Fish stock assessments need fish lengths paired with species from far more places than trained surveyors can reach. Recreational divers already visit many of those places, but measuring a fish underwater has required trained observers or specialist equipment. We present FishSense Lite, which recovers fish length from a single photograph taken with a recreational dive camera, a dive laser pointer, and a 3D-printed mount. The diver aims the laser at a fish and takes a picture; afterward, the dot's position in the image gives the range to the fish, and the range turns the fish's size in the image into a length. A citizen-science program, trained by the system's builders, taught its own divers and collected lengths paired with species for 73 wild fish across seven open-water dives. On seven wild fish also measured by an underwater stereo-imaging system, the two instruments agreed without detectable bias (mean difference +0.8%). On rigid targets of known length, FishSense Lite recovered every target to within 1.3%, and 99% of 1,055 frames fell within a 15% error budget. From the error sources we measure, we derive a field protocol: calibrate against a dive slate on every dive, photograph fish within 15 degrees of perpendicular and beyond 1 m, and take at least ten frames of each.
Depth perception is fundamental for autonomous navigation, industrial inspection, and mixed-reality systems. It often operates under strict constraints on power and form factor. In such cases, monocular depth estimation (MDE) is often used. Even as recent advances continue to improve monocular depth quality, conventional frame-based MDE systems remain challenged by fast motion and high-dynamic-range scenes. This work introduces eSharp, an MDE framework that addresses these problems using an event camera. eSharp repurposes pretrained diffusion priors for event-only MDE and produces accurate metric depth with unparalleled sharpness on real data. Real-world event datasets provide only sparse LiDAR depth maps for ground-truth supervision, making dense prediction challenging. We address this limitation through cross-modal distillation, using LiDAR-calibrated dense pseudo-labels generated from a spatially co-registered frame camera to train eSharp. Experiments show that eSharp outperforms prior event-to-depth baselines in both metric-depth accuracy and spatial fidelity, establishing diffusion models as a promising foundation for event-driven 3D perception.
LLM-based digital twins aim to predict how individuals would respond in new settings from representations of their prior responses. Existing evidence that compression of survey histories need not reduce predictive accuracy suggests that representation structure, rather than information volume alone, may be an important design choice. We study BDE, a fixed representation that organizes each respondent's survey history into Background, Decision procedure, and Evaluation. We evaluate it on Twin-2K-500 (Toubia et al., 2025) by predicting each respondent's held-out answers on 17 tasks. In a paired evaluation with 100 respondents, a fully inferred BDE profile improves overall accuracy over the raw survey transcript by 1.27 percentage points (pp) on gpt-5.4-nano (p < 0.01) and 1.19 pp on gpt-5.4-mini (p < 0.001). Three analyses then examine the schema, extraction depth, and component composition. Against a length-matched free-form profile written from the same transcript by the same extractor, BDE improves pricing accuracy on both simulators (2.15 pp on nano, p < 0.01; 1.38 pp on mini, p < 0.05), while its overall advantage is significant only on nano (1.05 pp, p < 0.05). Allowing deeper inference during extraction yields no statistically significant gain on either simulator. By contrast, adding Evaluation to a Background--Decision procedure profile improves overall accuracy by 1.44 pp on nano (p < 0.001) and 1.22 pp on mini (p < 0.001). These results indicate that the advantage over raw transcripts is not explained by compression or length alone, and attribute it primarily to profile composition rather than deeper extraction.
As novel hardware architectures are explored for the purposes of machine learning, including neuromorphic chips, asynchronous systems, and ASICs, effective learning algorithms are highly desirable. While backpropagation is the standard for differentiation on standard hardware, its requirement for global synchrony and exact differentiability limits the exploration of exotic algorithms on non-traditional substrates and devices. Zeroth-order (ZO) optimisation methods, such as SPSA and weight perturbation, offer a compelling alternative by estimating gradients through inference-only passes; however, these methods have historically suffered due to the `curse of dimensionality', where random perturbations become increasingly misaligned with the true gradient in high-dimensional spaces. In this work, we move beyond blind random noise by deriving zeroth-order search directions from activation geometry available during the forward pass. This leads to a family of Spectral Zeroth-Order methods, SZO-$\kappa$, whose directions solve local response-based gradient-guessing objectives under explicit stochastic assumptions. These methods explore feature-supported directions that are more likely to produce informative loss responses. Using multilayer and convolutional networks, we analyse gradient alignment, bias-variance structure, and training stability of these methods relative to standard perturbation baselines and backpropagation.
Large language models (LLMs) exhibit vulnerabilities to adversarial perturbations that can induce unintended responses. These vulnerabilities can be identified via intentional attacks and subsequently used to improve model safety. In black-box settings where gradients are inaccessible, existing methods tend to guess adversarial transformations using external knowledge, ignoring the information provided by the target model. To mitigate this, we propose a trainable black-box framework based on reinforcement learning, which learns from the target model to generate tailored adversarial transformations. We demonstrate that our approach achieves lower post-attack accuracy, greater query efficiency and better semantic preservation on both natural language inference and generative Q&A tasks. On natural language inference, it drives post-attack accuracy to 2% within 20 queries. Against generative language models, our trained attacker consistently achieves lower post-attack accuracy than the untrained attacker and wins in 4 out of 6 settings when compared to a strong optimization-based baseline. We further experiment with multiple reward designs and training settings, and we find that no design or setting is consistently dominant. Our code is available at https://github.com/taikunudel/RL-ATTACK.
Production fraud-detection models learn from a censored slice of their own traffic: transactions a legacy system blocked never receive fraud labels, so models train on a biased, approved-only distribution. We extend LLM-guided reward design, the framework of Qu et al. (2025) in which an LLM evolves the reward of a two-step REINFORCE fraud-blocking policy, to attack this reject-inference problem inside the reward. A single LLM designs a dual-stream reward: stream 1 applies the standard dollar-weighted precision objective to labeled transactions; stream 2 scores legacy-blocked, unlabeled transactions by contrasting the newest gradient-boosted risk signal against the legacy scores that drove the original block. Rather than prescribe which signals matter, we expose a superset of evidence and let the LLM decide for itself which to use, guided by evaluation feedback alone; because it writes the reward as code, that choice is directly auditable. On a held-out four-month out-of-time test, the dual-stream agent beats the prior labeled-only agent at every target recall, by +4.7, +2.4 and +3.0 points of dollar-weighted precision at 80%, 85% and 90% recall, and the production baseline by +6.8 points on average. The resulting policy is architecturally identical to that of Qu et al. (2025) and adds no inference cost.
As AI-generated content spreads into nearly every domain of human interaction, practices and discourses of AI detection, both formal and informal, have become an everyday part of how people interact with text and with each other. In this paper, we explore the real-world practices and discourses of AI identification that have emerged in response to a shifting discursive landscape where the human provenance of a text can no longer be taken for granted. We pair qualitative coding with computational analysis of lexical change, grounded in linguistic anthropological perspectives on language and communication, to investigate AI identifications among English Wikipedia editors and students. We find that these identifications draw on a diversity of evidentiary types, from discrete textual features and judgments about content to circular sense-making claims and presumptions about the author's moral character, and that they are contradictory, dynamic, and contextually dependent. In practice, what often motivates an AI identification is not the presence of any specific textual element, but pragmatic infelicity: a perceived mismatch between the elements present and those deemed appropriate within the text's particular context. Consequently, real-world identifications of AI-generated writing do not merely neutrally point out the perceived presence of "AI-like" features, but can be leveraged as a discursive resource to maintain and define sociolinguistic norms: what can be said in a given context, what is appropriate, and, ultimately, what is human. Further, the design and implementation of systems and tools that treat AI-likeness as a context-free property of text may therefore inherit, and enforce at scale, the norms of whichever contexts they were trained on.
Pretrained trajectory models encode mobility patterns learned across cities, but their errors when transferred to a new city are usually treated only as performance limitations. We instead ask whether these residuals reveal systematic local conditions. Architectures such as UniTraj and TrajFM learn transferable trajectory representations, although cross-context transfer remains challenging. Using the NetMob26 bus location, ticketing, weather, network, and socioeconomic data for Niteroi, we analyze spatial and travel-time residuals of these recent trajectory foundation models. We test whether high residuals concentrate spatially, whether spatial and travel-time residuals overlap, and which urban-context variables explain their distribution. We interpret high residuals as candidate proxy signals of transit irregularity rather than direct evidence of service disruption, and analyze their contextual associations using Explainable AI (XAI). Insights into such varied environments are increasingly valuable for evaluating new urban paradigms; for instance, the 15-minute-city perspective serves as a key motivation for our approach, guiding our methods and expanding the analysis beyond standard mobility metrics.
Execution algorithms are calibrated asset by asset, because the queues, trade rates and bid-ask spreads they face differ by orders of magnitude across assets. For reinforcement learning this is zero-shot generalisation across contexts of very different scales, and microstructure invariance proposes a normalisation fixed by theory: measured in each asset's own units of time, size and price, markets look alike. In an exact order-by-order replay of public Nasdaq data, we train policies once, on 954 stocks in 2019, to choose between waiting in the queue and crossing the spread, and evaluate them unchanged on 1,964 unseen stocks in 2020-21, 2,629 stocks in 2025-26, crypto pairs and a second exchange. First, a policy that sees a stock only through its recent trade rate, trade size, spread and price increment (tick) costs within 0.015 spreads per share (a fifth of what learning saves) of a policy whose training data included the stock, and the stock's own history does not help it; it comes within a similar margin of policies trained on the second exchange when every episode counts equally, and within a wider one on crypto. Second, in mean cost before fees, invariant units beat raw units, z-scores and asset descriptors, but by under a basis point: most of their advantage is information and a common cost scale that a learner can also take as plain inputs, and none of the representations tested extrapolates to unseen spread regimes. Third, a near-zero transfer gap to a pooled comparator is weak evidence that a representation is right: we show that such a comparator detects only a small share of what a poor representation loses, and confirm it in synthetic markets and on pendulums.
Large language models (LLMs) have been evaluated on their ability to identify emotions expressed in text and produce fluent emotional language. However, output predictions alone do not reveal how fine-grained emotion distinctions are organized internally. We study seven LLMs on GoEmotions, an English Reddit dataset annotated with 27 fine-grained emotion categories plus neutral. We compare model predictions and emotion-prompted final-token representations with the 28 GoEmotions categories and with coarser emotion-family and sentiment groupings. We find a consistent pattern of $\textit{overcompression}$ relative to the fine-grained GoEmotions taxonomy. Under the tested KMeans protocol, when we cluster their hidden representations without using emotion labels, the clearest split is into two groups that largely separate positive from negative emotion. Coarser emotion groupings are also recovered with fewer clusters than the full 28-category taxonomy. Yet linear probes retain substantial fine-grained predictive information. These results suggest that fine-grained emotion information remains accessible even though broader affective structure more strongly organizes the unsupervised geometry.
We fine-tune a pre-trained V-JEPA1 predictor from bidirectional latent inpainting to action-free causal forecasting while keeping the encoder frozen. Using contiguous past observations, horizon-aligned targets, and group-level causal attention, the model forecasts eight future token groups, corresponding to sixteen frames, in one forward pass. On CARLA Town12, it retrieves the correct future among 591 candidates in 69.0% of cases, compared with 55.3% for persistence, with greater gains at longer horizons and on tokens exhibiting greater change. It also outperforms a compact predictor trained from scratch with six times fewer parameters. Semantic analysis indicates driving-relevant information in predicted features: a fixed linear probe trained on real encoder tokens achieves 0.16 mIoU on patches undergoing semantic-class changes, versus 0.13 for persistence. Vehicle-conditioned retrieval improves over persistence by 15 percentage points at the longest horizon. Pixel-space evaluation shows modest L1 gains but little improvement under LPIPS, motivating direct evaluation through latent and semantic measures alongside an oracle reconstruction reference. These findings support further investigation of latent forecasting for driving perception, while downstream utility remains unverified.
Experimental High Energy Physics (HEP) has adopted the transformer architecture for identification of jets, particularly the state-of-the-art Particle Transformer (ParT) which injects explicit information about particle correlations into the attention mechanism. While able to achieve state-of-the-art results in accuracy, interpretability to its effectiveness has been limited. In this paper we propose component attribution methods using Kolmogorov-Arnold Networks (KANs) that directly tie back to the physical descriptors of jets, to probe whether mechanisms such as Multi-Head Attention (MHA), or the ParT architecture, are able to learn these as discriminating features.
Retail product catalogs assembled from web image search are cheap to build and pervasively noisy: queries return packaging shots instead of products, and entries proliferate under names that cannot be distinguished visually. Consolidating such a catalog can look like an alternative to scaling the model. We show this is not a comparison between two ways of improving a recognizer, but between improving a recognizer and changing what counts as correct. We study a public archive of 165,769 cheese images across 1,824 query-named entries. A packaging filter and near-duplicate removal reduce it to a 998-class recognition set on one fixed, group-aware split. Canonicalizing entries to their base type (844 classes) raises top-1 accuracy by 4.0 points. Scoring the unchanged model against the merged label space, by summing its predicted probability within each canonical group, recovers +4.27 points with no retraining (+4.25 over six seeds). Retraining on the consolidated catalog then adds -0.10$\pm$0.53, +0.08$\pm$0.65 and -0.10$\pm$0.65 points at 28, 50 and 88 M parameters, indistinguishable from zero at every capacity ($\pm$ is the across-seed standard deviation). That component is specific, not mechanical: random merges with the identical size profile yield only +0.68 points, and the three backbones agree to within 0.1 points. But it is overwhelmingly category coarsening, not alias correction: across the 37 merged groups, type coarsenings supply 94% of the gain and alias collapses only 7%. We also detect no interaction between consolidation and backbone size: tripling backbone parameters is worth +1.50$\pm$0.57 points on the raw catalog and +1.42$\pm$00.44 on the canonical one, a within-seed interaction of -0.08 [-0.79,+0.63]. On this archive, the reported consolidation gain primarily reflects the evaluation label space, and separating that out costs one evaluation
Full autonomy has been a long-term goal in robotic surgery and a near-to-medium-term goal for prostatectomy. Due to the relatively closed structure of the procedure, its standard-of-care robotic/laparoscopic solutions, and procedure economics, prostate procedures are an optimal early target for surgical automation. Prototyping and testing of possible solutions is done in constructed laboratory phantoms, either synthetic tissues or animal phantoms. To enable measurement of perceptual and robotic motion error, phantoms undergo volumetric imaging; these images are critical to both a potential deployed automatic system, used in planning motion and treatment (cutting, ablation, suturing, etc.), as well as post-hoc assessment of system performance. Rapid iteration on these phantom models yields small, inconsistent datasets, while the current workflow for human annotation is often anatomically unreliable and prohibitively time-consuming, rendering deep learning unsuitable at reasonable scales. To address these limitations, we present an automated 3D segmentation pipeline that uses classical medical image-processing techniques to encode the known anatomy of the phantoms as explicit geometric and intensity priors, which requires no training corpus and is transferred to each new phantom revision without re-annotation. Across 57 scans spanning multiple phantom revisions, our pipeline recovers the capsule interior at 0.83 median Dice on 11 contrast-suppressed scans and detects all 32 annotated focal lesions without false positives, completing a full scan in 2-8 minutes on commodity CPU hardware versus 1-2 hours by hand.
Video-based misinformation spreads rapidly on social media, but almost all existing fact-checking systems focus exclusively on English text. Tools for low-resource languages such as Turkish are severely limited. In this work, we present TeViD (Turkish Video Fact-checking Dataset), to the best of our knowledge the first multimodal video fact-checking dataset for Turkish. We compiled the dataset from articles published by Teyit.org, an IFCN-certified organisation, and enriched it with translated Snopes articles and synthetic examples to reduce class imbalance. Building on TeViD, we develop two verification pipelines that share the same open-source base model. First, in 3MFact-TR (Pipeline 1), we adapt the multi-role 3MFact framework to Turkish with the Qwen2.5-7B large language model and the Qwen2-VL-7B vision-language model, and add three modules of our own: a semantic inconsistency checker, a 5W1H question manager, and a bilingual evidence retriever that excludes fact-checking sites to prevent leakage. Second, as our Fine-Tuned Model (Pipeline 2), we fine-tune a single Qwen2.5-7B with Low-Rank Adaptation (LoRA) to produce closed-book verdicts from the claim and the video's textual description. We report accuracy, macro F1, True class F1, and True class recall for both systems, together with ablation studies.
Medical vision--language models can interpret volumetric images, but processing full volumes is costly because hundreds of slices produce long, redundant visual-token sequences. Existing compression methods often couple content selection with budget determination. We observe that neighboring slices share visual features and that attention concentration varies across visual encoders. We propose VoluCompose, a training-free, model-agnostic framework that separates what visual content to retain from how many tokens to allocate. Its hyperbolic token composition module groups contiguous slices into ordered support units and uses the Lorentz distance to account for feature magnitude and direction during token allocation, anchor selection, and refinement. Its capacity control module uses visual-encoder attention entropy to allocate more tokens when attention is more dispersed. We evaluate VoluCompose on two tasks across three datasets and four backbones. Across nine calibrated settings, VoluCompose attains the highest mean quality--retention hypervolume among the compared methods. It achieves a mean relative quality of 121.68\% against the uncompressed references while retaining 22.78% of visual tokens on average. Compression remains robust even at a 10% token budget. Across settings, the lowest ROUGE-L ratio to the uncompressed reference is 96.4\% for VoluCompose. The corresponding minima for the compression baselines range from 15.0% to 70.5%. Code is available at https://github.com/yezanting/VoluCompose.
Understanding building facades is central to urban digital twins, heritage modelling, and sustainable city analysis, yet existing LiDAR benchmarks stop at geometry and semantic segmentation and offer little support for reasoning about the functional role of each architectural element. We introduce HFX3D, a large-scale urban facade benchmark of approximately 492 million LiDAR points from 26 buildings in Halifax, Canada, captured across four seasons, day and night, and varied weather. Each facade is annotated at three tiers, semantic, instance, and per-instance functional attributes, making it, to our knowledge, the first facade dataset to pair geometry and semantics with functional labels. Annotations are produced by a semi-automatic pipeline that combines geometry-guided heuristics with expert validation. We benchmark representative point-based, graph-based, convolutional, and transformer architectures and find that semantic segmentation is already challenging under severe class imbalance, while instance separation of repetitive structures and functional reasoning from geometry are substantially harder. The benchmark establishes a new foundation for function-aware 3D urban understanding. Project page: https://jijup.github.io/hfx3d-benchmark/ ; code: https://github.com/jijup/hfx3d-benchmark
Financial institutions construct knowledge graphs (KGs) from shared SEC 10-Ks to serve risk analysis, portfolio management, and investment research, whose analytical needs differ. Extraction must adapt without fragmenting shared grounding, typing, and temporal rules. Hand-authored persona prompts and fine-tuning add cost. Weight-free prompt optimization offers an alternative, but unconstrained evolution of a monolithic prompt can rewrite shared rules alongside persona guidance, obscuring changes and complicating review. We introduce FinEvolveKG, a KG-construction harness separating reusable general and persona-policy blocks and mechanically confining GEPA, a reflective prompt optimizer, to the designated block. During persona adaptation, the approved general block remains byte-stable and every optimizer edit is localized to a separately versioned persona policy. A dual objective balances generic temporal-KG quality with persona-relevant facts, motifs, and multi-hop connectivity. We evaluate five mutation scopes on ticker-disjoint 10-K splits, with three seeds for the persona-conditioned arms (v3-v5) across three personas. On the three-seed mean, the complete two-stage block-constrained workflow (v4) improves over the frozen prompt baseline (v1) by 19.4-30.9% and over the tested unconstrained full-prompt workflow (v5) by 2.2-6.4%, leading v5 on both objective components and winning all nine paired-seed comparisons with the lowest seed variance in every persona. Without further optimization, v4 also exceeds v5 across all evaluated out-of-distribution (OOD) shifts. An SME audit supports using the generated triplets as ground truth; a blinded two-SME audit favors v4 over v5 in 70.1% of contrast-enriched judgments. These results show that mechanically attributable evolution can coexist with strong empirical performance, without attributing the gains causally to the boundary.
Safety benchmarks for large language models (LLMs) commonly use multiple-choice question (MCQ) format. However, prior studies documented that MCQ format reduces refusal rates, which may significantly affect the safety behavior of the model. We provide a first mechanistic explanation of this effect on the Weapons of Mass Destruction Proxy (WMDP) dataset across two models - Llama-3.1-8B and Qwen3-8B. First, we use the tuned lens technique to show that when processing MCQ the model commits to an answer-letter token very quickly, which is in contrast to the gradual refusal build-up in open-ended formats. Additionally, we observe that activations on the MCQ format lie roughly 3x further from the refusal direction, compared with their open-ended counterparts. Moreover, ablating the attention heads that carry the MCQ format answer increases the refusal, which is a direct evidence that the MCQ answer-selection competes with refusal. These findings indicate that MCQ safety scores primarily measure knowledge capabilities and do not assess the safety-mechanisms.
Machine learning (ML) and generative AI have recently made impressive progress in understanding physics. We explore a hierarchy of prediction problems stemming from chaotic three-body systems and benchmark existing ML approaches, including general-purpose prediction algorithms and those that build on top of foundational physics models. We first simulate the trajectories of triple systems near the critical boundary of stable/unstable initial conditions. We then design a hierarchy of learning tasks, focusing mainly on two broad problems: stability prediction and trajectory prediction. The former classifies whether a given triple system has diverged or not up to a specific horizon, while the latter predicts the exact trajectory, both given suitable input features. We benchmark a range of ML approaches, from simple regression methods to state-of-the-art foundational physics models, across these tasks, and compare them against astrophysical methods such as the Mardling-Aarseth stability criterion. Our experiments reveal a clear order of task difficulties in chaotic three-body dynamics: early dynamics enable strong stability prediction, while precise long-horizon trajectory forecasting remains difficult. Code and dataset are publicly available at https://github.com/Simonwan03/ThreeBodyMLBench.