Vision-language-action models (VLAs) encode images with 2D encoders that capture little 3D structure. A common approach is to feed the policy estimated depth, but a UAV chooses its own altitude, must judge distances tens of meters away, where monocular depth is less reliable, and must relate its forward and downward views. On long aerial routes, adding depth features directly brings no significant gain (42.5% vs. 42.6% success), so we study how 3D information should be given to the policy. Analyzing flight failures, we find that two forms of 3D grounding fail in opposite ways: explicit 3D lifting of depth collides less on the way but more often fails to finish at the target, while implicit alignment with a 3D geometry model finishes more often but collides more. We propose GRID, which combines both: (1) it lifts monocular depth into ray-corrected 3D positions in the UAV's frame for both cameras, and (2) during training only, it aligns its visual features with a frozen multi-view geometry model, at no inference cost. GRID adding explicit lifting to implicit alignment significantly increases how often the policy reaches the target ($61.5\% \rightarrow 69.5\%$, $p<0.001$), and adding implicit alignment to explicit lifting significantly increases how often it completes the task once there ($57.7\% \rightarrow 61.4\%$, $p<0.05$). On TravelUAV (3,005 test episodes), GRID achieves the best success rate on every split, with relative gains over the state of the art of 9.5% (seen), 17.1% (unseen objects) and 25.4% (unseen map).
https://arxiv.org/abs/2611.23491
Absolute 3D human pose estimation (HPE) aims to recover both a person's body pose and their global position in 3D space. Privacy-preserving millimeter-wave (mmWave) radar provides distance and direction measurement capabilities, but its sparse point clouds make HPE under-constrained. Low-resolution thermal imaging can provide complementary observations of body direction and structure. However, our evaluation shows that a strong feature-level fusion baseline loses much of its localization and HPE advantage when generalizing to unseen subjects, with radial localization becoming less accurate than a radar-only baseline. Motivated by this finding, we introduce mmTherm, a lightweight framework that explicitly incorporates the complementary geometric properties of thermal imaging and mmWave radar. It combines body structure and joint directions estimated from thermal images with range information from radar to localize the body and reconstruct its 3D pose. Experiments on OctoNet show that mmTherm reduces cross-user absolute joint error by 33% and root localization error by 38% relative to the strongest corresponding baselines, while requiring substantially fewer parameters and less computation at inference.
https://arxiv.org/abs/2611.23477
Post-training quantization is the standard way to serve large language models at low cost. However, the best quantization recipe differs across models and precisions. As a result, experts must redesign it for every new model, which is time-consuming and often sub-optimal. Fortunately, an LLM agent can take over this design by refining the recipe over rounds of building and evaluating candidates, so that more rounds lead to better recipes. However, evaluating one quantized reasoning model takes tens of GPU-hours, and combining methods from different codebases means porting each to the new architecture by hand. We present QDA, a system in which an agent designs the recipe for each model and precision. To make evaluation fast, QDA scores each candidate in minutes by its output KL divergence from the original model on reasoning traces that the agent never sees, and checks promising ones with long generations to catch failures such as repetition loops. To make building fast, QDA provides a library of public PTQ methods that are compatible with one another and with new model architectures. Across four reasoning models, QDA improves accuracy over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search. On the 320B-parameter MoE model GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of the FP8 model's accuracy, against 96.9% for the best vendor checkpoint. Every QDA checkpoint also runs on existing serving kernels with no inference overhead. We will release the library, the evaluator and the harness.
https://arxiv.org/abs/2611.23388
Interactive video world models enable exploration through text and camera controls, yet direct object interaction requires finer control. Hands provide a natural interface, but translating continuous hand movements into coherent visual responses remains challenging because of complex hand-object dynamics and demanding real-time computation. To address these challenges, we introduce LynnReal-Hand, a hand-controlled video world model that unifies precise interaction modeling with efficient streaming generation. We curate 100K interaction clips spanning diverse real and synthetic environments and develop hand-aware supervision to improve fine-grained hand motions and object interactions. Building on this foundation, we learn a causal flow map for few-step generation and introduce region-aware score distillation that combines complementary video and image priors to refine interaction dynamics and visual appearance. To further reduce inference latency, we propose Ultra-Fast-VAE, a compact video decoder that preserves high reconstruction fidelity with substantially lower computational cost. Extensive evaluations on HandBench demonstrate superior hand-following accuracy and visual quality compared with existing approaches. Our full-sequence model achieves an FVD of 181.67 and a MANO Dice score of 93.02, while the three-step causal model maintains a MANO Dice score of 92.49. For real-time interaction, 540p streaming on a single H200 reaches 18.03 FPS with 741ms latency and an optimized peak GPU memory footprint of only 34.89GB. We will release the datasets, model weights, and the full training and inference code to support research on hand-controlled video world models.
https://arxiv.org/abs/2611.21479
Goal recognition is the task of inferring an agent's intended goal from its observed behavior. Most existing work focuses on high-level symbolic recognition, abstracting away the continuous motion through which agents reach their goal. Methods that do reason over continuous trajectories rely on an explicit dynamics model that is hard to acquire and rarely transfers across settings. Neither approach explicitly learns how agents move toward each candidate goal, limiting accuracy on ambiguous trajectory prefixes where such knowledge is most needed. We present WoRD (Waypoint Recognition with Diffusion), which uses a learned trajectory prior to infer the agent's goal from partial observations. At each step, WoRD runs one warm-started denoising chain per candidate goal, steers each chain toward the observed trajectory through a closed-form log-likelihood gradient, and reweights the resulting samples by Bayes' rule into a calibrated posterior over goals. We evaluate WoRD on simulated robot settings, pedestrian trajectories, and vehicle trajectories using four metrics targeting distinct failure modes. Across all three settings, the same inference procedure preserves calibrated multimodal beliefs through ambiguous prefixes where filter- and classifier-based observers collapse to a single mode or grow overconfident.
https://arxiv.org/abs/2611.23192
We investigate how language models process negation in natural language inference (NLI). We propose the \textit{Polarity Composition Hypothesis}, a high-level causal model of this computation. Behavioral evaluations show that language-model predictions largely conform to the proposed hypothesis. Using Distributed Alignment Search, we identify causally effective subspaces aligned with the variables in the polarity-composition hypothesis. These representations transfer across negation forms and partially to a subset of MNLI. We release the dataset and code.
https://arxiv.org/abs/2611.23185
Empirical research is changing rapidly as AI coding agents are becoming increasingly proficient in every step of it, from writing code to manipulating data. We argue that while the potential of coding agents for science is vast, their use in hypothesis testing poses a substantial threat to classical inference, as these agents enable a supercharged form of selective inference. In particular, we argue that coding agents carrying out empirical research introduce a new type of post-selection inference that combines human-like, open-ended search over an implicit space of estimators with programmatic replayability of both the selection and the estimation process. Indeed, we find that on simulated data with zero effects, four frontier agents report significant p-values between 54% and 98% of the time when asked to produce a compelling finding. We propose a simple procedure for detecting and correcting these distortions. After an agent has completed its analysis, we reset and replay it on subsamples of the data to estimate the agent's relative probability of selecting that analysis when it looks statistically favorable. For agents with stable selection behavior, the resulting corrected test is asymptotically valid and consistent conditional on selection. Importantly, the correction is learned from the black-box agent's actual choices rather than a model of how it searches, so it remains valid even as AI technology continues to evolve rapidly. In simulations, the correction robustly lowers the high false rejection rates to near the nominal 5%. We illustrate the method on applications from two published papers.
https://arxiv.org/abs/2611.23174
Recent advances in text-to-image generation have substantially improved image quality, particularly through diffusion-based methods that require multiple inference steps. In contrast, Sphere Encoder offers an efficient generative framework capable of producing images with fewer forward steps. However, extending it to free-form text is nontrivial, because its class-conditioned design lacks cross-modal alignment. To address this challenge, we propose TextSphere, an efficient text-to-image generative framework that extends Sphere Encoder with language conditioning and cross-modal alignment. We incorporate pretrained CLIP text embeddings through adaptive layer normalization (AdaLN) to guide image generation while retaining the original spherical sampling mechanism. To bridge the cross-modal representation gap, we introduce a Spherical Semantic Alignment and Uniformity (SSAU) objective that jointly promotes semantic alignment and representation uniformity on the hypersphere. Experiments on CUB-200-2011, MS-COCO-2014, and Flickr30k demonstrate that TextSphere consistently reduces generation FID (gFID) over the text-conditioned baseline while achieving reconstruction FID (rFID) below 1.0. With only 4 and 6 sampling steps, TextSphere achieves lower gFID than the compared diffusion- and GAN-based models trained from scratch on CUB-200-2011 and MS-COCO-2014, respectively. We further observe that improved reconstruction fidelity does not necessarily yield better generation quality. In addition, we use Partial Least Squares Singular Value Decomposition (PLS-SVD) to analyze cross-modal alignment and modality gap, offering insight into the effectiveness of our approach.
https://arxiv.org/abs/2611.23155
Inference-time scaling improves the performance of large language models by sampling multiple responses and evaluating their collective utility. However, many existing approaches obtain these samples by repeatedly sampling from a single policy. In this work, we study if the sample budget at inference time can instead be allocated across a population of policies. We introduce stratified policies, where each of $k$ learned policies contributes one response to a $k$-sample inference-time objective. We realize this policy population using separate LoRA adapters on a shared language model backbone and optimize it for pass@$k$. We theoretically motivate the approach by analyzing the linear setting; a population of LoRAs can achieve zero best-of-$k$ approximation error in settings where every single LoRA has strictly positive prediction error. Empirically, we demonstrate that stratified policies consistently improve pass@$k$ over a single-policy baseline across multiple reasoning tasks. Overall, our results highlight how distributing samples across complementary policies can be a more effective use of sample budget than repeated i.i.d. sampling from a single policy at inference time.
https://arxiv.org/abs/2611.23154
AI-assisted RTL localization and debugging need ground-truth bug datasets with known locations and observable failures, while testbench grading uses injected bugs to measure detection capability. Creating such datasets at scale remains a major EDA challenge. BugGen 1.0 shows that Large Language Models (LLMs) can generate context-aware RTL mutations. Still, it relies on the model for both choosing where to mutate and generating the edit, which increases inference costs and can lead to invalid choices. BugGen 2.0 changes this division of labor. Deterministic algorithms handle structural analysis, target selection, and validation, while the LLM focuses on localized semantic rewriting. Across four OpenTitan IP blocks and FabScalar with three LLMs, BugGen 2.0 reaches every requested quota of compilable bug scenarios that cause at least two test failures, with 14-16 of 16 configured mutation types represented. In completed same-model comparisons, it reduces token use by 6.8x-15.4x relative to BugGen 1.0 and 1.5x-7.9x relative to Firefly. BugGen 2.0 also expands the dataset without simulating every new bug. After finding accepted scenarios with observable failures, it generates alternative mutations at the same locations, simulates only a subset, and estimates acceptance for the rest. This skips 63.6% of AES and 33.9% of FabScalar variant simulations. Full simulation finds that 77.7% of AES variants and 95.2% of FabScalar variants are accepted, with a 1.09% estimation error on the directly comparable FabScalar set. These results show that focusing LLMs on semantic rewriting lowers inference and simulation cost while improving consistency across models and designs.
https://arxiv.org/abs/2611.23091
Long-video question answering benchmarks like Video-MME-v2 typically evaluate systems under a uniform frame budget for every question, despite stark variation in the evidence each query actually demands. We analyze this mismatch on Video-MME-v2 and find that different question subsets prefer different frozen frame-budget policies, while uniformly increasing the frame count does not reliably improve all metrics. These findings motivate fixed-mean frame-budget routing, a controlled inference setting in which each question may receive a different frame budget while the average budget matches a uniform reference. We introduce FrameRouter, a training-free test-time framework that combines an evidence-demand router, a budget-constrained allocation rule, and a frozen bank of frame-sampling policies. By controlling how much visual evidence each question receives while leaving the underlying samplers unchanged, FrameRouter is orthogonal to query-aware frame selection. Across Video-MME-v2 and several long-video understanding benchmarks, FrameRouter improves long-video question answering by using the same average visual budget more selectively across questions. Code is available at this https URL.
https://arxiv.org/abs/2611.23055
Vision-language models (VLMs) can correctly interpret physics diagrams yet apply constraints outside the states or conditions in which they hold, causing premise errors to propagate through subsequent calculations. We introduce PhysScope, an inference-time visual harness that supports self-correction of these scope mismatches without updating model weights. PhysScope enables the model to link physical scopes directly to its visual observations and explicitly constructs derived views to support reasoning. The harness imposes explicit constraints on state distinctions, conditions, and physical scopes during model reasoning, and checks agent-proposed revisions against sources and scopes. Accepted revisions invalidate registered dependents, thereby maintaining scope-aware working memory. On PhysAlign, PhysScope achieves higher solving scores than direct solving across all six evaluated VLMs. With Gemini-3.8-flash, it outperforms Canvas-CoT by 9.34 percentage points and substantially improves solving accuracy on the out-of-distribution (OOD) PhysUniBench benchmark. Ablation studies support explicit scope management and dependency maintenance as design principles for multimodal physics reasoning agents.
https://arxiv.org/abs/2611.23010
In automated driving, world models provide future scene predictions for evaluating candidate trajectories, but fixed selection rules may not capture scene-dependent trade-offs. We propose a residual-guided reinforcement learning (RL) framework for lightweight trajectory selection with a frozen world model and a fixed candidate bank. The framework uses proximal policy optimization (PPO) to fine-tune a joint policy over candidate selection and bounded trajectory residuals. A context-conditioned ranking network learns candidate preferences from feedback obtained by evaluating local perturbations of the selected trajectories. At inference, the selector combines learned preference and predicted risk ranks using fixed weights and returns the selected candidate trajectory without modification. Fine-tuning updates 16,106 parameters across the actor and critic, while the deployed selector retains only the 11,725-parameter ranking network, with the residual branch and critic used exclusively during training. In open-loop evaluation on the nuScenes validation set, the method reduces average Euclidean trajectory error (L2) by 3.99% relative to the RiskWorld model baseline without RL-based selection. Under matched budgets for feedback accesses and actor updates, residual exploration reduces L2 by 0.83% relative to the categorical PPO variant trained without residuals. The method also maintains lower L2 than RiskWorld when risk features are removed or shuffled.
https://arxiv.org/abs/2611.22998
Is a face or a voice still heard once the words are known? Incomplete multimodal learning increasingly uses large language models (LLMs) as fusion engines: modality features enter as token blocks and missing modalities become placeholders. We find that LLM-driven fusion can override non-verbal evidence it can read. Contrasting its full-input prediction with counterfactual views, the same model shown one modality or no content, we identify language shortcuts. We build MR-Bench: MR-Bench-A types every test sample of four sentiment benchmarks; MR-Bench-H contains 900 human-validated clips (620 language-shortcut, LS; 280 non-verbal-shortcut, NS) from seven sources, and every model is scored on all of them. Zero-shot MLLMs solve at most 32.7% of the LS clips; adding the transcript to the muted video lowers vision-language accuracy by 46.1-53.1 points. Yet inverting a text-only model's answers solves 94.4%, so LS accuracy alone cannot establish use of non-verbal evidence. Inside the fusion LLM, fused states align increasingly with the text view late in the decoder. Counterfactual-View Reliability learning (CoVeR) trains views as unimodal judges and withholds fused supervision where they disagree. It achieves state-of-the-art results on five benchmarks without extra inference cost in accuracy mode. On MR-Bench-H, it reaches 61.7% LS and 93.6% NS accuracy; robust mode raises LS accuracy to 63.0% while lowering NS accuracy to 84.8%. Among LS clips solved without text, adding text causes errors on 30.1% with CoVeR versus 32.2% with LLM-Fusion. Project page: this https URL.
https://arxiv.org/abs/2611.20349
Agentic systems using retrieval-augmented generation (RAG) must repair their own answers by gathering more evidence, testing alternative policies and deciding if, when, and what repairs are needed. Yet how do we know if a system's repairs are working when metrics like accuracy can be misleading? We present RagHarnessDebugger, a framework that records every repair decision in a Jev-guided harness for contract inference and audits the record to facilitate complete debugging of all decisions taken during a RAG repair journey. RagHarnessDebugger checks that each trace is intact, that each validation applies to the candidate it is counted for, and which answers the acceptance rule can reach. We tested our approach on ContractNLI non-disclosure agreements with two reader models and a replication set of separate contracts. Our debugger's audit layer exposed failure patterns that were hidden under aggregate metrics, policies acceptance cost of repairs vs regressions and policy acceptance gap of 77% and 76% of available but unrealised repairs on development and replication sets. Our audit, together with a labeled error analysis and replay mechanism, allowed following the execution path all the way to the terminal decision - uncovering where missing repairs became lost. Our revised harness improved accuracy over the initial BM25-only baseline by 3.6 points on development contracts and 3.7 points on the replication set, and even improved our full contract access reference score by 1.1 on development and 2.4 on replication set. While our work here demonstrates how the debugger can conduct decision root-cause analysis for missed repairs, the true value lies with providing this debugger for judgment visibility serving continuous improvement of RAG decisions and their explainability. Keywords: agentic RAG, AI evidence repair, harness explainability, counterfactual replay, Jev, ContractNLI
https://arxiv.org/abs/2611.22991
Typed decision models combine natural-language understanding with a structured interface that returns a selected action and an option-level probability distribution without generating lengthy textual responses. Their low inference cost and response latency make them attractive for transportation problems with predefined alternatives. This paper investigates their effectiveness, robustness, and efficiency in two representative transportation tasks: individual travel-mode choice prediction and sequential traffic signal control. We systematically evaluate TypeSafe Jev and GPT-6 Luna Decisions against conventional transportation methods, reinforcement learning algorithms, and four chat-based large language models. Experiments on the Swissmetro dataset show that typed decision models achieve lower prediction accuracy than traditional discrete choice models and the strongest chat-based baseline, while their performance varies substantially with input representation and question framing. In closed-loop traffic signal control experiments, typed decision models demonstrate competitive performance against established controllers in several scenarios, but exhibit performance degradation under more challenging network conditions. Furthermore, the typed models achieve lower inference latency and computational cost than the evaluated chat-based large language models, highlighting their potential for efficient, zero-shot transportation decision-making. Overall, these findings demonstrate both the potential and limitations of typed decision models, emphasizing the importance of input design, task characteristics, and robustness evaluation for their effective application in transportation systems.
https://arxiv.org/abs/2611.22926
Abstract visual reasoning requires identifying relationships between few examples and applying them to novel inputs. Recent advances have been driven by large language and vision-language models or specialized domain-specific architectures. It remains unclear whether abstract reasoning can be supported directly by the relational geometry of pretrained visual representations, without relying on language-based inference or task-specific representation learning. We investigate this possibility using frozen self-supervised vision encoders. With latent vector arithmetic, we confirmed that the representational geometry of two self-supervised encoders support abstract concept discrimination on Bongard problems and encode systematic relationships between input-output pairs in ARC tasks. Simple vector differences between image embeddings, which we call intuition vectors, capture transformations that generalize across instances of the same task. These vectors distinguish among 794 abstract reasoning tasks with up to 87.2\% accuracy, preserve relational structure across task-invariant visual transformations, and guide output retrieval for unseen inputs. Given this, we tested the hypothesis that this relational structure can steer generative reasoning by conditioning a diffusion transformer operating in pretrained visual latent space. On held-out ARC tasks, intuition-vector conditioning outperforms conditioning on complete demonstrations while requiring nearly six times less test-time compute, and supports zero-shot inference. Taken together, our findings suggest that pretrained visual representations provide a transferable substrate for abstract reasoning, enabling a modular approach in which the relational structure inferred from demonstrations steer generative reasoning.
https://arxiv.org/abs/2609.32619
World-action models (WAMs) couple action generation with predictions of future states. An initial action chunk may be flawed. A future prediction that is not conditioned on the chunk is identical for every proposal and thus carries no proposal-specific signal, limiting its usefulness for refinement before execution. We introduce ThinkWAM, a latent WAM that establishes bidirectional feedback before execution through a proposal-rethink-refine loop: the policy first proposes a complete action chunk, the proposal corrects multi-horizon future latents through a learned residual, and the corrected latents are injected into a second pass of the same action head, which generates the executable chunk. An affordance-guided observation-action memory and a near-fine, far-coarse multi-horizon prediction scheme, both motivated by predictive coding, support this loop, and all prediction stays in latent space. ThinkWAM achieves 98.70% success on LIBERO, 84.76% on LIBERO-Plus without OOD fine-tuning, 85.75%/65.85% on RoboTwin 2.0 Clean/Random, and 65.17% on RoboCasa24. ThinkWAM achieves 64.4% mean success across three real-world OOD conditions. The mean two-pass inference latency per action chunk is 153.48 ms on LIBERO, 164.61 ms on RoboTwin, and 101.2 ms on the Galaxea A1Z. Ablations show complementary gains from rethinking and affordance guidance. Diagnostics indicate that proposal-conditioned corrections change both predicted futures and action trajectories, showing that the corrected future representations improve the final action.
https://arxiv.org/abs/2611.21258
Large language models (LLMs) in embodied robots create security dependencies beyond inference. We investigate a network-exposed configuration attack on the Pollen Robotics Reachy Mini Lite, allowing an unauthenticated attacker to replace personality settings and the cloud LLM's system prompt through a management interface, without adversarial conversational input. Eleven prompt-injection attempts across five categories were reported as resisted, whereas both configuration-layer attack paths produced observable behavioural changes on physical hardware. We demonstrate local and network-adjacent access, mapping the weakness to CWE-306 (Missing Authentication for a Critical Function) and CWE-1327 (Binding to an Unrestricted IP Address), with a reported CVSS v3.1 score of 6.5 (Medium). A configuration endpoint also accepts cross-origin requests without a CORS preflight (CWE-352; 6.5 Medium), while the persistent-memory sink stores content without enforcing its stated restrictions. A source review of 59 public repositories identifies missing authentication in 10 of 11 source-verifiable genuine applications, or 11 of 11 including a framework-default case. These findings indicate recurrence within the sampled repositories, not prevalence in commercial products. We characterise a deployment gap between developer-oriented network assumptions and consumer-facing embodied AI security: inference-layer refusal cannot prevent unauthorised changes to model instructions. Potential consequences include surveillance through onboard cameras and microphones and social engineering of nearby users. Proof-of-concept code, evidence and the survey are available at this https URL.
https://arxiv.org/abs/2611.22761
Promotional events shift traffic composition, purchase behavior, and the scale of conversion probabilities on different time horizons. We present \textbf{COTTI} (\textbf{CO}ntrastive \textbf{T}raining and \textbf{T}ime-aware \textbf{I}nference), an industrial recommendation framework that separates historical representation learning from event-time score adjustment. Cross-Period Contrastive Alignment (CPC) regularizes paired histories of the same user with adaptive consistency weights. Active Re-weighting (RW) controls the influence of promotional examples through historical-interest distance. Adaptive Boundary Alignment (ABA) combines stochastic prediction variability with a temporal event kernel to adjust logits while keeping predictor parameters fixed. We analyze the weighted objective and bound ABA-induced probability and ranking changes. On Fliggy data, COTTI attains an offline AUC of 0.8127, compared with 0.8101 for TEN and 0.8123 for TEN with ABA. Cross-category analysis reveals improved discrimination alongside declines in paid-item recall for novel categories. End-to-end measurements give a p99 latency of 38.9\,ms and throughput of 14,700 requests/s; enabling ABA adds 2.2\,ms to p99. A ten-day online experiment against the serving HDR model yields relative gains of 3.14\% in conversion rate and 2.68\% in revenue per mille. Together, these results characterize the trade-off between aggregate prediction, novel-category discovery, and serving cost in promotional recommendation.
https://arxiv.org/abs/2611.22597