Hao Wang, Max Cohen, Zeyuan Feng, Donggeon David Oh, Oswin So, Songyuan Zhang, Javier Borquez, Chuchu Fan 等 12 位作者
Safety is a central concern in modern autonomous systems. Safety certificates---functions whose existence guarantees that a system meets its safety specification---have long been used as a principled tool for rigorously reasoning about safety. Methods for synthesizing such certificates are scattered across research areas, with little guidance on when to use a particular approach. This tutorial provides a holistic treatment of safety certificate synthesis centered on two widely used families: Hamilton--Jacobi (HJ) reachability analysis and control barrier functions (CBFs). For each family of methods, we first introduce its general formulation, along with classical and learning-based tools for certificate synthesis. Next, we introduce two benchmark problems consisting of vehicle systems of differing complexity to assess the certificate quality of different synthesis methods and their effectiveness in safety-critical control. We use these benchmarks to compare various methods and provide practical guidance for selecting an approach.
Russell Yang, Ruishi Chen, Pierce Kelaita, Riya Ranjan, Sibo Ma, Charles Dickens, Matthew Guillod, Megan Ma 等 9 位作者
Two methodologies dominate current practices of benchmarking: rubric-based scoring evaluates items against predefined criteria, whereas comparative judgment elicits pairwise preferences between outputs. Although both methodologies are widely used, the choice between them is rarely justified. We release JudgmentBench, a benchmark of 30 real-world legal tasks, paired with 1,539 rubric scores and 1,530 pairwise preference judgments collected from practicing attorneys--including at major U.S. law firms--with substantial experience. The annotations constitute the first publicly available dataset in a high-expertise domain in which both supervision signals are elicited from the same pool of experts on the same tasks. Using LLM-generated outputs at three constructed quality levels, we provide an initial empirical comparison: comparative judgments recover the intended quality ordering substantially better than rubrics under both a per-task rank-correlation metric (mean Spearman's rank correlation of 0.908 vs. 0.150, estimated difference = 0.758 [0.494, 1.021]) and a per-judgment pairwise win-rate metric (0.669 vs. 0.542, estimated difference = 0.127 [0.067, 0.186]), while requiring less than half the annotation time. The patterns hold for human annotators and LLM autograders. Beyond this initial comparison, the paired structure of the dataset supports a broader research agenda on how expert judgment should be elicited, aggregated, and used as supervision in domains without verifiable ground truth.
Multimodal large language models (MLLMs) remain prone to visual hallucinations when correct responses depend on subtle visual details. One line of post-training methods constructs visual contrasts through image retrieval, generation, or editing to improve fine-grained perception. However, these contrasts are not built around the model's own hallucinations, making it difficult to ensure that supervision targets its actual perceptual confusions. Another line suppresses hallucinated outputs through response correction or preference learning under fixed image conditions, but leaves the design of supervision signals on the image-input side underexplored. Inspired by how humans correct misconceptions, we introduce DreamRealm, a post-training framework that explores diverse latent hallucinated responses, minimally edits original images to make them valid, and trains the model to assign each response to its supporting image. Our data construction pipeline, RealmForgePipe, mines observed and latent hallucinations from a target model and materializes them through image editing. Using hallucinations mined from Qwen3.5-9B, we construct the Realized Hallucination Dataset (RealmSet). Our training objective, RealmMatch, constructs a likelihood score matrix across image--response combinations and combines supervised learning with bidirectional contrastive losses to reinforce correct matches and suppress mismatches. Compared with the original Qwen3.5-9B model, DreamRealm improves performance across six visual perception and hallucination benchmarks, including gains of 7.26 percentage points in HallusionBench accuracy and 11.34 percentage points in MMVP pair accuracy. Together with ablation studies on data construction and training methods, these results demonstrate that hallucination assignment is a promising paradigm for improving visual perception and mitigating hallucinations.
LLM-enabled customer-service software in financial institutions must choose a downstream service before responding. Classifiers and LLM routers can identify a relevant label or destination, but a composed application must still verify dispatch permission and entity registration. We study this integration problem as boundary-aware pre-response routing. We present RouteBR, a controller in which one model call proposes a primary service action and registered object, then deterministic code checks identifiers and versioned rules before dispatch or a specified non-dispatch outcome. Lower-ranked candidates cannot replace a blocked primary route. We construct 600 privacy-preserving scenarios from recurring confidential-request patterns at a real financial institution; three engineers jointly reviewed the scenarios and labels. On the 200-case CONFIRM MAIN split, RouteBR reaches 97.5% final-path exact match. A matched one-call baseline given the same rules reaches 95.5%, whereas a diagnostic baseline without the rule program obtains 94.0% action accuracy but only 41.0% final-path exact match. Across 166 non-dispatch MAIN and FLIP cases, RouteBR produces no false dispatch. The artifact pairs all 1,960 prediction rows with raw receipts while excluding gold and reviewer fields from model requests. These results support separating model proposals from code-owned, replayable execution. Our implementation is available at https://github.com/YgcCoder/RouteBR.
Low latency is critical to user experience. As NVIDIA Blackwell and Rubin GPUs advance multi-die scaling, data placement and cross-die communication become increasingly important. However, the latency implications and underlying costs of cross-die communication and die-aware execution remain largely unexplored. In this report, we present a comprehensive evaluation of the Locality Domain (LD) feature introduced in CUDA 13.4. Conducted on a single node equipped with eight NVIDIA B300 GPUs, our study examines three aspects: memory management and access, computation, and communication, to identify the sources of latency and, more importantly, explore how LD can mitigate them. Our measurements show that the benefits of LD depend on data access patterns. Through die-aware adaptations and integration with widely used libraries and tools, we achieve speedups of up to 43.5% for NCCL 2.32.3 (AllGather), 111.3% for nvbandwidth (DtoD), and 37% for DeepGEMM (EP). We also identify non-negligible LD-related GPU memory allocation overhead when allocation sizes exceed 32 MB due to barrier and TLB invalidation, we propose a fix to the GPU driver kernel module (submitted to NVIDIA) and show that overhead could be reduced 7x ~ 10x without requiring any user-code changes.
Building age is fundamental for urban infrastructure safety, operation, and planning, as it directly affects structural safety, and sustainable development. Existing methods mostly rely on detailed building information such as street view, function, or materials, which are harder to obtain than geospatial data like satellite imagery and Digital Surface Models (DSMs). However, existing methods based on those geospatial data mainly rely on optical imagery and local texture features, limiting their ability to capture structural information and distinguish buildings from different periods. Moreover, the inherent temporal ordering among building age categories is often overlooked. To address these limitations, we propose the first deep-learning framework for multi-building age estimation using widely available multi-source geospatial data, including remote sensing data and their derived products, as well as building footprint information that estimates age at both pixel and instance levels within a unified pipeline. Specifically, a large pretrained backbone with fine-tuning adapters extracts representative features, and raster-derived and property-derived fusion modules support pixel-wise and instance-wise predictions. Pixel-level estimation employs a pyramid decoder with Cosine-guided Mixture-of-Experts (CMoE) to fuse multi-source raster data across scales, capturing fine-grained building age differences, while instance-level prediction uses attention pooling and Property-Conditional Adaptive Instance Normalization (PCAdaIN) to extract instance-level representations and integrate property attributes. A Temporal Ordinal Constraint Loss (TOCL) enforces the sequential ordering of predicted ages. We also construct a building age benchmark for Hong Kong to enable comprehensive evaluation. The source code and datasets are available at https://github.com/littlebeen/RS-buildingAge.git.
Guanren Qiao, Ruixiang Ouyang, Sheng Xu, Yueci Deng, Yunxin Tai, Kui Jia, Guiliang Liu
Contact-rich manipulation requires precise control under complex contact dynamics while remaining robust to diverse visual conditions. However, real-world reinforcement learning often overfits to the visual appearance of training scenes. We propose Semantics-to-Contact (S2C), a stagewise framework for visually robust contact-rich manipulation. The insight is to decompose the problem into two stages: semantic focusing and contact refinement. In the first stage, a vision-language model localizes task-relevant regions across visually diverse scenes, providing coarse spatial guidance and reducing the burden of exploration. In the second stage, a real-world human-in-the-loop residual reinforcement learning policy learns fine-grained contact behaviors within the localized region using force feedback. To further improve visual robustness, we introduce an object-centric visual augmentation strategy that randomizes background appearance while preserving the robot and task-relevant objects. Experiments on six contact-rich real-world manipulation tasks exhibit that S2C consistently improves success rates over strong baselines and maintains reliable performance under significant visual and positional variations. These results demonstrate that stagewise semantic focusing and contact refinement provide a practical path toward visually robust real-world contact-rich manipulation. Our demos can be seen in https://edem-ai.github.io/S2C/.
Fazong Wu, Xin Wang, Ming Yang, Zhenyong Zhang, Xiaoming Wu
As large language models (LLMs) are increasingly deployed as general-purpose conversational assistants, they must handle diverse user requests while reliably refusing harmful ones. Jailbreak attacks craft adversarial instructions or conversational contexts to elicit unsafe responses despite these safeguards. However, existing jailbreak benchmarks often rely on manually designed or collected prompts, making it difficult to systematically extend coverage to new threat conditions. Moreover, existing attack search methods typically optimize for attack success rate (ASR), providing limited insight into how model behavior changes locally between safe refusal and unsafe response. To address these limitations, we present JailBound, a unified jailbreak evaluation framework that combines automated benchmark construction with first-order loss (FOL)-guided dual-branch search. JailBound organizes threats by risk category, application domain, and attack type. Conditioned on this taxonomy, a fine-tuned meta-attack generator synthesizes evaluation prompts across diverse threat conditions. FOL serves as a local sensitivity signal for exploring high-value attack states and behavior-sensitive states. We evaluate JailBound across 46 LLMs from 13 model families under a unified experimental protocol. Under matched settings, the High-Value Branch improves average ASR over a ZOL-only baseline by 5.1 and 4.9 percentage points across the two evaluation settings. Under local perturbations, states from the Safety-Sensitivity Branch exhibit higher behavioral flip rates than those from the High-Value Branch across all tested radii for JailBound and S-Eval, indicating greater local sensitivity of judge-assigned safety labels. Warning: this paper includes examples that may be offensive or harmful.
An experiment can be informative about the parameters of a working model and uninformative about whether that model is wrong. Adaptive design can reinforce this mismatch by repeatedly selecting settings where an incorrect model fits. We bound the power of any valid adequacy test in terms of the experimenter's own sampling distribution, then distinguish information absent from the selected data from information lost by the detector. A computable diagnostic measures unresolved error relative to a declared expansion of the working model. It motivates a training-free rule that reserves part of the design budget for falsification. The empirical record is mixed: the independently specified primary study fails detection eligibility, and a later numerator diagnostic fails its improvement criteria. After the frozen rule fails on fresh continuous-null panels, a disclosed secondary horizon-calibrated revision detects 35.8% of misspecified campaigns versus 31.7% for a class-separation control at T = 240, but has no resolved advantage at T = 480. In a corrected 289-unit measured replay it improves on random design and trails class-discrimination EIG. A conditional futility bound explains why remaining time matters without proving the gate constant or its power. These results support separating parameter learning, model checking and detector performance; they do not establish a universally superior design policy or prospective scientific discovery.
LaCAM is a search-based solver for multi-agent pathfinding (MAPF) that runs a depth-first search over joint configurations, using PIBT to generate successors and lazily adding single-agent constraints to explore alternatives. Although LaCAM is complete, it can struggle on dense, corridor-heavy instances because extracting blocking agents from deadlocks requires coordinated moves of several agents, often away from their goals, which PIBT's greedy one-step successors rarely propose and which LaCAM can only reach by enumerating constraints one agent at a time. This paper proposes Goal-Branching, where blocking agents are given temporary goals, each tried as an additional branch of LaCAM's search. We examine random and corridor-aware rules for choosing these temporary goals and show that completeness is preserved for any such rule. On dense instances (60-90% of free cells occupied) of three maze and room maps, Goal-Branching solves 89% of instances against 44% for LaCAM*, and, on instances both solve, it finds a first solution 10-14x faster on maze-32-32-2 and room-32-32-4, and nearly 200x faster in median time at 75% density on room-32-32-4. Solution cost increases by 15-22% on instances both solve, while on open maps and the full MovingAI benchmark its solutions cost no more than those of LaCAM* and it is at most 3% slower.
Understanding how individual training observations influence model predictions is important for developing reliable and interpretable data-driven transportation systems. Existing influence-function methods, however, are primarily formulated for unconstrained learning and do not explicitly account for physical constraints that couple different components of a transportation network. Under such constraints, perturbing a training observation can affect model parameters and predictions only through physically feasible directions, potentially changing how its influence propagates across the network. This paper develops a hard physics-informed influence-function framework for constrained learning, in which training-data influence is characterized within the physically feasible space defined by network constraints. The analytical results reveal how physical constraints alter the magnitude, direction, and spatial distribution of training-sample influence relative to conventional influence measures. Using traffic prediction under flow-conservation constraints as a case study, we show that influential observations identified by the proposed framework differ systematically from those identified by classical influence functions. Through orthogonal decomposition, we quantitatively characterize how much of a perturbation in the training data is propagated through the network via the flow-conservation constraints and how much is removed by the constraint projection. The proposed influence framework provides a physically consistent approach for evaluating training-data influence in constrained transportation learning problems, with potential applications to data diagnostics, sensor evaluation, and targeted data collection.
Guofeng Zhang, Rong Han, Xiaoyu Wang, Zhiyun Li, Zongbo Han, Xiaohong Liu, Guangyu Wang
Cyclic peptides are emerging as promising molecular scaffolds in drug discovery due to their high binding affinity and structural stability. However, extending generative models from linear to cyclic peptide design remains challenging, as cyclization sharply restricts the feasible design space through coupled geometric and biophysical constraints. Moreover, limited training data has led existing approaches to rely largely on zero-shot generation or post hoc filtering, resulting in low yields of feasible designs and limited control over multi-objective trade-offs. To address these limitations, we propose FAR-DPO (Feasibility-Aware and Robust Direct Preference Optimization), an architecture-agnostic framework that steers generative models toward structurally and biophysically feasible cyclic peptide designs, particularly for challenging targets. FAR-DPO integrates feasibility-aware preference construction with difficulty-aware group-robust optimization. Specifically, it constructs within-target preference pairs through feasibility-gated multi-objective dominance and adaptively reweights predefined difficulty groups according to their current preference losses. Under a fixed generation budget, FAR-DPO increases overall success rate on the CPSea benchmark from 46.9% to 57.8% on PepGLAD and from 48.0% to 49.6% on PepFlow. These gains also extend to the hardest target quartile and are accompanied by more favorable best-per-target binding scores. Together, these results demonstrate FAR-DPO's effectiveness in improving feasibility and target-wise robustness.
Semantic information can enrich the identifiers and structured fields used by recommender systems. Semantic tokenization provides a structured representation of this information by converting LLM-derived user profiles into discrete tokens. In the common two-stage pipeline, the tokenizer is trained before the recommender. Its token assignments are then treated as fixed input features during recommendation training. As a result, the downstream recommender can learn how to use the assigned indices, but it cannot revise the semantic partition that produced them. Making this partition task-aware is difficult because nearest-code assignment is discrete, and unconstrained adaptation may also distort the geometry learned during pretraining. Moreover, a fixed token depth exposes the same semantic resolution to every interaction. We thus propose OptST (Optimizing Semantic Tokenization), which makes a pretrained tokenizer responsive to downstream recommendation. To revise token assignments without directly updating the pretrained encoder or base codebooks, OptST learns a low-rank deformation whose drift is limited by an anchor regularizer. Recommendation supervision reaches this transformation through a straight-through estimator, while prediction retains hard nearest-code assignments. Beyond adapting the assignments themselves, OptST selects an interaction-conditioned prefix to control how much residual detail is exposed to the recommender. Experiments on four public datasets show that OptST achieves the highest AUC among conventional and LLM-enhanced baselines. Its improvements also persist across different recommendation backbones. These results establish task-aware semantic tokenization as an effective interface between LLM-derived profiles and downstream recommendation models.
Yifan Chen, Jijun Lu, Ye Zheng, Kaichen Chi, Zhe Sun, Tao Chen
Underwater image enhancement (UIE) aims to recover clear images from observations degraded by wavelength-dependent absorption, scattering, and spatially nonuniform degradation. Although existing generative methods can handle complex degradations, they may still generate visually appealing results while compromising content preservation, thereby degrading downstream visual perception. To improve content preservation, we extract spatial priors characterizing appearance and photometric properties from ground-truth images to provide stable guidance for enhancement. To bridge the gap between training with ground-truth references and inference without references, we propose reference-guided prior learning for underwater image enhancement (RPL-UIE), a two-stage teacher-student framework. The teacher network extracts stable priors from ground-truth images, while the student network takes only degraded images as input and learns to mimic the teacher's prior extraction capability. We further design Residual Prior Refinement Diffusion (RPRD) and Frequency-Aware Prior Residual Calibration (FPRC) to reduce the prior discrepancy between the teacher and student. RPRD progressively predicts corrections in the residual space, while FPRC preserves stable low-frequency residuals and selectively modulates high-frequency details. Experiments on multiple underwater image datasets show that RPL-UIE achieves competitive enhancement performance. For the engineering application of underwater visual perception, downstream object detection and instance segmentation experiments show that RPL-UIE improves visual quality while better preserving object structures and semantic content needed for reliable perception. Tests on real-world data collected by a custom-built remotely operated vehicle (ROV) further demonstrate its effectiveness and practical applicability.
The advancement of Embodied AI necessitates high-quality simulation assets that faithfully mirror the real world. However, transforming raw visual observations into simulation-ready scenes remains challenging due to the lack of physical grounding and scene-level interactivity in current image-to-URDF methods. We propose NeoWorld-Pro, a framework that reformulates monocular scene reconstruction as procedural programming for interactive 3D environments. Leveraging the zero-shot reasoning and code synthesis capabilities of MLLMs, NeoWorld-Pro converts a single RGB image into executable programs specifying object geometry, articulation, and physical properties. A physics-in-the-loop mechanism then iteratively refines the generated programs by validating their execution in a physics engine, enforcing physically plausible articulations, valid object compositions and interactions, and accurate spatial relationships. Experiments show that NeoWorld-Pro outperforms open-loop and prior monocular reconstruction methods, while enabling complex downstream tasks such as stable stacking and fine-grained manipulation.
The scientific computation methods development in conjunction with artificial intelligence technologies remains a hot research topic. Finding a balance between lightweight and accurate computations is a solid foundation for this direction. The study presents a neural operator based on the dynamic mode decomposition algorithm (DMD), mapping functional spaces, which combines DMD and deep learning (DL) for spatiotemporal processes efficient modeling. Solving PDEs for various initial and boundary conditions requires significant computational resources. The method suggested automatically extracts key modes and system dynamics using them to construct predictions, reducing computational costs compared to traditional numerical methods. The approach has demonstrated its efficiency through comparative analysis of performance with closest analogues DeepONet and FNO in the heat equation, Laplaces equation, and Burgers equation solutions approximation, where it achieves high reconstruction accuracy.
Uncertainty in resource reopening times complicates train holding and rerouting decisions when resource closures coincide with arrival delays and processing-time extensions. We develop a rolling-horizon distributionally robust model that updates predictions of the remaining closure duration as observations of disruption resolution become available. The model jointly optimizes train routes, sequences, and timings. It minimizes the worst-case expected total completion delay over an ambiguity set of scenario probabilities. The solution algorithm, GNN-LBBD, uses logic-based Benders decomposition to coordinate routing and resource scheduling. Conditional cuts are constructed from resource-use patterns, and a graph neural network guides the order in which resource relaxation subproblems are solved. Across 28 test instances for a five-track station, GNN-LBBD achieved a mean total completion delay of 61.13 train-min. This was 16.63\% lower than tabu search and 20.06\% lower than variable neighborhood search, with a mean total computation time close to that of tabu search. Information comparisons on six representative instances showed that using observations of disruption resolution reduced mean total completion delay by approximately 21.7\%. Scenario-based and conditional-mean predictions produced similar overall results under the same information conditions. The benefits of information were more pronounced in some instances with alternative routes. This suggests a relationship between the use of observations, the station resource structure, and the scope for schedule adjustments.
Fan Yang, Tatsuya Kikuzuki, Jun Liang, Jun Liu, Wenkai Li, Ryosuke Kawamura, Ananya Hazarika, Prajit KrisshnaKumar 等 10 位作者
To serve a request in a human environment, a robot system must find an object that satisfies a functional requirement and a robot that can feasibly reach it. Scene graphs record which objects exist and where, but they often omit capabilities documented only in product manuals. Retrieval-augmented generation methods can read such manuals, yet they rarely reason jointly over object capabilities, physical instances, robot states, and dynamic connectivity. We introduce E2W-Bench, the Exists-to-Works Benchmark, which links dynamic hybrid scene graphs with real product manuals to evaluate capability-grounded robot--object assignment under availability, mobility, vertical-reach, and connector constraints. We also present Embodied RAG++, a spatially grounded retrieval framework that scopes retrieval to the manuals linked to candidate objects, caches extracted capabilities per product model, and resolves robot--object feasibility through deterministic spatial reasoning. On E2W-Bench, Embodied RAG++ reaches a pair accuracy of 0.90 with Qwen2.5-32B and 1.00 with Llama3.3-70B, compared with at most 0.30 and 0.55 for the baselines, while requiring the shortest preprocessing time among the retrieval-based methods and tying for the lowest total token count.
Kiruthiga Chandra Shekar, Aliasghar Arab, Jingang Yi
Learned controllers for vehicle maneuvers are usually validated under the conditions and objective used to train them, where they appear highly reliable, yet the parameters that change on real roads are often not those the model was trained on. We argue that validation must be independent of training: it should vary physical parameters that training kept fixed and use acceptance criteria absent from the training objective. A learning-based control policy, G-Maneuver, with 90.11% success is tested for stunt-level maneuvers. A looser terminal-heading score ranked the worst setting as the best. Because real stunt tests are safety-critical, a simulation-to-reality-to-simulation (Sim2Real2Sim) loop uses the simulation to select safe zones for real-vehicle trials with 16.9% acceptance. Automotive safety analysis frameworks are used to identify parameters for validation using coverage tests. We demonstrated that the kinematic stunt-driving indicators, namely heading change, sideslip, and speed-normalized yaw rate, match between simulation and the real world, while force and roll indicators do not, which identifies the indicators that can support safe deployment. A video of the experiments is available at https://youtu.be/dN1Oes8pNIo.
Xuenmin Hu, Qingxue Qian, Wei Luo, Benyi Ma, Bo Tang, Long Chen
Legged robots equipped with manipulators have shown significant potential in the fields of daily life and industry in complex environments. However, existing methods mainly focus on global navigation or manipulation independently, while overlooking the last-mile navigation problem, where the robot must jointly determine the base placement, approach direction, and grasp execution to achieve reliable object grasping, and different manipulation stages require distinct control behaviors. To address these challenges, we propose Last-Mile Grasp-Navigation (LMGN), a unified policy learning model for Quadruped robot loco-manipulation. An Explicit Distance and Directional Reachability (ED-DR) representation is first introduced to explicitly represent the features of grasp-compatible base placement and approach direction. We then develop a semantically aligned Hierarchical Mixture-of-Experts (HMoE) policy that models stage-aware behaviors across the stages of base adjustment, pre-grasp alignment and grasp execution, and object lift. Finally, a new optimization method that jointly aligns navigation and manipulation, stage-dependent expert routing, and action-level consistency is proposed to optimize the HMoE policy. Extensive experiments on the Habitat and Isaac Gym simulators demonstrate that LMGN consistently outperforms state-of-the-art methods, denoting that integrating grasp-aware reachability with stage-aware policy learning significantly improves the success rates of last-mile navigation and target grasp for robots.
Using dehydration signatures as an encoding, we train GPT-2-style transformers to generate triangulations of 3-dimensional manifolds. Our main runs each use approximately 3 exaFLOPs, spanning a 200 million parameter model trained on 60 million sphere triangulations to an 11 million parameter model trained on 1.1 billion sphere triangulations. The optimal model has 10 blocks, width 640 and generates dehydration signatures of closed manifolds of various topological types at a rate of 6.9%, 82% of which are spheres. To investigate how our method scales, we use the technique of IsoFLOP curves on 15-tetrahedron triangulations of the 3-sphere, over a compute range of 6 to 3000 petaFLOPs, and observe neural scaling behaviour. We extract a parameter scaling exponent of $α= 0.49$ ($[0.48, 0.50]$ stat, $\pm 0.02$ syst), and an optimal ratio of 234 tokens per parameter for training. In addition, we present a number of smaller experiments, including evidence that transformers are able to learn to count degree-one edges near-perfectly, but get progressively worse when counting edges with higher degrees.
Most vision-language-action (VLA) policies decouple manipulation from cognition by bolting an external action mechanism onto a vision-language model (VLM)--either quantizing actions into discrete tokens whose errors bear no stable correspondence to physical displacements, or delegating continuous trajectory generation to a separate action expert whose gradients compete with or are blocked from updating the VLM backbone. We argue that manipulation and cognition should instead be embodied in the same model, shifting from a decoupled VLA to an Embodied Vision-Language Model (eVLM). We introduce Concertina, an eVLM that places continuous trajectory representation, blockwise generation, and physical-coordinate supervision directly on the computation and parameter path of a shared diffusion VLM backbone. Each control timestep is encoded as one continuous latent action token in the backbone hidden space, and the shared Transformer reconstructs temporally ordered action blocks via the same corrupt-and-reconstruct operator used for multimodal pretraining. Active action projections account for only 0.21% of parameters in the 4.46B model with zero separate action-generator layers, compared with 16.5%-35.1% dedicated action-expert capacity in representative continuous VLAs. Using our proposed Action Interface Audit, we confirm that cross-time trajectory coupling emerges in the middle and late layers of the shared backbone rather than at the action interfaces. On standard benchmarks, Concertina achieves 98.15% average success on regular LIBERO, 85.9% on the 10,030-task LIBERO-Plus benchmark, and 4.142 average completed chain length on CALVIN ABC-D. Code: https://github.com/lose4578/Concertina Hugging Face: https://huggingface.co/collections/Traly/concertina
Zidi Xiu, David Q. Sun, Kevin Cheng, Maitrik Patel, Josh Date, Yizhe Zhang, Jiarui Lu, Omar Attia 等 12 位作者
Next-generation AI must manage vast personal data, diverse tools, and multi-step reasoning, yet most benchmarks remain context-free and single-turn. We present ASTRA-bench (Assistant Skills in Tool-use, Reasoning \& Action-planning), a benchmark that uniquely unifies time-evolving personal context with an interactive toolbox and complex user intents. Our event-driven pipeline generates 2,413 scenarios across four protagonists, grounded in longitudinal life events and annotated by referential, functional, and informational complexity. Evaluation of state-of-the-art models (e.g., Claude-4.5-Opus, DeepSeek-V3.2) reveals significant performance degradation under high-complexity conditions, with argument generation emerging as the primary bottleneck. These findings expose critical limitations in current agents' ability to ground reasoning within messy personal context and orchestrate reliable multi-step plans. We release ASTRA-bench with a full execution environment and evaluation scripts to provide a diagnostic testbed for developing truly context-aware AI assistants.
Eshan Chattopadhyay, Noam Ringach, Nicholas Spooner
Beginning with the work of Dinur, Lin, and Vidick (FOCS, 2024), tensor codes with constant product expansion have been foundational to recent advances in quantum locally testable codes (qLTCs) based on cubical complexes. Informally, product expansion says that any low-weight parity check of the tensor code can be written as the sum of few parity checks of each component code in each dimension. Recent work of Kalachev and Panteleev (FOCS, 2025) showed that tensors of random linear codes have constant product expansion. In this work we initiate the study of random linear codes with prescribed automorphism groups. For every sufficiently large prime $k$, we construct pairs of rate $1/2$ MDS codes of length $2k$, over finite fields whose size depends on $k$, that have constant two-sided product expansion and are invariant under the simultaneous cyclic shift of two cycles of length $k$. This shift is the local action in the cube complexes of Rungtanapirom, Stix, and Vdovina (Internat. J. Algebra Comput., 2019). Along the way we give a proof of the existence of two-sided product expanding codes which, unlike the earlier result of [KP25], does not rely on explicit constructions of asymptotically optimal locally testable codes. Our proof relies on several new techniques that we view as interesting in their own right. Instead of working directly over finite fields, we work over the reals and use random Rademacher matrices for the generator matrices of the component codes. From here, we show how to relate product expansion to the operator norm of sparse restrictions of carefully chosen Gram matrices, which we bound using tools from random matrix theory. More generally, we propose a method for studying properties of random linear codes by lifting to random codes over $\mathbb{R}$ and $\mathbb{C}$, which we anticipate will be applicable more broadly to existential problems in coding theory.
Decision-based sparse attacks must reduce perturbation support using only the victim's predicted label. The best restoration strategy varies as support contracts: coarse removal works early, while tail states need finer, history-aware tests. We propose RLGPA, an online reinforcement learning method that schedules support-reduction executors and their granularity under a hard query budget. A masked Dueling Double DQN selects among seven stable policy--scale actions; each action produces support-relative restoration groups that are rebuilt after every macro decision. A feasibility-preserving executor commits only adversarial restorations and feeds support gain and query cost back to the planner. Under the matched top-1-label oracle, RLGPA reaches 61\% targeted ASR at sparsity 0.01 on CIFAR-10 (vs. 25\% SparseEvo, 0\% Pointwise), and 75.5\%/73.0\% on ResNet50/ViT at sparsity 0.005 (vs. 29.0\%/15.5\%). A fixed-probe test reduces median \(L_0\) from 165 to 23 after 40 images without any update. State-blind and state-shuffled controls degrade significantly, confirming that the learned policy uses attack state; the main gains come from state-conditioned executor scheduling and dynamic repartition. The source code is publicly available at https://github.com/Acui001/RLGPA.git.
The rapid development of large language models (LLMs) for code, which closely aligns with the practical needs of real-world software development, has sparked growing attention in repository-level code completion. However, existing Code LLMs present challenges in focusing on suitable contexts from the repository and in deeply reasoning about cross-file dependencies. To address these challenges, we propose a novel reinforcement learning framework for repository-level code completion. To better understand both in-file and cross-file contexts, we employ identifier-driven intent recognition to capture completion intent, thereby improving the model's performance in real-world scenarios. To enhance cross-file reasoning, we propose a reward-driven completion learning to effectively reward in complex repository completion scenarios. To guide LLMs in completing code that aligns with intended functionality and repository context, we introduce a selective-exploration strategy that directs the model to focus on low-confidence, high-reward tokens, promoting the exploration of valuable, underexplored completion patterns. Experimental results show that our approach significantly improves the performance of Code LLMs. The code and the dataset are available at https://github.com/zhenlongDai/RepoCompletion.
Shaopeng Liang, Biao Zhang, Hongzhen Ye, Zeyang Hou, Ma Jiang, Jintao Tu, yunxi. hz, Zhijia Fang 等 11 位作者
Generative retrieval predicts semantic identifiers for candidate entities. In local retail and food delivery, candidates must also be serviceable for a particular user at a particular time. Location, weather, and traffic reshape merchant and product supply, so a catalog-valid identifier may be infeasible for the current request. Post-generation filtering removes such candidates only after beam capacity and inference computation have been spent. We study retrieval over an explicitly materialized, request-time supply snapshot, where short output sequences leave little opportunity to amortize online constraint preparation. Our RTP-LLM pipeline compiles eligible residual-quantization identifiers into a compressed sparse row prefix index, overlaps CPU compilation with model prefill, and shares the uploaded index across beams. Legal expansions retain original-model log probabilities, and constraint states follow selected beam parents. We formalize snapshot-relative identifier validity and distinguish it from entity-level eligibility and freshness. The implementation generates three-token merchant identifiers. Historical evidence covers T+1 daily-refresh catalog constraints and subsequent request-scoped serving evaluations. In one matched-beam daily-constraint case, filtered merchants decrease from 170 of 299 candidates to 29 of 299; residual distance filtering motivates request-specific constraints. A different-host comparison with 30,000 supplied entries reports a 0.6 ms service-side latency increment, while client measurements expose additional request costs. Captured requests from three geographic settings characterize behavior across supply sizes and output widths. These observations span development stages and measurement boundaries. They establish a practical implementation and identify bottlenecks, but do not establish compute savings at matched retrieval quality.
Vision-Language-Action (VLA) models exhibit remarkable action generation for embodied intelligence, but their heavy compute make deployment on edge platforms impractical. Aggressive, sub-4-bit weight quantization is the natural solution, yet existing post-training quantization (PTQ) methods suffer severe performance degradation in this regime. To address this, we introduce ActQuant, an action-guided mixed-precision PTQ framework that operates in two stages: (1) an inter-tensor bit allocator that assigns each weight matrix a single bit-width based on how much it contributes to predicting the agent's actions; (2) an intra-tensor scale optimizer tunes per-block quantization scales using action-aware curvature, so that dynamic range is concentrated on the weights most influential for control. To deliver the on-device benefits of our aggressive quantization, we further introduce OmniModel.cpp, an agentic conversion pipeline that ports architectures into a native C/C++ runtime with efficient low-bit kernels. We evaluate ActQuant both in simulation and on a real-world 6-DoF UR3 arm, with all models deployed through OmniModel.cpp. On the LIBERO benchmark, ActQuant is the only method that operates at or below 3 bits-per-weight, retaining 95.0% on OpenVLA-OFT and 94.8% on $π_{0.5}$. Pushed further, ActQuant reaches 2.5 bpw at 90.1% on OpenVLA-OFT, compressing the backbone from 14.3 GB to 2.7 GB (5.3$\times$). On the physical UR3 arm, $π_{0.5}$ quantized with ActQuant retains the baseline's success rate while reducing the memory footprint by 2.5$\times$.
Yiping Ji, Mahalakshmi Sabanayagam, Peyman Moghadam, Hemanth Saratchandran, Simon Lucey
Large-scale transformer training and deployment are increasingly constrained by the transfer of activations, gradients, and optimizer states across accelerators. Low-bit quantization offers a natural remedy, but transformer activations are often heavy-tailed and outlier-dominated, making simple quantization highly lossy. We show that this difficulty is not only a property of the quantizer, but also of the architecture. Specifically, residual connections can drive transformer activations away from Gaussianity during training. Using controlled comparisons between residual and residual-free transformers, we demonstrate that this effect leads to substantially higher quantization error and accuracy degradation at low precision in residual models. We explain the phenomenon through an excess kurtosis analysis, showing that residual mixing can amplify non-Gaussianity, whereas dense mixing in residual-free models contracts non-Gaussianity. We then show that residual-free transformers can be made trainable using orthogonal initialization, spectral or second-order optimization, and depth-aware scaling of attention temperature. In language tasks, while there is a small drop in full precision performance, these models retain near-Gaussian activations and exhibit significantly improved robustness to low-bit quantization. Our results identify an accuracy--compressibility trade-off in transformer design and motivate architecture-level approaches to quantization-friendly foundation models.
AI agents that work for an individual over time may also help that individual serve external clients. We define a personal external service agent as an agent authorized and continuously managed by a specific individual to accept client requests, deliver work, and provide ongoing support using that person's knowledge, experience, professional methods, and service boundaries. Such agents could allow individuals to serve more clients with less direct involvement and generate economic value at scale. We propose DelegateEval, a platform-level evaluation framework for agent systems that may combine large language models, reusable skills, protocols such as MCP, and executable tool environments. It covers eight dimensions: ease of creation, service discovery and matching, delivery quality, expertise inheritance, service continuity, learning and improvement, reliability and governance, and efficiency and value. These dimensions assess whether a platform enables people with different backgrounds to turn their expertise into ongoing services that deliver verifiable work, develop capabilities through experience, and remain economically feasible. We present the framework and an experimental design. Complete datasets, executable environments, a unified scoring protocol, and cross-model results remain to be developed and empirically validated.