Coding agents now write whole codebases in long autonomous runs, guided mainly by whether their tests pass. Yet passing tests alone does not reveal accumulating complexity or gaps in test coverage, which then persist as agents extend their own code. We introduce metrics-in-the-loop, an inference-time intervention that gives agents in-flight feedback: cheap, automatically computed code metrics injected into the agent's context during the trajectory. After each code-changing agent turn, we insert short feedback messages naming changed functions whose cyclomatic complexity exceeds a limit (complexity feedback) or changed lines that are skipped by tests (coverage feedback), giving the agent the opportunity to refactor code or strengthen tests before continuing. We run extensive experiments and ablations comparing several interventions across three long-horizon coding benchmarks and four strong models. We find that complexity feedback significantly reduces erosion, a measure of complexity concentration of code, without affecting held-out correctness of the code. Broadly, our experiments investigate a practical form of agent "re-railing", where timely, actionable measurements redirect agents before weaknesses become the foundation they keep building on. As agents scale to codebases with millions of lines, such interventions become critical to sustaining effective development and long-term code maintainability.
https://arxiv.org/abs/2611.23817
Rigid motion of a patient head during head cone-beam CT (CBCT) scan can degrade image quality and complicate recovery of accurate anatomical structures from the measured projections. We present a 3D flow-matching prior approach that learns the image changes induced by motion. Its geometry-bridge connects motion-corrupted and motion-free Feldkamp-Davis-Kress (FDK) reconstructed images of the same source volume through progressively reduced simulated motion. The forward projector's geometry derivative supplies analytic velocity targets along this path. A plug-and-play loop encompasses the learned image updates together with spline-based pose estimation, conjugate-gradient data consistency, and image total-variation regularization. Both the refined image volume and the FDK reconstructed image volume at the estimated poses are evaluated. On simulated 360-view scans from 30 held-out CQ500 patients with nominal per-component motion amplitudes within 10 mm / 10 degrees, the framework achieved state-of-the-art image reconstruction quality and motion accuracy in comparison with other methos such as learned autofocus and diffusion-based joint reconstruction. Mean peak signal-to-noise ratio (PSNR) reached 37.1 dB and mean detector-plane reprojection error (RPE) was 0.27 mm. With matched endpoints at the inference settings, the geometry bridge improved PSNR by 1.6 dB and reduced RPE by 33% relative to image-linear training. With 50 prior updates and coarse-to-fine motion estimation, total processing time averaged 9.1 minutes per patient on an RTX A6000.
https://arxiv.org/abs/2611.23788
Vision Transformers (ViTs) now underpin the visual encoders of most multimodal models, but this success comes at a cost: relying on third-party pretrained checkpoints leaves these systems open to backdoor attacks. Existing test-time detection methods rely on either the predicted label or intermediate embeddings, limiting their applicability across diverse tasks and incurring substantial computational overhead. In this work, we study how backdoor triggers affect the multi-head attention mechanism of ViTs and identify two distinct anomalous patterns: $\textit{complete attention hijacking}$, causing the attention to exhibit abnormally high inter-head spatial overlap, and $\textit{partial attention hijacking}$, causing an abnormally large variance in the concentration of the attention distribution. Motivated by these findings, we propose $\textbf{DetectViT}$, a test-time backdoor detection method that requires neither training data, model outputs, nor any prior knowledge of the trigger. DetectViT quantifies the two hijacking phenomena via an inter-head consistency score and an inter-head entropy variance score, with thresholds estimated from a small out-of-distribution (OOD) calibration set drawn from an arbitrary source. Across four representative attacks, diverse architectures (DeiT, CLIP, and LLaVA), and downstream tasks (image classification and captioning), DetectViT substantially outperforms all baselines with as few as 64--128 OOD calibration samples, achieving a true positive rate of $100\%$ and a false positive rate of $0.17\%$ on BadCLIP. Notably, it introduces negligible additional overhead by relying solely on attention weights already computed during inference. Code is available at: this https URL.
https://arxiv.org/abs/2611.23770
Streamline patterns are ubiquitous in nature. Prominent examples include white matter fibre tracts in the mammalian brain and arrangements of elongated cells, such as cardiomyocytes in the heart wall. It has been posited that such organisations foliate into locally smooth surfaces to optimise function. However, the recovery of these foliations from a single vector field aligned with the local orientation of the underlying sheets remains an open problem. We address this challenge by introducing a mathematical model of a sheet and developing an algorithm for sheet inference, from measurements of the oriented structures. Given an input vector field, our model estimates a second vector field that is orthonormal to it, while also requiring the Lie bracket of the two to vanish. To balance sheet likelihood against low surface area, we also reduce the mean curvature of the recovered local surface. We compute the first variation of an associated energy and provide a numerical method which admits an efficient, parallel implementation for sheet estimation. We demonstrate our approach on both synthetic and biological examples, motivated by applications to medical imaging. These include white matter fibre tracts from a brain atlas and confocal microscopy images of cleared heart wall tissue. For the latter we are able to reconstruct sheets aligned with local cardiomyocyte directions at unprecedented (micrometre) scale.
https://arxiv.org/abs/2611.23765
Language-model reasoning alternates between sustained activities and local operations. We study how this temporal structure predicts continuation and the computation needed to analyze it. T/E/S combines text, local events, and spans of activity. Across reasoning and information-seeking tasks, events and spans improve next-outcome prediction in 70 of 76 evaluation groups for a Transformer and 74 of 76 for Qwen with LoRA. Span duration and preceding events are associated with transitions, and retaining event history across span boundaries improves prediction. We formulate budget-allocation counting and action-conditional continuation queries, prove counting hardness under explicit encodings, and derive conditions for affordable decision certification. In finite-process experiments, bounds that cancel shared continuation mass recover the same optimal actions as full dynamic programming while reducing mean inference CPU time by 77.7%. On 71 held-out HotpotQA and MuSiQue questions, a compact structural observer improves prediction of whether a fixed inference procedure finishes within a CPU allowance: Brier loss falls from 0.0836 to 0.0756. An annotation audit identifies substantial errors in that observer, separating this computational result from semantic validation of its features.
https://arxiv.org/abs/2611.23749
Dynamic spatial reasoning from monocular video enables vision-language models (VLMs) to understand how 3D environments evolve over time, supporting physical world understanding and embodied intelligence. Existing efforts primarily address observed-dynamics understanding, whereas agents in dynamic environments also require future-dynamics prediction beyond the observation boundary. Moreover, external geometric processing and textual descriptions can increase inference complexity or lose information when representing continuous dynamics. These limitations motivate a shared internal state that preserves observed dynamics and remains predictive beyond the input video. Accordingly, we present Pros4D, a framework for modeling evolving 4D worlds in latent space that unifies retrospective understanding of observed-dynamics with prospective reasoning for future-dynamics prediction. Specifically, Pros4D-Engine derives geometry-grounded supervision and constructs paired examples for both tasks, while Pros4D-Bench evaluates them across eight spatial task categories spanning dynamic cameras, objects, and scenes. At the model level, Pros4D compresses observed dynamics into latent observed states and autoregressively rolls them forward into latent future states. 4D Latent Alignment (4DLA) grounds these states through complementary visual-semantic and geometric supervision, while joint language supervision connects them to task reasoning. Experiments across benchmarks show consistent improvements over strong baselines in both capabilities.
https://arxiv.org/abs/2611.19803
Public benchmarks support comparisons of large language model (LLM) capabilities, yet users can experience the same highly ranked model differently. Drawing on traditional software reliability, we argue that the reliability of an LLM service should be assessed for specified users and conditions of use. This requires more than a capability score: evaluation evidence must support a claim about failure during future operation. We distinguish testing to discover faults from testing to assess reliability and identify ten modelling considerations behind such claims. We propose a six-workstream research roadmap linking claim specification, dynamic operational profiles, graded outcomes and noisy judgements, local probabilistic robustness, interaction dependence, and statistical inference with reassessment. Local probabilistic robustness addresses the behavioural variation hidden when different concrete inputs are treated as the same task. Published findings and an illustrative customer-support example motivate the roadmap; its integrated validation remains an open research task. Reliability claims complement, but do not replace, separate safety and security arguments. Capability scores should therefore be complemented by versioned operational claims that state whose requirements are covered, under which conditions, and with what uncertainty.
https://arxiv.org/abs/2611.23732
Diffusion and flow-matching policies have become a standard choice for visuomotor imitation learning, owing to their ability to model multimodal action distributions. In memory-dependent manipulation, the robot must act on information that is no longer visible. Existing methods feed the observation history to the policy as an additional condition, while action generation still starts from Gaussian noise. However, the policy must then pick out the relevant target from a long and redundant history, leading to spurious correlations and unreliable decisions as the memory load grows. Additionally, the initial noise carries no information about what the robot has seen, leaving generation without guidance toward the remembered target. To address these issues, we propose RecallFlow, a flow-policy framework that builds the source distribution from episodic memory. Recall-Relocalization relates the memory to the current view through a few recall slots, so that the source reflects where the remembered targets are now. Recall Anchors then turns each slot into a Gaussian component of the source, enabling generation to start close to the intended action instead of from noise. The transport network then completes the action from the current observation alone, keeping inference nearly as cheap as a memoryless policy. We further build a suite of memory-dependent tasks grouped by memory load and compare all policies under a matched protocol. Extensive experiments show that RecallFlow achieves a state-of-the-art average success rate of 80.4% on the public MIKASA-Robo memory benchmark, and outperforms the strongest baseline by 20.3 points on our suite and by 18.8 points on four real-robot tasks.
https://arxiv.org/abs/2611.23731
Spiking Neural Networks (SNNs) promise event-driven and energy-efficient vision, yet existing spiking vision transformers (S-ViTs) largely confine spiking tokens to fixed and uniform firing schemes along the temporal, spatial and channel dimensions, whereas suppressing the computational redundancy thus incurred calls for adaptive firing mechanisms that operate jointly across all three axes at test time. To this end, we propose the Adaptive Efficient Integrate-and-Fire (AdaE-IF) model, which consults the attention score of each spiking token and routes salient regions to fine-grained firing patterns while assigning energy-frugal coarse patterns to the remaining regions. Building on AdaE-IF, we further introduce Spiking Gated Self-Attention (S-GSA) and Spiking Local-Gated Unit (S-LGU), which enforce selective spike emission at token and channel granularity. We integrate the aforementioned mechanisms and modules into AdaS-ViT, a test-time adaptive-routing architecture trained with a progressive spatio-temporal curriculum that stabilizes the co-adaptation of firing precision and routing masks. Extensive experiments across benchmarks spanning diverse data scales and modalities demonstrate that AdaS-ViT consistently achieves highly competitive performance. Notably, on ImageNet-1k it consumes only 1.02 to 2.45 millijoules per image and attains an accuracy-per-energy ratio of 32$\times$ to 72$\times$, operating in an economical inference regime and thereby pushing the energy-efficiency boundary further.
https://arxiv.org/abs/2611.23728
Once deployed on vehicles, end-to-end driving models continuously encounter challenging scenarios, generating valuable human takeover data that capture both model outputs and human behaviors. However, human interventions are often imperfect, as drivers may intervene too late or unnecessarily. Consequently, direct imitation can reinforce late responses or discourage otherwise appropriate maneuvers. To address this challenge, we propose an offline post-training framework that transforms imperfect takeover data into proactive supervision. We first construct diverse trajectory candidates by augmenting the plans of the deployed model and shifting recorded human corrective maneuvers earlier in time. We then train a scene-conditioned trajectory scorer to rank these candidates using expert demonstrations and unpaired preference signals from takeover logs. Using this scorer, we selectively replace pre-takeover training targets with higher-scoring candidates while retaining the original targets in normal driving scenarios. Finally, we distill the resulting targets into the driving model without changing its inference architecture. In closed-loop evaluations using real-world scenarios reconstructed with 4D Gaussian splatting, our method achieves the highest overall success rate, outperforming the deployed model, direct imitation learning and trajectory post-optimization. These results highlight a practical path toward efficient updates of deployed driving models using existing fleet logs.
https://arxiv.org/abs/2611.23719
The growing volume of topically similar papers makes it difficult to identify new work that can offer useful research insights or methodological guidance. We present KnowMe Paper Radar, a lightweight discovery tool that combines multiple sources of evidence into personalized reading priorities. Users specify topics, priorities, exclusions and dates; the system brings together topical relevance, recency, author research histories, collaboration paths and direction-specific institutional context. By exposing each score component and its supporting sources, KnowMe helps readers select promising papers for their current questions and discover relevant work beyond familiar authors. Automatic acquisition, version merging and graph-cache reuse keep this workflow convenient without large language model (LLM) inference during ordinary use. Experiments on 174,640 acquired records show more than threefold faster repeated context preparation, while cached abstracts improve normalized discounted cumulative gain at rank 10 (NDCG@10) from 0.412 to 0.598 on expert-judged retrieval tasks. Complementary analyses examine how the same accumulated records support semantic retrieval, research-history modeling and literature-grounded planning. A browser interface and a structured command-line interface (CLI) make these capabilities available to both researchers and research agents, positioning KnowMe as a practical upstream tool for literature-informed recursive self-improvement (RSI) workflows.
https://arxiv.org/abs/2611.23702
The rise of the agentic web has been accompanied by growing demand for on-device Vision-Language Models (VLMs) that perform diverse cross-modal web tasks with enhanced privacy and reduced latency. Mixture-of-Experts (MoE) models are widely adopted in modern VLMs for compute-efficient scalability, yet their multi-expert architecture poses deployment challenges due to substantial memory demands. Mixed-precision quantization is an effective compression technique that reduces bit-widths while adapting to the heterogeneous hardware budgets of edge devices, particularly well-suited for MoE due to its inherent structural sparsity, which causes imbalanced importance among experts. While prior work has applied quantization for efficient MoE inference at the edge, we point out that in MoE-based VLMs, the modality-specific characteristics of visual and text tokens further impact expert quantization. To address this issue, we present MoMoVL, a modality-aware adaptive compression framework for MoE-based VLMs that brings vision-language inference within reach of on-device resource budgets. Motivated by our empirical findings on modality-induced disparities across experts, MoMoVL jointly exploits the per-expert activation profile and per-modality quantization sensitivity to guide bit-width assignment. Experimental results show that MoMoVL improves the accuracy-compression trade-off relative to state-of-the-art mixed-precision baselines and reduces memory footprint by up to 15.44% with a modest drop in accuracy over its uniform-precision baseline. Our code and data are publicly available at this https URL.
https://arxiv.org/abs/2611.23695
The pursuit of improved performance in deep learning has often been accompanied by increasing architectural complexity and computational demands. We explore an alternative direction: achieving more with less through structured supervision. We introduce LightSEAD, a lightweight SE(3)-aware keypoint detector initialised with calibrated Harris pseudo-labels and enriched through viewpoint-aware and appearance-aware supervision and selective knowledge distillation. LightSEAD comprises only six convolutional layers and 84,289 learnable parameters, achieving competitive repeatability across HPatches, KITTI 06 and 4Seasons while demonstrating its downstream applicability to relative pose estimation on MegaDepth-1500. Further optimised for mixed-precision inference, LightSEAD Coarse achieves approximately 909 FPS on an NVIDIA RTX 4090 and 139 FPS on a Jetson Orin Nano Super.
https://arxiv.org/abs/2611.23661
Directly aligning trajectory-distilled video generators slashes rollout costs and optimizes the deployed few-step model without teacher fine-tuning or re-distillation. However, conventional flow-matching preference objectives assume single-head local velocity supervision, mismatching the finite-interval, block-average transition interface of trajectory-distilled generators such as Parallel Decoding Distillation (PDD) and Parametric Trajectory Distillation (PTD). Naively supervising individual unfused sub-interval velocity heads across all sub-steps rather than the fused block-average velocity at block starts (DirectAlign) induces a severe training--inference mismatch that degrades preference below the unaligned baseline. We introduce BlockAlign, an interface-consistent preference optimization framework that constructs states at valid block boundaries, queries the fused block-level velocity predictor, and supervises span-normalized transition residuals---seamlessly supporting DPO-, NFT-, and rollout-style objectives while preserving the native few-step sampler. We further build a unified post-training pipeline integrating offline pairs, reward-free semi-online continuations, and online rollouts with hierarchical preference selection and objective-aware block scheduling. Experiments on multi-reference video generation across MiniMax-H3 Ref2VA and an in-house model RefVid-V4 show that BlockAlign consistently outperforms unaligned distilled models and DirectAlign in structural plausibility, reference consistency, and visual quality. Notably, our approach unlocks PDD's 5x few-step rollout speedup for 3x faster end-to-end online training, achieving statistically significant human preference gains over PDD baselines (54.0% vs. 33.3% wins/losses, or 61.8% non-tie win rate on RefVid-V4; 51.3% vs. 34.7%, or 59.7% non-tie win rate on MiniMax-H3) and up to an 82.7% win rate against DirectAlign.
https://arxiv.org/abs/2611.20351
One-step flow maps are increasingly used as posterior samplers, where the point of sampling is the uncertainty that the samples carry. We find that the one-step sampler loses this uncertainty and becomes less stable. Compared with multi-step samples of the same network, its samples lose variance and decorrelate. Supervising the one-step jump directly makes this worse; the samples collapse further toward the conditional mean and break down when training is extended. We trace both failures to the locations where training places its pairs of start and end times. Only pairs on the diagonal of the time triangle receive targets from data, and the pair used for one-step sampling is either left unsupervised or supervised on an island, a set of pairs that the self-consistency identity does not connect to the diagonal, where the objective neither determines the one-step map nor anchors its training. We bridge the island with an isthmus, which places 5% of the training pairs along the whole edge of the triangle that joins the one-step pair to the diagonal. The error of the one-step map is then bounded by the self-consistency residual along this edge. With the network, the loss and the inference code unchanged and at no added cost, the isthmus improves one-step sampling across several flow-map objectives, in MRI reconstruction, face inpainting, and unconditional generation on CIFAR-10 and ImageNet 32x32. On brain MRI, it restores one-step calibration to 0.96 and the correlation with multi-step samples to 0.93; both vary far less across seeds, and the agreement holds when training is extended to three times its length.
https://arxiv.org/abs/2611.23616
How much the teacher talks and when are among the most informative and costly-to-measure signals of teaching. Yet, classroom audio is among the most sensitive data an AI system can process: recordings capture minors and fall under GDPR, preventing raw audio from leaving the recording device. We present a teacher-student segmentation pipeline designed to meet exactly this constraint: label-free at inference, running on a recording device, and requiring no per-session calibration. The pipeline chains a lightweight voice activity detector, compactly pretrained speaker embeddings, agglomerative clustering, and a small logistic regression that assigns the teacher role based on session-relative structure rather than voice identity. We show that our structural cue transfers across classrooms, languages, and acoustic conditions. In leave-one-teacher-out cross-validation on 292 recordings from 63 teachers across English and German corpora, our pipeline achieves teacher F1 of 0.81 and 0.87, respectively, outperforming both enrollment-based baselines and a 30-second supervised enrollment ceiling. On a commodity smartphone, its performance matches pyannote, a strong neural diarization tool, while running 4.8x faster at comparable peak memory at a real-time factor of ~0.18. We also find that transcribing filtered audio only marginally increases teacher's word error rate from $0.26$ to 0.28. Residual failures concentrate where teacher and student voices are not acoustically separable. However, we show how to flag these sessions label-free at inference so that they can be routed to human review, enabling selective, privacy-preserving deployment.
https://arxiv.org/abs/2611.23609
Unlabeled test data can improve the zero-shot classifier of a vision-language model, yet transductive and test-time adaptation methods often fall below the zero-shot baseline when only a few vocabulary classes are present in a batch or when streams are correlated. We trace this failure to a shared design: every class owns independent parameters estimated from its own samples, so absent or rare classes are left at biased text anchors or fitted to foreign samples. We propose Graph-Regularized Anchor-displacement Field Transduction (GRAFT), which models the adapted prototypes as text anchors plus a displacement field that is partially pooled over a semantic class graph through a Gaussian Markov random field prior, together with an explicit class-presence model. Inference is closed-form and gradient-free on cached features. Anchor-only shrinkage and a single global shift are the two extreme settings of the prior; absent classes are provably harmonic extensions of the present ones, and the pooling strength can be chosen by empirical Bayes. Under the realistic batch and online protocols on eleven benchmarks and four ImageNet shifts, GRAFT improves over the zero-shot model in every setting and over anchor-based transduction where few classes are present.
https://arxiv.org/abs/2611.23591
Food insecurity creates a pressing need for guidance that helps households find usable food resources and assess food choices in light of their circumstances. Such guidance must connect eligibility requirements, practical access constraints, and dietary needs rather than return isolated listings or generic nutritional advice. Multi-agent food assistants can coordinate these tasks, with uncertainty indicating when clarification or verification is needed before a recommendation. However, uncertainty can be lost across agent handoffs, allowing tentative interpretations to propagate as facts and obscuring opportunities for correction. Moreover, confidence that a service requirement has been interpreted correctly is not confidence that a recommendation satisfies the user's needs; local confidence cannot simply be reused as downstream reliability. We introduce an inference-time, receiver-conditioned clarification framework that preserves shared premise identities and replays consistent interpretations across dependent agents. Each receiver exposes admissible decisions and remaining evidence obligations, so question relevance reflects both decision changes and missing support. Information questions and user-confirmable fallback adjustments are evaluated together with their anticipated follow-up interactions under shared budgets. This structure connects clarification to jointly supported plans for resource navigation, nutritional reasoning, and supply-side planning.
https://arxiv.org/abs/2611.23572
When LLM agents pass intermediate judgments downstream without uncertainty, later decisions may treat tentative assumptions as facts. Effective uncertainty propagation must preserve information for correction and update inherited beliefs as new evidence arrives. Yet scalar scores omit information needed for revision, while detailed descriptions and reflection incur additional costs. We introduce Latent Uncertainty, a latent-space estimator that preserves uncertainty-relevant information without lengthy descriptions. Building on this, we propose LProp (Latent Uncertainty Propagation), which revises inherited answer distributions using latent representations and new evidence without additional trajectory sampling. We evaluate both methods across multiple-choice, yes/no, open-ended question answering, and mathematical reasoning tasks. Our approach outperforms baselines on most metrics, while LProp achieves the lowest mean inference-token usage in all eight propagation settings. Matched feature ablations further demonstrate that inherited latent content improves error recovery beyond answer probabilities or model capacity alone.
https://arxiv.org/abs/2611.23564
Large vision-language models (LVLMs) and language models (LLMs) often exhibit attention sinks, whereby particular positions, such as the beginning-of-sequence (BOS) token, receive a disproportionate share of attention. Under standard causal masking, however, the BOS query attends only to itself, so the representation that later tokens repeatedly access carries no information about the input. We introduce BOS Summary, a training-free inference-time method that repurposes this persistent pathway to carry input context. At a single transformer layer during prefill, BOS Summary relaxes the causal constraint for the BOS query row alone, letting BOS aggregate all visual and textual input tokens into a self-derived context representation while every other token retains standard causal attention. The enriched representation then remains accessible to the model throughout decoding, with no additional parameters, auxiliary modules, or decoding-time overhead. Across multiple VLM and LLM families and eight multimodal and text-only benchmarks, BOS Summary improves over direct and Chain-of-Thought prompting and outperforms recent sink-intervention methods on multi-step reasoning tasks. A mechanistic analysis shows that making BOS input-specific also weakens its sink dominance and redistributes attention toward content-bearing prompt tokens, indicating that changing what a sink carries reshapes how later layers allocate attention.
https://arxiv.org/abs/2611.23508