Scaling device-control evaluation requires scaling both interactive worlds and the tasks agents must solve. We introduce SynthDCB, a benchmark for environment awareness and device control built through world distillation (WD). World distillation transforms knowledge of device behaviors, constraints, and environmental effects into modular executable components. A unified capability representation connects tool discovery, world execution, and task synthesis. SynthDCB instantiates Home, Car, and Phone with 114 device classes, 1,166 functions, and 95 events, supporting five task families with 32 subtypes. Explicit states and tool-call traces support task validation and agent evaluation. We evaluate 22 models on 2,560 test tasks; the best achieves 88.8% mean Pass@1. Construction experiments demonstrate token efficiency and expansion of task coverage. At the richest knowledge level, WD fully covers 22/23 properties, 19/29 functionalities, and 30/39 event requirements specified in a real RobotVacuum protocol. Furthermore, the 1,000 sampled tasks semantically cover 81.6% of the 854 reference instructions in the real-world corpus. SynthDCB connects an expanding space of real-to-sim executable capabilities to verifiable agent evaluation.
https://arxiv.org/abs/2611.18639
Referring multi-object tracking (RMOT) takes a video and a language expression as input and tracks all referred objects. Many tracking requirements involve how an object moves rather than how it appears. However, in a video captured by a moving camera, a parked vehicle may appear to move, while a moving vehicle may show little displacement. Existing RMOT methods relate motion with text but do not explicitly remove camera-induced motion. In this paper, we propose extracting residual motion across frames by estimating camera motion in driver-view videos and compare motion characteristics with the query expression. We consider the motion-matching extent and integrate it with the RMOT method's prediction result through late fusion. In the evaluation, we verify the performance gain of taking the motion compensation module as a plug-in across different RMOT hosts.
https://arxiv.org/abs/2610.04185
For more than half a century, video games have pursued better graphics, more natural interaction, and more engaging gameplay. Generative video models (VMs) now bear on all three. Neural networks already draw most of the pixels in shipped games, and recent world models let a lightweight game engine keep the world state while a VM turns the engine's coarse proxy video into final frames. These systems, however, treat the VM as a renderer that only restyles motion the engine has already authored, which leaves most of its generative capacity unused and limits the diversity a game can offer. We present GameWM, a design for next-generation games built on a clear division of labor. The GameWM-engine keeps what must be exact: rules, state, camera, and the layout of the world, which also spares the video model from having to remember it. The GameWM-VM generates what is expensive to author by hand: the look of the world and the detail of the player's interactions. One game mechanic can then be realized in many visually and physically distinct ways, while its outcome stays exactly as the engine computed it. To realize this division of labor, we build GameWM as a complete system: a data engine centered on interactions rather than exploration, an interface through which authored character sheets, scene art, and skill descriptions enter the model, and a model family from 5B to 33B parameters with a real-time causal variant for play. Together they give every player a fresh version of the same game: each player's own character, world, and skills turn the same engine mechanics into a new experience, while every event keeps the timing and outcome the GameWM-engine computed.
https://arxiv.org/abs/2611.19247
In this paper we present the Formal Atlas, a system for the study of natural language semantics using proof assistants. Our aim is to formalize semantic accounts, check their predictions and compare the implemented theories. To this end, we connect source records and formal developments with an interactive application, where researchers can inspect proofs and modify supported Rocq files. Montague's PTQ and Gallin provide our main cases, followed by dynamic and type-theoretic comparisons, RSA and distributional semantics. We also discuss what it takes to formalize less explicit analyses. Lastly, we consider the use of AI, both for assistance in developing these resources and for inference systems that can make use of them. This work extends an earlier programme in proof-assistant semantics, with research and teaching as complementary uses.
https://arxiv.org/abs/2611.19509
Disentangling writing style from semantic content is a fundamental challenge in literary text modeling. Style-content entanglement causes models to rely on semantic content for authorship attribution (AA) tasks and memorize author-specific content for style imitation (SI) tasks. We propose EDITORS (EditingLoRA Subspaces), a diagnose-then-deflate framework that adapts pretrained large language models and prioritizes stylistic features, making it effective for both AA and SI tasks. EDITORS first trains a diagnostic LoRA adapter using style-neutral statements derived from the training corpus to identify a content-oriented subspace. It then trains the final style adapter under activation-space deflation that projects input activations away from the identified content-oriented directions. We empirically demonstrate that EDITORS is effective. On three AA benchmarks spanning literature, social media, and news, EDITORS achieves state-of-the-art accuracy, outperforming the strongest existing baseline by up to 29%. On SI, EDITORS generates stylistically faithful and diverse passages, outperforming baselines in achieving style similarity while adhering to content specifications.
https://arxiv.org/abs/2611.20846
Video deraining is challenging due to rain streaks, occlusions, and atmospheric effects. Existing methods mainly follow a deterministic restoration paradigm, assuming that clean scenes can be uniquely recovered from rainy observations. However, rain degradation is spatially heterogeneous. Reliable regions require faithful reconstruction, whereas severely occluded regions contain ambiguous content that require stronger prior-driven inference. Therefore, conventional deterministic restoration often produces over-smoothed results, while global generative restoration may introduce unnecessary hallucinations. We introduce a Selective Generative Restoration paradigm for video deraining, where generative priors are adaptively activated according to restoration uncertainty. Based on this principle, we propose Selective Generative Video Deraining (SGVD), an uncertainty-aware diffusion framework that enables adaptive fidelity preservation and generative recovery. Specifically, SGVD exploits stochastic diversity between dual one-step diffusion predictions to estimate region-level restoration uncertainty, which determines where observation fidelity should be preserved and where generative inference should be activated. Furthermore, an anchor-guided trajectory consistency mechanism is proposed to regularize stochastic diffusion evolution across frames, reducing temporal generation drift and improving video coherence. To achieve practical efficiency, SGVD adopts one-step diffusion inference while maintaining uncertainty-aware generative capability. Extensive experiments demonstrate that SGVD achieves superior restoration performance, perceptual quality, and temporal consistency compared with state-of-the-art video deraining methods.
https://arxiv.org/abs/2611.16310
Multimodal coding agents decouple 4D world construction from visual generation, repurposing video models from world generators toward generative renderers. This division of labor raises a fundamental question: if code builds the world, can video models just render it? A reliable generative renderer hinges on two complementary criteria: (i) world preservation, retaining prescribed 4D kinematics and camera paths; and (ii) visual realization, persistently binding text-specified target attributes over time. To evaluate this capability, we introduce WorldDraft, a benchmark of $432$ strictly paired test cases across $48$ executable 4D world families spanning six dynamic challenge categories. Under a $3 \times 3$ factorial design, WorldDraft orthogonally crosses draft granularity (abstract proxy to previs) and appearance alignment (neutral to conflicting) under identical dynamics. Extensive evaluations across $12$ representative models reveal that: (I) world preservation and visual realization remain in stark operational tension; (II) richer draft scaffolding anchors structure but paradoxically impairs appearance fidelity; and (III) conflicting source cues trigger severe source leakage (up to $64\%$), exposing fragile cross-modal disentanglement. To mitigate these bottlenecks, we propose WorldDraft-Primer, a plug-and-play visual prompting strategy that consistently improves both objectives across frozen video renderers. Together, WorldDraft and WorldDraft-Primer lay a foundation for bridging programmatic 4D simulation with video generation, advancing physically grounded generative rendering.
https://arxiv.org/abs/2611.16217
Rules govern how player actions change interactive worlds. Different rules create distinct gameplay experiences within similar scenes and action spaces. For customizable gameplay, generative world models must follow user-defined rules beyond real-world behaviors. However, realizing user-defined rules presents two challenges: Rule-State Persistence and Rule-to-Visual Realization. The former requires preserving rule definitions, dependencies, and world states across saving and loading, so that restored worlds follow the established rules. The latter requires accurately realizing rule-induced changes on designated objects and attributes, particularly when affected objects are unseen. In this paper, we propose WorldRuler, which organizes rule definitions, dependencies, and object states in an executable 3D rule graph and connects rule execution, video generation, and perceptual feedback in a closed loop. The graph supports joint saving and restoration of rules and world states, while dynamic object binding grounds rule-specified entities in visual instances and uses stored states to constrain generation from their first appearance. We further introduce object attribute state replacement and counterfactual reconstruction training, enabling accurate attribute control and compatible appearance and motion changes while limiting interference with unrelated content. We develop an automated data collection and annotation pipeline using GTA-V and Habitat to construct RuleVerse, a dataset of rule-driven interactions with detailed state and event annotations, including cases where rules take effect before affected objects become visible. Extensive experiments demonstrate the effectiveness of WorldRuler in preserving rule-state consistency across saving and loading and accurately realizing rule outcomes on designated objects and attributes.
https://arxiv.org/abs/2611.24869
Small language models (SLMs) can run on commodity hardware and support privacy-preserving, on-premises clinical deployment, but their medical explanations are often short and poorly structured compared with large language models. Recent supervised approaches such as Meerkat transfer chain-of-thought reasoning into 7B-8B backbones, yet their training objective does not directly target the style or organization of the generated rationales. Building on the instruction-tuned Mistral-7B backbone, we investigate whether a lightweight reinforcement learning (RL) step can improve explanation structure without human preference labels. We apply Proximal Policy Optimization (PPO) with LoRA adaptation to 2000 Meerkat-derived medical prompts, guided by a simple self-supervised reward that favors longer, lexically diverse, and well-punctuated responses while penalizing very short ones. Training for about 170 PPO steps on a single NVIDIA A100 GPU yields a stable increase in the average reward. Quantitative and qualitative comparisons show that the RL-tuned model produces longer, more organized, multi-point explanations while preserving the underlying medical content. We discuss the promise and the limitations of self-rewarded RL for shaping the style of medical explanations.
https://arxiv.org/abs/2611.24650
Advances in vision-language models (VLMs) have improved remote sensing image analysis and supported geospatial decision-making. However, fine-grained understanding of remote sensing imagery remains challenging for VLMs. Large variations in object scale, complex geospatial relations among instances, and limited visual detail hinder precise instance localization, visible attribute identification, and spatial relation understanding, potentially leading to descriptions that lack visual grounding. Existing datasets and benchmarks insufficiently assess end-to-end instance localization and captioning without target-specific textual cues. To support this evaluation, we introduce RICapBench, a dataset and benchmark that unifies instance localization and fine-grained captioning in remote sensing. RICapBench contains 245,466 instance-level annotations and 498,005 spatial relation tuples spanning general remote sensing, UAV remote sensing, and dense small-object scenarios. Each annotation pairs an object's oriented bounding box with a natural language description of its visible attributes and spatial relations to surrounding instances and environmental elements. To evaluate the accuracy and completeness of spatial relation captioning, we propose a remote sensing spatial relation parser and the RelCap-F1 metric. We combine RelCap-F1 with instance localization and attribute captioning metrics into a composite metric, RICScore. Our evaluation of 15 advanced open-source and closed-source VLMs identifies instance localization as a major bottleneck for these models. LocateAnything achieves the highest RICScore of 31.68, but scores lower on spatial relation captioning than other task-fine-tuned open-source VLMs. Code and dataset are available at this https URL.
https://arxiv.org/abs/2611.16413
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.
https://arxiv.org/abs/2611.11136
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.
https://arxiv.org/abs/2610.14335
Reference-to-video diffusion transformers pack the variational autoencoder (VAE) latents of all references and the generated tokens into one full-attention sequence, in which reference tokens make up about half. Because attention cost grows quadratically with sequence length, these tokens take up an even larger share of the attention computation, yet most of them are static, smooth or irrelevant to the output. To accelerate reference-conditioned generation, we propose pruning and caching of reference tokens for video generation (PCRG), a training-free, plug-and-play method that uses two signals the model already computes to assign each reference token to one of three tiers, namely pruned, cached or kept. The first is a low-level redundancy r from VAE latents, since a token whose latent barely changes over time or across its neighbors adds little new content. The second is a high-level importance g from the vision-language-model (VLM) conditioner, where the attention that the prompt and any first or last frames pay to a token tells whether the output needs to read it. Tokens that are not redundant are kept, and redundant tokens are cached if important and pruned otherwise. Pruned tokens leave the sequence before the first step with all rotary position embedding positions unchanged. Cached tokens have their keys and values computed once and reused, and kept tokens are processed as usual. On MiniMax-H3, already distilled to 8 timesteps and quantized to int8, PCRG is 1.57x faster end to end on our V2V-Bench and 1.49x faster on the public VACE-Benchmark, while the quality of VBench drops by only 0.24% and 0.17%, respectively. PCRG shortens the sequence rather than changing how attention is computed or which steps are run, so it also stacks on step caching and sparse attention. On V2V-Bench it reaches 1.86x with TeaCache, 1.82x with Spark at nearly unchanged fidelity and 2.15x with Veda.
https://arxiv.org/abs/2611.22717
LLM agents have made impressive progress in coding and mathematical reasoning, motivating growing interest in learning from long-horizon interaction. Memory supports this process, but plausible lessons can be redundant or reinforce inappropriate actions. Can a memory's source decision provide a practical test for whether to retain it? We localize consequential trajectory steps, extract candidate memories, and replay each source decision with and without its candidate under the same context. This reveals two properties: most candidates barely shift the probed distribution, and both successful and failed source actions can be reinforced or suppressed. These observations motivate Anchor, which tests shift magnitude and source-role alignment before retention. Admitted memories guide later trajectories, enabling experience-driven improvement under a frozen policy. On held-out ALFWorld and WebShop tasks, Anchor improves success rate and mean score by up to 19.9% and 17.2%, respectively, relative to the strongest baseline for each backbone. With Qwen3.5-27B, it retains 41% fewer ALFWorld cards than unfiltered retention and outperforms all four WebShop cost-comparison baselines while using fewer tokens.
https://arxiv.org/abs/2611.10382
Residual reinforcement learning corrects an imitation-trained base policy using a limited budget of environment interaction. In a task with several phases, this budget must be divided among the phases, and the phase in which the base policy fails is not necessarily the phase in which a correction most improves task completion. We formalize this distinction through the completion gain: the change in full-task completion probability when, during a single phase, a state-feedback controller replaces selected components of the base policy's action. We decompose the completion gain into the probability of reaching the phase and the conditional improvement in completion from the start of the phase, and we estimate it from paired episodes. The positive estimates determine which action components the residual policy corrects and how training episodes are distributed across phases. In stochastic point-mass docking experiments, most failures occurred in the final docking phase, whereas the largest completion gain occurred in the preceding channel phase; gain-based allocation improved completion over allocation proportional to failure counts and, in a variant with a longer approach, over uniform allocation at half of the budget. In simulated bimanual robotic assembly, the completion-gain estimates were positive only for insertion and release, although few failures occurred during release; in a second assembly configuration, a residual policy trained in the release phase improved completion from held-out states at the start of release by 16.9%.
https://arxiv.org/abs/2611.25320
GUI grounding maps a screenshot and a natural language instruction to the coordinates of the intended interface element. Existing approaches improve grounding through visual scaling, instruction scaling, and sampling scaling, yet it remains unclear which grounding bottlenecks these strategies address and where their gains saturate. To address this gap, we introduce ScaleDiag, a controlled diagnostic framework that isolates the effects of visual input, instruction information, and prediction sampling on GUI grounding by independently varying target scale and surrounding context, target-specific information and auxiliary descriptions, and candidate coverage and selection, respectively. We evaluate ScaleDiag across five benchmarks and up to eight vision-language models and find that each scaling strategy operates within a much narrower regime than commonly assumed. (1) Visual scaling is governed by target occupancy rather than a fixed crop ratio, with full-screen errors typically rescued once the target occupies only 0.32% of the cropped view, while further zooming eventually becomes harmful. (2) Instruction scaling is driven by information that identifies the target rather than description richness itself, as even ground-truth-derived attributes such as color, size, and shape are often neutral or harmful, while appearance and spatial descriptions account for over 90% of the achievable gain. (3) Repeated sampling exposes a different limitation, as correct coordinates are frequently generated but remain difficult to select, leaving 5.1--11.0 points of oracle headroom while geometric consensus provides almost no improvement. Together, these results show that test-time scaling is constrained by distinct operating regimes rather than benefiting monotonically from more visual focus, richer instructions, or additional samples. Our code is available at this https URL.
https://arxiv.org/abs/2611.25297
Diffusion Transformers generate high-quality images, but iterative sampling remains computationally expensive. Feature caching reduces the number of full network evaluations, yet its approximation error can become more severe under short sampling schedules, where fewer full-compute steps and a coarser sampling grid can make history-based extrapolation less reliable. In this paper, we propose CalibCache, a lightweight offline calibration method that corrects systematic discrepancies between cached and full-compute feature dynamics. Specifically, CalibCache calibrates feature acceleration, defined as the second-order finite difference across sampling steps. Using only a single pair of full-compute and naive-cache trajectories at each setting, CalibCache fits scalar acceleration maps offline by centered least squares. These maps are then applied during sampling without additional full-network evaluations or online optimization. Across FLUX.1-dev and Qwen-Image on DrawBench and DiT-XL/2 on ImageNet-256, CalibCache consistently improves fidelity or generation quality over TaylorSeer under 20-step sampling at cache intervals $N=3,4,5$. At $N=5$, it reduces FLUX.1-dev LPIPS from $0.563$ to $0.429$ while achieving a $2.86\times$ denoising speedup over full-compute sampling, and lowers DiT-XL/2 FID from $16.39$ to $10.37$ relative to TaylorSeer.
https://arxiv.org/abs/2611.25296
Zero-shot large language models (LLMs) are increasingly used as text classifiers, but their labels are only useful when we know to what extent we can trust them. We compare black-box methods (verbalised confidence and sampling) with white-box methods that read token probabilities (P(True), and multiple logit-scoring methods using constrained generation). We evaluate eleven open-weight LLMs (1B-32B parameters) on five classification datasets with 3 to 151 labels, measuring classification performance (accuracy, F1-macro), discrimination (AUROC), and calibration (ECE and Brier score). The choice of confidence method rarely impacts accuracy significantly, but logit-based scoring methods lead to much better discrimination (AUROC 0.74-0.77) than black-box methods (AUROC 0.57-0.66). Replacing labels with single-token identifiers preserves discrimination but costs 3-6 accuracy points on average. We also introduce decision-only scoring, which only scores positions where labels diverge in the prefix tree (trie). It reduces decode steps by 88-96% on the three largest label sets and performs comparably to full-label scoring. The best open-weight model-method configurations (Gemma 4 31B and Qwen 3.5 9B) are competitive with the commercial Jev 1.13.0 classifier.
https://arxiv.org/abs/2611.25289
We extend PoseTReID with online appearance-bank matching (Gallery) and a residual-conflict guard for registered SVM recognition. All methods use supplied ground-truth (GT) person boxes. We average all-box assessment IDF1/HOTA equally across videos. On six SportsMOT clips, StrongSORT leads standalone trackers (79.44/84.72%) and exceeds both Gallery variants. Over eleven videos, Gallery AIN (83.29/85.44%) and COPE (84.66/86.38%) exceed standalone trackers with complete results. This lead depends on two retrospectively added GTA development compilations and is absent on the original nine videos. Registered SVMs lead both cohorts among complete results, with AIN at 95.72/94.82% over eleven videos. They use every GT crop from approximately 20% of frames, including future appearances. On MOT20-03, registered AIN reached 96.80% IDF1 and assigned the correct enrolled name to 90.72% of assessment GT observations. Gallery and standalone trackers receive no per-video labelled enrollment. Both identity routes restore identities across internal title cards, but Gallery trails SORT on three videos (four in HOTA) and both SVMs trail SORT on MOT16-04/MOT20-03. Under six-clip persistent-only scoring, native Gallery gains 2.39 percentage points in mean IDF1 over a matched control, with lower correct-P coverage and 5.02 times as many encoder attempts. Pose gating is mixed; the guard's metric effect is not isolated.
https://arxiv.org/abs/2611.25283
Dense metric mapping from a single camera and IMU remains challenging: visual-inertial odometry (VIO) provides metric motion but sparse geometry, whereas feed-forward depth models predict dense structure without ensuring metric scale. We present Accord, which constructs an accurate online dense metric map from any camera with an IMU. A VIO filter publishes metric poses, a depth foundation model conditioned on them predicts depth for short keyframe windows, and a factor graph fuses the resulting submaps. Recent work has explored several approaches to this problem. What is essential is the multi-view agreement (Accord), which is considered in SLAM as a source of geometric self-consistency. Accord anchors each keyframe window to metric 3D points triangulated from the filter poses, then exploits bidirectional dense depth agreement between co-visible images to estimate and correct per-image scale errors while preserving the metric scale established by the anchors. Context keyframes further improve depth prediction, and the resulting submaps are fused in a factor graph. We evaluate a single mapping configuration across five public camera-IMU setups and a new Phone-rig/GoPro dataset. With nothing tuned per dataset/setup and every map aligned only by Accord's own trajectory, Accord achieves consistently best/second-best mean F-score at 5 cm on every platform. Accord is open-source along with the phone-rig dataset for the benefit of the research community
https://arxiv.org/abs/2611.25279