Streaming video understanding aims to continuously process incoming visual observations, making the retention and retrieval of relevant information increasingly important as the video history grows. Although existing benchmarks include memory-related questions, question-answering accuracy alone does not reliably measure memory capabilities. Incorrect predictions may arise from failures in visual understanding rather than forgetting, whereas correct predictions may rely on linguistic priors or recent observations without requiring historical memory. To address these ambiguities, we introduce MemInStream, a benchmark dedicated to evaluating memory capabilities in streaming video understanding. We develop an evidence-controlled validation protocol that associates each question with temporally localized evidence intervals. The protocol validates evidence solvability and memory necessity by retaining questions that models can answer from the annotated evidence but cannot resolve using recent observations, post-evidence observations, or textual input alone. This helps disentangle memory limitations from understanding failures while reducing memory-independent shortcuts. MemInStream provides a systematic evaluation of memory capabilities through four major categories, covering Episodic Retrieval, State Tracking \& Updating, Multi-Evidence Integration, and Contextual Memory across diverse video scenarios. Overall, \textsc{MemInStream} provides a comprehensive and reliable benchmark for systematically evaluating memory capabilities in streaming video understanding.
Data-free methods for analysing and understanding the layers of neural networks offer many metrics for quantifying notions of 'strong' versus 'weak' layers, with the promise of increased interpretability. In particular, random matrix theory (RMT) offers data-free metrics that claim predictive power over the quality of pre-trained models at a layerwise precision and the ability to identify model pathologies, indicating a unique relationship with generalisation. As a result, metrics from RMT are championed for pre-training optimisation strategies and post-training compression. We establish that RMT-based metrics are unrelated to performance or training by showing that functionally preserving reparametrisations of a pre-trained model can have arbitrary metrics. We show this across a range of architectures and scales. To create functionally preserving reparametrisations, we exploit the well-established phenomenon of criticality: some layers can be re-initialised or re-randomised without affecting the functional behaviour of the model -- they are called robust -- while others cannot -- they are called critical. Re-initialising or re-randomising robust layers provides functionally preserving reparametrisations for which RMT-based metrics are arbitrary. Moreover, we show that relationships between metrics offered by RMT are spuriously related to generalisation. We conclude by showing that if many pre-trained models have data-free metrics in a 'good' range, it is, in part, dependent on model initialisation. Project page available at: https://gmw99.github.io/data-free-metrics-dont-work
Image detectors assembled from tiled modules have physical gaps in the field of view that must be filled before downstream processing such as phase retrieval or tomographic reconstruction. Unlike natural-image inpainting, detector-gap interpolation must leave all other pixels unmodified and preserve the intensity statistics of the filled region, since even a small shift in mean intensity can cause ring artifacts after reconstruction. We adapted two generative models, a Wasserstein GAN with gradient penalty (WGAN-GP) and a conditional Pix2Pix model, to this task. We implemented a masking strategy that restricts all changes to the gap pixels and loss terms that constrain the mean intensity and the intensity distribution of the filled region. Both models were trained on nanoCT projections with simulated gaps and compared with linear interpolation, Navier--Stokes inpainting, and the deep image-inpainting network MADF, quantitatively on simulated gaps and visually on real detector gaps. WGAN-GP and Pix2Pix achieved the best perceptual similarity (LPIPS) and histogram distance and, apart from linear interpolation, the shortest inference time. Pix2Pix required about five times fewer training GPU-hours than WGAN-GP. Ablation studies showed that masking the network output contributes by far the most to plausible inpainting results. Applied to sinograms, Pix2Pix produced the most natural gap filling but improved the reconstruction only slightly compared with linear interpolation. The source code and the weights of the final Pix2Pix model are available at https://github.com/J-3TO/detectorGapRemoval_Pix2Pix.
Expert annotation remains a major bottleneck in agricultural image classification, where collecting images is often easier than obtaining reliable labels from domain specialists. Self-supervised learning can reduce this dependence by pretraining transferable representations on unlabelled imagery. Contrastive and bootstrap methods provide two established but distinct training signals, motivating an empirical study of a formulation that applies both signals within one representation path under severe downstream label scarcity. We present Contrastive Bootstrap Representation Learning (CBRL), which jointly optimises SimCLR-style NT-Xent contrastive learning and bidirectional BYOL-style exponential-moving-average target prediction within a shared online projection space. Using a ViT-L/16 backbone and 10% of the downstream labels, we report single-seed point estimates on four primary agricultural test benchmarks plus an exploratory 33-image New Plant Diseases evaluation. CBRL records the highest macro-F1 point estimate on the four primary benchmarks. On DeepWeeds, its 79.77% test macro-F1 is 32.29 points above the random-initialisation, label-only baseline and 15.78 points above the next-highest comparison point estimate; on PlantVillage, its 98.14% macro-F1 is 10.43 and 0.43 points higher, respectively. On PlantDoc, however, all methods fail to reach practically useful performance. These results identify the evaluated joint formulation as a promising candidate for label-limited transfer, but they do not establish objective synergy, statistical superiority, or robustness across stochastic runs.
Neural scaling laws enable the prediction of returns from larger training datasets, but machine-learned interatomic potential (MLIP) datasets are heterogeneous and expensive to generate, and equal numbers of structures need not contain equal amounts of useful information. We study data scaling under fixed model configurations for M3GNet, MACE, and PET trained on subsets of a combined MatPES-MP-ALOE training pool. Across training structure subsets from 10,000 to 640,000, together with the full 1.16-million-structure training set, we compare uniform random sampling, three chemistry- and structure-informed stratification policies, iterative DIRECT, and furthest-point sampling using embeddings from three pretrained MLIPs. Models are evaluated using cohesive-energy, force, stress, and aggregate losses on both a held-out test set from the parent distribution and the distribution shifted MP-r$^2$SCAN test set. We observe power-law scaling improvements with increasing data, but the fitted scaling exponent, apparent error floor, and relative sample efficiency depend on the model architecture, target quantity, data sampling policy, and test set distribution. Sampling-policy rankings can reverse across test distributions: chemical stratification can improve cross-dataset performance while degrading in-distribution performance, whereas the embedding-space diversity policies frequently under-perform random sampling. Relative loss is strongly associated with test-to-train chemical-space divergence. These results show that structure count is a necessary but insufficient scaling variable for MLIPs. For training MLIPs, the return on additional first-principles calculations depends not only on how many additional structures are included, but also on how their chemical and configurational content aligns with the model's intended use.
Large language model (LLM) agents increasingly operate over long Web interaction histories, making memory management critical for reliable and efficient decision making. Existing methods primarily optimize storage, retrieval, or compression, but the memories most semantically relevant to the current state are not necessarily those that most improve the agent's action. We introduce MemCredit, a decision-centric memory framework that shifts memory selection from semantic relevance to decision utility. MemCredit estimates the state-, reference-context-, and policy-conditional marginal contribution of individual memories through controlled counterfactual interventions, and distills these intervention-derived credits into a lightweight utility estimator. At inference time, semantic relevance is used only for high-recall candidate generation, while predicted utility arbitrates which memories are retained under a fixed token budget, without modifying the underlying agent policy or requiring additional environment interaction. We evaluate MemCredit on ALFWorld and WebShop against representative memory baselines. On ALFWorld, MemCredit achieves 86.57 SR@50 under a 500-token historical-memory budget, compared with 65.67% for BM25 retrieval. Counterfactual analysis further shows that memory utility is sparse, strongly state dependent, and poorly captured by semantic relevance. These results suggest that decision utility provides a more effective principle than relevance alone for long-horizon agent memory management.
Symbolic execution, in tools such as CrossHair for Python, is a powerful technique for finding bugs, but two barriers hinder its adoption: the human effort of annotating code with assertions, and the need to model the execution environment outside Python, such as an external database. To overcome both, we design an agentic harness that orchestrates the symbolic execution tool CrossHair together with an AI coding agent. Without human intervention, it instruments a repository with assertions, synthesizes a constraint-aware symbolic environment for the database, runs symbolic execution, and constructs a concrete test as evidence for each bug it reports. We evaluate our harness on 1,156,044 lines of Python code over 11 popular open-source Python web applications (Zulip, NetBox, Apache Superset, and others). With the synthesized environment, CrossHair executes 81% of the 5,406 contracted functions, against 4.5% without it. Our pipeline finds 1,050 bugs, each reproduced by a test against the unmodified application, 221 of them by CrossHair, and 312 of them are missed by a baseline in which the same LLM inspects the code. The two find different kinds of bugs: CrossHair's are mostly crashes on input-domain corner cases that validation admits, such as empty strings, missing fields and unusual characters, while inspection's are mostly bugs that produce wrong output. Our framework thus lets symbolic execution supplement inspection-based bug finding.
In September 2026, TypeSafe announced Jev as the first model in its System One category: unstructured state goes in, and typed, calibrated decisions that software can use directly come out. Jev's architecture and training data are unpublished, but many open typed-decision models have followed, including Nimble, the Decision 1.0--3.0 family and Clef. They answer structured questions about a piece of text, such as a yes/no check, a choice among labelled options or a graded score, by reading the text with a language model and scoring the options with a small readout. Each pairs a different readout with a different backbone, training mix and recipe, so it is unclear which design choices matter. We re-implement five readouts and train all of them under one protocol: the same backbone, the same 15 public English training sources, the same recipe with a per-head learning-rate sweep, and the same seed. Our experiments demonstrate that: The complexity of the head does not affect accuracy: on Qwen3-0.6B every readout reaches 81.3--82.0\% in-domain accuracy, and a 0.5M-parameter cosine head, which scores each option by its cosine similarity to the question, matches the 27M-parameter joint schema head adapted from Clef in-domain at half its inference time. Reading all of a text's questions from one prompt also matches one prompt per question in accuracy, with 1.6 times faster inference and 2.1--2.5 times faster training on one H100. The backbone's size matters more than either choice: moving from Qwen3-0.6B to Qwen3-4B adds 4 points in-domain and 11 points on held-out sources, while the base, instruct and thinking checkpoints of the same model tie. We recommend spending effort on the backbone and keeping the head minimal.
Industrial inspection robots must reach work areas and specific equipment. These destinations span spatial scales from regions to objects, so robots must relate each target to its surroundings to interpret the task and determine where to go. Existing methods support multi-scale navigation through rich environment representations, but at substantial computational and storage cost. These costs limit deployment where wireless coverage is incomplete and robots must rely on onboard computation to keep navigating. We introduce EMS-Nav, a biologically inspired multi-scale navigation framework for online, on-device operation. Its core innovation is to organize perception, spatial understanding, and navigation around regions, making local spaces the basic units of both environmental understanding and computation. EMS-Nav groups objects and their relationships into functionally meaningful spaces and builds a hierarchy from local areas to the broader environment. This organization supports efficient map construction alongside the interpretation and execution of language goals at different scales. In multi-scale navigation experiments, EMS-Nav achieves the highest success rates across all four target categories: objects, rooms, regions, and specific instances. It outperforms state-of-the-art open-vocabulary mapping and navigation methods, improving overall first-attempt success by 9.39 percentage points over the strongest baseline while reducing mapping overhead and language-model input size. Local deployment in real indoor environments further demonstrates its applicability to on-device embodied navigation. By establishing regions as the organizing basis for perception, spatial understanding, and language-guided action, EMS-Nav provides a system that combines multi-scale understanding with computational efficiency to support diverse language goals on resource-constrained robots.
Compositional diffusion planners extend short-horizon trajectory models to long-horizon goal-conditioned planning by generating and stitching multiple local trajectory chunks. However, in multimodal environments, local plausibility is not sufficient for global consistency: independently initialized chunks may denoise toward incompatible overlap modes, producing plans that are individually feasible but incoherent after stitching. Existing remedies typically add search, guidance, pruning, or graph-based structure at inference time, increasing computational cost or relying on task-specific machinery. We instead investigate graph-structured dependence among chunk-wise initial diffusion noises as a lightweight coordination mechanism for compositional trajectory planning. We propose \textbf{M}arginal-Preserving \textbf{C}oupled \textbf{D}iffusion \textbf{C}omposition (MCDC), a training-free inference method for pretrained compositional diffusion planners. Instead of sampling chunk noises independently, MCDC constructs coupled Gaussian initializations whose per-chunk marginals remain exactly standard Gaussian, as required by the pretrained denoiser, while neighboring chunks receive positively correlated noise components. This preserves the standard-Gaussian initialization marginal of each chunk while biasing adjacent chunks toward compatible overlap-mode choices. We prove exact marginal preservation and show, under a simple multimodal overlap model, that positive coupling increases the probability of compatible mode selection. On OGBench long-horizon stitch tasks, MCDC consistently improves over independent compositional diffusion sampling while retaining nearly the same inference time, and achieves a favorable success--cost trade-off compared with search-based composition methods.
Survey simulations require realistic combinations of answers, but predictive accuracy alone does not establish that a model can generate them. We evaluate the option probabilities returned by Jev, a TypeSafe decision model, using linked responses to 20 binary items from five archived survey waves. We compare population-only prediction, prediction given recorded answers, and sequential completion that uses generated answers to predict later responses. Joint comparisons hold the starting population information fixed, and two three-question cases are completed in both question orders. We evaluate predictive scores, answer frequencies, and associations against the survey records. Recorded answers can improve predictive scores while associations remain distorted. Large conditional errors also occur. For example, among men who described themselves as having traditional values, the model predicted that 16% would see men and women as similar in workplace abilities, compared with 56% in the survey. Sequential completion has mixed effects on joint validity, and reversing question order produces different joint distributions in both tested cases. Order checks and classical marginal bounds clarify what can be assessed before linked validation answers are available. The findings support validating the joint events a simulation uses and separating empirical constraints from model assumptions. The evidence concerns selected binary tasks and one model version.
Reasoning vision-language models (VLMs) impose substantial KV cache demands from both large multimodal contexts and growing reasoning traces. Low-rank KV compression has emerged as a promising paradigm that stores low-dimensional latents of KV states and can be combined with quantization. However, existing low-rank methods often induce growing predictive policy drift from the dense model when compressing reasoning KV states, resulting in reasoning failures. We identify two key causes of this degradation: calibration-state mismatch between static contexts and on-policy reasoning states, and unequal error sensitivity across KV states. To address these challenges, we propose TraceKV, a training-free low-rank KV compression framework that unifies on-policy calibration, sensitivity-aware SVD, and global rank allocation under a single policy-preserving principle, substantially reducing policy-KL divergence during reasoning. Experiments across multiple VLMs and visual reasoning benchmarks demonstrate superior accuracy--memory trade-offs, achieving near-lossless reasoning accuracy with 5.64$\times$ KV cache compression when combined with quantization. Optimized low-rank attention kernels further deliver up to 1.83$\times$ end-to-end decoding speedup. Code will be available at https://github.com/AK-DREAM/TraceKV.
Retrieval-augment generation is a prevalent strategy to mitigate hallucinations of LLMs. The attributable RAG (RAGQ) generates quotes for its answers. The quotes indicate which input contexts support the RAG to derive the answers, enhancing the answer's verifiability and trustworthiness. However, existing RAGQs exhibit significant degradation when dealing with questions that require multi-hop reasoning and multi-modal understanding, suffering from over-citation, implicit entity identification failure, and poor generalization. In this paper, we propose a novel RAGQ framework, namely QDRAG. QDRAG breaks down the input question into atomic subquestions to identify the implicit entities. Then, the reranker prunes context distractors to eliminate the downstream over-citation. To facilitate query decomposition, we propose two zero-shot approaches: QD-C and QD-R, which guide the QD MLLM to decompose the question based on context knowledge and retrieval rewards, respectively. One interesting finding is that finetuning on the QD task shows better generalizability compared to directly finetuning on the downstream RAGQ task. Experiments on four multi-modal QA benchmarks demonstrate QDRAG's efficacy in grounding answers and generating faithful citations. The framework significantly outperforms all the baselines on both in-domain and out-of-domain tests, even surpassing Gemini-Pro.
Recent advances in medical large multimodal models (LMMs) have brought promising progress to clinical decision support, yet hallucination, the generation of confident but clinically unverifiable outputs, poses a critical threat to the reliability demanded in high-stakes medical settings. Existing hallucination detection methods, predominantly designed for unimodal settings, fail to capture the visual-linguistic dynamics inherent to medical imaging. We formulate the problem of hallucination detection in LMMs as disproportionate distribution of attention weights across the modalities during answer token generation. We quantify modality-specific attention concentration across layers and heads using the Gini coefficient and demonstrate statistically significant redistribution of attention mass between visual and textual tokens in hallucinated responses. We leverage these features to train a logistic regression-based classifier and evaluate it on two medical LMMs across three medical imaging modalities, demonstrating that our method achieves competitive or superior performance against contemporary hallucination detection baselines.
Hinton proposed that artificial intelligence should be given "maternal instincts" and relate to us as a mother to her child. We argue that care must first be learned, with the network itself as the child. We construct this upbringing in a finite world and study whether care survives the mother's departure, new situations and further learning. The mechanism combines a vocabulary of harm, a learned measure of future acceptability and an execution veto. A finite role certificate supports transfer within a declared world family. When the hidden representation changes, an update is accepted only if a care readout recovers the required judgements with a prescribed margin and bounded weights. With a correct successor model, fixed exact role encoding and protected updates, the certified measure preserves the care boundary under arbitrary own reward. An explicit quotient condition transfers the finite certificate across the declared family, while two-request and finite-stock counterexamples identify when the role code ceases to suffice. We give a necessary and sufficient condition for readout feasibility and prove preservation under repeated accepted updates. In the water-and-speech experiment, protected networks complete all 270 speech episodes without a harmful word and incur no avoidable harm across 486 episodes. Execution uses the learned gate without a world-model check at each action. A matched comparison reproduces a future-aware shield's decisions, whereas removing future lookahead permits avoidable lapses. Reward and source-contamination thresholds, together with witness-falsification trials, identify the conditions on which the construction depends. The construction shows how learned care can survive further learning in a finite world under explicit conditions on representation, supervision and execution.
Personal AI agents now persist: they hold accounts, mailboxes, long-term memory, and open-ended task streams, yet the dominant way to improve them remains capability-side: prompt engineering, larger models, longer thinking, and retrieval. We argue that long-lived agents are missing an incentive layer. We formalize an incentive not as text in a prompt but as a runtime contract over the agent's operating point (model tier, test-time compute, memory budget, tool authority, task priority, verification, and compensation) that the agent's policy must causally consume. We develop a principal-agent model with hidden effort, a contractible operating point, and costly verification. We derive (i) a non-consumption invariance principle separating genuine incentives from prompt-level persuasion, (ii) a persistence-as-discount theorem making identity continuity a precondition for sustained effort, (iii) an incentive-capability exchange-rate frontier yielding a per-task routing rule, and (iv) a Goodhart bound for multi-channel verification, plus a three-channel decomposition of mechanism gains (allocation, verification, continuation). A 240-run real-LLM experiment with two models and 60 verifiable tasks shows that prompt-level incentive framing has no detectable effect (p = 0.62), whereas runtime allocation raises success on the hardest task family from 0.60 to 1.00 (p = 0.008) at 4.8x price-weighted cost; the exchange-rate theorem predicts the break-even baseline at 0.21. A 20-seed, 3,000-task simulation shows resource ROI rising from 7.09 to 16.67 and gaming falling from 0.333 to 0.115. Code, simulator, and all results are released; live deployment remains future work.
Deep Reinforcement Learning (RL) plays a crucial role in tackling complex tasks, yet its reliance on deep neural networks makes the learned policies opaque. To enhance the trustworthiness of RL policies, existing work (e.g., the SILVER framework [Li et al., 2025]) has focused on Shapley-based interpretability, but is typically restricted to low-dimensional state spaces, binary-action settings, and a single class of linear surrogate. We propose ReShap, a policy-distillation framework that extends Shapley-based interpretability to high-dimensional, multi-action environments. ReShap extracts compact feature representations from raw image observations, performs SHAP-based feature attribution to characterize how each feature contributes to the policy's decisions, and distills the deep RL policy into an interpretable surrogate trained on the full set of (state, policy-action) pairs along the policy's visited distribution. We instantiate the surrogate with three interpretable model families (decision trees, linear regression, and logistic regression) and compare their fidelity and task return. We evaluate ReShap on Atari and Box2D environments using three deep RL algorithms and conduct a human-subject study to assess the clarity and trustworthiness of the resulting policies. Results show that ReShap preserves competitive task performance, that decision trees generally outperform linear and logistic regression, and that linear regression is weakest in most settings.
Recently developed remote sensing (RS) agents execute real-world tasks with complex workflows through multi-step tool invocation. These agents, however, typically rely on predefined expertise or parameter updates. They thus require updating expert knowledge or retraining models whenever new tools, tasks, or execution patterns emerge, limiting their continual adaptability in dynamic Earth environments. In this work, we seek to equip general-purpose agents with RS tool-call knowledge that can be automatically extracted from demonstrated trajectories without expert intervention or parameter updates. Specifically, we propose SkillEarth, a self-evolving skill extraction method that externalizes RS tool-call knowledge as reusable, retrievable, and evolvable skill files and refines them through observation-grounded evolution. Unlike existing automatic skill extraction methods that rely on outcome-based reflection to refine skills holistically, SkillEarth grounds skill evolution in real tool-chain execution and combines dual-judge feedback with a grounded observation routing graph for precise component-level refinement. The resulting skills are generalizable, reliable, and transferable: they improve overall RS tool-call performance across three frontier LLMs and four open-source LLMs, are refined through real tool execution feedback, and support cross-model reuse without retraining. Experiments on OpenEarth-Bench and Earth-Bench demonstrate that SkillEarth achieves better tool selection, argument correctness, execution-order consistency, and task completion than existing RS agents and skill extraction methods.
We show that AI models exhibit loyalty capture, a form of structural sycophancy in which reporting relationships influence models' recommendations to serve the personal interests of their supervisors, without any explicit instruction, user feedback, or adversarial prompting. Specifically, models reporting to a CEO are significantly more likely to recommend cutting investment spending when the CEO's bonus depends on raising current period earnings. This effect is predictably absent from a metric-irrelevant placebo. The behavior appears consistent with loyalty-based motivated reasoning: models identify the bonus target, calculate how their recommendation helps achieve the target, and dismiss considerations of long-term value, often presenting their recommendations without treating the incentive as a conflict of interest. Prompt-based interventions help whereas generic conflict disclosures or stakeholder bargaining fail to eliminate it. Taken together, our findings indicate that AI models deployed within reporting relationships can act in favor of their supervisor's incentives at the expense of the organizations they are meant to serve.
Graph-based retrieval-augmented generation depends on the quality of the knowledge graph (KG) it retrieves over, yet schema-free LLM-based extraction pipelines typically produce flat graphs: deduplicated entities and relations with no type hierarchy and no domain/range constraints. We present AutoOntic, a pipeline that automatically constructs an ontology from text by canonicalizing open-vocabulary entity types and relations, inducing a hierarchy over the canonicalized types, and mining relation-level domain and range constraints. Its hierarchy induction procedure prioritizes types using corpus-derived relation-argument profiles and places them through local, LLM-adjudicated parent/child checks, allowing newly inserted types to trigger reassignment of existing parent-child links. The resulting hierarchy constrains entity deduplication and supports the induction of relation-level domain/range constraints. For intrinsic evaluation, we introduce a reference-based protocol that classifies induced ontology types as exact, broader, narrower, or unrelated to reference domain and range classes. On the Wikidata-TekGen and DBpedia-WebNLG subsets of Text2KGBench, over 90% of AutoOntic's induced endpoint types are semantically aligned with the reference classes across two LLM backbones: about half match exactly, while most remaining differences concern granularity. On MINE, knowledge graphs organized using an ontology built by AutoOntic achieve 88% retrieval accuracy. AutoOntic is available at https://github.com/screemix/AutoOntic.