacher modes at the expense of diversity across samples. We introduce Posterior Matching Distillation (PMD), a new member of the distribution-based distillation family that combines mode-covering posterior sampling with any-step generation. At zero population loss, PMD recovers the teacher's full posterior of clean images at each trained noise level. Its stochastic student supports any-step sampling budgets within the trained timestep grid without further training. To learn these posteriors, PMD matches exponentially reweighted averages of student samples conditioned on a fixed noisy state to corresponding teacher posterior means. Under Gaussian corruption, exponentially tilted posterior mean is a single teacher evaluation at a shifted input. We prove that, matching tilted means over an open neighborhood uniquely identifies the full posterior. PMD thus supervises the entire conditional distribution through a family of mean constraints, without any auxiliary student-score network. Experiments on SDXL and FLUX.2 [klein] 4B demonstrate that PMD achieves the best FID among 4-step methods on both models, while maintaining competitive or leading performance across prompt-alignment and perceptual-quality metrics.
Reinforcement learning (RL) post-training substantially improves the visual quality, text rendering, and compositional accuracy of diffusion transformers. Because capabilities such as rendering requested text or composing multiple objects are inherently prompt-dependent, a natural hypothesis is that improving them requires learning complex, input-dependent transformations. To isolate how much of these gains an input-independent shift explains, we probe each RL-tuned model with a deliberately simple parameterization: a fixed set of steering vectors added to the block outputs of the frozen base model. On SD3.5-M, these vectors hold 65x fewer parameters than the RL-tuned LoRA (0.29M against 18.8M). We fit them to match the velocity predictions of public RL-tuned checkpoints, with no reward in the objective. Although the vectors are identical for every prompt, token, and sampling seed, they recover 90.0-96.3% of the reward gains of six RL-tuned image models, and 89.8% and 111.9% on two RL-tuned video models. Initialized at zero and trained with online RL, the vectors alone exceed the reward gains of weight-based RL on text rendering (OCR) and human preference (PickScore) on SD3.5-M. A constant shift thus suffices for most of the gain of policy-gradient RL post-training: without new input-dependent transformations, it steers the pretrained generator toward what it can already compute.
Recent AI agents can generate playable games, but these games are not necessarily engaging. Mastery is a key source of engagement, yet designing for it requires extensive human play and tuning. We propose PlayCraft, an automated framework that uses a vision-based agent to play games and draws on its detailed play experience to revise them for mastery. Starting from an initial game, PlayCraft alternates between playing and crafting. In the playing stage, a player agent watches the game screen, uses the controls, and pursues the game objective over repeated episodes while retaining memory. After each episode, it reports its progress, outcomes, decisions, and reflections. In the crafting stage, a crafter agent uses these reports to revise the game. It estimates player abilities, predicts performance at different proficiency levels, and compares those predictions with a target learning progression. It diagnoses the game and addresses the gaps by giving players meaningful opportunities to improve through better decisions and execution. We evaluate games in fresh runs with player agents across three intelligence settings, measuring learning gain and learning slope. Compared with the initial games, PlayCraft yields larger learning gains and steeper learning slopes in most games. It also outperforms the baselines on the macro-average of each metric at every intelligence setting.
We characterise path-independent auto-deleveraging (ADL) rules as an extended log-exponential family. Immunity to account splitting, merging and reopening further restrict the class. Homogeneity (essentially) restricts the class to a one-parameter family ordered by both fairness and platform risk. Its split-proof members continuously interpolate between proportional division and water-filling. Among these rules, decreasing the parameter reduces risk at the cost of increasing inequality in how accounts are treated. This suggests selecting the parameter through an ex-ante calibration to a risk bound on historical data to choose a maximally fair rule subject to that bound. A counterfactual replay of Hyperliquid's ADL episode on 10 October 2025 finds that a range of intermediate parameters significantly improves on the observed allocation in both burden dispersion and subsequent risk statistics.
Reinforcement Learning (RL) has enabled Large Language Models (LLMs) to acquire increasingly complex reasoning and agentic behaviors. In this work, we propose two simple techniques to improve policy gradient algorithms for LLMs. First, we replace the fixed anchor policy during RL with an Exponential Moving Average (EMA), similar to a target network in deep Q-learning. Second, we introduce Top-$k$ KL estimator, which allows for flexible interpolation between exact KL and sampled KL. We derive the stability conditions for using EMA anchor; moreover, we show that our Top-$k$ KL estimator yields both unbiased KL values and unbiased gradients at any $k$, while bringing the benefits of exact KL. When combined with GRPO, the two techniques (EMA-PG) lead to a significant performance boost. On math reasoning, it allows R1-distilled Qwen-1.5B to reach 53.9% on OlympiadBench compared to 50.8% by GRPO. On agentic RL domains, with Qwen-3B base, EMA-PG improves GRPO by an average of 35.6% across 7 datasets of Q&A with search engines, including 29.7% $\rightarrow$ 45.8% on HotpotQA, 27.4% $\rightarrow$ 40.5% on 2WikiMultiHopQA. Overall, we show that EMA-PG is a simple, principled, and powerful approach to scaling RL for LLMs. Code: https://github.com/LunjunZhang/ema-pg
Hybrid architectures have become mainstream among open-weight language models. Their global-attention and bounded-state layers maintain heterogeneous caches whose demands vary with context length and concurrency. We present TideCache, a serving system on CPU--GPU superchips with three components for allocating GPU memory among caches and managing their residency across GPU and host memory. The Transparent Cache Pool redistributes physical pages among cache components behind fixed virtual addresses, preserving cache layouts and backend interfaces. The Cache Streamer uses computation intervals provided by bounded-state layers to prefetch the next global-attention layer's prefix into GPU slots shared across global-attention layers, reducing GPU residency while overlapping layer-wise transfers with computation. Generated KV remains resident. The Request Scheduler selects resident or streamed execution and triggers capacity adjustments according to request size and queue state. Together, they let HBM released by streaming accommodate other cache components and additional requests. Implemented in SGLang and evaluated on GB300 and GH200 across seven model/platform configurations, TideCache completes up to 2.04x as many agent turns, reduces overall P50 per-token latency by 20--59%, and improves agent P50 TTFT by 1.40--3.05x compared with SGLang with HiCache.
O²LaF 从同一个预训练视频 DiT 初始化逆动力学和正动力学模型,通过控制信息访问和端到端瓶颈优化,从无标签视频中学习可迁移的潜在动作。
cs.CV · cs.AI · cs.LG · cs.RO
查看原文摘要
Videos reveal how the world changes, but rarely the actions that cause those changes. Latent action learning recovers such supervision from visual transitions, yet often learns the underlying dynamics largely from scratch. We ask whether a pretrained video generator can instead provide the prior for both inverse and forward dynamics. We introduce O$^{2}$LaF, built on a simple principle: one prior, two dynamics. We initialize both models from the same pretrained video DiT and adapt them to their distinct roles: a feed-forward transition encoder with learnable action queries for inverse dynamics, and a causal, action-conditioned video generator for forward dynamics. By controlling information access and optimizing both dynamics end-to-end through a compact bottleneck, O$^{2}$LaF learns transferable actions from unlabeled videos. Across navigation and robotic manipulation, O$^{2}$LaF sets a new state of the art in few-shot action probing and enables stronger zero-shot action transfer across visual contexts and embodiments. Beyond generative video priors, we further extend our framework to predictive representation priors such as V-JEPA2, achieving competitive performance. Our findings suggest a broader paradigm for latent-action learning: repurposing pretrained video models to discover transferable actions and predict their consequences.
High-resolution generation enables video world models to render fine-grained scene structures and textures, but directly generating high-resolution video incurs substantial computational and memory costs. We present ScaleWorld, a causal video model that unifies low-resolution world prediction and high-resolution refinement within a single shared backbone. By refining its own predictions, ScaleWorld produces high-resolution world videos without requiring a separate refinement model. For each video chunk, the shared backbone first completes a low-resolution prediction and then performs one-step high-resolution refinement conditioned on that prediction. To support efficient streaming generation, we maintain separate causal states with a long low-resolution history for prediction and a short high-resolution history for refinement. To address task interference during distillation, we separate specialist optimization from capability consolidation and introduce task-asymmetric on-policy distillation (TA-OPD). This procedure transfers prediction and refinement behaviors from specialized teachers into a shared student through local velocity matching under their respective sampling and history policies. {Extensive experiments demonstrate that ScaleWorld achieves high-quality world generation with improved computational and memory efficiency.
Frontier multimodal language agents such as GPT-6 Astra have demonstrated strong capabilities in controlling robots directly via low-level numerical actions without high-level primitives. However, existing benchmarks report only episode-level success rates, which show neither how these agents succeed nor where they fail. In this paper, we present a systematic action-level analysis of frontier agents on direct robot control across 12 manipulation benchmarks. We introduce AgentRoboCritic, an intent-grounded critique framework in which the policy agent states a local intent alongside each action, and a critic evaluates the action's outcome and failure mode against this intent. Our action-level analysis demonstrates the promising capabilities of frontier agents on direct robot control, including closed-loop recovery from local execution failures, reliable semantic grounding across diverse scenes, and strong zero-shot manipulation without task-specific adaptation. It also identifies their current limitations, including task-relevant physical execution bottlenecks such as grasping and alignment, and uneven recovery across failure modes. Finally, we validate our critic on a 400-action human-annotated benchmark and release a dataset of 12,844 annotated actions alongside a fine-tuned open-source critic, Muse-Glimmer-30B, whose failure-mode accuracy rises from 32.58% to 74.24%, enabling low-cost action-level analysis.
Algorithms frequently assist, rather than replace, human decision-makers. However, the design and analysis of algorithms often focus on predicting outcomes and do not explicitly model their effect on human decisions. This discrepancy between the design and role of algorithmic assistants becomes particularly concerning in light of empirical evidence that suggests that algorithmic assistants again and again fail to improve human decisions. In this article, we formalize the design of recommendation algorithms that assist human decision-makers without making restrictive ex-ante assumptions about how recommendations affect decisions. We formulate an algorithmic-design problem that leverages the potential-outcomes framework from causal inference to model the effect of recommendations on a human decision-maker's binary treatment choice. Within this model, we make a connection between algorithmic recommendations and instrumental variables. Leveraging this connection, we introduce a monotonicity assumption that leads to an intuitive classification of human responses to the algorithm. Under this assumption, we can express the human's response to algorithmic recommendations in terms of their compliance with the algorithm and the active decision they would take if the algorithm sends no recommendation. We showcase the utility of our framework using an online experiment that simulates a hiring task. We argue that our approach can make sense of the relative performance of different recommendation algorithms in the experiment and can help design solutions that realize human-AI complementarity. Finally, we leverage our approach to derive maximin-optimal recommendation algorithms that can be implemented with machine learning using limited training data.
Dataspaces are an emerging concept to enable secure and controlled data sharing between multiple stakeholders. However, most existing dataspace architectures, such as those based on the concepts of IDSA (International Data Spaces Association) and Gaia-X, are heavily concept-driven and require significant effort to implement, making it difficult to quickly build a functional dataspace. Solid (Social Linked Data) offers a compelling alternative by providing a lighter and more practical approach. However, for Solid to function as a true dataspace, like traditional dataspaces, it must be built in a decentralized manner, i.e., the dataspace must consist of more than a single Solid server. However, this decentralization introduces a key challenge: Data Discovery. Unlike traditional dataspace architectures, which often include predefined components such as data catalogs to facilitate data discovery and access, Solid does not inherently provide such mechanisms. As a result, managing and locating relevant data across multiple Solid servers and pods becomes a significant barrier to usability. To address this gap, we present a FAIR-compliant data catalog for Solid. Our catalog follows the FAIR principles (Findability, Accessibility, Interoperability, and Reusability) by integrating semantic models that describe all available data sources within the Solid ecosystem. This approach enables enhanced data discovery, improves accessibility through standardized metadata, enhances interoperability between different data sources, and promotes data reuse across applications. By bridging this critical gap, our catalog improves the usability of Solid-based dataspaces, making them a viable option for decentralized and open data ecosystems, such as smart city applications.
Graph-based data structures such as DiskANN, HNSW, and NSG constitute the state of the art for large-scale approximate nearest-neighbor (ANN) search. Indyk and Xu introduced the notion of $α$-reachable graphs, which include graphs built by DiskANN's slow-preprocessing construction, and proved that greedy search achieves an approximation factor of $\tfrac{α+1}{α-1}$. Independently, Al-Jazzazi et al. proposed adaptive beam search (ABS) with a relaxation parameter $γ> 0$, showing a $\tfrac{2}γ$ approximation on navigable graphs ($α\approx 1$), where greedy search admits no guarantees. In this work, we unify and generalize these results by proving that ABS with parameter $γ$ on $α$-reachable graphs achieves an approximation of $\tfrac{α+1}{α(γ+1)-1}$ for any metric. Moreover, on sorted $α$-reachable graphs (introduced by Gollapudi et al.) under the $\ell_2$ metric, we obtain the tighter bound $\max\bigl(\tfracα{α(γ+1)-1}, \tfrac{α^{2}+1}{α^{2}(γ+1)-1}\bigr)$. Beyond these theoretical results, we design a beam predictor using features inspired by fine-grained analysis of ABS, trained via a constrained objective that minimizes mean distance computations subject to a target recall. On three 100-million-point benchmarks, using graphs built by DiskANN's practical construction, beam predictor lowers mean query latency on SSD-resident indices at matched recall by up to $29.6\%$ relative to classical beam search (CBS). Relative to ABS, it is faster in $11$ of $15$ settings, by up to $6.5\%$, and at most $1.4\%$ slower in the other $4$; at $95\%$ recall, its advantage over ABS grows from $0.8$--$4.2\%$ for $100$ neighbors to $4.6$--$5.6\%$ for $500$ neighbors. In memory, it needs $1.3$--$23.7\%$ fewer distance computations than CBS and $0.0$--$3.9\%$ fewer than ABS.
We study policy representation learning from unlabeled multi-policy behavioral data. Each episode is generated by a fixed policy, but policy labels are unavailable. This setting appears in robotics play, demonstrations, games, racing, and other datasets where heterogeneous behaviors are mixed without annotations. We introduce Behavioral INR, a self-supervised generative model that adapts implicit neural representations (INRs) from vision to behavior. Instead of mapping coordinates to RGB values, Behavioral INR represents a policy as a state-action function mapping states to subsequent actions. An episode-level latent modulates this function through FiLM layers, yielding a generative prior over policies and allowing policy identity to be inferred without supervision. Because INRs treat each datapoint as samples from an underlying function, the same model naturally accommodates variable episode lengths and different sampling granularities, as in vision INRs with different image resolutions. We also define policy-level out-of-distribution (OOD) shifts along state-distribution and action-distribution axes, which arise when policies overlap in states or actions but are not captured by standard behavioral OOD settings based only on new agents or environments. We evaluate on synthetic Gaussian random field data, MuJoCo demonstrations with controlled OOD splits, and real-world chess, Formula 1 racing, robotics, and Seek-Avoid datasets. Behavioral INR most consistently improves policy identifiability in the hardest continuous state-action settings, especially when longer episodes, more policies, and OOD splits reduce the usefulness of marginal shortcuts; amortized history encoders remain competitive when policy identity can be recovered from symbolic repetition or low-dimensional action statistics.
Collecting and labeling inertial measurement unit (IMU) data for human activity recognition (HAR) is expensive, and models trained on one group of subjects often fail to generalize to unseen individuals. Even with balanced data, classifiers struggle to distinguish dynamically similar activity classes and benefit from targeted augmentation. We explore whether synthetic IMU data generated in the representation space of a pretrained biosignal encoder, Bio-PM \cite{b1}, can serve as a practical augmentation source. Rather than synthesizing raw signals, we leverage Bio-PM's frozen encoder to convert 10-second IMU windows into token embedding sequences and train class-conditional decoders in this space: (1) an autoregressive transformer decoder that predicts tokens sequentially, and (2) a diffusion decoder that learns the token distribution through a denoising objective. We analyze generation realism through distributional metrics and embedding-space visualizations, evaluate downstream utility through leave-one-subject-out HAR classification, and present a PatchDecoder for approximate raw waveform reconstruction. Our diffusion decoder achieves a Train-on-Synthetic-Test-on-Real (TSTR) macro-F1 of 0.68, demonstrating that generated tokens are realistic enough to train a classifier that transfers to real data. Combining synthetic-token augmentation with a 320-dimensional gravity-aware feature representation raises LOSO macro-F1 from 0.729 to 0.809 over a real-only baseline across 51 subjects on the WISDM benchmark, with statistically significant gains concentrated in the most challenging activity classes. These results demonstrate that high-fidelity generation in a pretrained encoder's representation space is achievable and practically useful for cross-subject HAR augmentation.
Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth
Peiyun Hu, Marko Mihajlovic, Lyndon Duong, Lahav Lipson, Wei Dong, Amaël Delaunoy, Stephan R. Richter, Vladlen Koltun
AI 速读 基于摘要
Depth Pro 2 用一个可扩展的单一网络同时处理图像与视频,在原生分辨率下预测兼具度量准确、边界锐利与时序一致的稠密深度。
cs.CV · cs.LG
查看原文摘要
We present Depth Pro 2, an approach to estimating dense metric depth for images and video at full resolution. By analyzing prior methods for monocular depth estimation, we find that they commonly trade off temporal consistency and boundary sharpness: multi-frame methods favor consistency, single-frame methods favor sharpness. To overcome this trade-off, we train a single, scalable network that processes both images and video and predicts depth maps at native resolution. Depth Pro 2 produces depth maps that are metrically accurate, sharp, and temporally consistent all at once, outperforming prior work in all of these dimensions. We release a family of models that define a new Pareto frontier in terms of metric accuracy, sharpness, temporal consistency, and runtime. We release code and weights at https://apple-aiml-research.github.io/ml-depth-pro-2/
Pre-trained encoders fine-tuned on natural language inference (NLI) and other language understanding tasks keep much of their accuracy after the input is perturbed, e.g. by removing all nouns or shuffling the words. Does the same hold for large language models (LLMs), and if so, is it because they have seen the benchmarks during pre-training? In this paper we revisit these perturbation experiments with eight LLMs alongside fine-tuned encoders on GLUE and the three rounds of ANLI. Regardless of model size, every system that answers directly retains 36-72% of its above-chance performance without nouns on MNLI compared to only 14-29% on ANLI. Only a model that reasons before answering departs from this pattern. We also check contamination indicators (including exact $n$-gram counts in the public pre-training data of two of the models) and find no link between exposure and robustness. Our strongest evidence comes from 368 new, human-validated items that none of the models can have seen. All eight LLMs keep more of their performance without nouns on items written in MNLI's style than on items written without its cues. When the perturbation instead removes the function words carrying such cues, this advantage reverses. In light of these results, we argue that robustness to perturbation on familiar NLI benchmarks reflects dataset shortcuts rather than memorisation, and tells us more about the benchmark than about the model.
We design a polynomial-time algorithm for matroid cups with $O(1)$ discrepancy. Our matroid cups algorithm design consists of a sequence of reductions to simpler instances, culminating in $O(n)$ instances of matroid cups where the matroid has rank 1.
LLM-based agents are becoming increasingly capable at discovering, validating and patching vulnerabilities in real software, and vendors and researchers report that this is already changing how vulnerabilities are found and fixed. These claims, however, are scattered across papers, vendor reports and disclosures and are rarely comparable. We introduce a framework that rates these capabilities on an autonomy scale modeled on the levels of driving automation, and qualifies each claim by its correctness, its availability and the strength of its evidence. We use this framework to systematize the published evidence on LLM capability in vulnerability research. We find that agents discovered, validated, and patched vulnerabilities in real-world software without human intervention before the restricted frontier models of 2026, and that open-weight models subsequently attained comparable autonomy for discovery and validation. We then measure the consequences of evolving LLM security capabilities in a decade of public vulnerability records, advisories and exploitation data. In 2026 the number of disclosed vulnerabilities more than doubled, with an increasing share rated critical or high, and severe bugs that had survived a decade or more in widely used code were reported several times more often than before. Yet the time between the stages of a vulnerability's life, from report to fix and from disclosure to attack, has not shifted noticeably so far. We grade the mitigations proposed in the literature with an explicit rubric and draw guidance for each group of stakeholders from the best-supported ones.
Simulating rigid-body dynamics with contact is fundamental to robotics, from motion synthesis and control to mechanism design. Many widely used simulators rely on cone complementarity formulations, whose physical assumptions and numerical solution require substantial background. While excellent surveys, monographs, and course notes exist, they mostly address computational mechanics or computer graphics or document a single simulator. To the best of our knowledge, a self-contained pedagogical treatment for robotics, bringing these formulations and their solvers together with complete derivations and a unified notation, does not yet exist. This tutorial addresses this gap by connecting rigid-body dynamics, conic optimization, and frictional contact. In particular, we explain how complementarity conditions encode non-penetration and frictional dissipation, determining separation, sticking, and sliding without prescribed contact modes. We distinguish equivalent transcriptions from convex relaxations, showing how the latter can introduce artifacts such as lift-off and creep. We examine the existence and uniqueness of motion and contact forces, including the solver's role in selecting among multiple solutions. We then compare first-order and Newton-based methods, explain regularization and proximal strategies, and discuss trade-offs between computational cost, robustness, and physical fidelity. Connections to widely used simulators, optimal control, and differentiable simulation help readers interpret simulation results, choose suitable formulations and solvers, and develop their own contact dynamics routines.
Acquiring every feature of an instance is often costly: a measurement can carry a monetary price, a time delay, or a privacy cost. Active Feature Acquisition (AFA) addresses this by learning a policy that decides, per instance, which features to acquire under a budget. Existing AFA methods train this policy by optimizing a self-constructed utility target rather than distilling an explicit teacher. We construct this teacher explicitly. At training time the label and the unobserved feature values are available, and together they determine the teacher in closed form: an oracle that ranks every candidate by its effect on the downstream predictor's error. OracleImitator distills this ranking into a deployable policy by soft-target imitation. Across multimodal medical diagnosis and the AFABench benchmark, OracleImitator matches or outperforms existing AFA methods across the budget range, recovering 90% of the attainable accuracy gain at under a quarter of the acquisition cost. Because the teacher is an explicit algorithmic object rather than an implicit optimization target, we can also measure directly how much of its ranking is recoverable by a policy restricted to what is visible at deployment.
In-context learning (ICL) enables language models to perform new tasks from demonstrations without weight updates. However, every ICL inference requires processing the full set of examples, resulting in inefficient deployments, and how ICL works mechanistically is not fully understood. Prior work compresses ICL into fixed activation vectors extracted from specific layers or positions, but these input-independent interventions fail on complex tasks where the output depends on fine-grained interactions with the input. By analyzing the ICL forward pass, we show that each attention head's output is an affine transformation of its context-masked counterpart, and that the parameters of this transformation are empirically stable across samples for a given task. Building on this, we introduce Task Operator (TO), which replays this transformation as an analytically derived update to the attention output projection. Across lexical, algorithmic, and reasoning tasks, TO achieves the best overall performance among prior methods and substantially narrows the gap between zero-shot inference and ICL. We further show that the extracted knowledge concentrates in a task-specific sparse circuit across layers and positions, and that averaging operators from disjoint demonstration batches enables effective many-shot scaling without expanding the context window. Our code is available at https://github.com/gzxiong/task_operator.
Robotic vision-language-action (VLA) models can execute diverse individual manipulation skills, yet they still struggle with long-horizon loco-manipulation tasks, where temporal structure, skills, and precision are tightly coupled within a monolithic model. We propose Physically Grounded Execution (PGX), a framework that offloads temporal and precision control to two classical training-free components, allowing the VLA to focus on stage-specific skill dynamics. PGX decomposes long-horizon loco-manipulation into three layers: (i) a predicate-based state machine governs task progression, using physical predicates to gate phase transitions, trigger recovery, and coordinate navigation and waiting. Shared predicate semantics connect offline demonstration segmentation, online execution, and stage-wise evaluation; (ii) VLA skills operate within active manipulation phases, with flat or stage-trained models concentrating on task-relevant actions; and (iii) world-coordinate locking (WCL) maintains spatial consistency through base station keeping, object-relative positioning, and world-anchored contact targets. We demonstrate the effectiveness of PGX with extensive experiments conducted on three simulated tasks and one real-world task using a Unitree G1 humanoid robot for long-horizon loco-manipulation, showing improvements in stage progression and task success rates. Project page is available at: https://kl-research.github.io/pgx
Code2Video uses agent-authored 3D proxies to guide video generation, enabling explicit control over structure and dynamics. However, how to construct proxies and prompts for optimal downstream quality remains underexplored, as do the capabilities and limitations of this paradigm. Through controlled experiments and blinded human evaluation, we study proxy granularity, physics simulation, and prompting. The results reveal task-dependent trade-offs: coarse proxies favor natural motion, while medium-grained proxies favor consistency. Physics simulation and proxy-aligned prompting further improve average quality. Guided by these findings, we introduce Coda-DaVinci, a task-adaptive agentic framework for Code2Video generation that selects between proxy-guided and direct generation, adapts modeling granularity to each entity, and coordinates simulation with proxy-grounded prompting. Its multi-agent harness preserves scene assets and guides repairs through visual inspection and independent review. Experiments demonstrate leading instruction following on VTR and scene consistency on R2M, improved physical plausibility over direct generation, and higher visual realism than the compared Code2Video baselines. Long-video evaluations further show gains in event structure and consistency across successive interactions, while revealing limitations in transition stability and adherence to scene settings.
Self-consistency certification treats repeated sampling from a language model as sequential evidence accumulation and stops when an anytime-valid e-process crosses a threshold, guaranteeing finite-sample Type I error control by Ville's inequality. Prior constructions assume sampled rollouts are conditionally independent. We show this assumption fails in practice: reasoning streams are positively correlated through shared reasoning prefixes or across-prompt heterogeneity, changing not merely certificate efficiency but the hypothesis certified. Modeling dependence via a per-unit latent state (the random-effects / intraclass-correlation structure) induces exchangeable, equi-correlated increments with correlation $ρ$. Correlation splits the single null into a realized null about the drawn belief and a population null about average model behavior, which coincide only when $ρ=0$. We show: (i) the standard betting e-process remains anytime-valid for the realized null, not the population null; (ii) read as a population-null test, a suitable mixture has asymptotic crossing probability at least $π=P(μ(Z)>0)$, potentially exceeding nominal $ε$; (iii) the within-unit mean's effective sample size saturates at $1/ρ$, so extra within-unit rollouts cannot restore population validity; and (iv) a prompt-level e-process restores population validity, with optimal allocation at one rollout per prompt under a pure rollout budget. Simulations and Claude Haiku 4.5 rollouts confirm this regime dependence: on GSM8K ($\hatρ=0.435$, $\hatθ=+0.96$) the gap is dormant, whereas on AIME ($\hatρ=0.480$, $\hatθ=-0.15$) the single-unit rule crosses on 24% of prompts at nominal $ε=0.05$, while the prompt-level procedure declines (0.00). This correctness result identifies which reliability claim an anytime-valid certificate actually supports under dependence.
Large Vision Language Models (LVLMs) show promise in medical applications, but their inability to faithfully ground responses in visual evidence raises serious concerns about clinical trustworthiness. While visual attribution methods are widely used to explain LVLM predictions, whether these explanations actually reflect the visual evidence underlying the model's decision is largely unverified, since ground-truth annotations for internal model reasoning are typically unavailable. We address this question for chest X-ray (CXR) reasoning by developing a causal evaluation framework that retains only CXR-VQA samples for which the expert-annotated region is verified, via counterfactual editing, to be causally responsible for the model's prediction. Using this framework across 11 attribution methods, six open-source LVLMs, and two output modes (direct answer and step-by-step reasoning), we find that existing attribution methods often fail to identify the evidence used by LVLMs. To address this failure, we propose MedFocus, a concept-based attribution method that localizes clinically meaningful anatomical regions via unbalanced optimal transport and measures their causal effect on model outputs through targeted interventions. MedFocus produces spatial, concept-level, and token-level attributions and substantially outperforms prior methods, taking a step toward more trustworthy attribution for medical LVLMs. Our data and code are available at https://github.com/gzxiong/medfocus/.
Vision-language models (VLMs) integrate visual and textual information, but their outputs do not reveal internal concept representations. Global workspace theory (GWT) proposes that information is selected from parallel brain processes, integrated in a shared workspace, and made available to specialised processes. We introduce the Multimodal Concept Lens (MC-Lens), a multimodal Jacobian lens, to study these functions in VLM concept processing. Its vision encoder and language decoder lenses map intermediate states to a shared semantic space defined by the decoder's output vocabulary to compare concept accessibility across modalities. MC-Lens supports internal concept interventions to examine the role of accessed concepts in downstream computation. (i) Accessibility and selective access. Multimodal concept information is readable in intermediate states and selectively accessible as tasks require. We extend "readout carriers" from one name token to a token set to better capture and express concept information and support counterfactual experiments. (ii) Concepts in downstream computation. Counterfactual experiments show that accessed concept information participates in downstream computation and reasoning. (iii) Development of cross-modal alignment and selective access. We develop two J-space metrics tracking the development of cross-modal concept alignment more clearly than hidden state cosine in controlled VLM training. The experiment also shows that reliable selective access develops later than concept accessibility. (iv) Shared functions across models. Across VLM families and sizes, we find shared patterns of concept accessibility, query-dependent selectivity and cross-task intervention effects despite differences in internal organisation. These findings offer an experimental basis to study GWT in multimodal artificial systems and how it can guide cross-modal alignment improvements.
Verify the Verifier: Specification-Grounded LLM Cross-Checking for Bugs That Escape UVM Scoreboards
Ana E. Sanchez Villalobos, Michael D. Quinn, Aakash Tyagi, Minh D. Luu, David Kebo Houngninou
AI 速读 基于摘要
提出Verify the Verifier框架,用基于LLM的异常交叉检查器将通过的回归波形轨迹与规范对照,识别能逃过UVM记分板的规范违背型缺陷。
cs.SE · cs.AI · cs.AR
查看原文摘要
A passing UVM regression shows the design agrees with the reference model on what the scoreboard checks, not that it meets the specification. Bugs that pass every test yet violate the specification, termed scoreboard escapes, survive mature testbenches. We present Verify the Verifier, to our knowledge the first framework whose LLM-based Anomaly Cross-Checker examines a passing regression's traces against the specification. Of 564 bugs inserted into four OpenTitan IPs, 68 are candidate scoreboard escapes: non-equivalent, exercised by the tests, yet passing. The framework confirmed 41 as scoreboard escapes with assertions failing on the buggy design and proven on the original.
ProtoDiffusion 提出一种基于共享空间原型和每层候选向量的潜在扩散框架,以 HW+N 个令牌表示 N 层 RGBA 设计,支持图像到设计和文本到设计。
cs.AI
查看原文摘要
Graphic designs are ordered compositions of RGBA layers whose separation enables selection, movement, and reuse. Existing layered models represent each layer with a full-canvas or region-sized spatial latent grid, so representation size grows with the cumulative spatial extent of the layers. We present ProtoDiffusion, a latent-diffusion framework built on ProtoVAE, a multi-layer variational autoencoder that encodes a design into one shared spatial prototype and one candidate vector per layer. Each candidate conditions a shared decoder to reconstruct its corresponding RGBA layer. For $N$ layers, the representation contains $HW+N$ tokens before backbone-specific packing, compared with $NHW$ for full-grid-per-layer representations. ProtoDiffusion jointly denoises the prototype and a fixed set of candidate slots, conditioned on an image or text, with bounding boxes as optional layout guidance. The model leaves unneeded slots empty and determines the number of layers and, in absence of boxes, their layout, requiring neither as input. The same architecture supports both image-to-design and text-to-design. ProtoVAE achieves high-fidelity reconstruction from its compact shared latent representation, reaching 40.43 dB PSNR on layer crops. On the LayerD evaluation split, image-to-design achieves RGB L1 scores of 0.040 with bounding boxes and 0.057 without, with box-guided alpha soft IoU (0.850) comparable to the strongest fine-tuned baseline (0.871). On merged images, text-to-design reaches 66.41 FID versus 86.84 for FLUX.2 [klein], achieves a 0.860 text-accuracy score, and runs $19.1\times$ faster than the same-backbone generation-then-decomposition cascade.
Vision-language-action (VLA) models adapt web-pretrained vision-language models to output robot actions, so that web knowledge carries over to physical control. In under four years the field has grown from RT-2, the first VLA, to open models that fine-tune on a single consumer GPU and proprietary generalists that clean unseen homes. VLA research joins robotics, computer vision and language modelling, so this survey is written for researchers and practitioners from each of these fields. It reviews 89 works from December 2022 to October 2026, mainly on manipulation and navigation, plus humanoid control, world-action models and evaluation. Seven design axes follow a VLA's life cycle: model design (inputs and memory, architecture, action output, intermediate predictions), training (pretraining data, post-training) and deployment (real-time execution). We compare 22 models on these axes, with their training data, model size, inference speed and hardware. Three trends emerge. (i) Training data has broadened beyond teleoperated robot data (about 10,000 hours for $π_0$) to community-collected episodes and egocentric human video (20,000 hours for GR00T N1.7). (ii) Model sizes have split into two groups: most recent models have 3-7 billion parameters; a new group under one billion runs on consumer GPUs. (iii) Generalist VLAs handling both manipulation and navigation are emerging, such as UniVLA and Qwen-VLA. We also compare VLAs with world-action models, which predict future observations as well as actions, and review evidence that some VLAs with near-perfect benchmark scores fail after small scene changes. A practitioner's guide gives a default for each design decision, and a glossary supports readers new to the field.
Activation steering controls LLM behaviour towards target behaviour by intervening in internal representations, yet it often degrades reasoning and retrieval performance. We argue that a primary cause of this trade-off is attention rerouting: steering vectors alter query-key matching, shifting attention away from contextually important tokens toward less informative ones. To address this, we propose Steering via Key-Orthogonal Projections (SKOP), a steering method that constrains harmful attention rerouting without eliminating steering efficacy. SKOP achieves this by preserving attention patterns on a small set of focus tokens the model relies on for reasoning and retrieval, while allowing redistribution among less critical tail tokens. Across multiple steering benchmarks, we show that SKOP achieves the best joint steering-utility trade-off, reducing utility degradation by 5-7x while retaining over 95% of vanilla steering efficacy. Our results further suggest that, in long-context retrieval settings where vanilla steering approaches are ineffective, SKOP can maintain robust performance by avoiding attention rerouting.