The canonical $\mathbb{Z}_{2^s}$-linear Hadamard codes are the Gray images of the $\mathbb{Z}_{2^s}$-additive codes. Their rank, the dimension of the binary linear span, was known in closed form only for $s\le 3$, and for $s \ge 4$ only through computer computations for small lengths. We give a method that computes the rank for every $s$. We show that the rank is a binomial polynomial in the type parameters, with coefficients $ρ_s(\boldsymbol{k})$ independent of the length. We compute all the coefficients for $s\le 5$, which yields explicit rank formulas over $\mathbb{Z}_4$ and $\mathbb{Z}_8$ (recovering the known ones), and over $\mathbb{Z}_{16}$ and $\mathbb{Z}_{32}$.
Vision-language-action (VLA) policies implicitly assume that third-person cameras remain fixed at their data-collection poses; once a camera is bumped, remounted, or swapped at deployment, task success degrades precipitously even though the underlying task and robot remain unchanged. We present Homeward, a training-free, plug-and-play observation adapter that canonicalizes observations from an uncalibrated deployment camera back into the training view for frozen VLA policies. Our design stems from a closed-loop diagnostic paradox in simulation: while re-rendering perturbed views into the training camera with oracle geometry recovers much of the lost performance, substituting an off-the-shelf feed-forward 3D foundation model can leave the policy worse off than applying no correction at all. This degradation concentrates on the thin, articulated robot arm, which feed-forward depth prediction tears and scatters - yet the arm's exact 3D state is already known from robot kinematics. Homeward exploits this asymmetry: it estimates what is unknown - camera parameters and static scene depth - from the live RGB frame and a single training-view reference, and renders what is known - the moving robot links - directly from the kinematic model. Evaluated across $\sim$300,000 closed-loop episodes spanning four frozen VLAs, Homeward boosts success across all 1,599 camera variants of LIBERO-Plus from 62.5%, 76.5%, 68.0%, and 3.6% to 77.6%, 95.1%, 91.2%, and 26.4% for OpenVLA-OFT, $π_{0.5}$, $π_0$, and OpenVLA, respectively. Without any scene fine-tuning or deployment camera calibration, Homeward extends to 6-DoF random placements, focal-length changes, and, with per-step pose estimation, continuous intra-episode camera drift, while matching or surpassing a calibrated, scene-trained view-synthesis baseline on three of four policies on LIBERO-Plus.
Scene text recognition is an important and challenging task in computer vision. However, most prior works focus on recognizing pre-defined words, while there are various out-of-vocabulary (OOV) words in real-world applications. In this paper, we propose a novel open-vocabulary text recognition framework, Pseudo-OCR, to recognize OOV words. The key challenge in this task is the lack of OOV training data. To solve this problem, we first propose a pseudo label generation module that leverages character detection and image inpainting to produce substantial pseudo OOV training data from real-world images. Unlike previous synthetic data, our pseudo OOV data contains real characters and backgrounds to simulate real-world applications. Secondly, to reduce noises in pseudo data, we present a semantic checking mechanism to filter semantically meaningful data. Thirdly, we introduce a quality-aware margin loss to boost the training with pseudo data. Our loss includes a margin-based part to enhance the classification ability, and a quality-aware part to penalize low-quality samples in both real and pseudo data. Extensive experiments demonstrate that our approach outperforms the state-of-the-art on eight datasets and achieves the first rank in the ICDAR2022 challenge.
Tracking large human-motion datasets has become the dominant route to general, expressive humanoid whole-body control, but scaling across many motions is expensive. Current pipelines use model-free reinforcement learning, either merging per-motion specialists through lossy distillation, or brute-forcing a single policy with cluster-scale compute. SuperMimic amortizes learning across all motions with one differentiable world model of their shared dynamics, fit from logged transitions of a standard rigid-body simulator. The world model and the tracking policy train together in one loop, and backpropagation-through-time over short imagined rollouts gives the policy analytic gradients. Fast imagined rollouts and analytic gradients make policy learning efficient. In 3 hours on a single GPU, SuperMimic surpasses the return of a state-of-the-art motion tracker baseline trained for 7 days on 128 GPUs. It reduces GPU-hours by more than $1000\times$ and fails on only 25 of 129,514 benchmark motions, versus 179 failures for the baseline. After 9.8 hours on two GPUs, SuperMimic achieves zero failures. From 1 to 264 hours of motion, the world model's prediction error stays stable, and GPU-hours to a target return grow slowly. Our policy transfers to a physical Unitree G1 across diverse and acrobatic motions. SuperMimic removes compute as the bottleneck to scaling humanoid whole-body control, making it practical to measure motion-tracking scaling laws.
Deficits in reactive balance can increase fall risk among older adults and individuals with neuromuscular impairments, motivating the need for interventions that improve balance recovery. Wearable robotic exoskeletons offer a promising approach to addressing this need by augmenting balance control. However, developing controllers that deliver appropriately timed and scaled assistance remains challenging, as evaluating each candidate strategy requires repeated exposure to destabilizing perturbations. We propose a simulation-to-real framework that combines motion imitation and reinforcement learning to develop balance-assistive ankle exoskeleton control. A muscle-driven neuromusculoskeletal model was trained using a small set of motion-capture data from a single subject to reproduce kinematic and kinetic responses to backward-translation perturbations. This model served as a virtual human for training a reinforcement-learning exoskeleton controller. The resulting controller generated bilateral ankle assistance using only inertial sensor measurements, without explicit perturbation detection. We implemented the controller on-board a bilateral ankle exoskeleton and evaluated it in five healthy subjects whose data were not used for training. Assistance increased the perturbation magnitude that subjects could withstand without stepping by 11.8% and reduced peak center-of-mass displacement and velocity by 13.5% and 28.2%, respectively. Individual benefits varied with baseline recovery strategy and body mass. These results indicate that a single virtual human can yield transferable balance assistance, and that more diverse virtual humans could further improve its effectiveness.
Mingyang Lyu, Yinqian Sun, Gongce Cao, Chujie Tang, Yuhao Zhang, Lu Jia, Qian Zhang, Bo Wang 等 9 位作者
原文摘要
Query-based camera-radar detectors choose where and over which frames each object query samples from the query's position or feature. We test whether radar evidence around the query should also set these choices at the 150 m range of MAN TruckScenes, where recall differs across object-level radar-return counts within every range band. EviScope is a minimal test: the current frame's returns form a non-learned bird's-eye evidence field, which each query reads at its decoded centre to scale its radial sampling extent and to weight eight frames. It adds 5,450 parameters (0.009%) to RaCFormer, with no latency increase in our timing. Our module-free recipe exceeds the published TruckScenes mAP (25.9 over seven runs against 24.1), so we measure the module only against its own baseline at matched seeds. The evidence a query reads rises with its object's return count (Spearman 0.51) but captures zero-return objects only in part, and the learned radius tracks it ($ρ\approx-0.91$ in four runs). Over seven four-epoch pairs the module changes NDS by $+0.31\pm0.36$ and mAP by $-0.27\pm0.22$ (mean $\pm$ SE). Against a geometry-only conditioner with the same parameter count, trained at three matched seeds, it is higher by $+1.25\pm0.72$ NDS and $+0.87\pm0.21$ mAP, on mAP in every pair, and the control's mean NDS is near the baseline's. An exploratory far-range gain failed an internally pre-specified held-out test, and large inference-time substitution penalties also occur without query-local evidence: neither learned allocation nor inference-time dependence substitutes for matched end-to-end comparisons.
Pre-trained models provide transferable representations for class-incremental learning (CIL). However, sequentially adapting these representations introduces optimization costs and may disrupt old-class feature structure. This motivates us to explore whether the discriminative potential of these representations can be further exploited by continually accumulating class information while keeping the encoders frozen. To this end, we propose TAP, a new and novel framework for tapping the discriminative potential of frozen representations in class-incremental learning. It accumulates visual and semantic class information in the frozen CLIP embedding space with only lightweight optimization. TAP comprises three modules: class-conditional visual distribution modeling, class-adaptive semantic calibration, and visual-semantic posterior fusion. The first module retains class means and within-class scatters, combining class-specific and shared covariances to stabilize distribution estimation. The second module learns class-specific weights over text templates from current-task images, adapting text representations to downstream classes. The third module combines the visual and semantic posteriors through a weighted product of experts for prediction over all seen classes without retaining images from previous tasks. Across nine datasets under two incremental protocols, TAP achieves state-of-the-art performance, attaining the highest average incremental and final-stage accuracies among the compared methods. The code is provided as an ancillary file.
BraTS-GoAT Task 3 evaluates brain-tumor segmentation across heterogeneous entities using shared ET, TC, and WT regions. We present a 3D ResidualEncoderUNet trained with quality-controlled consensus pseudo-labels, a training-only nested-region objective, and a four-expert bottleneck adapter. Five-fold ResEnc-M/L teachers gener ated predictions for 1,138 unlabeled cases; fusion, hierarchy enforcement, and confidence-and-agreement screening retained 983 for training with 1,351 annotated cases. The final five-fold ResEnc-M entry achieved mean DSC 0.834, mean NSD 0.510, and Acalc = 0.6720 on the validation server. Component effects were modest and region-dependent, with the clearest gains on ET. Source code is available at https://github.com/HeeSeung-Shin/Brats2026-Goat-HS.
Transformer-based language models for code have shown remarkable performance in various software analytics tasks, but their adoption is hindered by high computational costs, slow inference speeds, and substantial environmental impact. Model compression techniques such as pruning, quantization, and knowledge distillation have gained traction in addressing these challenges. However, the impact of these strategies on the robustness of compressed language models for code in adversarial scenarios remains poorly understood. Understanding how these compressed models behave under adversarial attacks is essential for their safe and effective deployment in real-world applications. To bridge this knowledge gap, we conduct a comprehensive evaluation of how common compression strategies affect the adversarial robustness of compressed models. We assess the robustness of compressed versions of three widely used language models for code across three software analytics tasks, using six evaluation metrics and four commonly used classical adversarial attacks. Our findings indicate that compressed models generally maintain comparable performance to their uncompressed counterparts. However, when subjected to adversarial attacks, compressed models exhibit significantly reduced robustness. These results reveal a trade-off between model size reduction and adversarial robustness, underscoring the need for careful consideration when deploying compressed models in security-critical software applications. Our study highlights the need for further research into compression strategies that strike a balance between computational efficiency and adversarial robustness, which is essential for deploying reliable language models for code in real-world software applications.
Bo Li, Taoyong Cui, Bob Zhang, Mengran Li, Chengyang Zhang, Yang Zhang, Pheng Ann Heng, Wanli Ouyang 等 11 位作者
原文摘要
Neural networks encode learned capabilities in their weights, yet extending those capabilities with explicit task knowledge remains a fundamental challenge. Here we show that Weight-Space Programming (WSP), in which executable programs read, construct or modify a network's weights directly, can install new capabilities within trained neural networks. WeightCode, a WSP framework, uses a frozen language model to write executable weight programs that install new computations into trained neural networks using task knowledge and their existing weights, without gradient-based training. Once installed, the new capabilities run independently of the language model. Given a grammar or eight demonstrations, weight programs enable previously unseen command compositions (0% to 99.8% exact match) and a new pickup action (0% to 98.4% success), while preserving the original command and navigation capabilities. On ImageNet, four weight programs achieve net gains of 8.84-17.62 points while preserving clean accuracy, outperforming the best of eight fine-tuning methods under matched examples, checks, and retention constraints. In a Vision Transformer, causal interventions show that installed noise-correction gates drive robustness gains, while their selective activation preserves clean accuracy. Rerun on five independently trained Vision Transformers without further language model calls, ten weight programs retain 83-110% of their source-network gains on average, whereas copying their fixed weight edits yields negligible gains. Six programs written for an already edited network add 5.23-14.50 points of contrast robustness while retaining the previously installed noise gain. Together, these results establish WSP as a new route to acquiring and accumulating capabilities in trained neural networks through reusable weight programs.
Deploying fruit instance segmentation in real agricultural settings requires generalizing across unseen species, orchards, and imaging conditions under strict onboard compute and annotation budgets. We introduce SegAnyFruit (SAF), a data construction and progressive training framework that distills foundation-model priors into lightweight, prompt-free segmenters. SAF combines generic mask pre-training, broad mid-training on automatically annotated web imagery from our curated WebFruits dataset, and human-verified agricultural refinement post-training on PaperFruits, together spanning 733K images and 9.66M whole-fruit masks across 165 fruit types. Across seven benchmarks, zero-shot SAF-RF-DETR-XL reaches 69.06 average mask AP50:95 (+4.76 over SAM 3) while remaining deployable on edge devices. SAF initialization also boosts few-shot to full-shot adaptation across five agricultural domains while mitigating catastrophic forgetting. Finally, we systematically analyze training stages, data scale, model capacity, and generative augmentation, and validate our pipeline's transferability to underwater, cell, and human segmentation. We will release our datasets, code, and SAF models.
Nanxiang Jiang, Zhaoxin Fan, Baisen Wang, Daiheng Gao, Junhang Cheng, Jifeng Guo, Yalan Qin, Yeying Jin 等 11 位作者
原文摘要
Concept erasure serves as a vital safety mechanism for removing unwanted concepts from text-to-image (T2I) models. While extensively studied in U-Net and dual-stream architectures (e.g., Flux), this task remains under-explored in the recent emerging paradigm of single-stream diffusion transformers (e.g., Z-Image). In this new paradigm, text and image tokens are processed as a single unified sequence via shared parameters. Consequently, directly applying prior erasure methods typically leads to generation collapse. To bridge this gap, we introduce Z-Erase, the first concept erasure method tailored for single-stream T2I models. To guarantee stable image generation, Z-Erase first proposes a Stream Disentangled Concept Erasure Framework that decouples updates and enables existing methods on single-stream models. Subsequently, within this framework, we introduce Lagrangian-Guided Adaptive Erasure Modulation, a constrained algorithm that further balances the sensitive erasure-preservation trade-off. Moreover, we provide a rigorous convergence analysis proving that Z-Erase can converge to a Pareto stationary point. Experiments demonstrate that Z-Erase successfully overcomes the generation collapse issue, achieving state-of-the-art performance across a wide range of tasks.
Physical AI requires models that understand spatial relationships and dynamic environments and translate observations into actions under limited onboard compute. We present TuringPAI, a family of models for on-device Physical AI comprising TuringPAI-VL for shared vision-language understanding, TuringPAI-Drive for autonomous driving, and TuringPAI-Embodied for robotic manipulation. TuringPAI-VL is a 0.65B model built on HybridViT-MLQ, which combines convolutional local aggregation with Transformer-based global modeling. Multi-scale Latent Queries integrate spatial features and observation history within the same visual Transformer, providing a compact, structured interface to the language model. Training combines general vision-language data with geometric, cross-view, and temporal supervision. TuringPAI-VL supports general multimodal understanding and improves quantitative geometric perception over Qwen2-VL-2B, while providing a shared initialization for the two downstream VLA models. For autonomous driving, TuringPAI-Drive maintains all evaluated ADE/FDE metrics within 1.2\% of a full-token HybridViT baseline using the MLQ interface and achieves comparable aggregate scores in an in-house simulation evaluation. Using XPeng Native Runtime (XNRT) on a single NVIDIA Orin-X SoC, TuringPAI-Drive generates trajectories at 12\,Hz in on-vehicle operation. For robotic manipulation, TuringPAI-Embodied achieves a 96.8\% average success rate across the four LIBERO suites. These results demonstrate a shared vision-language foundation for driving and robotic manipulation under practical on-device constraints.
Hongbin Ye, Haixu Liu, Shenwei Kang, Wenxiu Wei, Tinglan Wang, Hongdong Zhu, Fan Tang, Qi Gao 等 16 位作者
原文摘要
Existing 3D Gaussian Splatting (3DGS) pruning methods primarily rely on global Top-K selection based on independent importance scores, lacking regional budget control and ignoring joint deletion effects. Consequently, they cannot ensure balanced reconstruction quality across a scene. We propose SplatQ, a combinatorial pruning algorithm that coordinates regional budget allocation and joint Gaussian selection under a fixed budget. SplatQ partitions Gaussians by spatial proximity and appearance similarity into size-limited blocks, decomposing global selection into subproblems of a specific size. It then smooths importance-based block retention quotas to prevent excessive budget concentration. Finally, it constructs a quadratic approximation of joint deletion effects using the responses of representative pixel colors to changes in Gaussian opacity. Individual responses, pairwise interactions, and quota penalties are incorporated into a quadratic unconstrained binary optimization (QUBO) objective, solved using simulated annealing (SA) or a coherent Ising machine (CIM) to jointly select Gaussians. Experiments on 5 benchmarks demonstrate that SplatQ outperforms state-of-the-art pruning methods in aggregate reconstruction quality, spatial quality balance, and rendering speed. Ablations confirm that budget smoothing improves local detail reconstruction, while interaction modeling improves Gaussian selection quality under a fixed budget.
Training vision-language-action (VLA) models to follow gameplay instructions across diverse games is limited by the cost of collecting demonstrations. We introduce a framework that automatically generates game environments and expert demonstrations. An LLM-powered agent uses game-building tools to construct executable environments and tasks targeting specified skills. Programmatic experts then use privileged simulator state to generate aligned instructions, rendered observations, and keyboard and mouse actions. We train our game policy, Script2Play, by post-training a 4B vision-language model with behavior cloning on 150.3 hours of synthetic gameplay in Minecraft and Roblox, without human gameplay demonstrations or reinforcement learning. Evaluation on multiple public game benchmarks demonstrates that Script2Play follows gameplay instructions effectively and generalizes across game engines, highlighting the effectiveness of targeted synthetic supervision.
The rapid expansion of the Internet of Things (IoT) has introduced complex and evolving cybersecurity challenges. This paper presents a Large Language Model (LLM)-based framework for intelligent threat detection and prevention in IoT networks. The proposed architecture integrates two core components: a Detection Component that leverages fine-tuned models trained on IoT-specific datasets (IoT-23 and TON_IoT) to identify anomalous network behaviours, and a Prevention Component that generates adaptive response strategies to mitigate detected threats. The framework incorporates prompt management, local adaptation, and global refinement modules to support context-aware decision-making and continuous model-guided adaptation. Experimental validation in a Docker-based edge-cloud environment shows that BERT Small achieves 99.75% accuracy, F1-score, precision, and recall for threat detection, while the Prevention LLM achieves 98.3% accuracy in selecting suitable mitigation actions. These results highlight the potential of LLM-driven architectures to provide explainable and adaptive security solutions for next-generation IoT ecosystems.
Younghwan Joo, Jeonghoon Han, Sanghyun Oh, Soosik Bang, Sung-il Kim
原文摘要
Physics-based digital twins of air conditioning, a principal energy consumer in manufacturing and buildings, can be generated automatically from a declarative specification of its equipment, connections, and sizing. Large language models (LLMs) now write such specifications from plain-language descriptions. Evaluations of these specifications stop at parse success, completed runs, or textual similarity, and cannot reject a model that converges in simulation but is physically wrong. In this study, a four-level protocol is proposed that tests whether an LLM-written specification not only executes but also yields a physically valid digital twin. Each specification is executed through static validity, model build, simulation convergence, and behavioral equivalence with a hidden reference model within bands calibrated by $\pm$20% sizing perturbations. The protocol was applied to eight LLMs and fourteen HVAC topologies, five describing an existing low-humidity test chamber. To test for authorship bias, engineers and a second model family also described six topologies. Within three repairs, seven of eight models built in 98.6 to 100% of trials, with simulations converging in at least 80.0%, but reached behavioral equivalence in only 50.0 to 98.6%. About half of the failed trials converged, but not to the reference steady state, mostly in desiccant rotor systems. Pooled pass rates varied by under five percentage points across description authors. These results imply that the physics hardest to model in each field, here the desiccant rotor, must enter the generation framework as validation levels. The protocol can be adapted to other fields that generate executable digital twins from declarative specifications.
Zihao Dan, Jinyu Zhang, Jinzhao Lian, Yun Chen, Shinan Liu
原文摘要
Large language models (LLMs) are increasingly deployed as agents that translate user requests into sequences of local tool invocations. While prior cache-based side-channel attacks target the locally deployed models, they typically require the adversary to share model-related components or hardware resources with the victim's inference process. These assumptions do not apply to the predominant agentic settings, in which commercial LLMs execute on vendor-controlled infrastructure. We show that agentic execution exposes a more practical attack surface: although inference remains remote, tools are executed on the user's machine and commonly rely on shared system libraries. In this paper, we first systematically evaluate Flush+Reload and Prime+Probe for monitoring agentic tool invocations. Our evaluation across three hardware platforms shows that Flush+Reload enables accurate, real-time tool monitoring, whereas Prime+Probe, despite supporting cross-VM attacks, incurs prohibitive monitoring overhead and suffers from cache-set aliasing. We present ToolPulse, a cache-based side-channel eavesdropping framework against tool-using agents. We first design the Tool Recovery Attack to reconstruct an agent's tool-execution workflow. We evaluate the Tool Recovery Attack on 11 LLMs across Codex and Claude Code. It consistently reconstructs tool workflows, with the lowest F1 score exceeding 93%. We further demonstrate two attack scenarios enabled by the recovered traces. Task Profiling Attack infers user intent across two Terminal-Bench task categories and the Trigger Localization Attack identifies sensitive tool calls in encryption/decryption and model training/inference workloads. Even under substantial system noise on a server shared by approximately 20 users, both Task Profiling Attack and Trigger Localization Attack maintain high accuracy.
We consider the probabilistic nonadaptive group testing problem, in which a uniformly random defective set of size $d$ is to be identified among $N$ items using a binary test matrix. For the explicit Kautz--Singleton construction, it is known that order-optimal scaling $T=Θ(d\log N)$ for high-probability recovery can be achieved in the regime $d=Ω(\log^2 N)$. However, the leading constant in this scaling remains uncharacterized even for standard low-complexity decoders. In this work, we revisit the probabilistic analysis of this construction and improve the leading constant for the COMP decoder, as well as establishing the same bound for the DD decoder. Our approach relies on a sharper analysis of the number of distinct roots associated with the defective items, yielding tighter bounds on the decoding error probability. In particular, we show that both COMP and DD achieve vanishing average error probability with $T\le (1/\ln 2+\varepsilon)d\log N$ tests for any $\varepsilon>0$, in the admissible regime $N=q^k$ for some prime power $q = Θ(d)$ and integer $k \ge 2$. For COMP, this leading constant matches that of the random near-constant column weight design, whereas for DD, a gap remains except in the sparse regime $d = N^{o(1)}$.
Vision-language models (VLMs) play a pivotal role in enabling autonomous vehicles and robots to perceive and reason about the physical world. Edge computers in autonomous systems typically serve concurrent VLM inference requests with various deadlines, each mixing a compute-bound encoder and prefill stage with a memory-bound decode stage. GPU serving optimizes for high throughput in datacenters, with per-iteration host scheduling, which can cause an urgent request to wait for a long time, and small-batch decode costs much energy per token. Prior accelerators on Versal AI Engine (AIE) SoCs map one model onto a fixed AIE dataflow graph for one input stream, without decode or request scheduling, which limits their capability in serving multiple requests. We present HETEROVLM, which is optimized for multi-request VLM serving on a resource-constrained heterogeneous platform. It keeps the AIE graph static and makes it schedulable. The AIE array runs the encoder, prefill and attention, and the programmable logic (PL) implements GEMV engines for small decode steps and a scheduler that orders short work units by class and deadline. Concurrent requests share one decode step that reads the weights once. An urgent request preempts at a work-unit boundary, and the paused request resumes from its results in DDR memory. On an AMD Versal VC1902, text and camera questions get their first tokens in 140 and 303 ms. HETEROVLM decodes a single request 1.3x and 4.0x faster than NVIDIA Tesla T4 and Jetson Orin Nano Super GPUs, consuming only 61% and 73% of their energy per token, or 51% of the energy per token of an NVIDIA GeForce RTX 3090 Ti. A single bitstream runs multiple VLMs, switching between them at run time. A demo is available at https://anonymous.4open.science/r/demo heterovlm-EFM2/.
Latent world-model planning has shown promising results in goal-conditioned robotic manipulation, but remains challenging for long-horizon planning due to limited temporal guidance and inefficient action search. A distant goal image offers little guidance for preparatory actions, capturing intermediate subgoal images by hand is impractical, and uninformed sampling makes action search inefficient. We propose PILOT (Policy-Initialized Local Optimization with Temporal subgoals), which brings two foundation models into latent world-model planning: a video model turns a language instruction into visual subgoals, and a vision-language-action policy initializes the cross-entropy method (CEM). CEM refines action sequences toward each subgoal with an action-conditioned world model, and no goal image is needed at test time. With both models fine-tuned on 40 demonstrations, PILOT achieves 68.8% success on four real UR5e tasks, compared with 55.0% without policy initialization, 13.8% for policy-only execution, and 0.0% for direct final-goal planning. In simulation, PILOT exceeds direct execution of the same policy by 30 percentage points on long Push-T tasks and by 22 percentage points on LIBERO-10. Subgoals and the policy prior are complementary: better subgoals and stronger priors each raise success, and even a weak policy that succeeds alone on only 3.3% of medium Push-T tasks is an effective prior.
Madeleine Nouri, Gelareh Mohammadi, Francisco Cruz
原文摘要
Hierarchical reinforcement learning enables sequential tasks to be represented as compositions of separately trained skills. This paper evaluates demonstration-free hierarchical pipelines for MiniGrid navigation and robotic manipulation, comparing their training requirements and task performance with flat reinforcement learning baselines. The evaluation concerns complete pipelines with task-specific observation, reward and control structures. In navigation, two parameterised skills are reused across unseen FourRooms layouts under manually specified high-level plans. In manipulation, reaching, grasp alignment and lifting are trained separately, with downstream skills initialised from states collected during predecessor rollouts. Across unseen navigation layouts, the hierarchical pipeline substantially outperforms the evaluated flat DQN baselines while requiring fewer training interactions. In manipulation, the hierarchical chain requires fewer training steps than flat SAC and exhibits lower observed variation across seeds. The frozen Pick and Lift skills are further reused for stacking by training an additional Place skill.
Vision-language models (VLMs) reason about long videos over hundreds to thousands of tokens. Every decoding step re-reads the key-value (KV) cache of the entire video. Must the model keep looking while it reasons? In LookAway, the model first writes timestamped notes on what the question needs and then puts the video away. Once the notes close, the inference engine hides the video's KV cache for the rest of the response. Across both hybrid and pure-attention VLMs, dependence on the video is concentrated early in the response. On Qwen3.6-27B, writing notes preserves accuracy better than free reasoning at the same withdrawal step. Replacing them with another video's notes harms accuracy, whereas keeping the video visible alongside the model's own notes adds no significant benefit. Without additional training, LookAway nearly matches full-video thinking accuracy at 0.12-0.13 times its modeled decode cost. To shorten this initial look, we fine-tune the model on its own correct LookAway responses. Reinforcement learning then rewards correct answers while penalizing KV reads. The trained model puts the video away after a median of 66-68 tokens. Across three held-out benchmarks, it writes more tokens than an identically trained full-video model yet reads only 27-35% as many KV positions at matched accuracy.
Pengyu Li, Zhitao Gao, Lingling Zhang, Wenxuan Huang, YiCheng Xiao, Chen Zheng, Jun Liu
原文摘要
Streaming video models must answer questions about a live video under strict memory and latency limits. Feeding only the latest few frames keeps present perception sharp, whereas adding past frames dilutes attention on the current scene and triggers hallucinations. Recent cascade methods retrieve past frames when uncertain, but they still fail when current frames are blurry or when decisive future clues have not appeared yet. We introduce ActiveForager, a streaming video agent that maintains a clean short working window and gathers extra evidence only on demand. For each query, the agent either answers directly, zooms into high-resolution recent frames, recalls past events from memory, or waits for upcoming frames. When recalling history, a shared-weight sub-agent verifies candidate clips in isolation and returns a compact evidence summary. A stream-decoupled attention mask then keeps historical evidence from interfering with present visual tokens. We train this policy using counterfactual information gain, which rewards a tool call only when the gathered evidence makes the true answer more likely and prevents tool collapse across all actions. Across Qwen2.5-VL-7B and Qwen3-VL-8B on OVO-Bench, ActiveForager improves real-time perception, historical recall, and proactive response at once, lifting overall accuracy on Qwen3-VL-8B from 47.6% to 65.9%. It also doubles zero-shot accuracy on the multi-hop Video-Holmes benchmark while adding minimal visual tokens per query.
Junzhe Dang, Mengmeng Zhang, Yuxiang Zhang, Wei Li
原文摘要
We study how to obtain a task-specific Vision-Language Model (VLM) for language-guided path generation between start and end points on remote sensing images. We explore this question from three aspects. First, we implement three widely used architectural approaches and compare their performance. Second, we design a scalable method for constructing pre-training data and build an effective pre-training dataset with a degree of diversity. Methods trained on this dataset achieve promising zero-shot performance. Third, we construct three benchmarks that support stratified evaluation and design three metrics for path generation. We investigate this problem through extensive experiments. The results show that the zero-shot performance of methods trained on the constructed pre-training dataset can exceed that of some advanced commercial VLMs. Code and Dataset can be available at https://github.com/jzdang/RemoteVLPG
Tianwei Ye, Yifan Mao, Minwen Liao, Jian Liu, Chunchao Guo, Dazhao Du, Quanxin Shou, Fangqi Zhu 等 9 位作者
原文摘要
Embodied AI and robotic systems increasingly depend on scalable, diverse, and physically grounded 3D content for simulation-based training and real-world deployment. While 3D generative modeling has advanced rapidly, embodied applications impose requirements far beyond visual realism: generated objects must carry kinematic structure and material properties, scenes must support interaction and task execution, and the resulting content must bridge the gap between simulation and reality. This survey reviews 3D generation for embodied AI and organizes the literature around three roles that 3D generation plays in embodied systems. In Data Generator, 3D generation produces simulation-ready objects and assets, including articulated, physically grounded, and deformable content for downstream interaction; in Simulation Environments, it constructs interactive and task-oriented worlds, spanning structure-aware, controllable, and agentic scene generation; and in Sim2Real Bridge, it supports digital twin reconstruction, data augmentation, and synthetic demonstrations for downstream robot learning and real-world transfer. We also show that the field is shifting from visual realism toward interaction readiness, and we identify the main bottlenecks, including limited physical annotations, the gap between geometric quality and physical validity, fragmented evaluation, and the persistent sim-to-real divide, that must be addressed for 3D generation to become a dependable foundation for embodied intelligence. Our project page is at https://3dgen4robot.github.io.
Post-training text-to-image generators with reinforcement learning usually relies on external reward models that need annotated preferences or auxiliary evaluators and are prone to reward hacking. We show that the generator's own denoising accuracy, used by a recent intrinsic reward, is a poor proxy for prompt fidelity. Within a same-prompt group, this absolute self-confidence is dominated by image complexity and ranks correct samples above incorrect ones near chance. We introduce CANDOR (Contrastive Alignment from Null-referenced Denoising lOg-Ratio), the log-ratio of unconditional to conditional denoising errors under shared noise probes, which measures how much the prompt helps the model denoise its own sample. It cancels prompt-independent difficulty that scales both errors and ranks the better of two same-prompt samples higher in 60% and 63% of pairs on compositional generation and text rendering. As the only Flow-GRPO reward for a flow-matching model, CANDOR improves text rendering, aesthetic and human-preference scores over the base model and a prior absolute self-confidence reward, with text-rendering and preference gains that stronger classifier-free guidance does not reproduce. On public Flow-GRPO adapters trained on external rewards, CANDOR further raises PickScore and ImageReward, so it can complement those rewards. Code is available at https://github.com/jhlee0619/candor.
We study the problem of training image diffusion models while preserving pixel-exact reconstruction. Prior works typically address this problem by directly modeling images in pixel space, requiring modifications to the diffusion training objective or model architecture. In this work, we introduce a different paradigm: rather than modifying diffusion models, we construct a lossless transformation that maps pixels into a representation more amenable to diffusion modeling. We term our approach the Lossless Latent Diffusion Model (LLDM). We construct LLDM using residual latent transformation and uneven latent scaling, and explain its effectiveness through a proposed metric, semantic rank priority (SRP). Empirically, LLDM substantially outperforms prior pixel-space diffusion models in generation quality and even surpasses several widely used lossy VAEs. Furthermore, we show that SRP is strongly correlated with downstream generative performance.
Hao Zhang, Yumin Huang, Linfeng Tang, Junjun Jiang, Jiayi Ma
原文摘要
Image fusion often lacks ground-truth fused images, forcing existing deep methods to indirectly define fusion objectives by requiring the fused image to simultaneously resemble multiple source images. However, the distinct imaging characteristics of different modalities can impose competing constraints on the fused result, potentially limiting the effective integration of their complementary information. To address this long-standing challenge facing the image fusion community, we propose ApexFuse, a novel generative image fusion paradigm that moves beyond source-resemblance constraints toward learning a general distribution of high-quality scene representations. Specifically, we develop a human-LMM-in-the-loop pipeline that leverages large multimodal generative models and iterative human feedback to translate abstract fusion objectives into reliable visual references, thereby defining the target distribution of ideal fusion outcomes. Through target-distribution-guided adaptation, we transfer the rich generative priors of pretrained FLUX to image fusion, enabling high-quality scene reconstruction conditioned on multimodal observations. Extensive experiments demonstrate that ApexFuse harmoniously integrates salient thermal targets, natural colors, and fine textures, achieving superior visual quality as well as semantic segmentation and object detection performance over previous source-resemblance methods. Moreover, ApexFuse exhibits remarkable cross-task generalization, extending from infrared-visible fusion to multi-exposure fusion, multi-focus fusion, and visible-depth fusion without additional fine-tuning.
Latent crystal generative models typically combine a variational autoencoder (VAE) for learning atom-wise representations with a flow model that transports a simple source distribution to the learned latent distribution. However, previous studies have mainly focused on the quality of generated crystals, leaving the structure and generative utility of the atom-wise latent space largely unexplored. In this work, we show that even a vanilla crystal VAE implicitly organizes atom-wise latent representations according to elemental families, with different families occupying largely separated regions of the latent space. We term this phenomenon emergent chemical semantics. This structure suggests that sampling from a family-specific latent region can bias the decoder toward atoms from the corresponding elemental family. Building on this insight, we propose FamilySpace, a framework for controlling the elemental-family composition of generated crystals. FamilySpace assigns each elemental family a distinct source component, forming a family-indexed multi-source distribution, and trains a flow model using source-target pairs matched by family. In inference, a user-specified family template selects the source component for each atom-wise representation, enabling direct control over the family composition of generated crystals. Experiments on MP-20 and MPTS-52 show that FamilySpace achieves competitive crystal-generation performance while matching prescribed elemental-family compositions. Moreover, selected family templates increase the fraction of generated crystals predicted to be stable or metastable, demonstrating that latent-space composition control can support targeted exploration of promising chemical regions without additional optimization techniques such as reinforcement learning. Code is available at https://github.com/hhw11/FamilySpace