Vision-language-action (VLA) models enable unmanned aerial vehicles (UAVs) to follow natural-language instructions, yet still fail frequently in closed-loop flight. Correcting these failures requires repairs, i.e., actions that avoid them at the states the policy itself reaches, which human pilots cannot label at scale and rule-based planners cannot supply reliably. In this paper, we present AeroRSI, a recursive self-improvement framework in which an aerial navigation policy iteratively learns from its own failures. Our key insight is that a failing policy should judge its own repairs by flying on from them. Judged in this way, its sampled actions and even rule-based proposals that hurt it when imitated directly become reliable supervision. Moreover, an episode terminates at its first failure, concealing subsequent failures until the earlier ones are repaired. AeroRSI thus renews its supervision in every round, anchoring the failures of the current policy, verifying repairs by its continuation, and distilling them while keeping its behavior elsewhere. On TravelUAV, three rounds raise the success rate of a supervised 2B-parameter policy from 46.4% to 67.5% and reduce collisions by half, surpassing the previous state of the art by 10.6 points from the same initialization without extra human labels or inference cost. Code and models are available at this https URL.
https://arxiv.org/abs/2611.24077
Pretrained Time-Series Foundation Models (TSFMs) provide strong representations for downstream classification tasks. Yet adapting a TSFM under changing operating conditions remains challenging. A model must improve on each newly observed condition without sacrificing conditions it has already learned, often from only a small labeled calibration set and without knowing the operating condition at inference time. We find that this difficulty has an important geometric structure. Under an operating-condition shift, the embedding space of a TSFM does not move uniformly: some local regions change substantially, while others remain nearly unchanged. Consequently, a global adaptation can unnecessarily move embeddings that were already correctly classified. To address this, we propose the novel PACT (Piecewise Adaptation via Chart Transport). PACT partitions the frozen source embedding space into local charts and learns lightweight residual transports for the regions affected by each new condition. An explicit identity path is always retained, and a learned gate softly selects among the identity and accumulated transports for each input. The pretrained TSFM encoder, classifier head, and previously learned transports remain fixed. With only 5% labeled calibration data per arriving condition and no condition identity at test time, PACT maintains an average macro-F1 above 0.95 across a four-condition bearing-fault stream. Under homogeneous tasks and label spaces, similar retention-adaptation behavior is observed across public benchmarks. Code is available at this https URL.
https://arxiv.org/abs/2611.24075
Reasoning traces improve vision-language models on complex multimodal tasks, but forcing every input to follow a long reasoning process leads to overthinking and unnecessary inference cost. We study adaptive fast/slow thinking for VLMs, where one model answers directly for simple inputs and reasons step by step when needed. Fast answers and slow traces often come from distinct behavior policies, making a single-reference offline objective misweight one mode. We propose DUAL-REFERENCE BOLT (DR-BOLT), which distinguishes the desired fast/slow mixture reference from the sampling distribution and corrects their mismatch in reward-weighted training. Fast/slow partition competition explains routing in the optimal target; an efficient source-wise correction approximates the exact objective when response formats are well separated. Matched Qwen3-VL comparisons use identical candidate banks to isolate the objective from slow-teacher quality. DR-BOLT improves score-length trade-offs over the evaluated offline and online baselines on ScienceQA and three held-out transfer benchmarks, with a 24.1% cross-benchmark slow ratio and a 2.90-fold end-to-end training speedup over reproduced GRPO. These results support adaptive reasoning in the studied two-reference setting, while broader robustness remains open. Our code is available at this https URL.
https://arxiv.org/abs/2611.09154
High-resolution visual question answering (VQA) requires models to identify question-relevant small objects in high-resolution images and select the correct answer based on what they recognize, demanding fine-grained perception. Previous methods focus on obtaining better visual evidence (image crops), but obtaining it alone is insufficient, using it effectively is equally crucial. Visual evidence should therefore be utilized as effectively as possible throughout inference. We argue that small-target evidence is obscured by unconditional information that remains observable in subsequent processing even after the condition on which it should depend is removed. Such information hinders the effective use of small-target evidence. We propose R$^{3}$Sieve (Residual-Region-Response Sieve), a training-free framework. We divide inference into three stages---localization, cropping, and answering---and introduce distinct counterfactual references to suppress interference at each stage. During localization, the Shared-Subspace Canceller (SSC) uses a generic image description to construct a counterfactual reference and obtain question-relevant attention maps. During cropping, the Unique-Contribution Spotter (UCS) constructs counterfactual references from visual and non-visual features to identify distinctive, complementary crop regions. During answering, the Content-Aligned Voter (CAV) constructs an image-free branch to calibrate answer responses. Experiments on three high-resolution VQA benchmarks show that R$^{3}$Sieve achieves the highest overall accuracy among compared methods across diverse settings, improving base models by up to 21.5 percentage points.
https://arxiv.org/abs/2611.24060
Frontier post-training follows a common recipe: train one expert per domain with reinforcement learning (RL), then consolidate the experts into one model via multi-teacher on-policy distillation (MOPD). Such consolidation is expensive: the student samples in every environment again, an expert scores every sampled token, and the student inherits the drift of the final experts. Yet we argue that the experts' RL runs have already produced what consolidation needs: rollouts with their sampling probabilities and rewards, in the order the experts learned from them. We propose multi-expert replay policy distillation (MRPD), which replays these training logs through the student in the experts' order. At each step, the policy gradient is recomputed on these scored trajectories, and the final experts' distributions are distilled at the logged prefixes. Following the experts' own trajectory reduces the training-inference mismatch and keeps the student within the clip range, so the replay is off-policy in origin but quasi-on-policy throughout consolidation. MRPD samples nothing and calls no environment or reward model: the trainer is the sole process, and consolidating the experts costs only about a seventh of the compute used to train them. On three long-horizon agentic domains (reasoning, search and SQL), MRPD yields the best student model, substantially outperforming online RL baselines, and surpassing MOPD from the same experts at a fraction of MOPD's cost. Our analysis on the same logs shows that the replayed trajectories carry most of the gain, and that both the policy gradient and the distillation term add to it. Most interestingly, the order of the logs matters more than their amount or rewards, suggesting that expert training logs are a curriculum, not a data mixture.
https://arxiv.org/abs/2611.24059
Test-time compute has emerged as a powerful way to improve the accuracy of language models at inference time, with scaling laws quantifying how accuracy improves with additional forward passes. Existing scaling laws, however, focus on repeated sampling from a fixed prompt, even though many tasks admit multiple equivalent representations of the same underlying problem. A graph does not depend on how its nodes are numbered, and a program does not depend on the names of its variables, yet language models can change their answers under such rewritings. In this paper, we study test-time symmetrization, which uses forward passes on rewritings of the input and aggregates the resulting answers over the group $G$ of rewritings. We prove that no amount of sampling and repeated decoding from a single representation can replace rewriting, while only $O(\log |G|)$ random rewritings suffice to recover the information provided by all $|G|$ rewritings. The same set of $O(\log |G|)$ rewritings provides simultaneous approximation across models, decoding rules, and sampled reasoning traces, and the logarithmic scaling is optimal in general. We validate the theory with experiments on Qwen2.5-7B, Phi-4-mini, Llama-3.1-8B, and Qwen2.5-14B, with chain-of-thought reasoning, on graph reasoning, logical deduction, and multiple-choice question answering. At the same number of forward passes, spending them on rewritings of the input is substantially more accurate than spending them on repeated sampling from the original prompt whenever the model's answer depends on how the input is written, and the number of rewritings needed grows only logarithmically with the number of possible rewritings, as the theory predicts. Our results provide a principled way to enforce symmetries at inference time when they cannot be built into the model, which we call model-agnostic symmetrization.
https://arxiv.org/abs/2611.24032
Arm-free quadrupeds can perform loco-manipulation, walking on three legs while manipulating with a front leg, yet no existing policy for it follows language instructions. World action models (WAMs) have emerged as effective language-conditioned manipulation policies, but are pretrained for robot arms and sample slowly. WAM-Policy turns one WAM into a quadrupedal loco-manipulation policy via two stages. The first, Cross-Morphology Policy Adaptation and Acceleration, adapts the arm-centric WAM to quadrupeds and distills the transferred policy into a student with one video and two action steps by on-policy self-distillation (OPSD). This student, however, is bounded by its teacher, which only imitates demonstrations. The second stage, Task-Specific Improvement and Merging, therefore tunes the teacher by reinforcement learning (RL) and merges the slow specialists into one few-step student by multi-teacher on-policy distillation (M-OPD). On QLM-Bench, ten simulated tasks with 1,000 demonstrations synthesized without teleoperation, the transferred policy reaches 77.0%, and the OPSD student keeps 76.7% while sampling 14 times faster. Task-specific RL and M-OPD lift it to 85.0%. We also deploy the OPSD student on a real Unitree Go2 Air. WAM-Policy thus achieves a stronger quadrupedal policy at a fraction of the inference cost.
https://arxiv.org/abs/2611.24011
Feed-forward 4D reconstruction enables efficient creation of renderable driving scenes for autonomous driving simulation. However, most existing methods rely heavily on dataset-provided cameras, LiDAR depth, or 3D box annotations, limiting their applicability to unannotated driving videos. Pretrained visual models can predict cameras, depth, and object tracks from RGB alone, but their geometry is mutually inconsistent: multi-view predictions come in short windows with independent scales and frames, while monocular depth is metric only per image. Our key insight is that, once reconciled into a single metric frame, these priors can replace dataset annotations for feed-forward 4D reconstruction. We present \textbf{FreeForward4D}, which calibrates overlapping prediction windows against monocular metric depth and registers them into one frame across time. On this geometry, 2D tracks are associated across cameras in 3D, and objects move through hierarchical motion that combines a shared object velocity with local Gaussian corrections, so that both training and inference use only RGB videos. On Waymo, FreeForward4D achieves 27.66 dB PSNR and 0.836 SSIM while recovering metric-scale cameras and depth without scale alignment. It surpasses STORM, SLARM-F, UFO, DGGT, and StreetForward, which rely on dataset annotations, in full-image PSNR and attains the best RGB fidelity on 8-second windows. RGB-only fine-tuning further adapts the framework to nuScenes and Argoverse 2. Code will be released.
https://arxiv.org/abs/2611.24007
Existing latent reasoners typically run for a preset latent horizon during inference. Intuitively, a longer horizon should provide more opportunity for reasoning. However, in this work, we find that extending latent Chain-of-Thought (CoT) to the full preset horizon may in fact degrade reasoning performance. We observe that latent CoT evolves through distinct functional stages, with earlier computation forming frontier reasoning states, while later computation mainly reorganizes existing candidates and can weaken gold candidate support. These findings motivate a rethink of fixed-horizon latent inference. To address this, we characterize the reasoning benefits of latent CoT by realized and remaining constructive progress. Empirically, we find that higher realized constructive progress corresponds to higher current reasoning accuracy, while more remaining progress corresponds to larger future gains from continued latent CoT. To this end, we introduce a lightweight method that estimates the remaining reasoning progress and adaptively adjusts the latent horizon during inference. Across representative latent CoT reasoners and multiple arithmetic benchmarks, our method improves reasoning accuracy by up to 7.2 percentage points while reducing latent computation by up to 29.0%, without retraining the LLM.
https://arxiv.org/abs/2611.24001
Generative novel view synthesis (NVS) employs a diffusion model to render a static or dynamic scene from a target camera view. Explicit warp methods complete a reprojection of the source with video diffusion models, and implicit methods encode the camera parameters or 3D cues. Both add inference overhead, and the generated layout often drifts from the ground truth. We present WaveView, which injects the warped source into the phase of the initial noise of a pretrained generator. The Dual-Tree Complex Wavelet Packet Transform (DT-CWPT) allows local and anisotropic phase injection. The generator is thus constrained less in disoccluded regions and along the epipolar line. WaveView keeps the architecture and adds negligible overhead. On the DL3DV benchmark, it outperforms all baselines in every overlap bin. In the hardest bin, it gains 0.84 dB PSNR on the covisible region over MetaView, which shares its generator, while running 2.1 times faster. For dynamic NVS, fine-tuning AnyView with the same noise raises its zero-shot PSNR on AnyViewBench from 11.58 to 12.35 dB.
https://arxiv.org/abs/2611.23994
In latent diffusion models, expanding the latent dimension of variational autoencoders (VAEs) is essential to push the upper bound of reconstruction quality. However, without explicit structural constraints, such unregularized expansion easily yields a rugged and ill-conditioned latent manifold, causing generated latents to drift off the manifold during generation (i.e., off-manifold problem). Recent works smooth this manifold by using external foundation models (e.g., DINOv2) as a topological regularizer. However, we argue that tying the latent space to frozen external priors inherently restricts generalization: bound by their pre-training distributions and semantic biases, they may fail to capture faithful visual details on out-of-distribution domains. Observing that low-dimensional latents naturally form a smoother, generation-friendly manifold, we propose dual-manifold VAE (DM-VAE) to leverage internal representations as a topological regularizer. Specifically, DM-VAE pairs a high-dimensional reconstructive branch with a low-dimensional branch that extracts a clean topological skeleton, anchoring the high-dimensional manifold to this smooth subspace. Building upon this framework, we further introduce topological representation alignment (TRA) to supervise intermediate generator features during training, and internal guidance with topological skeleton (IG_TS) to guide structural components while preserving high-frequency details at inference. On ImageNet-1K (256x256), our foundation-model-free DM-VAE achieves a state-of-the-art rFID of 0.12 and a generation FID50K of 1.28 (w/o CFG) and 1.20 (w/CFG), and further reaches 1.11 when orthogonally combined with DINOv2.
https://arxiv.org/abs/2611.23964
Proactive video understanding requires vision-language models (VLMs) to decide not only what to answer as a video unfolds, but also when newly arrived evidence warrants a response and when to remain silent. Although modern VLMs excel at offline long-video understanding, they struggle in proactive streaming settings, leading prior work to introduce task-specific reinforcement learning or auxiliary triggering modules. We find that explicit reporting rules and structured streaming context substantially improve a frozen VLM's proactive response performance. We exploit this observation to build SIMPLEPRO, a training-free baseline for proactive video understanding. At each decision step, the model combines recent high-resolution frames, bounded historical observations, and prior answer records to determine whether new information warrants a response and, when needed, generate a concise answer. The system then filters consecutive duplicate answers while retaining the raw outputs as context for subsequent decisions. Without any parameter updates or auxiliary triggers, SIMPLEPRO with a 4B backbone consistently outperforms specialized proactive systems across OVO-Bench, StreamingBench-PO, and ProactiveVideoQA. SIMPLEPRO shows overall performance gains as backbone size increases from 2B to 32B. We hope this simple yet effective method will serve as a strong baseline for proactive video understanding, and we will release our code and inference pipeline to support future research.
https://arxiv.org/abs/2611.23954
Classifier outputs provide a fundamental and univer- sally available knowledge interface for knowledge distil- lation (KD). However, standard logit objectives entangle transferable semantic relations with architecture-dependent parameterization artifacts, causing a classifier geometry mismatch that makes direct knowledge transfer unreliable. To address this, we propose Angular Conditional Distilla- tion (ACD). First, the Angular Coordinate Interface (ACI) reconstructs the student classifier using angular coordi- nates, suppressing magnitude and bias effects to provide a geometry-compatible space. Then, Scale-Invariant Condi- tional Distillation (SICD) standardizes outputs and exclu- sively transfers conditional non-target relations, filtering out untransferable absolute confidence. Together, ACI de- fines where to transfer, while SICD defines what to transfer. Extensive experiments show that ACD achieves state-of-the- art performance over existing KD methods. Moreover, as a universal drop-in replacement, it consistently improves existing distillation pipelines while requiring no feature adapters or additional inference-time cost. Code is avail- able at this https URL.
https://arxiv.org/abs/2611.23941
Noise search at inference time improves text-to-image diffusion but requires many multi-step teacher rollouts per query. Distilled few-step models are now routinely released, yet their potential as proxies inside noise search is uncharacterized. We measure proxy-teacher rank agreement across three distillation regimes and three backbones (DiT, MMDiT, UNet) and find a sharp asymmetry. On our main pool, closed-form image statistics transfer from the proxy to the teacher (Spearman $\rho$ up to 0.80), while learned semantic scorers do not ($\rho \le$ 0.12). Paired null controls trace this rank to the first denoising step for coarse statistics and to distillation for finer ones. We use the asymmetry for noise search under a joint constraint: a semantic reward specifies content, and a numeric target specifies image statistics. In PROXYSEARCH, the proxy filters candidates by statistical match in 1 step, and the teacher returns the survivor with the highest semantic reward. No weight trades one objective against the other. On FLUX-1024, PROXYSEARCH tightens the statistical target by 21-23% at the same number of teacher trajectories, with no detected loss on the optimized reward.
https://arxiv.org/abs/2611.23929
Kernel Ridge Attention (KRA) is a recurrent sequence mixer that combines learned nonlinear features with covariance-guided memory updates. It stores associations through recursive ridge regression, using correlations among past keys to determine how new observations modify memory. A content-dependent Ornstein--Uhlenbeck transition controls retention and restores uncertainty for future writes. This construction admits an exact temporal-kernel interpretation and maintains an inference state whose size is independent of context length. Associative block maps and explicit adjoints support parallel training while preserving the recursive update. In controlled language-modeling experiments, KRA improves prediction quality over attention and recurrent baselines, strengthens natural associative recall over the evaluated recurrent models, and achieves practical distributed training throughput. These properties provide a principled foundation for nonlinear, adaptive sequence memory.
https://arxiv.org/abs/2611.23896
Visual proxies decouple explicit world-state simulation from photorealistic video synthesis, offering a controllable interface for both offline clip creation and real-time interactive streaming. Yet a practical renderer must address two challenges. Upstream proxies vary from coarse layouts to fine articulated geometry depending on the interaction, and streaming inference must operate in a few causal steps without accumulating drift across self-generated history. We present Invoker-Render, pairing a bidirectional creation model with a causal four-step streaming model under a shared conditioning interface over structural proxies, an appearance reference, and language. Trained on an aligned approximately 200-hour multi-granularity corpus, Invoker-Render honors fine geometric constraints when supplied while synthesizing plausible local motion under coarse proxies, enabling runtime selection of input detail. To bridge bidirectional creation and real-time streaming without compounding step compression and rollout drift, we introduce Progressive Forcing (ProForcing). ProForcing first trains a causal predictor and distills it into a four-step renderer using histories derived from training videos. Training then shifts to continuous self-generated rollouts. A pipelined serving stack supports efficient streaming inference. Experiments show versatile control across proxy families, granularities, and camera paths, consistent gains over recent streaming baselines, and stable 30-minute racing rollouts that preserve scene and subject appearance.
https://arxiv.org/abs/2611.22925
Multi-step visual reasoning suffers when early visual misinterpretations persist as incorrect premises in subsequent reasoning. We argue that how visual evidence is represented during reasoning matters as much as how much of it is gathered. Rather than appending observations and inferences to an ever-growing dialogue, we introduce a Visual Board-centered harness that separates visual observations, uncertain interpretations, and derived reasoning within a structured, capacity-bounded, and revisable workspace. This separation makes each premise explicit and auditable, so that conflicts between what is seen and what is derived become detectable even by a text-only critic and can trigger targeted reinspection of the original image. Without task-specific training, our harness consistently outperforms standalone models across 4B, 8B, Qwen3-VL-Flash, and 32B configurations on Euclid30K, SolidGeo, and ChartQA bar charts, with gains of up to 11.68, 18.42, and 24.82 percentage points, respectively. Under identical tools, critic, and interaction limits, the structured Board outperforms shared free-form notes by 6.67 to 13.00 points and no-Board retries by 7.35 to 18.80 points while using 7.9\% to 38.2\% fewer tokens. Supervised fine-tuning on harness trajectories further improves direct answering and critic-free tool use, with positive transfer to chart reasoning.
https://arxiv.org/abs/2611.23872
Audio artificial intelligence (AI) now recognizes speech, sounds, and music remarkably well, yet voice agents still act when they should not: they execute commands addressed to others, act on misheard content, and commit before the user finishes speaking. This position article argues that the bottleneck in agentic audio AI has shifted from recognition to decision-making. We propose a reference architecture with an explicit decision layer, separate from perception and reasoning, that chooses whether to act, wait, ask, or remain silent from evidence on addressedness, speaker identity, content confidence, tool reliability, utterance completeness, and action risk. We motivate the design through six recurring decision failures and a decision-theoretic argument that transcript confidence alone is insufficient to justify an action. We implement the layer with fuzzy inference, evolutionary optimization, and sequential decision-making, and evaluate it through simulations, ablations, and recorded audio, including egocentric recordings. Addressedness and speaker evidence matter most, risk-aware gating suppresses costly errors, and well-tuned policies perform similarly, with fuzzy gates more robust to overconfident addressedness estimates. On recorded audio, explicit gating cuts executed side talk from 41\% to 7--10\% while serving 83--85\% of valid requests under successful clarification, outperforming prompting-based baselines. These results support explicit action gating as an architectural requirement for reliable voice agents.
https://arxiv.org/abs/2611.23847
Advances in deep generative models now enable the synthesis of realistic imagery. However, aerial imagery is dominated by tiny objects, such as vehicles, ships, and aircraft, which occupy only a few pixels owing to the overhead, wide-area nature of aerial imaging. Existing generative models struggle substantially more with tiny objects than with larger ones. This limitation has been overlooked, as prior work on generative evaluation emphasizes prompt following or human preference rather than whether small objects are rendered faithfully. To address this gap, we introduce TOGBench, a quantitative, object-level benchmark for evaluating tiny-object generation fidelity in aerial imagery. We further propose two simple train-/inference-time techniques that improve the ability of layout-to-image generators to render small objects, raising their performance on TOGBench.
https://arxiv.org/abs/2611.23840
Reinforcement learning from terminal answer rewards enables search-augmented language models to interleave reasoning and retrieval without step-level annotations, yet outcome supervision assigns identical credit to every search in a trajectory regardless of whether it retrieved useful information. This limitation is especially severe when all sampled rollouts for a question fail, collapsing group-centered outcome advantages to zero even when trajectories gather markedly different evidence. We present Evidence-based Credit Assignment (ECA), which augments terminal outcome rewards with a training-only auxiliary signal measuring how retrieved documents improve the likelihood of a reference answer. Rather than scoring the full dialogue prefix, ECA queries a frozen snapshot of the current policy under an evidence-only prompt containing solely the question and the deduplicated documents observed so far, isolating the utility of retrieved passages from surface variations in generated thoughts or search queries. Within this framework, we compare two return formulations, ECA-Step (which accumulates step-clipped potential increments) and ECA-Horizon (which preserves the telescoping horizon gain in potential), paired with cross-fold group normalization, novelty gating, and search-local token assignment. Across seven Search QA benchmarks, ECA-Horizon improves strict macro exact match over an outcome-only GRPO baseline by 1.03 and 4.80 percentage points on Qwen2.5-3B- and 7B-Instruct, respectively, outperforming both RLOO and GiGPO. Zero-shot evaluation on MultiHop-RAG further demonstrates strong transfer to multi-hop inference queries.
https://arxiv.org/abs/2611.23824