Vision-Language-Action models (VLAs) enable general-purpose embodied intelligence, but their performance is brittle under distribution shift, where unanticipated failures can damage equipment or endanger people. Deploying VLAs in real-world systems therefore requires a runtime monitor that is timely, calibrated, and informative while being efficient enough to run on the edge. Existing methods fail to fulfill these requirements. We present VLA-Tug, a lightweight probe over the latent layers of a VLA backbone that improves early failure detection and bounds the missed-detection and false-alarm rates, with minimal inference overhead. We show that supervising on the pre-failure segment of a trajectory increases early failure detection at matched false-positive rates by up to 30% over the strongest learned baseline. Under shift to unseen tasks, the bounds on missed-detection and false-alarm rates remain valid in up to 86% of runs, versus at most 21% for baselines. Unlike existing methods, VLA-Tug predicts the failure mode to enable corrective actions, raising top-3 accuracy on unseen tasks by up to 44% over a majority baseline. VLA-Tug enables real-time monitoring where the well-timed and informative probe supports lightweight steering that improves task success. We release a dataset of 3,180 annotated manipulation episodes to support future work on latent failure detection in VLAs at this https URL.
https://arxiv.org/abs/2611.08425
Vision-Language-Action (VLA) models enable end-to-end robot control, but inference is increasingly expensive. Each decision processes hundreds of visual tokens, often across multiple cameras, leading to substantial computation while allowing irrelevant background content to distract action prediction under distribution shift. We argue that these issues stem from two complementary forms of redundancy: many visual tokens are spatially irrelevant, while attention over the remaining tokens is often misallocated. Based on this observation, we propose T3R, a Training-Free Two-Stage Token Refinement framework. In the first stage, segmentation-guided token pruning removes irrelevant visual patches to reduce computation and eliminate distracting context. In the second stage, saliency-guided attention biasing reallocates attention toward action-critical visual regions and language instructions within the retained tokens. Together, pruning produces a cleaner candidate set, while attention biasing further emphasizes the most decision-relevant evidence, yielding a better accuracy-efficiency frontier than either stage alone. On LIBERO-Spatial, T3R removes 85% of scene-camera visual tokens while essentially matching the unpruned success rate, and improves success by up to 7 percentage points under out-of-distribution object-placement perturbations. The timing comparison attributes 7.8 ms per inference call to pruning and attention-bias application. As input resolution increases, LLM-forward profiling shows up to a 3.3x speedup over the unpruned baseline. Evaluations across LIBERO, SIMPLER, and RoboTwin with three VLA backbones show the strongest accuracy-efficiency benefits in multi-camera settings. These results demonstrate that training-free token refinement is a practical approach to efficient and robust VLA inference.
https://arxiv.org/abs/2611.08413
Self-consistency (SC) is a widely used decoding strategy for improving chain-of-thought reasoning, but it requires sampling many complete reasoning trajectories with a preset budget. Its efficient variants reduce the cost mainly by stopping full-trajectory sampling early or distributing the number of samples across questions. In this paper, we propose StepBack, a step-level error-localized resampling method for efficient test-time reasoning. Given a small set of draft trajectories, StepBack localizes likely first error positions using lightweight signals and then resamples the detected error step and its suffix, followed by majority answer voting. Empirically, StepBack matches the accuracy of SC on average across seven mathematical reasoning and multi-hop QA benchmarks, while reducing token usage by up to 41.26%. In addition, by nature, StepBack can be combined with adaptive inference methods such as Adaptive Consistency, further improving their inference-time efficiency.
https://arxiv.org/abs/2611.08403
Oral cancer and the oral potentially malignant disorders (OPMD) that can precede it are visible on the oral mucosa, which makes smartphone photographs a promising tool for screening. Deep learning models for this task classify individual photographs. However, screening refers patients, not photographs: a patient is usually photographed several times, and the referral decision must miss no more than a set share of patients with disease. We propose OraRefer, a patient-level screening model that reads each photograph with a segmentation-guided foundation model that conditions its class decision on the predicted lesion, combines all photographs of a patient with risk factors, and refers through a threshold with a finite-sample sensitivity guarantee. The model localizes the mouth itself, so inference needs no annotation. On 3000 photographs of 714 patients, patient-level fusion reaches an AUROC of 0.919 and outperforms attention-based multiple-instance learning. At a 90% sensitivity target, OraRefer detects 95% of patients with OPMD or cancer and misses 3 of 61 patients with cancer, at a specificity of 0.610. Because the threshold is calibrated without retraining, each clinic can set its own guaranteed operating point on a local calibration sample.
https://arxiv.org/abs/2611.08360
Sentiment Analysis (SA) is a fundamental NLP task concerned with the identification and categorization of subjective information in text. While lexicon-based methods offer transparency, they often struggle with linguistic phenomena such as negation and intensification. Syntactic parsing has been integrated into SA pipelines to address this, but conventional parsers introduce significant computational overhead. Recent work shows that reformulating dependency parsing as a Sequence Labeling (SEL) task can improve efficiency. We propose and compare two strategies for integrating syntactic knowledge into a neural sentiment analysis model under unified Multi-Task Learning (MTL). The first, Standard-MTL, jointly learns dependency parsing (via sequence labeling) and sentiment polarity classification using a shared encoder with two task-specific heads. The second, Contrastive-MTL, provides implicit syntactic guidance through a contrastive auxiliary objective and a syntax-biased attention pooling mechanism, without requiring explicit label prediction. Both rely on syntactic annotations during training and eliminate need for hand-crafted compositional rules or external parsing at inference. We evaluate eleven datasets spanning seven languages---the MARC corpus (English, Spanish, French, German) and seven datasets from SemEval-2022 Task 10 (English, Spanish, Norwegian, Basque, Catalan)---using four established SEL encodings. With the relative encoding, Contrastive-MTL outperforms a syntax-naive baseline on 10 of 11 datasets and surpasses the external rule-based pipeline (SELSP) that parses at inference on all six datasets where the latter applies, while being 1.5--9$\times$ faster at evaluation. Standard-MTL is more dataset- and encoding-dependent. These results show implicit syntactic supervision during training can match or exceed explicit parsing at inference without computational overhead.
https://arxiv.org/abs/2611.08323
Automatic sleep stage classification from electroencephalography (EEG) signals is a critical task in clinical sleep medicine. While deep learning models have shown promising results in this domain, their performance is fundamentally constrained by the quality of training labels, which human experts manually annotate and that are subject to inter-rater disagreement and systematic labeling errors. In this paper, we propose a novel LLM-assisted label-noise filtering pipeline that operates exclusively during training to identify and remove unreliable sleep-stage annotations before retraining the classifier. Specifically, TinySleepNet is first trained on the original dataset to extract per-epoch SoftMax confidence scores, then low-confidence epochs are submitted to LLaMA 3.1 along with their EEG spectral features and surrounding sequence context. The LLM evaluates each epoch against AASM sleep-scoring guidelines and assigns a verdict of RELIABLE or SUSPICIOUS. Suspicious epochs are removed from the training set, and TinySleepNet is retrained on the cleaned data. Evaluated on the Sleep-EDF Expanded dataset using 20-fold cross-validation, our method improves overall accuracy from 81.8% to 84.6% and Macro F1-score from 75.6% to 77.4%, with particularly notable gains in REM detection. Crucially, because the LLM operates only during training, the lightweight TinySleepNet architecture (1.3M parameters) is preserved at inference, making the proposed pipeline suitable for real-time, wearable sleep monitoring applications.
https://arxiv.org/abs/2611.08311
Manim supports precise, programmatic video generation, but a successfully rendered educational animation can still have visual defects, missing explanations, or incorrect content. Rendering does not directly identify the code changes needed to correct these visual failures, so feedback-driven revision is needed. Existing agentic frameworks use iterative rendering and revision but typically rely on large models, often proprietary. Existing Manim datasets primarily contain final programs and often lack quality verification and the trajectories of feedback and revisions needed to learn this refinement process. We introduce ManimData and ManimCode, a data and training framework for feedback-driven educational animation generation. We construct ManimData through a render-guided pipeline that iteratively generates, executes, reviews, and revises Manim programs with retrieved API documentation. Its 24,034 render-verified, quality-filtered records preserve intermediate programs, execution errors, visual critiques, and corrective revisions. We use role-conditioned supervised fine-tuning on these trajectories to train ManimCode, a single 4B model, to write, review, and revise animation programs. To improve correction of errors that differ from those in the collected trajectories, we then train the Reviser with reinforcement learning on self-generated states, using rendering outcomes and reference-free quality feedback. ManimCode achieves 96.6% and 97.0% render success on ManimData-test and ManimBench, compared with 52.6% and 82.0% for the untrained 4B model under the same inference setup, while also improving video quality scores from 2.48 to 2.94 and from 3.00 to 3.44. Integrating ManimCode into two long-form educational video generation systems increases completed videos from 19/117 to 107/117 and from 53/239 to 239/239, alongside improvements in educational effectiveness and visual quality.
https://arxiv.org/abs/2611.08294
Neural theorem proving increasingly uses retrieved premises and project context to support proof generation. Evaluating these resources requires measuring the advantage of target-selected declarations over comparable alternatives and how this advantage depends on training. We introduce LeanContext, a benchmark of 129,826 Lean whole-proof generation problems, to investigate these questions. We group selected complete preceding mathematical declarations as Mathematical Context (MC) and the remaining selected local source commands as Environment Context (EC). Four aligned input views support crossed training and inference evaluations with shared targets and original Lean verification environments. For both diffusion- and autoregressive-based models, selected declarations outperform comparable alternatives by a larger margin when training includes MC. Final-input audits identify omitted declarations referenced by the target. Prioritizing these dependencies improves Pass@32 by 1.40% and 1.70% for LLaDA and Llama, respectively, on 2,000 targets. The selection experiment operates without access to target reference proofs and holds checkpoints, EC, and inference budgets constant. The code and dataset are available at this https URL.
https://arxiv.org/abs/2611.08277
Reinforcement learning (RL) has become a standard post-training method for diffusion models, but most RL objectives roll out the current policy at every iteration, and with a many-step sampler these rollouts typically dominate the cost of training. Few-step models cut this cost, yet they model an average velocity over an interval, whereas existing RL objectives are built for an instantaneous velocity with a many-step sampler. We adapt RL objectives to few-step models systematically in three phases, namely a transformation, a decomposition and an adaptation. The transformation to $x$-space expresses each objective we consider through a common policy prediction, at the clean end by default and at an explicitly marked level for transition or flow-map objectives, so that objectives stated over densities, velocities or rewards become comparable. The decomposition writes the gradient of each of these objectives as the sum of a self-regression toward a reference and a reward augmentation, so that a method reduces, at the gradient level, to a choice of anchors and weights. The adaptation computes each anchor from a few-step model at the end of its transport or along its tangent, with the instantaneous velocity recovered from the network itself or from an auxiliary network. This yields few-step versions of FlowGRPO, AWM, DiffusionNFT, DiffusionOPSD and ReFL in one modular codebase. In our experiments, reward-gradient-based methods are the most effective when the reward is differentiable and its gradient is affordable; otherwise DiffusionNFT-style methods generally reach high rewards in relatively few iterations despite a sometimes higher per-iteration cost. Extra inference steps show no systematic aggregate benefit in our experiments, although selected outputs visibly improve with two to four times as many sampling steps as in training.
https://arxiv.org/abs/2611.08240
Self-attention carries a quadratic cost in token count, which limits how far Vision Transformers can scale with input resolution. We introduce SSoG, a separable sum-of-Gaussians spatial operator that replaces dense self-attention with a learned mixture of axis-factorized Gaussian kernels, reducing the per-layer cost from $\mathcal{O}(N^2)$ to subquadratic in $N$. Unlike dot-product self-attention, whose Lipschitz constant is unbounded outside a compact input domain, SSoG with fixed kernels is globally Lipschitz continuous, which, via the Picard--Lindelöf theorem, guarantees that the continuous-depth variant has a unique solution on the whole integration interval, a global guarantee that dense-attention counterparts lack in general. On CIFAR-100 at $224\times224$, in a matched discrete-Transformer setting and in the continuous setting, SSoG improves Top-1 accuracy by 2.2--4.1 points across three model widths while using 8.8--16.6\% fewer parameters, with depth and training recipe held fixed. In a looped-Transformer setting, SSoG improves Top-1 accuracy by 5.4--8.1 points. We further probe inference-time scaling by evaluating on inputs up to $1024\times1024$ at matched batch size: in the continuous-depth (Neural ODE) architecture, SSoG reaches up to 154\% higher throughput and over 60\% lower peak memory than dense attention, with the gap widening as resolution grows, consistent with the difference in asymptotic cost. In the discrete and looped settings, SSoG also achieves over 30\% higher throughput than dense attention at this resolution. Together, these results establish SSoG as a practical, parameter-efficient, and theoretically grounded alternative to dense attention across discrete, weight-shared, and continuous-depth Transformer variants.
https://arxiv.org/abs/2611.08236
Person identification from millimeter-wave (mmWave) radar point clouds has primarily focused on gait. However, since indoor walking is typically brief and fragmented, gait-based identification is largely constrained to designated passages that support continuous walking. Meanwhile, non-gait activities of daily living (ADLs) are pervasive and also encode discriminative personal signatures. In this study, we evaluate the feasibility of ADL-based identification from sparse mmWave point clouds using mm-ADL, a dataset comprising seven core ADLs collected from 11 subjects under controlled protocols. Moving beyond uniform gait introduces significant activity-dependent variations that confound identity representations. To address this challenge, we systematically benchmark three activity-identity learning paradigms: monolithic learning, joint multi-task learning, and two-stage activity-conditioned routing, under both closed-set identification (ID) and subject-disjoint re-identification (ReID) biometric settings. In parallel, we propose Dual-Stream Static-Dynamic PointNet (DS-SDPNet), an architecture tailored to radar point clouds that captures complementary spatial morphology and motion dynamics. Across seven ADLs, DS-SDPNet achieves 70.2% mean closed-set identification accuracy and outperforms existing baselines on both mm-ADL and mmGait. Moreover, benchmarking the three paradigms without ground-truth activity labels during inference yields up to 68.0% closed-set accuracy and 75.4% ReID mAP, providing a systematic baseline and confirming the feasibility of point-cloud-based biometric identification across diverse daily activities.
https://arxiv.org/abs/2609.08818
World-action models adapt pretrained video generators, and vision-language-action models build on pretrained vision-language backbones; both tie policy architecture and inference cost to large transformers designed for generation rather than control. It remains unclear how much these generation-oriented choices affect control. In this work, we introduce Latent Action JEPA (La-JEPA), a compact world-action model whose policy transformer is trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder. Following Fast-WAM, La-JEPA learns to predict future features jointly with actions during training and generates only actions at inference. Among policies without embodied pretraining, La-JEPA achieves state-of-the-art performance in clean and randomized RoboTwin 2.0 scenes and on LIBERO-Plus with correct instructions, while using far fewer parameters than the strongest competing policies. It runs substantially faster than Fast-WAM at comparable performance and outperforms the closest JEPA-based methods, LeapBot-WA and JEPA-WAM, on LIBERO-Plus and RoboTwin 2.0 without their pretrained policy backbones. On four real-robot tasks, its average success is comparable to pi0.5 and above non-embodied-pretrained baselines. Through a design study of the choices behind La-JEPA, we show that JEPA features outperform video VAE latents, most clearly out of distribution, and that training-time future prediction choices significantly impact the policy even when future prediction is skipped at inference. We correct the LIBERO-Plus instructions, as the officially released evaluator corrupts them for most variants, and re-evaluate several strong open baselines. Without a generative pretrained policy backbone, La-JEPA is better and faster than far larger policies, suggesting that robot foundation models may benefit from architectures designed for control rather than inherited from generation.
https://arxiv.org/abs/2611.08156
Tokamak experiments record diverse measurements, including waveforms, profiles, fields, and scalars, which are used to reconstruct the plasma, forecast its evolution, recover associated control signals, and estimate disruption risk. Most models are built separately for one task and a fixed set of inputs and outputs, and each relies on a prescribed subset of diagnostics rather than structuring the full multimodal observation set to extract global, cross-signal information about the discharge. These limitations motivate a reusable foundation model that adapts to each task's input-output schema and models heterogeneous target signals jointly, rather than retraining a separate model for every configuration. In this work, we extend the TokaMark benchmark with two groups of tasks for actuator action inversion and disruption prediction, yielding 20 MAST tasks in six groups with explicit schemas and evaluation protocols. Then, we introduce GenToka, a generative tokamak foundation model that jointly generates schema-selected targets through a shared conditional diffusion Transformer, with signal-specific codecs preserving native time series, profiles, fields, and scalars; foundation pretraining and task-specific adaptation transfer cross-signal structure across all six groups. Evaluation on MAST measurements shows that GenToka reduces macro-average normalized root mean squared error relative to TokaMind by 4.7% on Groups 1-4 and by 17.7% on Group 5, and improves macro-average window-level F1 on Group 6 by 8.4% at a common validation-selected threshold. GenToka provides a reusable, generative, and multimodal pathway for experimental tokamaks, linking plasma-state reconstruction, actuator inference, and disruption warning from shared measurements to support more integrated, data-driven fusion plasma prediction and control research.
https://arxiv.org/abs/2611.08153
Tree-based speculative decoding accelerates large language model inference by verifying multiple candidate branches in parallel. However, in practical serving scenarios, similar requests repeatedly induce overlapping candidate branches. While retrieval-based methods mitigate this by reusing historical drafts, the resulting candidate trees often contain many low-quality branches as the candidate tree grows, which reduces verification efficiency and wastes target-side computation. To address this, we propose ValuSpec, a plug-and-play candidate valuation mechanism that filters low-quality branches from a supplied candidate tree before expensive target-side verification. ValuSpec introduces transfer-based supervision to align filtering with the target model's behavior. A filtering threshold is estimated from the target model's greedy trajectories on a calibration split, without any parameter training. The target model then verifies the filtered tree under the output-identical greedy rule, so accepted tokens are guaranteed to match standard autoregressive decoding exactly. Experimental results on HumanEval, GSM8K, Dolly, and SpecBench show that, in our main evaluation, ValuSpec achieves up to 2.38x end-to-end speedup in retrieval-based tree settings. Furthermore, integrating ValuSpec into the EAGLE-3 dynamic tree framework yields up to 2.78x end-to-end speedup without modifying its tree construction strategy, demonstrating that ValuSpec serves as a generally applicable and plug-and-play candidate valuation module.
https://arxiv.org/abs/2610.31338
LLM routing aims to reduce inference costs by assigning easy queries to less expensive models and reserving stronger models for cases in which they are needed. However, we find that existing routers often violate this principle under loose cost budgets: as the budget increases, they increasingly select the strongest and most expensive model even when less expensive feasible candidates provide comparable or identical utility. We refer to this behavior as strong-model over-selection under loose budgets. Our analysis shows that ambiguity alone does not explain this phenomenon. Substantial overpayment can arise when differently priced models offer equivalent or near-equivalent utility and the router retains a persistent model-level preference for costly models; budget relaxation exposes this preference by making premium candidates feasible. Motivated by this diagnosis, we propose EquiRouter, a decision-aligned router that learns query-dependent model rankings, preserves meaningful utility differences, and explicitly prefers lower-cost models among equivalent candidates through a Cost-Aware Ranking Objective and a lightweight Query--Model Interaction Representation. Experiments across diverse text, multimodal, continuous-utility, and large-model-pool settings show that EquiRouter consistently improves cost efficiency while maintaining competitive peak and whole-curve performance. Code is available at this https URL.
https://arxiv.org/abs/2602.03478
Self-hosting AI is often presented as a way to reduce inference costs and increase control, but deployment introduces its own costs. We report our experience deploying a hybrid AI stack at Grayhat Developers, a small software company with an AI-intensive development workflow. Our initial motivation was to counter the growing cost of commercial AI usage: Claude/Codex incurred approximately \$325 per month, with more than 25.93 billion recorded tokens. We experimented with local open-source models as replacements for commercial coding agents, but after learning that their performance and inference speed were insufficient, we adopted a hybrid approach in which commercial agents remained responsible for complex tasks while local models handled lower-complexity tasks, organizational information retrieval, and synchronization with self-hosted tools including Outline, Plane, and ERPNext. We estimated that approximately 30--40\% of our commercial AI workload was associated with context transfer, documentation synchronization, task management, and related operational coordination. By moving this work to local models, we estimate that approximately 90\% of this coordination workload could be handled without commercial inference, corresponding to an estimated 27--36\% reduction in overall commercial-AI workload or cost-equivalent usage. Our first internal deployment nevertheless failed to achieve adoption because it introduced UX friction. Replacing it with a centrally configured deployment substantially improved feedback and adoption. We conclude that local inference is most useful not as a wholesale replacement for commercial AI, but as part of a hybrid architecture that assigns tasks according to their capability requirements, cost, context demands, and operational sensitivity.
https://arxiv.org/abs/2611.08116
World-action models (WAMs) jointly model future traffic states and driving actions for end-to-end autonomous driving in intelligent transportation systems. However, their action learning remains dominated by expert-trajectory imitation, leaving unsafe alternatives close to expert trajectories insufficiently constrained. To improve WAM robustness in safety-critical scenarios, we propose NegWAM, a safety post-training framework for unified WAMs that learns explicit safety boundaries from near-expert failure trajectories. NegWAM constructs negative trajectories that are geometrically close to the expert trajectory yet carry safety risks, uses them as repulsive anchors to steer predictions away from dangerous behavior, and encodes them as scene-conditioned rider tokens to learn scene--trajectory compatibility. During training, NegWAM jointly optimizes the flow-matching and near-expert negative objectives, while the auxiliary rider branch is discarded at inference, incurring no additional overhead. On NAVSIM v2, NegWAM achieves 40.4 EPDMS on navhard, a 4.1% relative improvement over the strongest compared scorer-free WAM baseline, and achieves 93.3 EPDMS on navtest. These results demonstrate that NegWAM achieves strong performance in safety-critical scenarios while preserving routine-driving performance. Zero-shot closed-loop evaluation on HUGSIM further demonstrates cross-benchmark generalization. The code is available at this https URL.
https://arxiv.org/abs/2611.08111
Diet is a major determinant of health, yet accurate dietary assessment from free-living meal images remains challenging. Existing retrieval-based Vision-Language Model (VLM) approaches to nutrient estimation typically rely on textual food representations, while, to our knowledge, direct grounding on visually similar annotated meals has not been investigated. We present NutriVisRAG, a retrieval-augmented generation framework that uses a vision encoder to retrieve similar meals with known nutritional annotations and provide them as multimodal in-context exemplars for VLM nutrient estimation at inference time, without task-specific fine-tuning. We evaluate NutriVisRAG across four datasets, four VLMs, three retrieval encoders, and two prompting strategies, including free-living meals from adults with insulin-treated diabetes, where carbohydrate estimates are compared against two dietitians' annotations. Across all 96 configurations (4 datasets x 4 VLMs x 3 retrieval encoders x 2 prompting strategies), NutriVisRAG reduced the primary estimation error in 56, with improvements more frequent under whole-meal (35/48) than per-item prompting (21/48). Across the evaluated configurations, the lowest NutriVisRAG error was below the best Vanilla error in 13 of 16 VLM--dataset pairs. For GPT-5.4-mini, under whole-meal prompting, the best-performing retrieval encoder for each dataset reduced the primary error on all four datasets relative to Vanilla, with reductions of 1.7--35.7%. Overall, our results show that grounding in multimodal exemplars comprising images and respective annotations can improve nutrient estimation at inference time without task-specific fine-tuning, although the benefits remain configuration-dependent.
https://arxiv.org/abs/2611.08054
Geospatial foundation models are a widely adopted starting point for remote sensing tasks. Most are pretrained on Sentinel-1/2 and Landsat imagery at a ground sampling distance (GSD) of 10m or coarser. Many applications instead rely on very high resolution (VHR) imagery, where individual buildings, roads, and amenities are resolved. We study which pretraining regimes transfer well to this scale. We introduce VHR-Bench, a benchmark suite of ten datasets at 1.5m GSD or finer spanning five task families, eight application domains, and both satellite and aerial platforms. We evaluate twelve encoders in three pretraining tiers with frozen backbones at two training-data fractions. We find that both pretraining on large natural-image collections and VHR-specific pretraining transfer effectively to VHR tasks, with complementary strengths. SkySense leads when labels are scarce; at the full label fraction the web-pretrained DINOv3 ViT-L edges ahead, with the two close together there. Encoders pretrained only on Sentinel-scale imagery transfer least well under RGB input. Guided by these complementary results, we continue the self-supervised pretraining of DINOv3 ViT-L on 1.02 million held-out VHR tiles and evaluate the resulting encoder without changing the downstream protocol. The continued model gives a clear improvement over the released backbone when labels are scarce and a modest one at the full fraction, at no increase in inference parameters or compute. These findings support a generalize-then-specialize strategy: already strong, broad natural-image pretraining can be further improved by a short VHR continuation stage while preserving its general visual features.
https://arxiv.org/abs/2611.08003
Reasoning trajectories generated for the same problem often arrive at conflicting answers. These disagreements reveal alternative intermediate decisions that can be reused to explore new solutions. We introduce Adaptive Semantic Branching (ASB), a training-free inference procedure that turns cross-trajectory disagreement into targeted candidate generation. ASB first groups independently sampled trajectories by extracted final answer, selects a representative from each group, and aligns their reasoning steps with a monotonic dynamic program. A persistent-divergence objective identifies the point at which initially similar reasoning paths develop different continuations. For each selected directed pair, ASB retains the source trajectory's textual prefix, extracts a short reasoning window from a competing donor trajectory, and generates a fresh solution conditioned on both. A diversity-aware policy allocates up to three additional generations according to the answer-cluster structure. In a 500-problem MATH evaluation with Qwen2.5-7B-Instruct, ASB increases oracle candidate coverage from 445/500 (89.0%) with seven initial samples to 455/500 (91.0%), recovering correct candidates for 10 of the 55 initially uncovered problems. The method uses 576 additional generation calls across the evaluation set, averaging 8.152 calls per problem. We provide a precise specification of the branching operator, alignment objective, generation policy, and experimental configuration.
https://arxiv.org/abs/2611.07918