Fine-grained visual reasoning is essential for remote sensing in urban management, yet remains underexplored in the evaluation of multimodal large language models (MLLMs). We introduce UMBench, a fine-grained urban remote sensing benchmark built from unmanned aerial vehicle (UAV) images for multiple urban management tasks such as traffic monitoring, urban construction, environmental quality, and disaster warning. Evaluations on UMBench reveal that current MLLMs often rely on scene-level priors and overlook fine-grained visual evidence. To address this issue, we further propose a novel evidence-first chain-of-thought method that directs models to inspect fine-grained visual evidence in a training-free manner. Experiments show that our method can improve the performance of both open-source and commercial MLLMs on UMBench.
We study logistic regression for continual learning with linear predictors in a high-dimensional statistical setting. Existing theoretical work has largely focused on continual linear regression with variants of L2 regularization, motivated by their analytical simplicity and connection to the kernel regime of neural networks. We first derive the optimal L2 regularization strength for continual linear classification and characterize its generalization performance, recovering behavior analogous to prior results for continual linear regression. Next, we examine the effect of L2 regularization in deeper models using a minimal two-layer diagonal linear network. In the single-task setting, L2 regularization on the weights of such models is known to promote sparsity in the resulting predictor, yielding a rich, sparse feature-learning regime beyond the kernel regime (a similar effect arises in two-layer ReLU networks). However, in continual learning, where L2 regularization is typically centered on previous-task weights, this sparsity bias is no longer preserved. We therefore propose a weighted variant of L2 regularization that extends this beneficial sparsity bias to continual learning. We show theoretically and empirically that, for a sufficiently sparse teacher, our variant converges substantially faster than directly learning a linear predictor with conventional continual L2 or L1 regularization, trading some long-horizon bias for markedly improved performance over moderate task horizons.
Strong single-turn editing performance does not guarantee reliable multi-turn interaction: as edited outputs are repeatedly reused, small unintended changes can accumulate into severe visual drift. This long-term drift is often mitigated through training on multi-turn data or cross-turn rollouts, potentially altering pretrained priors, or through specialized inference-time mechanisms that are difficult to scale to foundation models. We introduce TailSteer, a final-stage training framework that improves multi-turn robustness while retaining the generative and editing capabilities acquired through earlier training. TailSteer follows a simple principle: keep early denoising dynamics close to those of the base model and concentrate preservation corrections near the trajectory tail, where fine-grained deviations can be suppressed. These corrections are learned from the editor's own on-policy denoising rollouts, using unchanged source content as a preservation reference without requiring curated ground-truth editing pairs. Experiments across the FLUX, SenseNova, and Qwen families show that TailSteer substantially reduces cumulative drift over 15-20 editing turns while largely preserving base-model performance on extensive text-to-image and single-image editing benchmarks. These results demonstrate that long-horizon editing robustness can be unlocked in foundation models after TailSteer, rather than by relearning their existing capabilities.
Scene Graph Anticipation (SGA) predicts object relationships in future video frames. Continuous-time SGA models evolve each subject-object pair with an independent Neural ODE, so relationships sharing an entity cannot influence one another during integration, and nothing prevents mutually exclusive predictions, such as "sitting on" and "standing on", for the same pair. We introduce CoupledSGA, which links such relationships through a Relationship Interaction Graph with graph attention inside the ODE vector field, and adds a Structural Consistency Loss ($\mathcal{L}_{SC}$) based on a matrix of mutually exclusive predicates. Developing CoupledSGA revealed a limitation of the evaluation protocol. With-Constraint evaluation keeps one predicate per group for each subject-object pair, so its R@K and mR@K cannot capture the within-group violations targeted by $\mathcal{L}_{SC}$: we prove they are unchanged whenever each group's winning predicate and its score are held fixed. We further show that the gradient of $\mathcal{L}_{SC}$ depends on predicate scores rather than ground-truth labels: for two-predicate rules it amplifies the ordering the model already holds, and when all scores in a group are equal it suppresses predicates in more exclusion rules more strongly, explaining why the penalty can lower recall while improving consistency. On Action Genome, activating $\mathcal{L}_{SC}$ after two warm-up epochs reduces R@20 from 42.21% to 35.18% while lowering the No-Constraint contradiction rate from 45.92% to 4.07% (matched-budget R@20: 40.48% CoupledSGA vs. 44.78% SceneSayerODE). CoupledSGA is a testbed, not a state-of-the-art claim: With-Constraint recall captures the cost of enforcing structural consistency but not the improvement. We recommend reporting No-Constraint results alongside a contradiction rate for methods that constrain predicate co-activation.
Small language models offer a compact basis for shell simulation without executing submitted commands, but full-output decoding requires them to predict recurring utility structure alongside content that varies across executions. We introduce template gap reconstruction (TGR), which supplies reusable output literals interspersed with sentinels to be interleaved with autoregressively generated gaps. A template selector chooses which template to be reconstructed, with the empty template retaining full-output decoding. We construct a template bank from 18.7 million outputs collected across heterogeneous execution environments and train a template selector and five reconstruction models initialized from various small language model checkpoints on a target environment excluded from bank construction. Evaluation demonstrates that TGR improves output edit similarity (O-ES) over full-output decoding (FOD) when the optimal template covers more than 25\% of the template, while performing similarly for lower-coverage outputs. Additionally, TGR outperforms FOD for shorter gap sequences, with O-ES values of 0.964 and 0.812, respectively, while maintaining similar performance for longer sequences. TGR reduces the number of tokens generated by 26.3--38.3\% across used tokenizer families. These results support using a reusable output structure as a practical design choice for small shell simulators, improving response fidelity and shortening generation targets while keeping template selection and residual prediction central to response quality.
Streaming video understanding requires models to repeatedly decide whether to respond or continue observing. Recently, Group Relative Policy Optimization (GRPO) has been applied to this setting, with existing approaches emphasizing reward design while retaining a single trajectory-level advantage across turns. Such shared credit can reinforce redundant replies after earlier successes and discourage useful follow-ups after earlier failures. We propose TurnWise, a turn-level GRPO framework that aligns learning with individual streaming decisions. TurnWise addresses the mismatch between trajectory-level feedback and turn-level interaction, promoting informative responses and reducing redundancy. Specifically, turn-level credit assignment uses silence-completed prefix scoring to construct group-relative advantages from each decision's remaining return, separating its learning signal from earlier rewards. Turn-mean policy optimization averages token losses within each turn and then across turns, giving response and waiting decisions equal weight regardless of output length. These complementary designs improve streaming policy learning without a learned critic or additional rollouts. Extensive experiments show our method outperforms prior RL approaches for streaming video understanding.
Continuous symmetries in neural networks lead to conserved quantities that are invariant under gradient flow, such as layer balancedness in linear and ReLU networks. Finite-step gradient descent need not preserve these laws, particularly at the edge of stability (EOS), where the iterates oscillate around a smoother trajectory, which can be described by the central flow dynamics (Cohen et al., 2025). We study how gradient flow conservation laws evolve under the central flow description of this regime. We show that the curvature correction induces a finite-dimensional drift determined by the oscillation covariance along critical parameter directions. We thus take the first steps to analyse which conservation laws remain and are broken at the EOS according to the central flow dynamics by analysing a scalar two-layer model, two-layer matrix factorisation, and a two-layer ReLU network, where the known conservation laws are shown to be complete (Marcotte et al., 2023).
Retrieval-augmented generation (RAG) assumes that language models can reason over relations stated in their context rather than fall back on associations memorized during pretraining. We study this evidence-conditioned relational reasoning in isolation, with the facts supplied directly in the prompt rather than retrieved. \textsc{Rel-Bloom} is a controlled multilingual diagnostic of 450 multiple-choice questions built from 30 Wikidata relations; the same source relations are instantiated at five Bloom-inspired levels (Remember, Understand, Apply, Analyze, Evaluate) and under three veracity conditions (Factual, Counterfactual, Neutralized). Across 43 models, six languages, and two multilingual settings (226,800 predictions), three findings stand out. First, accuracy drops sharply from Apply to Analyze in every evaluated open-weight model series, reaching 47.2 percentage points for the Large Granite bucket, but not for either Gemini model. Second, Evaluate items remain a bottleneck even for models near ceiling on the other task types: Gemini~3.1 Pro averages 99.5\% on the other four levels but 60.7\% on Evaluate. Third, item- and relation-level difficulty rankings are consistently preserved across languages, indicating a difficulty structure partly shared across languages within this benchmark. Across knowledge conditions, Counterfactual contexts are generally harder than Factual ones, while removing entity identities can improve higher-level performance in some model regimes. Greater model capability therefore does not by itself ensure reliable reasoning over supplied relational evidence. Relevant resources are available at: https://huggingface.co/spaces/QCRI/non-parametric-memory-relational-reasoning.
Deep generative models, particularly Flow Matching frameworks, have emerged as powerful tools for synthesizing complex physical systems. However, standard architectures remain purely statistically driven and agnostic to the underlying governing equations, often violating fundamental physical constraints during inference. Existing zero-shot approaches designed to enforce these constraints typically rely on rigid algebraic projections. While effective for sparse boundary conditions, applying such discrete geometric corrections to dense, full-domain PDEs may disrupt the probability path, \rev{potentially deviating from the learned manifold and introducing substantial computational overhead}. To address these challenges, we introduce Energy-Guided Transport, a novel training-free framework that formulates non-linear physical constraints as a smooth potential energy landscape. Rather than interrupting the ODE integration, our method applies a continuous dissipative force directly to the intermediate generative states, steering them toward the physical manifold. By evaluating this force implicitly via reverse-mode automatic differentiation, our framework preserves the natural dynamics of the flow while maintaining an efficient $\mathcal{O}(N)$ algorithmic complexity. We extensively evaluate our framework across various 1D and 2D physical systems, including the chaotic Navier-Stokes equations. Our method achieves stable generation without structural divergence, significantly improves the stability-accuracy trade-off compared to state-of-the-art methods, and substantially accelerates inference with respect to exact projection strategies. Code available at: https://github.com/ispamm/Projection-free-energy-guided-flow-matching.
Training an agent to play Pac-Man with only up, down, left and right actions raises a basic question: when performance is poor, what should be improved? Difficulty can arise from reading the screen or from choosing actions over a long sequence. We examine these problems in a Qwen3.5-9B-based agent through separate perception checks and primitive-policy reinforcement-learning studies. In the policy study, which uses textual boards to remove image transcription from the decision interface, completion on eight held-out mazes increased from 0% before this optimization to 80.2% at a selected checkpoint, with three extra lives. The perception checks show that accurate static screenshot reading can coexist with errors during interaction. A separate policy comparison also shows that further optimization can accompany worse action choices. Together, these findings motivate concrete checks for post-training: evaluate perception in use, distinguish training success from full-game validation, and examine decisions and later performance rather than relying on one selected score. The study provides empirical lessons for improving future agent training and evaluation.
Deep learning methods for pathology are constrained by the limited availability and high cost of large-scale pixel-level annotations. Synthetic data can alleviate this limitation, but existing pipelines often lack reliable quality control and remain weakly connected to downstream analysis. To address these issues, we propose a three-stage framework that integrates automated data screening, generative representation learning, and efficient adaptation to downstream tasks. First, we develop a pathology agent system to assess image quality and filter samples, resulting in TCGA-200K, a dataset containing 217,030 pathology patches selected from the public TCGA repository. Second, we train a generative model on the screened data. In addition to image synthesis, its backbone learns structural and semantic representations that can be transferred to downstream tasks. Third, we adapt the pretrained generative backbone to semantic segmentation using a lightweight task-specific head. We further introduce PathSeg-380K, a semantic segmentation dataset with expert annotations for approximately 380,000 cells across 12 cell types. Experiments on multiple pathology datasets show that the proposed framework achieves state-of-the-art segmentation performance. Our code is publicly available at https://github.com/script-Yang/Agent-GenPath.
Post-hoc out-of-distribution (OOD) detection is mature on far-OOD yet remains weak on near-OOD. These two settings differ in their use of class-characteristic feature directions: far-OOD inputs largely fail to activate those of their predicted class, whereas near-OOD inputs activate them much as in-distribution samples do. Because the axis separating far-OOD leaves near-OOD almost indistinguishable from ID, a score optimized for one setting is misaligned with the other; where a detector already saturates far-OOD, added near-OOD sensitivity is bought partly against it. This divergence motivates a targeted near-OOD signal designed to complement existing scores. We propose PLUS (Per-class Low-dimensional sUbspace Signal). For each class, PLUS selects the $m$ most class-active neurons from a sparse autoencoder, orthonormalizes the corresponding decoder directions to form a low-dimensional subspace, and scores an input by its per-class whitened squared Mahalanobis distance $z$ within the subspace of its predicted class. Designed as a dedicated complement rather than a standalone detector, the signal $z$ is fused additively with any calibrated base score at a single, fixed weight. This weight is a single global constant: no value is adjusted per backbone, per OOD dataset, or per deployment, and all calibration statistics are estimated from ID validation data alone. When applied to 18 existing detectors spanning diverse method families, $z$ consistently improves near-OOD AUROC in all 126 detector--backbone combinations. Far-OOD performance is broadly maintained or improved, with only marginal declines where the base score is near-optimal, making $z$ a robust, detector-agnostic complement across backbones.
Online reasoning video segmentation (Online RVS) requires a model to segment query-referred objects as a video streams in, using only the frames observed so far. Existing formulations largely assume a fixed target set throughout the video, overlooking a fundamental property of streaming reasoning: as new events occur, the objects that satisfy a query may change over time. To address this, we propose a new task, Generalized Reasoning Video Segmentation (GRVS), where the target set depends on query-relevant events accumulated over the observed video history. As new evidence accumulates, the target set can evolve from empty to one or multiple objects, requiring a model to continuously maintain target validity while localizing the valid objects currently in view. To evaluate GRVS, we build Isekai, a diagnostic benchmark of 60 videos and 197 queries across 11 temporal query families, with 12,569 scoring points whose target sets are causally annotated over time. We further propose Sekai, an online framework with two complementary components: an Event Memory Reasoner (EMR) that maintains query-relevant events and evolving target states in structured memory, and a Memory-Grounded Decoder (MGD) that identifies valid targets from this memory and localizes them in the current frame. Experiments show that Sekai significantly outperforms existing online RVS methods, achieving 50.33 J&F and improving over the strongest baseline by 3.2 points. Code, example annotations, and evaluation scripts are available at https://github.com/zenastangbi/Sekai.
Machine unlearning is commonly evaluated with standard post-deletion metrics that compare an unlearned model to one retrained on the remaining data. Such snapshot evaluations can conceal consequential differences in the features the two models encode. We prove that if an unlearned and a retrained model agree on their outputs but encode different features, a single gradient step on their prediction heads can break this agreement. We further show that deleting an arbitrarily small fraction of the data can cause retraining to select qualitatively different features, so matching retraining can require learning features that the full model never had. We then introduce a measure of feature recovery that separates missing from extraneous features. Across eight unlearning methods, models that perform well on standard metrics often fail to recover the retrained features. This mismatch predicts deviations from retraining under subsequent relearning and downstream adaptation, and directly correcting it systematically reduces these deviations.
We study the classical apple-tasting setting, a partial feedback model in which the learner observes the true label only when it predicts $1$. A combinatorial characterization for the rates of this problem in the online adversarial setting has been studied by Raman et al. (2024). We study the self-directed setting, in which the learner can adaptively choose the order of the points it predicts on. In the realizable setting, we establish a tetrachotomy on the expected minimax number of mistakes made by any learner. Depending on the VC dimension, the $\emptyset$-centered star number, and the self-directed dimension, we show that the rates could possibly be $\Theta(1)$, $O(\log n) \cap \omega(1)$, $\tilde \Theta(\sqrt{n})$, or $\Theta(n)$. In the agnostic setting, the minimax regret is $\tilde \Theta(\sqrt{n})$ or $\Theta(n)$, depending on the VC dimension.
Automatic cardiac motion tracking from cine cardiac magnetic resonance (CMR) imaging is essential for quantifying myocardial motion abnormalities associated with cardiovascular disease. Existing methods fall into two groups: those that represent motion without anatomical correspondence across subjects and therefore limit inter-subject motion assessment; and those that represent motion with globally corresponding points, which typically require a time-consuming and error-prone semi-automatic step to fit a cardiac atlas to the subject space. To address these limitations, this study proposes CardioDrift, a fully automatic, atlas-based probabilistic framework that represents subject-specific cardiac motion on a 3D atlas mesh with globally corresponding points. The Coherent Point Drift (CPD) formulation is generalised to jointly register the atlas surfaces of multiple cardiac regions to the anatomically corresponding contours extracted from cine CMR segmentation maps throughout the cardiac cycle. Slice-shift correction is embedded in the same probabilistic model, estimating slice-wise translations that suppress the misalignment artefact limiting atlas-fitting accuracy. Evaluation on 1301 randomly selected UK Biobank subjects and the M&M dataset, CardioDrift attains Dice similarity coefficients (DSCs) of $92.23\pm8.70\%$, $82.11\pm7.56\%$, and $88.69\pm4.13\%$ for the left ventricle (LV), myocardium (Myo), and right ventricle (RV), exceeding the strongest baseline on every distance and overlap metric of both datasets.
API distillation campaigns try to hide model-stealing traffic by splitting requests across thousands of load-balanced accounts, which standard content classifiers often miss. However, request timestamps, specifically arrival time and stream completion time, reveal the underlying dispatching strategy, such as round-robin counters or key-shuffling pools. Using a 20 percent seed set, this timing analysis can forensically rank the remaining accounts of a campaign with 0.92 to 1.00 AUC within 20 minutes, even across thousands of real production accounts. While adaptive attackers struggle to hide rotating dispatchers using jitter, pinning keys successfully evades detection, though it can severely throttle throughput when quotas bind. Ultimately, timing exposes shared infrastructure patterns even when content filters fail.
Public bodies increasingly rely on AI-produced evidence, including classifications, summaries and assessments. These outputs can inform policy decisions and service delivery, making their reliability a matter of public accountability. Establishing whether this evidence can be relied upon requires testing its outputs and examining how organisations decide whether those tests are sufficient. We combine technical evaluation with analysis of AI governance requirements, informed by measurement theory and management-based regulation. We compare eight GPT-4o configurations classifying 1,200 UK Prime Minister's Questions into 21 policy topics, against human-coded labels and across unchanged repeats. Accuracy ranges from 68.2% to 70.3%, yet 24.3% of questions receive different labels, with changes in topic shares and rankings. Configurations agree more with one another than with human coding: agreement does not guarantee correctness. Under the implemented score-extraction procedure, the proportion flagged for human review ranges from 0.9% to 41.0%. We analyse 37 UK and EU policy documents. Both sets expect testing and oversight, but reporting expectations do not consistently cover changing answers, shared errors, selective evaluation, substantive conclusions or review routing. Some UK guidance explicitly addresses stability. The UK sample emphasises public bodies checking their own use; the EU places substantial testing duties on providers while leaving them important choices about measures and acceptance thresholds. Without appropriate checks, evidence could misrepresent policy issues, redirect attention or resources, and leave errors undetected. Public policy should require organisations to justify reliance through checks beyond aggregate accuracy, clear responsibility, access to expertise and scrutiny of results.
A robot action can satisfy shape and range constraints yet carry the wrong semantics: axes may be permuted, signs inverted, rotations rescaled, or gripper polarity reversed. We study learned state-action-effect relations as offline contracts for detecting these mismatches and ranking their likely fields. Clean held-out transitions calibrate alarms; a fixed inverse vocabulary supports closed-set localization. On ten LIBERO Spatial tasks comprising 432 source episodes, a capacity-controlled comparison at 2.5% empirical clean false-positive rate gives 46.1% recall for a ridge forward residual, 61.0% for a four-component joint Gaussian mixture, and 65.7% for the exact conditional of that same mixture. The joint score retains a 14.9 percentage-point advantage over ridge, while conditioning improves recall by a further 4.7 points. Joint scoring instead has higher localization: 95.8% versus 94.6%. These matched-rate comparisons are retrospective diagnostics, not deployment calibration guarantees. Earlier Object/Goal results broaden tasks within the LIBERO convention; DROID, all three HYDRA tasks, short temporal faults, small offsets, exact repair, and simulated mitigation fail their respective gates, while xArm contributes one task only. The evidence supports a bounded offline diagnostic and a detection-localization tradeoff, rather than joint-density superiority or cross-robot safety.
Diffusion models for temporal prediction typically corrupt each future frame with independent noise, leaving temporal dependence entirely to the denoiser. We propose Dynamical State-Space Diffusion (DSSD), which learns the temporal structure of the corruption process itself. DSSD builds a temporal kernel from the damped oscillatory modes of a stable diagonal state-space model, whose learned decay rates and frequencies let the corruption capture multiple timescales and oscillatory correlations, while the denoiser architecture remains unchanged. Running this model on a finite clip, however, requires the state before the first frame, which is unobserved; starting from a zero state leaves the early frames with less noise than the diffusion schedule prescribes. We avoid this by using the Cholesky factor of the model's stationary covariance as the temporal operator. We prove that this operator is causal, exactly invertible, keeps each frame's noise variance equal to that of standard diffusion, and reduces to the identity at the clean end of the process, and that it is exactly the Kalman innovations form of the state-space model. The denoiser is trained to predict the temporally correlated noise, and sampling removes and reapplies the temporal mixing at each step. On scientific-field and robot-video datasets, DSSD outperforms diffusion baselines that share its denoiser, and the evolution of its learned modes during training shows that the temporal operator adapts to each dataset.