Reliable tissue detection is essential for whole-slide image analysis, yet high average segmentation accuracy can conceal failures involving pale tissue, small fragments, and heterogeneous staining. Pathology foundation models offer transferable representations, but exhaustive inference is computationally demanding and may introduce false-positive tissue predictions. We propose SCOUT (Selective Correction of Uncertain Tissue), a risk-guided cascade that selectively refines an initial thumbnail-based tissue mask. A lightweight audit network estimates the spatial distribution of segmentation errors and prioritizes candidate regions for processing by a frozen pathology foundation model. An asymmetric consensus rule integrates the resulting predictions with the initial mask to recover missed tissue while limiting false positives. A calibrated slide-level review score additionally controls the marginal rate of unflagged case-group failures under the calibration assumptions. Evaluation on a public multi-source, multi-stain benchmark shows that SCOUT reduces the proportion of case groups requiring mask correction from 38.1% for the best released detector to 18.6%, with particularly pronounced improvements in immunohistochemistry. Source-held-out experiments demonstrate better generalization than a segmentation network trained on the same annotations, while selective inference achieves an observed failure rate no higher than processing all candidate regions. Downstream, lymph-node classification with SCOUT masks is non-inferior to that obtained with two of the three released detectors evaluated. A variant built on an openly licensed foundation model is publicly released as an installable Python package with trained weights. These results support risk-guided foundation-model inference as an effective approach to tissue detection across heterogeneous histopathology data.
https://arxiv.org/abs/2611.24366
Traditional long-horizon LLM agents often hand off context through full-history prompting, retaining all previous action--observation interactions. As execution proceeds, this history keeps growing, increasing inference cost and potentially degrading accuracy. Existing methods control this growth by preserving compact conclusions while removing earlier observations, but doing so can discard evidence that later becomes relevant. We introduce Semantic Context Handoff (SCH), whose central abstraction is state--evidence separation: a compact semantic state carries the agent's current interpretation, while supporting action--observation evidence remains separately addressable and can re-enter context when later execution makes it relevant, without separate summarization or retrieval-model calls. Empirically, SCH improves long-horizon agent performance while substantially reducing context cost. In a controlled 150-task SWE-bench study, SCH's evidence recovery improves Pass@1 from $80.0\%$ to $86.9\%$ at essentially unchanged token use and execution time. In the complete-system comparison, SCH achieves 86.9% Pass@1 while using 43% fewer tokens and 41% less wall-clock time than the strongest baseline by task performance. Across models and benchmarks, SCH consistently reduces token use relative to full-history prompting while remaining competitive with or improving upon the evaluated baselines. Trajectory analysis confirms that SCH recovers evidence from substantially earlier parts of the interaction history rather than merely preserving a recent context window.
https://arxiv.org/abs/2611.24357
Trajectory forecasting at urban intersections is often studied assuming that reliable bird's-eye-view (BEV) motion histories are available, which in practice requires costly resources such as LiDAR, aerial views, HD maps, or calibrated infrastructure. We address this upstream perception-to-forecasting problem by reconstructing and predicting agents' BEV trajectories only from paired opposite-corner RGB cameras mounted on existing urban infrastructure at variable heights at inference time. We introduce the Opposite-Corner Trajectory Reconstruction and Anticipation Architecture (OCTR2A), a modular pipeline integrating per-view detection with class refinement, geometric cross-camera matching, transformer-based BEV projection, and egocentric GRU forecasting, trained in cascade and mainly supervised with a soft multi-scale spatial-overlap loss. Since no public benchmark provides this sensing configuration, we introduce R3ST++, a Blender-based synthetic benchmark built from SIND trajectories that provides synchronized opposite-corner views, cross-view correspondences, and BEV ground truth, while also serving as a deployment-support simulator for site-specific fine-tuning. Experiments on these datasets show that, with 12 observed and 12 predicted frames at 10\,Hz, OCTR2A achieves projection~-~forecasting ADE/FDE of 0.144/0.140~-~0.154/0.241~meters in real time on a single consumer GPU. Evaluated on ground-truth BEV histories, the forecasting module improves over the strongest state-of-the-art methods on SIND by 66.7\%/41.7\% ADE/FDE. Transfer experiments on TUMTraf-V2X show that R3ST++ pretraining supports adaptation to real-world roadside domains with different sensing configurations.
https://arxiv.org/abs/2611.24352
AI coding agents have become one of the most popular applications of large language models. Yet, agentic AI comes at the cost of a tremendous amount of GPU compute resources, which limits its profitability and its viability at scale. In this paper, we investigate whether smaller, cheaper models can cut this cost without sacrificing accuracy. We evaluate 36 popular large and small open-weight models released since 2024 on a software engineering benchmark, and find that recent small models are closing the accuracy gap compared to their large counterparts. We investigate the structure of this gap and find that the gap is highly concentrated around hard tasks that small language models fail to solve while achieving remarkably good accuracy for easy and medium complexity tasks. Following this observation, we show that by utilizing a hybrid inference strategy that tries a small model first and escalates to a large one only when needed, we can achieve up to 33% reduction in total GPU-hour costs and 16.4% on all tasks from SWE-bench Verified, without sacrificing accuracy. Strikingly, this simple cascade nearly matches an oracle-style task router, highlighting a practical path for deploying cost-aware coding agents in production without complex learned task routers and with minimal engineering complexity.
https://arxiv.org/abs/2611.24324
Scientific problems have no answer key, so evaluating an agent means deciding what counts as a better solution before any agent is run. Existing benchmarks either match a symbolic answer against a key or, where they require executable work, score it by unit tests, by a rubric, or by position on a competition leaderboard. None of these is the criterion by which a field decides that one published method beats another. STEMBench adopts that criterion directly. Each of its twelve research problems in high-energy physics, astrophysics and neuroscience is scored by its own field's figure of merit, and each comes with two measured anchors: a submission that uses no domain knowledge, and an expert reference reproducing a published method, both obtained by running code shipped with the benchmark. The interval between them states how much of a problem is crossed by scientific insight rather than by engineering. Across eight frontier models, task-level wins and domain-level leadership dissociate: a model that wins five of the twelve tasks leads no domain mean. In the high-effort cross-model comparison, at least one model's mean score exceeds the expert anchor on seven of twelve tasks. The remaining five span uncertainty estimation, latent and causal inference, and cross-site clinical classification. Numerical scores and official-submission validity capture different aspects of performance and are reported separately. We release the tasks, graders, anchors and harness.
https://arxiv.org/abs/2611.24321
Domain adaptation for semantic segmentation is challenging when target-domain images are unavailable during training. Zero-shot domain adaptation addresses this by adapting models using only textual descriptions of the target domain. However, existing prompt-learning methods are not well suited for ResNet-based CLIP vision backbones or pixel-level cross-modal feature alignment. We propose CROMPT, a prompt-learning framework for zero-shot domain adaptation in semantic segmentation. CROMPT introduces Cross-Modal Feature Augmentation (CROMA), combining Random Visual Prompting with Hybrid Textual Prompting (HTP) to jointly adapt visual and textual representations. We further propose CROMA-VL, which leverages vision-language models to generate domain-conditioned prompts, and a progressive easy-to-hard adaptation strategy for improved domain invariance. Experiments on two benchmark datasets demonstrate consistent improvements over strong baselines and state-of-the-art methods, achieving 1.63-2.63% mIoU gains while maintaining inference-time complexity. These results establish CROMPT as an effective and efficient framework for zero-shot multi-domain semantic segmentation. OpenReview: this https URL
https://arxiv.org/abs/2611.24284
Vision-language models (VLMs) reason about long videos over hundreds to thousands of tokens. Every decoding step re-reads the key-value (KV) cache of the entire video. Must the model keep looking while it reasons? In LookAway, the model first writes timestamped notes on what the question needs and then puts the video away. Once the notes close, the inference engine hides the video's KV cache for the rest of the response. Across both hybrid and pure-attention VLMs, dependence on the video is concentrated early in the response. On Qwen3.6-27B, writing notes preserves accuracy better than free reasoning at the same withdrawal step. Replacing them with another video's notes harms accuracy, whereas keeping the video visible alongside the model's own notes adds no significant benefit. Without additional training, LookAway nearly matches full-video thinking accuracy at 0.12-0.13 times its modeled decode cost. To shorten this initial look, we fine-tune the model on its own correct LookAway responses. Reinforcement learning then rewards correct answers while penalizing KV reads. The trained model puts the video away after a median of 66-68 tokens. Across three held-out benchmarks, it writes more tokens than an identically trained full-video model yet reads only 27-35% as many KV positions at matched accuracy.
https://arxiv.org/abs/2611.24259
Filming a scene from an additional angle after production usually requires a costly reshoot. We present One Shot Any View, a system that takes a recorded clip and generates a reshot video of the same performance along a user-specified camera trajectory. As in recent geometry-grounded methods, the system conditions generation on the source video and an explicit 3D prior, namely a point video rendered along the requested trajectory from a feed-forward reconstruction of the source video. We target human-centric cinematic videos, which are often shot with long lenses, shallow depth of field, and low-key lighting, all of which degrade feed-forward reconstruction. Existing geometry-grounded methods tend to inherit the degraded point video, and their output quality drops visibly. Two designs make such point videos usable. First, in addition to rendered videos of 3D assets, we train on real-world multi-camera videos, whose point videos are produced by the same reconstruction pipeline as at inference. This training-inference consistency enables the generator to take only the composition of people and scene elements from the point video, and their appearance from the source video. Second, a geometry-anchored RoPE offset keeps the generated video temporally synchronized with the source while largely preventing the output from copying the degraded appearance of the point video. Meanwhile, since the generator is a pretrained reference-to-video (R2V) model, the system also accepts optional character and scene reference images, which supply the appearance of regions that the source never shows. In both objective and human evaluations, One Shot Any View outperforms existing reshooting methods.
https://arxiv.org/abs/2611.24253
Language-goal aerial navigation requires a UAV to reach a destination specified in natural language using observations collected during flight. A key challenge is linking changing egocentric perception with spatial evidence anchored to stable locations for long-range navigation. We propose CityNav-VLA, a dual-frame vision-language-action model for urban navigation integrating dynamic egocentric context with persistent allocentric spatial memory. The UAV incrementally accumulates visual observations on a fixed north-up map initialized with geographic and landmark priors, preserving evidence across viewpoint changes. A dynamic query conditioned on the instruction, current observations, UAV state, and motion history retrieves relevant memory evidence to predict a map-frame target. A deterministic controller converts this target into body-frame commands using the UAV's current pose, connecting global reasoning with egocentric execution. To reduce target fluctuations over long horizons, early goal commitment (EGC) fixes the execution target to the coordinate-wise median of the first $K$ predictions without extra trainable parameters. On Test Unseen, CityNav-VLA reduces NE by 21.2% and improves SR and SPL by 15.8% and 30.9%, respectively. It also reduces NE by 6.2% and improves SR and SPL by 5.6% and 23.2%, respectively, relative to the former SOTA after dataset refinement. Cumulative ablations show that navigation-map features yield the largest gains, while metric coordinates further improve endpoint accuracy. With a 4B-scale backbone, inference averages 0.68 s per decision with 8.6 GB peak memory on a single A100, presenting remarkable model efficiency.
https://arxiv.org/abs/2611.24247
Action-conditioned world models let robots imagine the outcomes of their actions for planning and offline policy evaluation, but these uses need many rollouts, which multi-step diffusion makes expensive. Drifting models move the iteration from inference to training and generate samples in one network pass. DriftWorld brought this paradigm to action-conditioned world models; to strengthen action following, it repels predictions from the current frame, a construction borrowed from classifier-free guidance. We show that this accentuation has a target bias: the push persists even when a prediction matches the recorded future, in kernel, balanced, and unbalanced Wasserstein drifting fields alike. We introduce Mass-Share Accentuation (MSA) to remove this bias. MSA treats the current frame as a counterfactual destination competing with the recorded future in a semi-relaxed optimal transport problem; subtracting the closed-form share of the correct future makes the guidance vanish at the target, and the remainder is shared by all predictions. We prove that the recorded future is the unique fixed point along the motion direction for collapsed predictions, and that the guidance fades with prediction error. On Push-T, Robomimic, and BridgeV2, where trained models under-react rather than overshoot, MSA improves action ranking, response slope, and moving-region error over latent-space baselines at no inference cost, though these benchmarks do not isolate bias removal as the cause.
https://arxiv.org/abs/2611.24222
Energy-saving mechanisms in RANs must balance energy reduction and network performance under varying traffic conditions. Conditional generative models support what-if analysis by generating plausible power-consumption trajectories for radio configurations and operating scenarios, helping operators evaluate energy-saving policies before deployment. However, representative measurements are costly and operationally constrained, so historical data cover only limited configurations and conditions, leaving some contexts poorly represented or unseen during training. In this paper, we study conditional generation of RAN power-consumption trajectories under decreasing training coverage, from well-represented to rare and unseen configurations. We first introduce a context-normalized conditional generator combining temporally aligned conditioning with configuration-dependent target normalization. It improves generation for well-represented configurations but becomes less reliable as training exposure decreases because target consumption is harder to estimate. We therefore introduce a prediction-based generative framework that decomposes power consumption into global prediction, residual level and scale correction, and temporal residual generation. Experiments on a commercial 5G RAN and a public dataset show that hardware metadata provides transferable information about residual scale, while limited target observations at inference improve generation through residual-level correction without retraining. An ablation study confirms that residual-level correction provides most of the gain. Both approaches outperform state-of-the-art methods for well-represented configurations, while the prediction-based approach also generalizes to rare and unseen contexts and, with one target observation, reduces median MAE by about 50% compared with TimeWeaver.
https://arxiv.org/abs/2611.24208
World Action Models (WAMs) connect future scene prediction with action planning for autonomous driving. Each driving record pairs an executed ego trajectory with its observed future, providing a factual outcome for learning this connection. We present TrajWAM, a trajectory-conditioned WAM that improves vision-language driving through auxiliary bird's-eye-view (BEV) future learning. An ego-trajectory query conditions DINOv3-aligned BEV predictions from shared planner representations. Training combines factual reconstruction under the executed trajectory with one-sided feature separation under an unlabeled perturbed trajectory. Together, these objectives train the shared planner to generate driving waypoints. Native inference directly generates waypoints and bypasses the auxiliary future-prediction branch. On all 220 Bench2Drive routes, TrajWAM achieves 91.53 Driving Score and 75.45% Success Rate, exceeding a schedule-matched planning-only fine-tuning baseline by 5.31 Driving Score points and eight successful routes.
https://arxiv.org/abs/2611.24192
Reliable uncertainty estimates are important for adaptive execution in vision-language-action (VLA) models. Existing approaches such as Adaptive Action Chunking (AAC) estimate action entropy from multiple sampled action chunks, introducing additional inference cost and capturing both uncertainty within an action trajectory and diversity across alternative strategies. We introduce Divergence-based Uncertainty-aware Adaptive Action Chunking (DUAC), a post-hoc method that extracts trajectory-conditioned predictive uncertainty from the vector field of a pretrained VLA. Building on a closed-form flow-posterior covariance formulation, DUAC derives action-level uncertainty scores from a single generated chunk using Jacobian-vector products, without retraining or multiple endpoint samples. DUAC strongly correlates with AAC, while capturing distinct information: AAC measures global dispersion across sampled chunks, whereas DUAC measures local uncertainty around the trajectory selected for execution. Consequently, DUAC is less sensitive to diversity among distinct but valid action strategies. Across LIBERO, LIBERO-Pro, LIBERO-Plus, and RoboCasa, DUAC-based adaptive horizon selection performs favorably among the evaluated methods, combining strong task success with low uncertainty-estimation latency. These results establish the flow vector field as an efficient, trajectory-conditioned source of predictive uncertainty for robot decision making.
https://arxiv.org/abs/2611.24178
Vision-Language Models increasingly rely on textual reasoning to solve advanced multimodal tasks. Despite its effectiveness, this paradigm faces two fundamental limitations. First, relying on textual reasoning alone can limit perceptual ability in vision-centric questions, where accurate answers depend on interpreting task-relevant visual evidence. Second, applying textual reasoning uniformly across inputs overlooks question-dependent variation in both reasoning modality and depth, often producing verbose and unnecessary traces when little logical inference is required. To this end, we propose Adaptive Visual-Textual Reasoning (AVTR), which introduces visual imagination as a complementary reasoning modality to textual thinking, enabling the model to revisit its visual observations and distill the salient evidence through generative visual reflection. Furthermore, we formulate reasoning-modality selection as a learned routing problem, allowing a single policy to adaptively harness the complementary strengths of visual reflection and textual reasoning in response to the perceptual and inferential demands of the input. Across 16 benchmarks, AVTR improves the average score by 4.5 points over Qwen3.5 baseline, while reducing inference latency by 33\%. It also scales effectively into a smaller size backbone, outperforming the evaluated 2B-scale baselines.
https://arxiv.org/abs/2611.24161
Estimating how an AI system will perform on a new benchmark need not require another large model. We use public evaluation counts to choose a weak Beta prior, then update it exactly from at most 31 revealed outcomes. The predictor uses no item semantics, pretrained weights, or inference API. Its single online evaluation in the AIMS competition returned a team-reported Brier ALC of 0.188477, a 24.6% reduction from the constant-0.5 reference. The complete submission is 7,754 bytes; a server microbenchmark measures 12.71 microseconds per prediction with 31 matching labels. We describe how source evaluations are compressed into 301 prior means and show why a concentration of two lets new evidence rapidly determine the prediction. The resulting posterior gives exact uncertainty about the modelled success rate without numerical integration or neural inference. A public evaluation covers 79 system-benchmark pairs and eight seeds, reaching Brier ALC 0.200250 for the submitted method. Accompanying code and sufficient statistics reconstruct the fitted prior byte for byte and replay all 632 evaluation trajectories. The result demonstrates how much predictive value a compact Bayesian model can extract from sparse target feedback at very low online cost.
https://arxiv.org/abs/2611.24159
Multi-modal image fusion combines complementary observations to improve scene visibility and perception. However, in real-world acquisition, differences in sensing principles and configurations introduce geometric discrepancies, while degradations further compromise source quality. Existing methods typically rely on explicit registration, yet large parallax complicates correspondence estimation, while mismatched fields of view leave some regions without cross-modal counterparts. Meanwhile, independent restoration underuses cross-modal information, whereas restoration-fusion methods assuming alignment remain vulnerable to misregistration. This study revisits misaligned and degraded image fusion in the era of generative foundation models, in which cross-attention enables position-agnostic complementary information aggregation, while generative priors support degradation recovery and inference in missing regions. Building on this perspective, we propose GenRF, which adapts pretrained FLUX through LoRA to jointly enhance degraded observations, infer missing cross-modal context, and fuse complementary information without explicitly estimating deformation fields. Extensive experiments on real-world infrared and visible captures show that GenRF delivers high-quality fusion directly from unregistered, degraded images without preprocessing, while improving downstream task performance under misalignment. Additional experiments demonstrate that GenRF can generalize to misaligned multi-exposure and multi-focus image fusion without task-specific fine-tuning. Project: this https URL.
https://arxiv.org/abs/2611.24157
Organizations increasingly deploy downloaded open-weight GUI agents. A backdoored agent passes acceptance tests on clean data and misbehaves only when a visual trigger appears on screen, turning an unvetted checkpoint into a supply-chain risk. However, existing sample-level detectors target image classifiers and CLIP-style encoders, leaving step-level backdoor detection for GUI agents unaddressed. On clean steps, the instruction drives the action; when an unreferenced screen region hijacks the step, the instruction loses control over the decision state or that region dictates it instead. We term this signature intent--evidence dominance inversion and propose IDEA (Intent-Dominance Evidence Audit) to detect backdoor activation at inference time for each step without trigger knowledge or attack data. Specifically, Intent Counterfactual Probing (ICP) measures how far redirected instructions move the decision state, and Region Causal Audit (RCA) quantifies how far ablating a candidate region moves it. A gated conformal score compares both displacements against clean steps at the same distance from the clean manifold, providing a finite-sample false-positive guarantee. Extensive experiments on Qwen2.5-VL-3B across five attack families show that IDEA consistently detects backdoor activation under a controlled false-positive rate and outperforms state-of-the-art detection methods, while localizing the trigger region for every alarm and adding only modest inference overhead.
https://arxiv.org/abs/2611.24102
In continual learning, regularization-based methods aim to mimic neuroplasticity by updating network weights to learn new tasks while constraining changes to parameters that are important for previously learned tasks through penalty terms. However, as the number of tasks increases, preserving knowledge from earlier tasks becomes less effective. This is mainly because updates for new tasks are distributed across the entire network, where parameters are highly interdependent due to the dense connectivity between neurons. Inspired by the Exponential Distance Rule (EDR), which states that the number of neural connections decreases with increasing physical distance, learning is naturally locally correlated while inter-area updates remain less task-dependent. Motivated by this principle, in this paper we demonstrate that applying weight regularization to multiple small and independently learned complementary representations is more effective than regularizing a single large shared representation. By decomposing the input into independently learned components, learning new tasks induces only localized parameter updates within specific representation, which limit interference and improve knowledge retention across tasks. We employ independent autoencoders that learn different parts of the input independently through unsupervised regularization. At inference time, a linear classifier is used for evaluation. The experimental results on a sequence of tasks using 5-Splits MNIST, sequence of 5-datasets, 5-Splits CIFAR-10, 10-Splits Tiny ImageNet shows competitive results with supervised and replay-based methods. Code is available.
https://arxiv.org/abs/2611.24093
Evaluating distributed large language model (LLM) serving on heterogeneous edge devices requires exploring diverse model partitions, device placements, and network conditions, making repeated physical experiments costly. An efficient simulator for edge LLM serving must support diverse, fine-grained placement configurations, capture runtime behavior, and accurately predict execution times while keeping simulation overhead low. We present LEDGE, a lightweight simulator for fast, fine-grained evaluation of distributed LLM serving on heterogeneous edge devices without executing tensor operations. LEDGE constructs event graphs supporting non-uniform tensor and pipeline parallelism, estimates execution costs using reusable performance models calibrated from end-to-end measurements, and simulates shared-link contention. We evaluate LEDGE on a heterogeneous testbed of 10 Jetson and Raspberry Pi devices, covering 17 model--precision combinations across 4 LLM families. On Qwen deployments, LEDGE achieves request-level mean absolute percentage errors below 3.6% and a 135.3 times aggregate speedup over physical inference. Performance model calibrated on one LLM family yield 7.44% zero-shot end-to-end latency MAPE across three other families. Further simulations show how shared-link contention affects placement decisions and network topology constrains scaling. Across four edge environments, LEDGE completes 26,773 simulation runs in 83.8 minutes, enabling efficient exploration of deployment strategies under diverse resource conditions and serving objectives. The source code is publicly available at this https URL.
https://arxiv.org/abs/2611.24085
Magnetic resonance imaging (MRI) is an essential medical imaging technique, but its high costs and long acquisition times can limit accessibility. Accelerated MRI reduces acquisition time through k-space undersampling, yet conventional methods often rely on predefined sampling patterns and uniform slice budgets, overlooking anatomical variations and inter-slice relationships. In this paper, we propose a volume-adaptive k-space acquisition framework that jointly learns where and how much to sample across an entire MRI volume. Using only acquired auto-calibration signal (ACS) measurements, our method exploits volume-wide anatomical information to adaptively allocate sampling budgets across slices while identifying informative k-space lines. A learned joint slice--frequency refinement module coordinates sampling decisions while preserving an exact total volume-level acquisition budget. To reconstruct the acquired measurements, we design a sensitivity-aware synchronous cross-slice reconstruction network that estimates coil sensitivities from ACS measurements and progressively refines neighboring-slice reconstructions while enforcing hard k-space data consistency. Both networks are jointly optimized using reconstruction-based supervision without requiring fully sampled k-space at inference. Experiments on knee and brain MRI across multiple contrasts and acceleration factors demonstrate the adaptive behavior of our acquisition framework and evaluate its reconstruction performance.
https://arxiv.org/abs/2611.24081