Papers

Although post-training significantly enhances model performance, the use of massive undisclosed data raises concerns about benchmark contamination and copyright infringement. This drives growing interest in training data detection for post-trained models. Existing evaluations for training data detection often leverage benchmarks constructed post hoc, which can introduce distribution shifts. Alternatively, controlled evaluations rely on small-scale training setups that diverge from realistic post-training settings. As a result, their reported performance may not faithfully reflect the effectiveness of detection methods in real-world scenarios. To this end, we introduce \method, a reliable benchmark designed to evaluate detection methods in realistic, large-scale post-training settings. Built on a multi-stage pipeline, \method\space encompasses stage-specific models throughout the post-training lifecycle, spanning SFT, DPO, and RL. To perform a reliable evaluation, we construct model-specific evaluation sets across four diverse tasks, featuring distribution-matched members and non-members with verifiable membership labels. Our results reveal that even the best-performing method achieves an average AUC of only 0.529 on our benchmark, which approaches random-guessing level (0.5). Furthermore, our analysis suggests that evaluation protocols can significantly affect measured detection performance. These findings highlight the significant potential for developing effective, practical, and robust detection methods under realistic scenarios.

Learning equilibria from data is difficult when several decision makers interact indefinitely and the system state has no prescribed bounds. We examine discounted general-sum stochastic games with compact action sets and an unknown nonlinear drift observed through Gaussian transition noise. The proposed procedure uses a single uninterrupted trajectory to update Gaussian-process models of the drift and to solve successively refined finite games. We prove a perturbation result showing that stability of equilibrium selection, together with uniform KL accuracy of Gaussian transition models, is sufficient for regional uniform convergence of the induced values and approximate Nash behavior in the original game. The region specifies where the guarantees are evaluated; strategies and dynamics continue to act on the entire Euclidean state space. For GP estimators, simultaneous confidence bounds yield a rate controlled by posterior uncertainty, kernel information gain, and joint-state dimension. Numerical studies illustrate how kernel choice and dimension appear in the predicted convergence behavior.

A ghost-cell immersed boundary method is proposed for the two-dimensional shallow water equations with a variable bed (topography), solved with an explicit Godunov-type finite volume scheme (unsplit second-order MUSCL, HLLC flux, hydrostatic reconstruction) on Cartesian grids. Ghost-cell methods leak mass through the wall. To find a robust and conservative treatment, eight constructions of the ghost values and a cut-cell form of the continuity equation were applied to seven test problems, which include wetting and drying on sloping beds and a laboratory dam break. In the resulting scheme, the ghost values are interpolated bilinearly at the image point; solid and dry nodes are excluded, and the weights of the remaining nodes are renormalized. Mass is conserved by solving the continuity equation in cut-cell form, while the momentum equations are left unchanged. The proposed scheme is about as accurate as the ghost-cell method without this modification and preserves a lake at rest whose shoreline crosses an immersed wall. Because only the continuity equation is modified, it avoids the treatment of the momentum equations on cut cells (wall pressure, cell merging) that a full cut-cell method requires.

Dynamic sparse training (DST) aims to reduce training costs while maintaining accuracy. Existing DST methods mask out parameters but keep matrix dimensions unchanged. Training speedups therefore depend on specialized support, such as sparse kernels. We introduce BARTER, a structured DST framework that reduces training cost using standard dense kernels by keeping training tensors physically compact while dynamically reallocating capacity across layers. Unlike existing works that require computing gradients for all parameters, BARTER instead uses statistics collected only from active units during ordinary backpropagation to prune and clone attention heads and feed-forward channels under separate fixed global budgets. At each topology update, it resizes the affected parameter tensors and remaps their optimizer states, so subsequent training steps operate directly on the compact model. Training and fine-tuning experiments with Transformer-based vision models on two image classification benchmarks at densities from $0.1$ to $0.7$ show competitive Top-1 accuracy relative to full-tensor DST baselines. On an NVIDIA L40, BARTER achieves speedups of up to $6.29\times$ in measured training segments and reduces peak allocated GPU memory by up to $81.3\%$ relative to these implementations. An iPhone 13 Pro Max training benchmark additionally shows a $2.78\times$ speedup over the full-tensor DST baseline.

We present Unravel, a training-free framework that turns a single image into an interactive 3D scene, where every object is an independent and movable asset. Existing generators either produce one entangled geometry in which no object can be moved, or predict all objects in one feed-forward pass, where a wrong guess cannot be detected or repaired. Unravel instead takes a different approach: it first takes the scene apart, and then puts it back together. Starting from the input image, an agentic harness that iteratively peels the image layer by layer, and adopts point cloud as major modality for spatial relationship recovery. Specifically, a vision-language model plans which objects are fully visible, an image editor removes them and restores the occluded regions. Every edit is verified before being incorporated into the reconstruction. This process gradually turns occluded parts of the scene into direct observation, allowing objects to be reconstructed when they are best exposed rather than hallucinated from the original image alone. The scene is then rebuilt in reverse removal order, so that the supporting objects are naturally placed before the objects they support. Since a single-view image only reveals part of an object, we treat the observed points as a lower bound on its size: an asset may grow to cover its points, but can not shrink away from them. Finally, we physically settle the reconstructed scene to remove floating objects and interpenetrations, followed by an agentic audit that corrects remaining placement errors. Extensive experiments on diverse indoor scenes show that Unravel recovers more objects, places them more faithfully, and produces scenes that are more physically-plausible than existing methods.

Agent frameworks can persist checkpoints, trace tool calls, and limit resource use, but these facilities often assign different identities to execution, governance, and replay. We present Pollard, a framework-independent runtime based on governed semantic execution trees. A step identity commits to canonical input, kind, attempt, and parent identity; the same tree records completion, refusal, resource settlement, and branch history. Strict replay returns a verified stored result without invoking the wrapped operation, while a transactional store coordinates reservations across concurrent workers. Under a signed deterministic freeze, two state-model replicas each executed the evaluation transitions without an invariant violation, retained-lease PostgreSQL groups were tested against exact limits, strict replay avoided external calls, and all semantic mutations were detected with no false alarms on the controls. A multi-pair framework matrix preserved reference behavior and verified the trees. A separate cost-capped hosted study exercised live provider paths, and every completed Pollard trial produced a verified tree. Together, the studies support the core mechanisms under explicit lease, boundary, and seal-custody assumptions.

Serving stacks for large language models manage context through four recurring mechanisms: incremental edits to a growing context, branching and snapshotting for tree search, identification of reusable prefixes for KV-cache reuse, and encoding of repeated segments. Each mechanism is served today by a different specialist structure. We show that all four reduce to a single persistent weight-balanced rope whose every node carries a polynomial hash of its subtree over a Mersenne prime, so that the structure supports logarithmic edits and logarithmic verified content-identity queries at the same time. We validate each mechanism with pre-registered, replicated experiments, then compare against a deployed specialist per mechanism: the Ropey editing library, a faithful PagedAttention copy-on-write block table, and the vendored SGLang radix matcher. On workloads that interleave edits with identity queries, Hashrope is net-superior to Ropey by a factor that grows without bound (369.7x at one million characters). Against the radix matcher on a live vLLM stack serving Qwen2.5-7B-Instruct-1M on four A100 GPUs, Hashrope identifies a one-million-token shared prefix exactly and 1.20x faster, at a context length the dense engine cannot even prefill.

SLAM (Simultaneous Localization and Mapping) is a core technology in autonomous systems that enables them to estimate their own position and build a map of unknown environments simultaneously. SLAM systems in dynamic environments are of great practical importance. However, existing dynamic SLAM approaches still have several limitations like over-removal and high computational costs. To address these limitations, we propose a real-time RGB Dynamic SLAM framework that precisely removes moving objects at the pixel level without relying on predefined labels or segmentation models. Our system combines pixel motion predicted by ego-motion using DUSt3R with point tracking from All-Tracker, removing only actually moving pixels by analyzing discrepancies between the two motion sources. Unlike object-level masking methods, our approach eliminates only currently moving regions, allowing temporarily static parts to be utilized for tracking. Operating at approximately 10 FPS, the system achieves real-time performance on standard 30 FPS video streams by subsampling every 3 frames. Relying solely on RGB inputs, this significantly improves upon state-of-the-art dense dynamic SLAM approaches (0.5-3 FPS), presenting a highly practical solution.

This paper develops a unified framework for distributional regression with arbitrary response variables. Rather than relying on scalar thresholds or Euclidean structure, we model conditional event probabilities indexed by measurable sets, which extends distributional regression to general response spaces while retaining a direct connection to classical event-based formulations. Within this framework, we study two complementary estimators: a preliminary estimator based on separate regression of event indicators, and a constrained estimator that enforces inclusion-consistent predictions through a directed acyclic graph (DAG) over a finite set family. We establish general error bounds for both estimators and further specialize the results to dense ReLU networks, yielding explicit convergence rates overcoming the curse of dimensionality. Simulation studies and real-data experiments further demonstrate the practical effectiveness of the proposed framework.

LLM agents increasingly operate over multiple entities, where reliable response generation requires not only retrieving relevant evidence but also preserving its association with the correct entity. We identify Phantom Merge (PM), an entity-evidence binding failure in which an agent attributes unsupported claims to its committed entity, potentially even when the task is completed correctly. We formalize PM at the claim level and characterize its prevalence and error subtypes across four agent domains and four backbone families. To address this problem, we propose a two-stage framework for claim-level risk estimation and correction. First, Anchor-Grounded Risk (AGR) combines a representational readout of the agent's pre-emission hidden state with Jacobian-based functional information, providing complementary slot-level and anchor-value risk scores. Second, Fixed-Anchor Correction (FAC) audits flagged claims and selectively revises them using observed anchor evidence or removes them when no supported replacement is available. On the held-out Shopping detection cohort, AGR-slot achieves an AUROC of 96.2% and an average precision of 97.3%, outperforming the evaluated detection baselines. In the end-to-end Shopping mitigation evaluation, FAC reduces trajectory-level PM from 51.4% to 6.2%, with an output-to-original claim ratio of 80.7%. These results demonstrate the prevalence of entity-specific grounding failures and the effectiveness of combining internal risk estimation with evidence-constrained claim correction. Our code is publicly available at https://github.com/LabRAI/PhantomMerge.

Transportation planning has traditionally required substantial expertise in network modeling, traffic assignment, and optimization. As a result, strategic analyses such as network design, capacity expansion, and model calibration remain inaccessible to non-expert decision makers. Recent advances in large language models (LLMs) provide an opportunity to bridge this gap by enabling natural-language interaction with complex transportation planning tools. However, existing LLM-based systems primarily serve as interfaces to simulation software and lack the ability to efficiently search, evaluate, and optimize transportation solutions. This study proposes an integrated human-AI transportation planning framework that combines a large language model, a model context protocol (MCP) interface, an open-source transportation planning engine (AequilibraE), and a differentiable user equilibrium (UE) assignment layer. The framework allows non-expert users to execute network analyses and improve solutions for optimization. A differentiable UE layer is developed by directly differentiating the Karush-Kuhn-Tucker (KKT) conditions of the Beckmann traffic assignment problem. The proposed framework is demonstrated through two representative planning tasks: calibration of travel time functions from observed traffic counts and network design under budget constraints. The examples illustrate how differentiable transportation models and LLM agents can jointly democratize transportation planning and substantially reduce barriers between domain expertise and transportation decision making.

Learned world models for autonomous driving operate in unconstrained latent spaces, yet the physical systems they represent are governed by strict non-holonomic constraints: ground vehicles cannot realize instantaneous pure lateral motion. We study long-horizon latent rollout modeling for autonomous driving and ask whether rollout quality can be improved by assigning a motion-aware geometry to latent transitions, rather than by redesigning the predictor architecture. Starting from an unconstrained latent rollout baseline, we introduce Directional Pullback Regularization (DPR), a training time method that defines a kinematics-aware anisotropic Riemannian metric in decoded state space and pulls it back through the decoder Jacobian to the latent manifold. The resulting latent-space metric assigns higher cost to latent directions that decode to lateral vehicle motion, while preserving the original encoder-dynamics-decoder backbone. This yields a clean three-way comparison among an Unconstrained Latent Rollout model (ULR), an Euclidean Pullback Regularization ablation (EPR), and the full directional method (DPR). We evaluate all models on open-loop ego-motion rollout over a common 10-second test horizon, using training rollout lengths of 1, 3, 5, and 10 steps. Across all four training regimes, DPR achieves lower 10-second final displacement error than both ULR and EPR. The strongest relative accuracy gain appears for models trained with 3-step rollout supervision, where DPR reduces the 10-second final displacement error by approximately 44 percent relative to ULR. The strongest overall control-relevant tradeoff is obtained for longer rollout-trained models, especially at 10 steps, where DPR simultaneously improves displacement error, lateral motion consistency, and forward motion preservation relative to both ULR and EPR.

Memory ranking is the stage in personalized Large Language Models (LLMs) that decides which of the candidate memories returned by a retriever enter the limited context, so that the answer is grounded in the right parts of the user's history. Existing memory rankers operate in a single pass, scoring each candidate once against the raw query and cutting the ranked list at a fixed top-K. The single pass misses memories that the answer needs but that look unrelated to the query on their own, and the fixed cut ignores how many memories the query needs. To address both problems, we propose ReMind, a memory ranker that runs as a recurrent and self-terminating loop, following the cognitive theory of memory search: each recalled item joins the cue for the next recall (compound cueing), and recall stops once enough has been collected for the task (metamemory monitoring). For compound cueing, ReMind adds the top-ranked memory to the query context at each step and reranks the remaining candidates under the expanded cue, so that each recall makes the cue more specific and later steps look for what is still missing rather than for what merely resembles the query. For metamemory monitoring, a sufficiency predictor estimates at each step whether the collected memories are enough to answer the query and stops the loop once they are, so that the number of selected memories adapts to each query. Experiments on public benchmarks show that ReMind achieves the best results compared to state-of-the-art memory rankers.

Omni-modal large language models (Omni-LLMs) extend LLMs to process heterogeneous modalities (e.g., vision, audio, and IMU signals) alongside language, enabling a broader range of multimodal tasks. Yet we find that these models often fail to express information from non-textual modalities even when they perceive it correctly. We term this disconnect knows-but-cannot-say. To investigate this phenomenon, we introduce Token-level Modality Alignment (TMA), which measures the alignment between individual modality representations and response representations. Our analysis shows that non-textual representations are less aligned with response tokens and receive less attention than textual representations, suggesting a potential explanation for uneven modality preferences. Motivated by these findings, we propose OmniCatalyst, a simple perceive $\rightarrow$ infer strategy that complements the original modality inputs with self-generated text, making non-textual knowledge more accessible. Across seven Omni-LLMs, OmniCatalyst consistently improves performance (+7.29%) and yields more balanced modality preferences, with further gains from diverse training strategies. Further analysis demonstrates that the knowledge extraction phases increase TMA, suggesting that improved alignment makes non-textual information more accessible during response generation. In summary, our findings uncover how Omni-LLMs can better express modality knowledge through language. Our code is available at https://github.com/naver-ai/OmniLLM.

Dataset distillation (DD) aims to retain the training utility of a large dataset in a much smaller set of examples. For skeleton-based action recognition, this requires capturing discriminative motion variation without disrupting the geometry and temporal continuity of articulated bodies. Directly optimizing joint coordinates can violate these constraints, while plausible motion generation alone does not ensure useful training examples. We propose Prototype-Conditioned Motion Dataset Distillation (ProMoDD), which conditions a motion diffusion model on class-wise prototypes of temporally resolved recognition features. We further tailor diffusion training to motion data, enabling the generator to model both skeletal configurations and their temporal evolution. We then construct each compact training set through budget-aware hybrid selection: representative real sequences cover densely populated regions of the pooled recognition space, while selected generated candidates complement them in less-populated regions. This couples prototype-guided generation with feature-space selection, balancing the reliability of observed motions with the variation of synthesis under a fixed per-class sequence budget. Across five difficulty-stratified subsets of the single-person classes in NTU RGB+D 60, ProMoDD achieves its largest gains under tight distillation budgets, reaching 40.32% top-1 accuracy on Full49 with three sequences per class, 4.57 percentage points above class-wise k-means. These gains reflect the complete framework. The same distilled sets train ST-GCN, MS-G3D, and CTR-GCN without regeneration.

Puzzles offer a natural testbed for Large Language Models (LLMs) reasoning because they pose verifiable goals under partial information and require multi-step decisions to reach a solution. However, existing puzzle benchmarks typically focus on narrow environments with rigid mechanics and tightly constrained solution targets. To bridge this gap, we introduce the Active Probing Benchmark (APB), which generalizes puzzle-solving by casting concept identification as an interactive reasoning task that integrates specialized knowledge assessment into the puzzle setting. Specifically, APB evaluates whether models can acquire information via multi-turn active probing and inference to pinpoint 960 target concepts of specialized difficulty across eight diverse academic domains. Our evaluation reveals a counter-intuitive phenomenon: most open-source reasoning models significantly underperform their standard instruction counterparts. This finding suggests that in interactive settings, these models struggle to leverage procedural reasoning to reliably access and apply specialized domain knowledge. Beyond these empirical insights, APB provides a scalable pipeline for generating benchmarks that probe knowledge reasoning under interaction across diverse settings.

Semantic-ID-based generative recommendation unifies retrieval and ranking through autoregressive decoding but differs structurally from discriminative models, which can directly access rich target-item side features. The attributes are compressed into a finite codebook during quantization, while the target item remains undetermined during decoding and its features cannot be looked up. This target-side feature vacuum weakens generation at each semantic-ID level through high conditional entropy and cascading errors, and final ranking through missing target-side feature crossing. Existing methods interact with candidate representations or generate discrete attribute tokens, but none reconstructs a continuous target-side representation inside decoding to support both stepwise generation and ranking. We propose TacGR (Stepwise Target-Side Feature Anticipation and Crossing for Generative Recommendation). Before decoding each codeword, TacGR anticipates a continuous target-side feature vector and crosses it with the semantic prefix and user behavior sequence inside the decoder. The crossed representation conditions current-level generation and is directly used for ranking. Continuous-vector regression risks a learnable target collapsing with its prediction and prediction noise propagating into generation through attention. TacGR addresses these challenges with a semantic-anchoring auxiliary tower and a unidirectional visibility mask, respectively. For extremely long industrial behavior histories, it introduces gated convolutional denoising to suppress noise and adapts the pyramid memory funnel to reduce inference cost. Offline experiments on a large-scale Lazada dataset show significant gains over OneRec, UniRec, and UniSGR. Online A/B testing improves IPV, Order, and GMV by 0.79%, 5.55%, and 3.09%, respectively.

Financial forecasting has traditionally relied on numerical time-series (TS) signals, while textual information such as news remains largely underutilized. We introduce XForecast, a benchmark and competition for news-grounded financial forecasting, where the task is to predict stock price movements four weeks ahead using both historical price data and paired news signals. XForecast combines stock price data with curated financial text generated using FinTexTS, a multimodal dataset generation method. It covers 100 large-cap stocks from 2019–2023, with training on 2019–2022 and evaluation on 2023. To better evaluate models that leverage textual signals, we adopt a weighted hit rate metric that places greater emphasis on instances where news information is expected to have stronger influence. The XForecast Challenge, hosted on Kaggle as part of KDD 2026 Finance Day, aims to benchmark methods for integrating financial TS and textual information.

Deploying LLM agents in non-stationary environments requires continually revising their instructions as the task drifts, yet rewriting after every reported mistake can turn noisy feedback into persistent errors. This raises an often-overlooked scheduling question in prompt adaptation: when should feedback trigger a lasting instruction update, and at what editor-call cost? Taking inspiration from proportional-integral-derivative (PID) control, we introduce a family of lightweight gates that react to the current error, accumulate evidence over time, and roll back deterioration after an edit. We benchmark these gates against both simple baselines and advanced prompt optimizers (TextGrad and GEPA), over a diverse suite of drifting tasks, LLM parameter sizes, feedback-noise levels, and random seeds. PID-gated continual prompt adaptation yields a more favorable Pareto trade-off between task performance (high accuracy, low error) and editor-call budget than the baselines considered, pointing to control-theoretic scheduling as a simple, general tool for robust prompt adaptation under unreliable feedback.

Most existing artificial intelligence systems for breast ultrasound rely primarily on image-driven prediction, with limited incorporation of the knowledge from the Breast Imaging Reporting and Data System (BI-RADS) and the intermediate assessment processes underlying radiologists' workflow, thereby constraining their interpretability and alignment with clinical practice. We curated, to our knowledge, the largest multicenter collection of breast ultrasound images paired with structured BI-RADS descriptors, comprising 9,941 patients across 20 independent datasets. Across 14 testing datasets, IBUS demonstrated consistent generalization performance across different clinical centers and imaging conditions. Incorporating BI-RADS descriptors enabled fine-grained risk stratification, while cross-checking predicted BI-RADS grades against malignancy probabilities identified internally discordant predictions that may warrant additional clinical review. In a reader study involving eight radiologists, IBUS assistance increased diagnostic sensitivity among both junior and senior readers, with the potential of guiding radiologists to make targeted revision in risk stratification. Overall, IBUS exhibits its potential as a second-stage decision-support system that complements radiologists by providing additional risk estimates and interpretable evidence for diagnostically challenging cases. The demo video of IBUS platform is publicly available at https://github.com/zsz1997/IBUS.

Large Language Models (LLMs) are increasingly explored in the Architecture, Engineering, and Construction (AEC) domain, yet their reliability in regulatory compliance remains uncertain. In this study, we evaluate the performance of LLMs on Hong Kong fire safety codes. In collaboration with domain experts, we create a localized benchmark dataset comprising both multiple-choice questions (MCQs) and open-ended questions (OEQs) to evaluate 12 widely used proprietary and open-source models. Our findings show that top-performing proprietary models achieve 80-90% accuracy on MCQs, whereas open-source models generally lag behind. However, on open-ended evaluation, expert human assessments reveal an accuracy range of only 50-70%, demonstrating that LLMs still have substantial room for improvement in specialized compliance tasks.

An agent can retain successful trajectories without learning which part of an experience should repair a new discrepancy. We formulate passive imitation (PI) as an architectural loop that localizes a violated contract, extracts a local goal, proposes an effect-matched skill, tests a small patch, and retains evidence after full-task verification. We instantiate this loop as a deterministic symbolic controller and analyze its finite repair-query complexity under explicit identifiability and noninterference assumptions. Across ten simulated tool worlds, PI solves all 600 cold held-out composition tasks with three allowed small trials, compared with 6.17% for a whole-trajectory control, while an equally expressive compositional repair control matches PI. On an unseen tool revision, PI achieves 70.0% success versus 24.0% without small trials and 79.0% with self-trial enumeration; an exploratory 24-task adaptation stream subsequently raises first-submission success to 100%, supporting a limited correction-and-retention mechanism rather than a unique incentive advantage or general intelligence. The CPU-only implementation, integrity checks, and reproducible results are available at https://github.com/FriendOfYishan/PI.

3D instance segmentation requires representations that are both geometrically structured and semantically discriminative. Pure 3D encoders model spatial structure well but often lack strong appearance semantics, while directly lifting 2D priors into 3D by concatenation injects rich semantics without ensuring that they align well with the underlying 3D structure. In this paper, we address this problem by formulating 2D-to-3D semantic transfer as a masked semantic reconstruction task. Specifically, we mask a subset of background lifted 2D priors and train the 3D encoder to recover the missing semantic priors, providing explicit supervision for distilling rich 2D semantics into the 3D backbone. We further propose geometry-modulated semantic injection, which uses raw 3D geometry to modulate lifted 2D priors before sparse encoding, so that geometry explicitly controls semantic injection and alleviates the distribution gap between 2D semantic features and 3D geometric features. Experiments on two baseline frameworks demonstrate that the proposed method achieves state-of-the-art results in both full-data and few-shot settings on ScanNetV2 and ScanNet200. Code is publicly available at https://github.com/yharoldsu0627/Geometry_Guided_Semantic_Reconstruction

Linear complementary dual (LCD) codes are linear codes that intersect their dual trivially. We study double circulant (Euclidean) LCD codes over the finite chain ring $R=\mathbb F_2+u\mathbb F_2$, where $u^2=0$. Using the Chinese Remainder Theorem and the factorization of $x^n-1$, we derive an explicit enumeration formula, with a separate treatment of the constituent corresponding to $x-1$, which is not obtained by a formal specialization of the previously known counting formula over $\mathbb F_q[u]/(u^k)$. Assuming Artin's primitive root conjecture for $2$, we show that the binary Gray images contain an asymptotically good family of rate $1/2$ whose relative distance can approach $H_2^{-1}(1/4)$ from below. More importantly from the constructive viewpoint, by combining Gray images with a known extension construction, we obtain a binary LCD code with parameters $[29,15,7]$ and construct two explicit permutation-inequivalent binary LCD codes with parameters $[45,23,9]$.

Existing large language model (LLM) based multi-agent systems have been shown to scale poorly, and we identify three problems: reading burden that grows with the group, useful information diluted by shared work of unknown quality, and fragility to failed agents. We argue that the problems are largely due to agent-centric designs: the coordination structure is wired into fixed workflows, central coordinators, or peer-to-peer messaging, and recent blackboard-style methods that move the coordination into a shared medium relieve only the last problem. To close the gaps, we build ScaleMAS, where asynchronous agents interact with a shared datastore of linked messages via three operations: publishing messages with optional references to earlier ones that they build on, searching for messages by content, and visiting messages by address. Search returns at most $k$ messages, so an agent's read stays bounded at any group size; results are ranked by relevance and by how often they are cited, which we find to be an effective proxy for quality; and messages persist after their producers leave, so the group carries on when individual agents fail. Across four agentic tasks, ScaleMAS outperforms agent-centric and blackboard-style baselines at lower inference compute, continues to scale up to 128 agents, and tolerates losing agents midway.

Differentiable simulation enables efficient policy optimization through analytic gradients, yet safety is commonly incorporated as a weighted penalty, requiring manual tuning of the trade-off between task performance and constraint satisfaction. We instead formulate obstacle avoidance as an explicit constraint and train quadrotor policies using primal–dual gradient descent–ascent through a differentiable simulator. We introduce two key components to enable effective constrained training. First, we propose a bounded clearance-based constraint function that provides informative gradients around the safety boundary while preventing large violations from destabilizing primal–dual optimization. Second, we introduce log-mean-exp (LME) aggregation to smoothly emphasize critical per-step safety violations across a trajectory and training batch within a single constraint governed by one dual variable. Together, these components enable the desired safety level to be specified directly while balancing constraint satisfaction and task completion. We theoretically prove that the proposed optimization with differentiable simulation converges to feasible policy under mild assumptions. Experiments demonstrate improved safety–performance trade-offs over fixed-penalty and existing policy-learning baselines, as well as generalization to denser out-of-distribution environments. Finally, the learned policy transfers zero-shot to Gazebo/PX4 and a physical quadrotor without additional fine-tuning.

Hour-long videos, streaming feeds, and movie-length content are pushing vision-language models (VLMs) past their breaking point – the standard recipe of encoding each frame into visual tokens yields sequences that grow linearly with video length, causing prohibitive memory consumption and inference latency. Existing approaches, such as token reduction and keyframe selection, shorten the token sequence but leave the O(n) scaling intact, failing to resolve the fundamental bottleneck. We propose FramePack-U, a strikingly simple, training-free frame packing framework that compresses arbitrarily long videos into a fixed size visual representation at constant memory cost. FramePack-U requires no fine-tuning or architectural modification – it plugs directly into off-the-shelf VLMs at inference time. Inspired by FramePack from the generation literature, we propose a content-aware compression strategy for visual understanding under a fixed token budget. Specifically, FramePack-U (1) scores each frame by its semantic similarity to the text prompt and ranks frames in a diversity-driven manner, and (2) packs the selected frames at varying resolutions into a fixed-length context window. This concentrates the memory budget on frames that are both instruction-relevant and visually diverse. While FramePack-U can in theory pack arbitrarily many frames into the fixed budget, we identify an empirical instantiation that generalizes robustly across settings. Extensive experiments on LongVideoBench, MLVU, and Video-MME show that FramePack-U consistently outperforms full-resolution baselines and prior token reduction and keyframe selection methods across diverse model families, parameter scales, and frame budgets. Notably, on Qwen3.5-35B-A3B with a 128-frame budget, FramePack-U yields gains of +12.9%, +10.6%, and +9.7% over the official model on the three benchmarks, respectively.

LLM-based multi-agent collaboration is increasingly used to solve complex tasks by dividing work among specialized agents and integrating complementary expertise. Yet existing systems and evaluations largely rest on an implicit perfect-colleague assumption: collaborators provide correct and complete information and respond promptly to requests. In realistic teamwork, however, collaborators may return information that does not answer the intended question, provide insufficient detail to support a decision, or fail to respond to a request. To separate task-solving competence from the ability to recover an unreliable evidence-gathering process, we introduce the Unreliable Colleagues Benchmark (UnCoBench), which organizes behavioral failures through Misaligned, Incomplete, and Missed settings. The main agent need to verify contextual correctness, diagnose information gap and track unresolved dependencies in order to obtain the correct evidence for solving the task. We also introduce a verifiable and scalable automated generation pipeline that pairs LLM-based scenario generation with executable checks of answer uniqueness and role necessity. Results across ten models show that similar reliable-setting performance can result in different robustness to unreliable evidence. Trace analysis reveals frequent omissions of corrective follow-up, and the recovery behavior involves inefficient waiting and redundant queries. These findings show that robust collaboration requires not only sound reasoning, but also targeted action to diagnose evidence failures, recover relevant information, and integrate it into decisions. Our data and source code are available at https://github.com/iNLP-Lab/UnCoBench.

Ultra-short-term forecasting of densely distributed photovoltaic (PV) systems remains challenging because external meteorological and satellite products may not adequately resolve rapid local variations, while heterogeneous PV systems exhibit different generation responses under similar atmospheric conditions. Meanwhile, dense PV fleets provide spatially distributed measurements that reveal regional generation dynamics. Here, we reformulate fleet-level PV forecasting from modeling discrete station time series to predicting explicit, spatially structured regional PV generation fields. We develop SatPV-Field, a regional-to-local forecasting framework that transforms distributed PV measurements into image-like generation fields and models their spatiotemporal evolution. By combining historical PV fields with satellite-derived future shortwave radiation guidance and meteorological forecasts, the framework predicts future regional generation states, which subsequently provide location-specific guidance for individual-station forecasting. Experiments involving 496 distributed PV systems, including 446 source stations and 50 held-out target stations, demonstrate that these regional fields can be effectively forecast and used to improve station-level predictions. The proposed approach reduces station-level MAE by approximately 10.0% relative to the best-performing single-station temporal baseline and by 2.5% relative to a graph-based regional forecasting approach using the same source fleet. These findings establish an explicit spatial-field modeling paradigm that connects regional generation dynamics with individual-site forecasting.

Maintaining up-to-date clinical practice guidelines (CPGs) is essential for evidence-based medicine, but rapid evidence growth and complex expert consensus formation limit timely updates. We developed GUIDE, a multi-agent copilot system for human-AI collaborative guideline generation and updating. GUIDE delegates labor-intensive tasks to specialized AI agents while preserving expert judgment and final arbitration. In 12 NICE ophthalmology datasets, GUIDE achieved 88.85% literature recall, 86.06% alignment in certainty-of-evidence ratings, and 91.67% recommendation agreement, requiring 162.17 minutes per dataset on average (28.68% human intervention). Performance remained stable across 20 additional datasets from cardiology, endocrinology, gastroenterology, and respiratory medicine. Ablation analysis showed that removing expert intervention reduced performance in judgment-intensive steps such as Risk-of-Bias assessment. In six post-cutoff datasets, GUIDE maintained robust end-to-end performance on unseen evidence. Prospective real-world validation showed that GUIDE reduced expert workload and produced more clinically acceptable recommendations than the original decade-old guideline (mean score, 3.73 vs. 1.53; agreement, 66.7% vs. 13.3%), with outcomes comparable to human-only updating (P = 0.588) but substantially less time and effort. GUIDE offers a practical framework for timely, trustworthy, expert-centric guideline development and validates computationally augmented consensus.