Multi-turn conversations with multimodal large language models (MLLMs) reuse input context, but long multimodal inputs increase time to first token (TTFT). Query-aware hard pruning reduces prefill tokens but permanently excludes information that later questions may require. We propose Raincheck, a framework for deferred multimodal prefill that decouples initial token selection from later information availability. An existing pruning method selects tokens for initial prefill; Raincheck preserves and progressively prefills unselected tokens in source order using unused compute capacity during decode. Unfinished prefill continues after answer return during available scheduling gaps. Attention visibility control and affine state mapping enable context updates after generation begins in full-attention and hybrid recurrent models. Prefill-decode fusion reduces repeated weight reads. We evaluate Qwen3-VL-8B and Qwen3.5-9B with 30% initial multimodal-token retention. Across both models and four pruning methods, Raincheck recovers on average 77.7% and 60.4% of the quality lost to matched hard pruning on MultiVerse and ConvBench, respectively. On MultiVerse, the SGLang implementation reduces mean first-turn TTFT by up to 24.1% relative to full prefill, with less than 0.5% increase in mean time per output token (TPOT).
Predicting where to grasp and how to move is fundamentally a geometry problem. Vision-language models treat it as language generation instead. They tokenize spatial coordinates into text and decode them autoregressively. This mechanism ignores spatial structure and truncates continuous precision. Errors compound across coordinate tokens during decoding. In this paper, we show that compact off-the-shelf VLMs with 0.5-2B parameters already encode rich spatial information in their intermediate hidden states. Autoregressive decoding is the dominant bottleneck for trajectory prediction, while affordance prediction depends additionally on feature depth and decoder architecture. We expose the latent spatial information with a lightweight, geometry-compatible prediction head that attaches to a frozen backbone. It reads features from an intermediate hidden layer, conditions a learned query on the task instruction and refines it via cross-attention over the visual tokens to regress affordance boxes and trajectory waypoints directly as continuous values, without modifying the backbone. We evaluate on ShareRobot-Bench and AGD20K. Our 2B variant reduces trajectory error by over 60% and improves ShareRobot-Bench affordance mAP by 85% relative to the autoregressive baseline. It approaches RoboBrain on trajectory prediction (0.251 vs 0.191 DFD) at 3.8x higher throughput. These results demonstrate a favorable accuracy-efficiency trade-off for instruction-conditioned 2D geometric prediction.
Despite recent advancements in ASR, identifying effective architectures, fine-tuning approaches, and data adaptation strategies for low-resource languages remains challenging. We present a systematic comparison across five low-resource African languages and Swahili as a mid-resource contrast, fine-tuning three models from distinct architectures: Phi-4-multimodal-instruct, Whisper-large-v3-turbo, and MMS-1B-all. We analyze data sufficiency, monolingual versus multilingual adaptation, domain fine-tuning, and language-adapter transfer. Phi-4 achieves the best overall accuracy, MMS shows comparable CER, while Whisper exhibits high CER, consistent with known hallucinations and repetition loops. Performance saturates at different data scales, offering practical data-efficiency targets, while typological complexity consistently shapes error rates. Domain fine-tuning benefits Phi-4 and Whisper but degrades MMS, revealing limitations of adapter-based adaptation. Proxy language adapters outperform training MMS adapters from scratch.
Training-time multi-agent system (MAS) design seeks to learn reusable orchestration policies, yet prevailing approaches remain largely one-shot: an Orchestrator generates a system but only observes and learns from its terminal outcome. This paradigm discards execution trajectories that expose system-level failures and contain fine-grained information beyond the terminal outcome, leaving no direct basis for revising the source MAS, evaluating the updated MAS, or learning from the resulting contrast. To this end, we introduce MAS-Loop, an execution-grounded framework that closes this feedback loop by coupling an Orchestrator with a trajectory-conditioned Refiner. For each task, the Orchestrator constructs and executes a source MAS to obtain an outcome score and an intermediate trajectory. Conditioned on the trajectory, the Refiner localizes an actionable failure, formulates a trajectory-grounded diagnosis, and proposes a bounded patch to the agents, information flow, or execution configuration. The executable updated MAS is re-executed on the same task to measure its realized outcome under stochastic execution. This pair-MAS comparison provides complementary learning signals: the Orchestrator learns to generate strong systems with fewer recoverable failures, whereas the Refiner learns to expose and repair observed weaknesses, thereby driving policy-level co-evolution across training iterations. Experiments show that MAS-Loop achieves the strongest overall performance among the evaluated baselines, with consistent gains from refinement across all evaluated benchmarks. The co-evolution process also yields MAS-JudgeCorpus, a dataset of execution-grounded source-updated pairs with orchestration- and trajectory-level supervision, which can provide the community with support for MAS reward modeling and trajectory-aware investigation.
Zero-shot vision-and-language navigation (VLN) requires an agent to use past observations without navigation-specific training. We study what information an 8B-scale navigation system should retain and how it should be organized for subsequent decisions. Our agent combines instruction progress, hierarchically compressed observation history anchored to poses, and accumulated geometry in complementary task, temporal, and spatial memories. An 8B observer summarizes observations, and an 8B reasoner uses the memories to select waypoint intents or request additional views, a geometric controller executes the selected targets. A semantic information perspective motivates the distinction between retaining evidence, making it accessible to the reasoner, and acquiring further observations. On R2R-CE val-unseen, the complete system achieves the highest among the evaluated Qwen3-based zero-shot baselines. Removing temporal memory reduces SR and sharply increases average path length from 16.9 to 39.1m, removing the spatial-map and planning component reduces SR as well. Real-world experiments on a Unitree Go2 show that our zero-shot method achieves higher success rates than the post-trained baseline in both corridor and office environments, suggesting better transfer to the evaluated real-world settings. As a complementary evaluation, LAW2 uses paired instruction interventions to examine behavioral responses, instruction following ability beyond endpoint success. These results support preserving task progress, observation history, and geometry in structured memory for zero-shot navigation with smaller open models.
Coding agents work in an environment built for them: a compiler or a test checks each change and points to the fault, and version control keeps every change in a project people can continue. Video editing agents lack one: an edit's correctness shows only in the costly rendered video, which does not point to the decision at fault, and existing systems offer either editing operations that cover only anticipated edits or a script whose output is a video the user cannot keep editing. We present Cassette-Editor, a complete agent harness for video editing that lets the agent edit the way a person does: knowing what the project allows, seeing what each change affects, and revising from there. The agent writes patches of typed changes to the editor's timeline document. A schema check and a rulebook shared with manual editing check each patch before it commits; a rejected patch leaves the project unchanged and returns each problem located with repair options, and an accepted one becomes a step of the history the user also edits. Reads map source analysis to the frames of the current edit. Because manual gestures are writes in the same language, folded by the same executor and rulebook, the agent can express every change a manual edit makes to the project. On 31 editing requests with the model held fixed, the complete harness passed 89% of trials against 47% for the same model with ten basic editing operations and without the rulebook, at less than half the agent time and cost per successful request; the loss fell on requests at rule boundaries, composite edits, and undo, not on finding content.
Dexterous manipulation with multi-fingered hands is the next frontier of robotics, yet learned hand skills remain limited to simple grasps or basic repeated motions, such as rotating an object. As robots attempt more demanding in-hand maneuvers, safety, such as not dropping the object, becomes critical, yet it is rarely measured. We study in-hand object pose reaching, a task that demands diverse and difficult grasps: moving an object to almost any position and orientation within reach of the fingers, at any wrist orientation, through long sequences of consecutive goals. We present a recipe for training such policies for a LEAP hand in simulation without human data, built on a range of insights, from diverse resets and goals via generated grasp datasets to an adaptive friction curriculum. With careful system identification, the policies transfer zero-shot to the real hand, using either motion capture or a single wrist-mounted depth camera, and prove remarkably capable and robust. On top of the task policy, we add a learned residual safety filter with adjustable conservativeness, letting us tune the balance between performance and safety and greatly reduce object drops. In 13.3 hours of real-robot evaluation over 4,800 unseen goals across 4 objects and diverse wrist orientations, the state-based policy reaches 95.2% of goals within 2 cm and 0.4 rad, and the filter cuts its drops by 75% while still reaching 91.5% of goals. The policies can also spin a lightbulb held in an arbitrary pose, reach out-of-distribution goals, withstand fast arm motion and perturbations, and, needing only a wrist camera, serve as portable skills such as screwing in a lightbulb. We release the full recipe as a platform for studying performance, reliability, and safety in high-dimensional dexterous systems. Videos and more are available at https://ian-chuang.github.io/dexsafe/.
Convolutional neural networks (CNN) are widely used for medical image segmentation. However, they lack an inherent measure of confidence. Existing methods for estimating prediction uncertainty often require a custom network architecture and/or dedicated training setups and need to be trained from scratch. In addition, the deduction of a single confidence score per input from the uncertainty maps is not always straightforward. In this study, we explore voxel-based uncertainty estimation methods that can be added to existing segmentation pipelines without training from scratch. We also propose a novel method to estimate a single confidence score from the uncertainty maps. Our evaluations on different ROIs and modalities suggest that our confidence score can predict segmentation performance in unseen data.
We introduce the Axiomatic Learning Structures (ALS) and Axiomatic Probabilistic Structure (APS) in order to study the interactions of Machine Learning, Optimization, with research frameworks- which, we take the view, can be characterized at the core by the corresponding axioms. The ALS explores the interplays between structures that one is interested in Learning Theory such as Loss functions and the axioms of the objects and transformations involved in these functions. The APS studies the interaction between structures of interest in Probability Theory such as Gibbs distributions, equations with noise, etc. and the axioms of the corresponding objects and transformations. We also introduce in Section 6 Transformations-Axiomatic Based Learning (TABLe) on Mathematical data-sets. One feature of this approach is the ranking of transformations from a table for a certain task based on scores which reflect which transformations work better than others, similarly to when human investigators try various methods on a task. We also study structure recovery from noisy data. In particular, to exemplify the ideas, we consider loss functions which contain derivations and derivations-like transformations. As a first application we study recovery in the context of polynomials with integer coefficients using standard derivatives along with other types of derivations which appear in Arithmetic: the p-derivations which satisfy \textit{different axioms} from the usual derivatives. As a first application of the APS we study a class of noisy (partial/mixed) derivations equations. In the last sections, to provide further potential applications of this framework, we give an overview of an initial simplified Opinion Dynamics model and its (simultaneous) study of rules of updates.
Recent long-tailed recognition (LTR) methods incorporate neural collapse (NC) and align features to a fixed Equiangular Tight Frame (ETF) as a uniform target for all classes. However, forcing tail classes to fully collapse over-contracts their samples and weakens generalization. To address this, we propose a controlled partial neural collapse, realized through two coordinated controls on a backbone: the Class-Adaptive Sparse Activation (CASA) classifier and the ETF-constrained adaptive margin (ETF-AM) loss. Rather than forcing a uniform full collapse, this regime allows head classes to collapse along the ETF while tail classes retain within-class diversity. This prevents the tail from being forced into full collapse, which is suboptimal under imbalance. To set the dimensionality of collapse, CASA induces class-related node specialization through feature-conditioned gating so that tail classes occupy their own feature directions. For the clearance axis, ETF-AM applies a frequency-adaptive margin, with smaller margins for head classes and larger for tail classes. On CIFAR100-LT, ImageNet-LT, Places-LT, and iNaturalist-2018, our method achieves the best overall accuracy over prior methods. In the empirical analysis, within-class spread increases monotonically from head to tail classes, consistent with tail classes retaining variability rather than fully collapsing.
Frontier models such as GPT-6 Astra now directly control robots and can generalize better than vision-language-action models (VLAs). We argue that this ability could arise from two distinct, composable skills: semantic reasoning, to understand what a task asks for in a scene, and motion control, to move the robot to do it. We show that small open VLMs are already capable of the first and mainly lack the second. We introduce Zero-real Agent Policy (ZAP), a recipe that teaches this missing motion skill, entirely in simulation on simple primitive tasks and from the model's own successes, with no frontier teacher or distillation and no real-world robot data. ZAP reaches 96.7% success on real-world tasks, matching GPT-6 Astra in our harness while running 3x faster and 17x cheaper, and 74.0% on a bimanual adaptation of RoboLab-120, on par with GPT-6 Astra's 72.2%. ZAP also generalizes to new objects, goals, and strategies, and in simulation to a robot embodiment on which it never trained.
Attention over tens of thousands of visual tokens dominates the inference of video diffusion transformers (DiTs). Existing methods reduce this cost through sparsity, low precision, or both, but use a single precision for all kept tiles. We observe that the output error introduced by approximation varies widely across tiles. Many tiles can be skipped with little error, others must be computed but gain little beyond low precision, and a few retain substantial error even at low precision. We present TriTile, a training-free attention method that assigns each tile one of three approximation tiers: skipping it by replacing it with its block means, skimming it in NVFP4, or refining it with NVFP4 residual correction. A lightweight router predicts each tile's error under each tier from block statistics and spends extra compute on a tile only where it removes enough error; a single knob, a price on compute, controls the whole speed-quality trade-off. A fused kernel executes all three tiers in a single pass on the FP4 tensor cores. Over dense FlashAttention, TriTile speeds up generation by 1.38-1.83x on Wan2.2 at 23.0-27.6 dB PSNR, 2.28-2.32x on HunyuanVideo at 34.5-36.0 dB, and 1.84x on LTX-2.5 at 34.1 dB, and in every setting it Pareto-dominates every baseline in speed and PSNR.
Web services increasingly expose backend interfaces that let users delegate tasks to AI agents, which invoke service operations directly without navigating websites. This shift changes how existing Web vulnerabilities can be exploited. We investigate these changes through VELA, an executable benchmark covering 44 vulnerabilities across twelve Web application families. The benchmark preserves vulnerable backends and derives agent-facing interfaces from legitimate workflows. We evaluate three commercial agents across multiple interface configurations, using controlled comparisons to separate interface effects from agent behavior and independent oracles to verify attack outcomes. Our results show that the agent realizes every attack that direct interface invocation realizes in the reference configuration, so exploitability is set mainly by the interface: 91% of vulnerabilities remain exploitable under the reference design. Interface restrictions prevent some attacks by removing required operations or limiting input control, while retention remains high across vulnerability classes. No case-level loss in the reference configuration is attributed to agent refusal or input normalization, and in one case the agent completed an attack by supplying calls the attacker omitted. Across the tested configurations, interface design affects exploitability more than agent choice, moving retention from 39% to 95% against at most 4.5 points across agents, but some restrictive interfaces also limit legitimate workflows. Our study provides a systematic empirical account of how agent-mediated access reshapes existing Web vulnerabilities, grounding future Web architecture design in measured security outcomes and trade-offs between reducing attack exposure and preserving legitimate functionality in agent-facing services.
Unsafe deserialization turns untrusted data into attacker-controlled execution, and a single vulnerability can admit diverse exploit variants that evade defenses tailored to a known exploit. Existing defenses typically operate at a single enforcement point with limited visibility into the full attack. Combining controls across layers offers complementary protection, but coordinating them to preserve legitimate functionality and adapting them to new variants requires substantial expert effort. We propose JANUS, a red-blue adversarial framework for automated cross-layer defense synthesis. Starting from a known exploit, \sys iteratively refines deployment-specific policies in an isolated offline environment. A black-box attacker probes the current policy for exploit variants using only externally observable feedback. The defender summarizes each verified variant into its security-relevant effects, uses counterfactual replay to separate complementary coverage from redundant enforcement, and jointly selects and scopes controls across layers. Each candidate policy is validated against accumulated exploits and benign workflows. We evaluate JANUS on sixteen vulnerability deployments across six Java and PHP applications against gateway, application, system, and cross-layer defenses. JANUS reduces held-out attack success to 6.2%, compared with 18.8-55.0% for the baselines, fully mitigates 13 of 16 deployments, and preserves 98.4% of benign functionality tests at a 6.9% targeted P95 latency increase.
A long-lived assistant that brings up memories without being asked must decide at every turn whether anything in its store is worth an interruption. We show that this decision is a multiple-testing problem whose difficulty grows with the store size N. On replays of LongMemEval, a similarity threshold tuned to 5% false alarms at N=8 reaches 84-100% at N=2048 across embedding models and cue types; LoCoMo and LLM judges show it too. We prove that a gate which thresholds the best of N independent irrelevant scores at false-alarm rate alpha operates at a per-item false-positive rate of about -ln(1-alpha)/N, which pins its hit rate to the far left of the ROC and makes it regularly varying in N under regular variation. Replayed hit rates stay within its bounds for mixed cues up to 0.3 points, and hit rates carry over to held-out questions to within 6.0 points. At matched false alarms the deployed score recalls 14-26 points more than raw cosine. A cue-swap conformal gate, which ranks the current cue among other cues on the same store, has a finite-sample false-alarm guarantee under exchangeability. With a background from the user's unfiltered history it keeps false alarms at 3-6% on LongMemEval at the hit rate of a label-tuned threshold; on LoCoMo it overshoots or turns conservative depending on the source of its reference cues. In a chronological replay with the user's own earlier messages as references, it fires on 5-8% of questions on stores without their answer and on 6-8% of session-opening turns, against 57% for the deployed gate at N=2048, at a median 0.4 s CPU per decision. A cross-encoder pays the same price of scale, and on a prospective-memory benchmark, similarity gates recall at most 9% of constraints at controlled false alarms. One human and two LLM annotators judge most false alarms unhelpful for a response. Code and data: https://github.com/RiverChenStillCompiling/price-of-scale
Camera motion recognition requires separating viewpoint changes caused by camera translation, rotation, or zoom from object motion and scene dynamics. Video-language models often rely on semantic cues and struggle with the geometric evidence needed for this task. Joint-Embedding Predictive Architectures such as V-JEPA offer strong spatio-temporal representations, but our experiments suggest that they underrepresent geometric cues in some regions. Under temporal-reversal supervision, V-JEPA improves overall but degrades on several motion families, whereas a weaker geometry-oriented VGGT predictor improves across all families. We refer to these geometrically underrepresented regions as geometric blind spots and propose Fovea-JEPA to identify and address them. Using V-JEPA as the base model and VGGT as a geometry expert, we construct intervention pseudo-labels from their prediction errors to weakly supervise a blind-spot predictor. The predictor produces token-wise response maps, which are aggregated into a clip-level intervention score to guide adaptive fusion of visual and geometric features. This selective intervention introduces complementary geometric information while preserving reliable visual predictions. Experiments on CameraMotionDataset, CameraBench, and ShotBench demonstrate the effectiveness of Fovea-JEPA, with gains of 9.67, 8.67, and 24.38 percentage points over previously reported state-of-the-art baselines, respectively. Code and models will be released upon acceptance.
We study the last iterate of clipped SGD, and of every rule that queries its current iterate and steps along the resulting stochastic gradient, when training is stopped at a data-dependent time, so that guarantees must hold at all times simultaneously. On smooth convex objectives under heavy-tailed noise with a bounded central $p$th moment, $1<p\le2$, we determine the optimal time-uniform rate of these rules. A clipped SGD schedule without horizon or confidence inputs attains $\tilde O(t^{-\beta_p})$ at all times, with $\beta_p=2(p-1)/(3p-1)$, also almost surely, and no such rule attains $\tilde O(t^{-r})$ for any $r>\beta_p$, even with adaptive step sizes and thresholds. This rate lies between the optimal rate $O(T^{-\alpha_p})$ for a known horizon $T$, $\alpha_p=(p-1)/p$, and the previous time-uniform rate $\tilde O(t^{-2(p-1)/(3p)})$ (Chezhegov et al., ICLR 2026); under finite variance, the exponents are $1/2$, $2/5$, and $1/3$. Their conjecture that a better schedule reaches the optimal rate thus holds with a known horizon (Liu; Jiang et al.) but fails without one. Querying gradients at the running average attains $\tilde O(t^{-\alpha_p})$ at all times, whereas heavy-ball momentum with geometric forgetting and polynomially bounded steps cannot improve on $\beta_p$.
Using quarterly FRED and CSO Ireland data over 1995Q1-2025Q4, we examine whether U.S.\ IT investment and Irish data-centre electricity are recomposing the Kitchin cycle around energy- and compute-backed assets. Three findings stand out. First, a Quandt-Andrews sup-Wald test locates a significant break in the U.S.\ IT-GDP correlation at 2014Q3 (sup-W $= 9.69$; $r$ falls from $+0.45$ to $-0.12$); the Fisher $r$-to-$z$ test survives Bonferroni adjustment ($z = +3.11$, $p_{\text{Bonf}} = 0.006$), whereas the analogous test on the imposed Cloud-vs-AI boundary does not ($p_{\text{Bonf}} = 0.164$), so ``Strategic Decoupling'' pre-dates the AI Era by approximately six years. Second, in Ireland the synchronous AI-Era correlation between metered data-centre electricity and Real GDP is large ($r = +0.41$, $N = 16$) but \emph{collapses to $r \approx 0$ when Real GDP is replaced by Real GNP}, suggesting the headline is largely an artefact of the shared multinational tax-structure footprint between hyperscaler operations and Irish gross output. Third, the Irish cross-correlation profile is asymmetric in the opposite direction to a leading-indicator reading: the largest positive coefficients cluster on the $k < 0$ side (GDP precedes data-centre load), so compute capacity is better read as coincident or moderately lagging. Given the small AI-Era samples and the absence of a quarterly multinational-adjusted aggregate such as GNI*, all coefficients are exploratory; subject to these caveats, the U.S.\ Kitchin cycle is best dated to 2014Q3, and the Irish Energy-Data Nexus is dominated by multinational firm activity.
If two minds were linked, would each still know which thoughts are its own? Language models are increasingly linked in this way: instead of exchanging text, one model reads another's key-value cache. Systems that join a foreign cache to the receiver's own use one of two layouts: the cache lies over the receiver's own context, at the same token positions, or in front of it. We give two copies of one model different private priorities and link them in both layouts. With the cache over its own context, the receiver gives a fact found only in its partner's cache in every answer. Yet in about half of its answers it gives the partner's priority as its own, in both Qwen3-4B and Llama-3.1-8B: as often as a receiver that ignored whose priority was whose, although the text in the cache names its owner. With the same cache moved in front, this falls to 5% in Qwen but only to 43% in Llama, although both models nearly stop naming the partner's priority as the one they are using. Where the cache sits matters, yet neither layout in use protects both models. Fewer such errors do not mean better attribution: once the cache is moved, the receiver names the partner's priority as the partner's less often, not more, and its report of an inserted memory cannot tell the partner's cache from a copy of its own. In short, shared caches deliver content but not ownership.
Quality-Aware Decoding (QAD) improves machine translation through test-time scaling, generating multiple candidate translations and selecting the final output with a quality-aware decoder. Usually the same candidate budget N is spent on every source sentence, assuming that all inputs require the same amount of search. We show that this assumption does not hold and that the quality gain obtained from additional candidates varies across sentences. We study search utility through three quantities: total scaling gain, how much quality a sentence can gain from a much larger candidate pool; marginal scaling gain, the expected gain from generating one additional batch of candidates; and the sampling saturation point, the candidate budget required to realize most of the total scaling gain. We investigate two sources of information for predicting these quantities: signals available before candidate-set expansion, and statistics computed from the candidate pool accumulated during search. We find that sentence-level signals correlate with total scaling gain but are less informative about the sampling saturation point. In contrast, properties of the current candidate pool provide useful signals of marginal scaling gain. Finally, we show that these signals can identify sentences that obtain larger gains when additional search is selectively applied.