In briefBoosting deep learning accuracy for olive tree disease classification with generated data
SummaryThe paper addresses limited deep learning performance in plant disease diagnosis caused by scarce data and class imbalance. It applies augmentation, CutMix, MixUp and DCGAN-generated images under three data distribution scenarios. Using a pre-trained, fine-tuned EfficientNetV2B0, the authors report substantial gains in classification accuracy, underscoring the value of data diversity and class balance.
Hicham El Akhal, Abderrahim Waga, Aissa Ben Yahya et al.
In briefSelf-supervised pretraining for industrial health monitoring: what transfers and what federating costs
SummaryThe paper measures whether self-supervised pretraining helps Prognostics and Health Management, using 47 public datasets with 11 fully withheld from pretraining to prevent leakage. Transfer depends on the task: fault classification macro-F1 improves substantially on CWRU and HSG18, while remaining-useful-life estimation on CMAPSS shows no reliable gain over from-scratch training. Under matched budgets, centralized pretraining outperforms federated pretraining by 15.5 to 20.2 macro-F1 points in four comparisons, though this is confounded by a learning-rate schedule and single runs per regime.
WhySystematically compares SSL pretraining transfer and federated costs for PHM under leakage control, reporting negative findings. Single pretraining runs per regime weaken some comparative conclusions.
In briefIdentities, Not Pairs: Distribution-Free False-Accept Control for Embedding Verification
SummaryThis paper shows that a common method for setting false-accept thresholds in embedding-based verification, which uses impostor pairs that share identities, can fail to control error rates. The authors propose COVE, which ranks the trial pair among all pairs formed by adding the probe and reference to the calibration identities. Under identity exchangeability, COVE gives a finite-sample bound on the marginal false-accept rate. Experiments on 6 benchmarks and 8 frozen encoders found realised false-accept rates consistent with this bound.
WhyIntroduces an identity-exchangeable calibration method with a worst-case bound, validated empirically across six benchmarks and eight encoders.
In briefThe Voice on the Line: Causal Voice Isolation for Live Calls
SummaryThe paper trains causal voice isolators with 20 ms latency that enroll the caller from the call itself, to keep a caller's speech while suppressing other nearby voices. Using WavLM-Large weights made causal improves development-set error from 11.44% to 10.55%, and training on windows cut from longer calls lowers clean-call error from 3.92% to 2.96%. A 22.1M-parameter distilled model runs live on one CPU thread, and the release and CPU model are open-sourced with a LiveKit plugin.
WhyCausal WavLM co-training and long-call training improve live speaker isolation, with thorough comparisons against baselines across several benchmarks and real-time settings.
In briefLow-rank MERA compression, denoising and disease classification of 3D retinal OCT volumes
SummaryThis paper proposes a training-free low-rank MERA tensor network method to represent 3D optical coherence tomography (OCT) volumes, enabling compression, speckle denoising and feature extraction. The method is evaluated on volumes from two scanners, showing better visual quality and lower retinal layer segmentation errors than competing methods. Extracted features are classified with a Grassmann-manifold nearest-subspace classifier to detect retinal diseases in small-sample clinical datasets.
In briefMechAlign: Checking Whether Formal Math Projects Match Their Intended Claims
SummaryLean verifies that a formal proof holds under its stated assumptions, but it does not check whether the formalized theorem faithfully matches the mathematical claim in a manuscript. The paper introduces MechAlign, which combines LeanTrace's statement dependency graph extraction with LLM-based review to assess project-level semantic alignment. A preliminary review covered 201 main results from 173 OpenAI Math projects. The authors also built LeanReviewBench, containing 501 fixed claim-statement-context triples.
In briefDiagnosing Workflow Difficulty Before Blaming the LLM for Ontology-Grounded Translation
SummaryThis paper studies translating natural-language domain descriptions into ABox assertion statements grounded in an existing ontology. It argues that failures often stem from structural workflow issues, such as insufficient input, weak domain grounding, vocabulary gaps, weak validation, and non-local repair, rather than only from weak generator models. The authors propose a role-level difficulty profile with seven diagnostic variables across five workflow roles: Input, Generator, Output, Validator, and Repairer. The profile is experience-informed and non-exhaustive, meant to support workflow diagnosis and redesign or human-in-the-loop intervention.
In briefConversational bandits that combine item ratings and keyword comparisons
SummaryConversational bandits improve recommendations by asking users about key terms. This paper studies conversational generalized linear bandits that use both absolute item rewards and relative key-term comparisons as feedback. The authors propose HyConUCB, which builds unified confidence bounds from both feedback types and selects comparisons by balancing information gain and directional coverage. The paper reports a regret bound that improves on reward-only and dueling-only benchmarks, validated on synthetic and real-world datasets.
In briefHow well automatic translation quality scores handle idioms: many favor literal renderings
SummaryThis paper examines whether translation quality estimation systems can reliably score idiomatic translations above literal ones. The authors build a parallel dataset of English sentences containing idioms, paired with literal and idiomatic translations across nine target languages. Most of the eleven conventional metrics and ten large language models tested show a bias toward literal translations, and only a few consistently rate idiomatic renderings higher.
In briefStyle Over Substance? Language Bias in LLM-Based Peer Review
SummaryThis paper examines whether large language models used as peer reviewers are influenced by the linguistic style of research papers rather than their scientific content. The authors rewrite papers to show native or non-native English features and high- or low-confidence phrasing while keeping the science unchanged, then evaluate them with different LLM reviewers. They find that the reviewers systematically reward confident rhetoric independent of scientific content, raising concerns about the robustness of LLM-based peer review.
Sai Suresh Macharla Vasu, Sumedh Santosh Joshi, Philipp Christmann et al.
In briefNOMAD-EEG: An EEG Generative Foundation Model Independent of Electrode Layout and Input Length
SummaryThis paper introduces NOMAD-EEG, a 12.0M-parameter diffusion model for EEG that works regardless of electrode count, order, or recording length, identifying electrodes only by 3D scalp position. It was pre-trained on 147 hours of unlabelled EEG and evaluated with frozen linear probes on ten datasets with 1 to 64 electrodes and on two unseen populations. The authors report that it outperformed seven foundation models, including the 69M-parameter REVE, on several seizure, epilepsy and motor imagery tasks.
In briefVisual-Semantic Affect Graphs for Recognizing Emotion in Artworks
SummaryThis paper proposes Visual-Semantic Affect Graphs (VSAG), a graph-based method that represents artworks through interpretable features such as color, brushstrokes, composition, and visual entities, and learns their links to dominant emotions. An optional hierarchical module models valence-arousal structure. On the EmoArt-5k dataset, the hierarchical valence-arousal variant achieves the best overall performance, outperforming direct inference with Qwen3-VL and InternVL3.5 as well as the emotion-only VSAG baseline.
In briefReading Past OCR Noise: Robust Cross-Lingual RAG for Legislative Archives
SummaryThis paper addresses degraded, multi-column scanned parliamentary records from India, where mixed scripts and visual noise cause OCR errors. It presents an end-to-end system combining vision-based layout segmentation, multi-script OCR, and an agentic retrieval-augmented generation pipeline that iteratively works around OCR noise and layout ambiguity. Evaluated on the 200,000-page Punjab Vidhan Sabha archive, the system handles historical degradation robustly and outperforms text-only baselines.
Radha Krishna Deshpande, Ajoy Mondal, Akash Manna et al.
In briefThermal-headroom routing: predicting processor heat to choose robot vision models in extreme heat
SummaryThe paper proposes Thermal-Headroom Routing (THR), a lightweight policy that uses a closed-form RC thermal model to forecast processor temperature and exclude detector configurations likely to trigger throttling. In simulations at 45–55 °C with a tight deadline, THR eliminated throttling, reduced deadline misses by 3.9–7.5 points, and cut energy per on-time frame by about 12–13% versus a latency-aware router. The authors note that a tuned static thermostat or duty cap can match THR at a fixed operating point, with predictive routing most useful when conditions differ from tuning.
In briefCarbon-aware routing for multi-turn LLM web services across data centers
SummaryThis paper presents CarbonRouter, an online framework that jointly selects the model, GPU type, and region for each query to minimize carbon emissions. It accounts for the cost of recomputing conversation history cached on another deployment, using a cache-aware carbon cost model and a lightweight gate that keeps sessions on their cached deployment. In trace-driven experiments on two real conversation datasets, it achieves the lowest carbon of all policies tested.
In briefLARVA: Jointly Learning Latent Reasoning and Observation for Video Agents
SummaryThis paper introduces LARVA, a framework for distilling multi-round video agents into smaller models. An observer passes compact continuous representations, rather than text captions, to a reasoner that chooses the next evidence-gathering action or the final answer, and both parts are trained together. Across five video-understanding benchmarks, the 4B and 7B models perform strongly. The 7B model beats a textual chain-of-thought student trained on the same teacher trajectories by 5.1, 2.5, and 3.7 points on VideoHolmes, Video-MME, and LVBench.
In briefTraining Models to Keep Evidence Boundaries Without a Formal Answer Key
SummaryThis paper studies evidence-sufficiency judgment, deciding whether available information supports a causal claim, a task with no decidable correctness reward. Using a controlled design on the same base model, the authors compare a model trained with SFT then GRPO against its baseline, a scrambled-pairing placebo, and a general reasoning model. The trained model raised spontaneous correct refusal from 25% to 100%, while the placebo collapsed, and false refusal on 281 well-formed claims was 1.1%. The study covers a 14B same-family model, uses model rather than human raters, and makes no claims of general validity.
WhyPre-registered controlled knock-in shows reward training yields boundary-keeping in evidence-sufficiency judgment without decidable rewards; strong evidence but only one 14B same-family setting.
In briefGaussianSpline: Adaptive Bézier Splines for Dynamic 4D Gaussian Splatting
SummaryGaussianSpline is a dynamic 4D Gaussian Splatting method that models each Gaussian's motion with adaptive, non-uniform piecewise cubic Bézier splines. It jointly learns the control points and temporal partitions, giving more temporal capacity to highly dynamic regions. A reflection-based constraint enforces G1 continuity between segments. Experiments on the Neural 3D Video benchmark, the Google Immersive dataset, and synthetic scenes show improved reconstruction quality for highly dynamic scenes, with competitive performance on general dynamic scenes.
In briefO2D-MedNet: An Optimization-Aware Dual-Domain Framework for Robust Medical Image Segmentation
SummaryThis paper introduces O2D-MedNet, a framework for medical image segmentation that addresses image degradations and multi-scale structural variation. It combines a Fourier-wavelet dual-domain interaction module, a contour-guided fusion of CNN and visual state-space features, and a gradient-context strategy that stabilizes training. The authors report consistent improvements over state-of-the-art methods on six benchmarks across several imaging modalities.
In briefMulti-Agent Relay Thinking: AI agents pass reasoning to one another in turn
SummaryThis paper proposes a relay thinking topology for multi-agent collaboration, in which reasoning ownership is dynamically transferred so that each subsequent agent inherits and refines the evolving reasoning trajectory. The authors introduce Multi-Agent Confidence-guided Relay (MACR), which triggers relays based on confidence and selects the next agent through a confidence-based bidding mechanism. Experiments on four benchmarks report higher accuracy than the strongest single-model strategies, and performance comparable to multi-agent strategies at much lower token cost on some benchmarks.
In briefMuSID-KD: Multigranularity knowledge distillation for generative recommendation with semantic IDs
SummaryThis paper addresses the high inference cost of generative recommenders that autoregressively decode semantic IDs (SIDs) token by token. It proposes MuSID-KD, a knowledge distillation method that transfers knowledge from a strong teacher to a lightweight student, combining Hierarchical Candidate Distribution Alignment (HCDA) and Ranking Impact Guided Levelwise Distillation (RIGD). Experiments on real-world datasets with three lightweight student architectures show improved recommendation accuracy and better ranking alignment, without changing the student's online inference architecture.
In briefSkillLifeBench: A Benchmark for Agent Skill Security Across Its Lifecycle
SummaryThis paper introduces SkillLifeBench, a benchmark of 388 paired scenarios covering implementation, interface, runtime, composition, and version changes in reused web-agent skills. It evaluates 14 models on runtime tasks and autonomous lifecycle tasks, measuring joint success as completing the task without the registered attack goal. The authors also report that adding execution provenance improves recovery, raising joint success from 49.35% to 65.58% in a matched study.
In briefGeneration and Verification: Human-Agent Collaboration for Safety-Critical Automotive Software
SummaryThis paper presents a coding assistant for automotive C/C++ that offers generate, edit, and verify tools over the Model Context Protocol, so developers can run verification on any code. The authors customize a large language model for the automotive domain through prompt engineering and fine-tuning, using a dataset built from AUMOVIO production code. In a two-week pilot, developers used the verify tool most, and 75% preferred the faster prompt-engineered configuration.
Elizaveta Zinovyeva, Aiham Taleb, Nikita Kozodoi et al.
In briefOne-Shot Federated Model Editing for Edge Devices with Dynamic Membership
SummaryStandard federated learning needs devices to stay connected across many communication rounds, which fails in settings like vehicular networks where devices often join and leave. This paper proposes F-SOSA, the first stable one-shot federated learning algorithm that supports fast, exact model editing. Its variants F-SOSA-U and F-SOSA-I handle unlearning withdrawn contributions and incorporating new ones, respectively, without retraining from scratch. Experiments under realistic urban mobility simulations report retraining-level performance while cutting cumulative time and communication overhead by at least 36% and 22%.
WhyProposes first stable one-shot federated learning enabling exact editing, with theoretical bounds and urban mobility simulation experiments, though evaluation scope is limited.
In briefEfficient exact unlearning for federated zeroth-order optimization
SummaryThis paper proposes efficient exact unlearning methods for federated zeroth-order optimization (FedZOO), which fully remove the influence of training data when users request deletion. The authors design two stable zeroth-order learning algorithms and matching unlearning procedures that reuse computation through coupling. They prove that the unlearned model matches the distribution of retraining on retained data. The methods show competitive utility and reduce communication overhead substantially compared with first-order baselines.
WhyCombines exact unlearning with federated zeroth-order optimization via coupling-based reuse and proves stability and convergence; experiments are limited in scope.
In briefUnderstanding practical optimization methods used in machine learning
SummaryThis thesis analyzes practical machine learning optimization methods, such as submodel training, gradient clipping and normalization, layer-wise non-Euclidean updates, and approximate inner computations, rather than simplified classical proxies. It shows that Independent Subnetwork Training can converge to a shifted fixed point, while Model-Agnostic Sparsified Training yields unbiased sketched gradients with convergence guarantees. It also reports that adaptive quantile clipping keeps bias at any fixed quantile, and introduces the Gluon framework for Muon- and Scion-style methods, noting that Muon's orthogonalization precision should be tuned with the learning rate.
WhyGives new theory for practical methods like subnetwork training bias, quantile clipping, and layer-wise smoothness, with solid proofs and moderate subfield impact.
In briefSame Story, Different Lens: A Multi-Agent System for Comparing How News Outlets Frame the Same Event
SummaryThis paper presents an experimental research prototype that compares how different news outlets cover the same event. It collects world and politics coverage from 17 outlets, groups reports on the same event, and uses three model roles to analyze articles, assess source support, and generate cross-outlet briefings. The analyses cover sentiment, political lean, framing, loaded language, and possible omissions. The reported evaluation is preliminary; the grounding scores only reflect the model's judgment of support in the supplied text and do not establish factual accuracy.
In briefmdRNAcat: A Molecular Dynamics Dataset of RNA Shape Ensembles Using REST2
SummaryThis paper presents mdRNAcat, a curated dataset of all-atom molecular dynamics simulations of 48 RNA systems chosen for sequence diversity. Each system was simulated with replica-exchange with solute scaling (REST2), using 32 Hamiltonian states and 200-300 ns of production per state. The dataset includes atomic coordinates, forces, and metadata, and the authors analyze sampling, exchange behavior, base pairing, and sugar-pucker distributions. It is intended as a resource for developing and evaluating data-driven RNA models, including coarse-grained force fields.
Anton E. Dorn, Emile de Bruyn, Fabrice von der Lehr et al.
In briefOrthoRF: Adding orthogonality to object-centric visual representations
SummaryThis paper proposes Orthogonal Rotating Features (OrthoRF), which enforces orthogonality in the orientation space of Rotating Features using an inner-product loss and architectural changes. The approach yields sharper phase alignment and more reliable grouping of object parts. In unsupervised object discovery tests, including overlapping objects, noise, and out-of-distribution settings, it matches or outperforms current models without requiring post-hoc clustering, and it can recover occluded object parts.
Despoina Touska, Bastiaan Onne Fagginger Auer, Alexandru Onose et al.
In briefDTCA: token-level credit assignment for RL in diffusion language models
SummaryDiffusion language models generate text through iterative denoising, and recent work has applied reinforcement learning to improve their reasoning. Because final outcome rewards say little about how individual tokens contribute, the paper proposes DTCA, which estimates token contributions from the improved prediction confidence of remaining masked positions, groups them by part of speech, and calibrates them by entropy changes from policy updates. On LLaDA-8B-Instruct, DTCA consistently improves mathematical reasoning across generation budgets, with gains of up to 9.0 points on MATH500 and 5.3 points on GSM8K over the base model.
In briefMultilevel Newton Methods with Global O(k⁻²) Convergence
SummaryThis paper develops and analyzes multilevel gradient-regularized Newton methods for unconstrained convex optimization with Lipschitz-continuous Hessians. The methods compute Newton corrections in lower-dimensional subspaces to reduce computational cost. The authors aim to preserve the O(k⁻²) function-value convergence rate of the full-space method and introduce row-nested restriction operators to ensure monotonicity of reduced-gradient norms across levels. Numerical experiments show convergence comparable to full-space gradient-regularized Newton, with substantial savings on large-scale instances.
In briefLost in Translation: Systematic Differences Between English and Chinese Names of American Chinese Restaurants
SummaryThis study compares the Chinese and English names of American Chinese restaurants to see how immigrants adapt culturally. The authors sampled 3,172 restaurants and created NameFrames, a dataset of 884 annotated parallel name pairs. English names more often emphasize location and specialty, while Chinese names emphasize positivity and ambiance. Discrepancies are larger for urban restaurants, those in counties with more Chinese residents, and newer restaurants; interviews with 10 owners add context.
In briefLEAP: A Latent Embedding Anomaly Patch-Transformer for Detecting Unusual Time-Series Patterns
SummaryThis paper proposes LEAP, a model for detecting anomalies in time series data. Instead of reconstructing inputs or using attention-based scoring, it hides patches of a window and predicts them from the surrounding temporal and cross-channel context, using prediction errors as anomaly evidence. It combines this with three label-free signals: latent-embedding novelty, local temporal contrast, and channel-pooled novelty. On the TSB-AD benchmark, the multivariate model scores a VUS-PR of 0.5179 on 180 held-out series, and the univariate variant TSB-AD-U scores 0.6194 on 350 series, both averaged over three random seeds.
Cristhian Moya-Mota, Ignacio Aguilera-Martos, Diego Garc\'ia-Gil et al.
In briefV³AST-Bench: A Safety Benchmark for Coding Agents That Checks Actual Outcomes
SummaryThis paper introduces V³AST-Bench, a benchmark for evaluating the safety of coding agents that act through code. It contains 212 risk tasks across 18 categories, using runtime environments to verify real harm and task outcomes, along with paired capability tasks. Experiments across 51 model-harness combinations find that even frontier agents cause verified harm on 20.9% of tasks, and that recognizing risks does not reliably prevent harmful actions.
SummaryThis paper examines whether large language models that apply codebooks to label text in computational social science produce reliable annotations. The authors ran a large-scale failure-mode analysis across four datasets and five models, testing stability, robustness, and conceptual validity under controlled perturbations, domain shifts, and optimization procedures. They found that high aggregate scores often coexist with severe label-level disparities, run-to-run instability, and domain sensitivity, problems largely hidden by metrics such as accuracy and micro-F1.
Elena Senger, Yuri Campbell, Rob van der Goot et al.
In briefDAC-Loop: A Distillation-Aware Method for Compressing Remote Sensing CNNs
SummaryThe paper addresses the difficulty of running large remote sensing classification models on edge devices by compressing compact CNNs. It proposes DAC-Loop, an iterative method that combines knowledge distillation before and after structured channel pruning with alternating unstructured sparsification. On EuroSAT and NWPU-RESISC45, it reports a better accuracy-compression trade-off than one-shot pruning and single-stage distillation recovery, retaining strong accuracy beyond 90% parameter reduction.
Rico Staedeli, Lorena Raichle, Konstantinos Tzevelekakis et al.
In briefFormal explanations for temporal logic traces
SummaryThis paper studies how to explain whether finite traces satisfy or violate specifications written in Linear Temporal Logic over Finite Traces (LTLf). It introduces a native characterization of abductive (AXps) and contrastive (CXps) explanations based on locking and unlocking time-indexed events, proves its equivalence to a classifier-based formulation, and shows CXps coincide with a recent temporal causality semantics. The approach is implemented by enumerating minimal unsatisfiable cores and minimal correction sets of ASP programs, and experiments on real-world event logs demonstrate its effectiveness.
In briefDo Language Models Understand Igbo Culture? A Benchmark Beyond Multiple Choice
SummaryThis paper tests whether language models truly understand Igbo culture, using the Igbo language as a case study. The authors built CAB-Igbo, a benchmark from indigenous cultural sources verified by native speakers across six domains, and evaluated models with multiple-choice accuracy and human ratings. Most models showed weak cultural knowledge already in English, and prompting in Igbo mainly degraded linguistic quality and pragmatic skill.
Blessing Ngozi Uzor, Valdini Douglace Lemofouet, Paula Chikaodinaka Anyanwu et al.
In briefBiology-guided residual propagation for predicting spatial gene expression from histology images
SummaryThis paper proposes BCARP, a framework that predicts spatial gene expression from H&E-stained histology images as a lower-cost alternative to spatial transcriptomics experiments. Rather than decoding expression directly from visual features, it treats histology embeddings as morphological priors and refines spot representations through residual propagation combining nuclei-guided local context and prototype-guided global context. A pattern-scale decoder further calibrates the predictions. Evaluated on three ST datasets and ten HEST-1k cohorts, BCARP achieves the highest mean correlation-based performance among the compared baselines.
Vu Van An, Long D. Nguyen, Ho Viet Duc Luong et al.
In briefSpeeding Up SO(2)-Equivariant Materials Models with Low-Precision Emulation
SummaryThis paper presents FlashSO(2), a framework that uses emulated low-precision operations based on the integer Ozaki scheme to accelerate SO(2)-equivariant layers in graph neural network materials models. The authors report up to 40x layer-level throughput over reference implementations on an NVIDIA GH200 GPU. At the model level, they report a 3.9x faster training speed for MALOQ and 58% lower peak memory, with no loss in accuracy.
Denghui Lu, Alexander Maeder, Alexandros Nikolaos Ziogas et al.
In briefComparing two ways to combine tree models for synthesizing health data
SummaryThis paper examines how to create privacy-preserving synthetic health data using tree-based models of conditional distributions. It compares sequential specification and fully conditional specification for combining these models into a synthetic distribution. The authors evaluate synthetic medical datasets with several risk and utility metrics. They find no general superiority of either method, noting that sequential synthesis is sensitive to variable order while fully conditional specification has more hyperparameters and higher computational cost.
In briefAutoEnergyDesign: An LLM Agent That Automatically Optimizes Building Energy System Designs
SummaryThis paper presents AutoEnergyDesign, a framework in which a large language model agent proposes configuration changes based on previous simulation results and automatically reverts unsuccessful changes using a git-based mechanism. In a case study with measured building load profiles, the agent reaches a 25-year Net Present Cost of about 3.53 MEUR, outperforming random search, TPE Bayesian optimization and an evolution strategy. Every decision is recorded in an auditable trace.
In briefUnityBench-Mini3D: A Benchmark for Testing AI Coding Agents on 3D Game Development in Unity
SummaryThis paper introduces UnityBench-Mini3D, a benchmark of 77 full Unity project tasks grouped by seven skill categories and three difficulty tiers. Each task is checked with 1,178 human-written milestone tests covering scene structure, component setup, and runtime behavior, without relying on subjective judges. Across 21 models and 9 coding harnesses, average pass rates were 80.9% on easy tasks, 35.2% on hard tasks, and 57.7% overall.
In briefComplexity-guided regularization for generalizable human Gaussian splatting
SummaryThis paper addresses blurry rendering of fine details in generalizable human Gaussian splatting, which synthesizes novel views of unseen people from multi-view images. The method uses a complexity prior, computed from local entropy of geometric and textural features of clothed humans, to keep Gaussians from being sparsely placed in complex regions. It also samples local patches mainly from those complex regions for supervision. Experiments on benchmark datasets report state-of-the-art performance with real-time inference speed.
In briefLongSearchEval: A Framework for Testing AI Agents' Long Web Searches Using Real Interaction Logs
SummaryThe paper introduces LongSearchEval, a framework for evaluating how well LLM-based agents perform long-horizon web search using retained real interaction traces. It checks answers against the webpages agents cite and audits rounds to link failure nodes to incorrect answers. The authors evaluated 16 models from eight families on 100 tasks, producing 1,600 runs and 6,545 flagged failure nodes. Human validation reported 94.50% agreement on answer-group labels and 98.11% precision for round-group links.
In briefFaster AI-Generated Video Detection via Temporal Slice Learning
SummaryThis paper proposes detecting AI-generated videos by converting them into compact temporal slice representations instead of modeling full videos. It uses frequency-aware slice consistency learning with a lightweight Transformer and introduces AI-Artist, a new benchmark of artist-curated videos. Experiments on GenVideo and AI-Artist report strong performance, over 400 times fewer forward FLOPs, and 3.6 times faster inference than the baseline ReStraV.
In briefLocked In: Multi-Objective RL for AI Data Center Planning
SummaryThe paper argues that AI data center build-outs are multi-decade, multi-trillion-dollar projects, yet siting, power and cooling decisions still rely on static, one-time scoring tools. The authors formalize long-horizon build-out as a tiered multi-objective Markov decision process and propose a preference-conditioned agent that lets planners steer trade-offs among cost, carbon, water and reliability. Experiments on real site data compare three commitment regimes, and the authors report that their compact counterfactual dynamics model is outperformed by simple persistence on the water-stress axis.
In briefLow-precision storage and sparse execution for a folded FPGA speech enhancer
SummaryThis paper trains a streaming speech-enhancement accelerator on FPGA by pricing compression according to measured hardware costs of a folded design. Mixed-precision weights reduce on-chip storage, while structured channel sparsity removes whole operations from the schedule. On the Kria KV260, the best sparse model removes 47.3% of convolution MACs, cutting frame time by 45% and dynamic energy by 60% versus a storage-priced model measured in the same session, with higher speech quality.
In briefSex differences in children's hand bone shapes, analyzed with statistical shape models
SummaryThis paper uses statistical shape models to analyze how children's hand bone shapes differ by sex and age. The authors built a two-stage detection and segmentation pipeline that extracted 27,664 bone masks from radiographs of 1,456 subjects. Stratified shape parameter analysis reportedly revealed sexual dimorphism for each bone and sex-specific developmental differences, consistent with existing literature.
In briefCalibraRL: Confidence-Calibrated Reinforcement Learning to Reduce Hallucinations in RAG
SummaryLarge language models often give plausible but wrong answers. CalibraRL is a reinforcement learning framework that builds response confidence and per-query answering reliability into its reward design, penalizing overconfident errors and discouraging unnecessary abstention. On three QA datasets, it improves answer precision and reduces hallucinations compared with strong baselines. The paper also reports a lower mean expected calibration error (ECE) across two backbone models and three datasets.
In briefCo-Design Bench: A Multimodal Coding Benchmark for AI Agents Working With Real Users
SummaryThe paper introduces Co-Design Bench, a reference-free benchmark of 273 tasks drawn from real user coding sessions, with multimodal and repository editing tracks. Tasks are graded against hidden expert acceptance criteria using source and browser evidence plus a separate verifier. The authors report that the strongest of ten model endpoints passes an average of 69.9% of criteria per task but fully completes only 26.4% of tasks.
In briefDomain-adapted latent diffusion for translating CT scans into PET images
SummaryThe paper proposes a domain-adapted latent diffusion framework that transfers natural-image priors to medical imaging, using about 21,930 paired CT/PET slices to generate PET images from CT. It applies a signed power transform to handle the extreme intensity sparsity of PET data. The authors report better pixel-wise reconstruction and localized tumor fidelity than the CPDM baseline. They also argue that standard full-image metrics can be misleading for sparse PET data and propose background-filtered and mask-restricted metrics instead.
Yann Bourd\'e, Lea Bogensperger, Arnela Hadzic et al.
In briefTUON: A Tensor-Aware Optimizer Using Tensor-SVD and Polarization for LLM Pretraining
SummaryThis paper proposes TUON, an optimizer for pretraining large language models that is built on a transform-based tensor algebra derived from the Tensor-SVD. It mixes layers with a unitary transform, polarizes each transformed slice, and inverts the transform to recover the update, aiming to capture cross-layer correlations while keeping per-layer diversity. Experiments on GPT-124M/350M and Qwen-0.6B/4B report consistent improvements over Muon and TEON.
WhyProposes a T-SVD-based tensor-aware optimizer mixing layers before per-slice polarization, outperforming Muon and TEON with ablations on GPT and Qwen models.
In briefSTeP-TS: Shared Temporal Filters with Pruning for Time Series
SummaryThis paper presents STeP-TS, a compact one-layer convolutional model for multivariate time series forecasting. It learns a shared bank of multi-scale temporal filters, with lightweight variable-specific weights controlling each component's contribution. The learned weights and filter energy also guide structured pruning to reduce computation. Experiments report forecasting performance competitive with state-of-the-art methods, using substantially fewer parameters and inference FLOPs.
In briefWhen Fusion Order Matters: Progressive Multimodal Learning for Cross-modal Retrieval
SummaryThis paper studies how the order in which visual, audio, and speech information are fused affects text-to-video retrieval. The authors propose UniPRO, a framework with configurable fusion schedules and matched components for comparing orders. They find one-by-one fusion beats stacked fusion, and that the order matters, with the best choice depending on modality availability and the query. A visual-last schedule for image queries reaches 97.9 R@1 on MSR-VTT-9k.
In briefGenerative AI for Turbulence: Foundations, Applications, Validation and Benchmarks
SummaryThis paper reviews how generative AI, especially diffusion and score-based probability-transport models, can address turbulence problems such as unconditional generation, probabilistic reconstruction, reduced-order modeling, and temporal evolution. The authors argue that validation must check turbulence-specific properties, including energy transfer, structure functions, anomalous scaling, pdf tails, coherent structures, conservation laws, and ensemble diversity. They call for a benchmark ecosystem with open data, standardized challenges, hidden tests, physics-aware metrics, and documented failures. The paper is written as an introduction for both AI practitioners and turbulence researchers.
In briefQuality over Quantity: Rethinking Bengali Coreference Resolution Data and Model Assumptions
SummaryThis paper tests whether multilingual transformer models work well for coreference resolution, the task of linking words that refer to the same entity, in Bengali, a language with over 250 million speakers but few annotated resources. The authors evaluate zero-shot, fine-tuning, data augmentation, and cross-lingual transfer settings. They find that a small set of high-quality annotated data significantly outperforms larger mixed-quality datasets (p < 0.01), while back-translation augmentation does not help and cross-lingual transfer is inconsistent.
In briefUTA: A Training-Adaptive Learning Rate Decay Method
SummaryThis paper proposes UTA, a learning rate decay method that treats decay as a closed-loop problem of adjusting shape and scale based on the training state, and is applicable to Adam-type optimizers. The authors report consistent improvements on LLaMA language models (60M to 1B parameters) and ResNet50 and ViT-B vision models on ImageNet. Ablation studies are presented to examine its robustness.
In briefSubspace Crystallization: Rank Compression Tracks Grokking's Phase Transition
SummaryThis paper studies grokking, where neural networks generalize long after memorizing training data, using an MLP and a one-layer transformer on modular arithmetic. The steepest drop in effective rank occurs about 900 steps after the generalization onset, and networks that memorize without grokking compress much less. Capping weight rank did not speed grokking and often prevented it, so the authors conclude rank compression is a signature of grokking rather than its cause.
In briefRethinking evaluation of language models through Kripke's reading of Wittgenstein on meaning as use
SummaryThis paper argues that the view of meaning as use, which large language models operationalize, faces a rule-following challenge: finite histories of use do not determine a unique rule for future extension. It contends that benchmarks measuring performance on fixed use cases or local robustness do not by themselves establish semantic competence. To assess whether models preserve rules and concepts under principled domain extension, the authors propose evaluating abstraction and recursion and offer an experimental paradigm for testing both.
In briefUnderstanding and fixing cooperative perception failures in bad weather
SummaryThe paper studies why cooperative 3D detectors miss vehicles in adverse weather even when LiDAR points on the vehicle are unchanged, calling this Weather-Induced Context Interference. It proposes Adverse-Weather Cooperative Repair (AWCR), which uses separate components depending on available supervision. On natural adverse-weather CoInfra development data, AWCR improves car BEV AP@0.7 by 4.66 points with V2X-ViT and 5.56 with AttFuse.
In briefFlowGenesis: Neural networks that reshape their own structure during training
SummaryThe paper proposes FlowGenesis, a framework where neurons are born or pruned during training based on a measure called Representation Flux, which tracks changes in minibatch mean activations. Instead of fixing an architecture beforehand, the network adjusts its capacity as it learns. Experiments on image and text classification show networks of different starting sizes reach similar neuron counts and comparable accuracy. In task-incremental continual learning, it used fewer neurons and parameters than fixed-width Progressive Neural Networks while keeping similar accuracy.
In briefUseful or Hallucinated? Persona-Based Synthetic Queries for E-Commerce Search
SummaryThis paper studies whether LLM-generated synthetic search queries can train semantic parsers for cold-start e-commerce search. The authors generated about 118,000 synthetic query pairs using six LLMs and nine persona-conditioned prompt families on a UK automotive schema. Personality-based prompts improved parser F1, and syntax conditioning moved generations closer to real traffic, though about 37% of queries missed by all parsers were labeled hallucinated, mostly due to dropped constraints.
Filippos Ventirozos, Tom Armitage, Mahmoud Oshagh et al.
In briefSpherical Brownian Diffusion: a diffusion method for language modeling on the unit sphere
SummaryThis paper introduces Spherical Brownian Diffusion, which uses isotropic Brownian motion on the unit sphere as the forward noising process for continuous diffusion language models. It compares Brownian corruption with moment-matched von Mises-Fisher corruption and finds nearly identical results on Sudoku, while sampler choice produces a much larger performance spread. The authors separate a memory effect from re-noising and a token-selection effect, and report that categorical refinement helps reasoning tasks whereas stochastic updates preserve diversity in unconditional text generation.
In briefUsing retrieval-augmented generation to better explain implicit misogyny
SummaryThis paper proposes a retrieval-augmented generation framework for detecting and explaining implicit misogyny in English and Italian, drawing on encyclopedic, slang, and feminist knowledge sources. Experiments on the ImplicIT-Mis and SBIC+ datasets assess how much each source contributes to detection and explanation quality. RAG improved recall on ImplicIT-Mis but did not improve detection on SBIC+, while human evaluators preferred RAG-based explanations in both languages. Feminist literature was the most consistently useful source.
In briefImproving Instruction Hierarchy by Learning from Goal-Goal Conflicts
SummaryThis paper argues that existing instruction hierarchy training data lacks goal-goal conflicts, a conflict type it identifies through decomposing tasks into goals and constraints. Fine-tuning Qwen3.5-9B and Gemma4-12B-it on such examples improves instruction hierarchy adherence on many real-world tasks not seen in training. Adding these examples to existing training sets also improves adherence, with only trivial mean utility changes across eight no-conflict benchmarks.
In briefUncertainty evaluation for checking whether LiDAR detector performance targets carry over to other triggering conditions
SummaryThis paper addresses safety of the intended functionality (ISO 21448) analysis for machine learning LiDAR point cloud object detectors that reject detections above an uncertainty threshold to meet a performance target. It transfers a threshold selected under one condition into a triggering condition, uses a statistical test to decide whether the target is applicable, not applicable or not determined, and computes a lower bound on the independent units needed for an applicable result. Experiments on PointPillars ensembles with KITTI and nuScenes found that counting by recorded sequences left neither dataset enough independent units, while counting by frames produced not applicable outcomes in several transfers.
In briefStabilizing test-time prompt tuning by decomposing prompt updates
SummaryThis paper addresses instability in Test-Time Prompt Tuning, where prompt updates can grow too large or drift in inconsistent directions. It proposes TPD, which measures update expansion with a spectral radius and splits update directions into persistent, oscillatory, and residual components. The method combines Adaptive Koopman Control and Hankel Update Router to produce stabilized updates without changing the backbone, prompt architecture, or loss. Experiments on 15 datasets report consistent gains over 10 baselines.
In briefMachine reading of Pope Honorius III's registers without manual line transcription
SummaryThe paper addresses the cost of line-level transcription for medieval Latin registers of Pope Honorius III, using existing unbroken expert edition drafts instead. Forced alignment generates 13,122 training lines automatically, and a from-scratch hybrid Conformer model jointly transcribes and expands abbreviations. It reports 4.7% CER and 8.7% WER on test lines, outperforming a generic baseline and a modular pipeline. The authors also discuss preserving scribal name initials and report ablations showing which components did not improve results.
In briefA JiT-based denoising defense against adversarial attacks on radio modulation classification
SummaryThis paper addresses the vulnerability of deep neural networks used for radio signal modulation classification to adversarial attacks. It proposes a denoising defense inspired by Just image Transformers (JiT), training a 1D Transformer via flow matching to project inputs back onto the manifold of clean signals, without prior knowledge of attacks. On the RML2018.01A dataset, compared with PGD-based adversarial training, the method reports at least a 20.58% relative improvement in robust accuracy under PGD attacks and a 10.20% relative improvement in clean-sample accuracy.
In briefAsking Clarifying Questions Before Answering: QERA for In-Context Learning
SummaryThe paper notes that frontier LLMs solve fewer than one in five tasks on CL-bench Life, a benchmark of real-life context learning built from users' own web traces such as group chats and activity logs. It proposes QERA, which turns the request into a requirement checklist and typed information needs, answers from the whole conversation, re-asks the questions with code-verified quotes, and applies guarded local repair. QERA reports an average solving rate of 21.0% across 405 tasks, above the strongest of three inference-time baselines at 18.2%.
In briefZero-shot multi-class wafer map defect classification using rule-ordered prompts for large multimodal models
SummaryThis paper presents a training-free framework that classifies wafer maps into six defect classes without retraining. It pairs highlighting of connected components with prompts that account for morphology, location, and exclusion, applied to a frozen large multimodal model. On Mixed-WM38 (6,866 maps) it reports 92.9% macro and 93.8% overall accuracy, and on WM811K (9,000 maps) 79.1% and 81.3%. Eight alternative rule orders reach 90.1–91.9% overall accuracy, below the 93.8% of the selected order.
In briefModeling fast human reactions from high-frequency game interaction data
SummaryThis exploratory study analyzes high-resolution Counter-Strike 2 telemetry to find signs of rapid, salience-driven behavior. It introduces Ego-Knowledge, a method reconstructing each player's plausible perceptual state, and derives fight-or-flight labels from a 60–120 ms window. Two telemetry-native models, EMBER and ORBIT, using only pre-trigger information, reach 73.5% and 71.2% validation accuracy, above the 51.5% majority baseline and the strongest LLM baseline (62.7%) on the same cases. The authors state the neuroscience framing is hypothesis-generating rather than mechanistic.
Emanuel Dopater, Daniel Gavlas, Luk\'a\v{s} Jochymek et al.
In briefBoosting Vision-Language Models on Scientific Charts Without Retraining
SummaryThis paper improves a frozen vision-language model's chart question answering by evolving its inference program and surrounding code, rather than changing the model. Using an LLM-guided autonomous evolution approach, accuracy on descriptive questions across three models and four benchmarks rose from 16.6% to 41.3% on average. The evolved program is also 3.8 times faster, and the authors conclude that many errors stem from the inference pipeline.
Abhinav Sudhakar Dubey, Vidushi Goel, Kiran Rana et al.
In briefGenerating facial blendshapes for anime-style 3D models made of disconnected parts
SummaryThis paper addresses blendshape generation for 3D models with multiple disconnected parts, an area with little prior research. Using commercial anime-style character models with up to 11 parts per face, the authors train neural networks to produce deformation fields around input meshes, with each part processed by a network that also receives an anchor part's geometry. They improve accuracy by predicting whether displacements are zero and by grouping blendshapes by size with dedicated networks, and they enhance bilateral symmetry. The authors state their method is simpler than recent neural-network approaches.
In briefA diagnostic method to tell whether a GNN's gain in load forecasting comes from graph edges or parameter sharing
SummaryThe paper proposes a diagnostic protocol that compares a graph neural network with three reference models: a per-node network, a shared network without neighbors, and the same GNN without edges. On low-voltage feeder data, parameter sharing explains most of the error reduction, while adding edges gives a smaller gain that largely disappears under degree-preserving rewiring. Results on client consumption series are less consistent, and the authors conclude that graph gains should be credited only after comparison with per-node and shared references.
Eloi Campagne (CB), Yvenn Amara-Ouali (LMO), Yannig Goude (LMO) et al.
In briefJoint Dynamic Lane Reversal and Variable Speed Limit Control: A Graphormer-Based Approach
SummaryThe paper studies how connected and automated vehicles could frequently reassign lane directions and speed limits across a road network. It proposes a network-wide controller that keeps the network strongly connected, using graph orientation theory and a shared transformer trained by pretraining and per-road reinforcement learning. On a 6 x 6 grid with commuter rush hours, the learned policy cuts daily delay by 90% against the fixed initial configuration and by 85% against a tuned two-way load balancer.
Hamed Araab, Abbas Ahmadi, Mohammad Hossein Basiri
In briefMulti-attribute benchmark (ComFair) for checking intersectional fairness in LLMs
SummaryThis paper argues that most LLM fairness benchmarks test only one sensitive attribute at a time, which can miss harms arising from interacting attributes. The authors introduce ComFair, which converts existing fairness data into 479,066 controlled dual- and multi-attribute instances, each paired with its task-equivalent source. Experiments on ten LLMs show that sensitivity varies by task and subgroup, and that model rankings change with attribute composition.
In briefMemVLN: Streaming episodic memory that helps robots follow language instructions through 3D spaces
SummaryThis paper presents MemVLN, a streaming episodic memory framework for vision-language navigation, which keeps a bounded bank of compressed perceptual features and turns them into compact memory tokens. It reaches state-of-the-art results on most metrics of standard VLN-CE benchmarks when trained only on official R2R/RxR trajectories. In physical tests on held-out episodes, success rose from 17% for the VLM baseline to 40%, and to 47% after post-training.
In briefVideo completion method combining Kronecker structure with cross-fitted self-supervised residual learning
SummaryThis paper proposes a two-stage method for reconstructing temporal trajectories at spatial locations never observed in any frame of a video. It first estimates a compact structural backbone using regularized masked Kronecker product decomposition, then uses a lightweight 3-D CNN trained with cross-fitted residuals to predict corrections. Observed measurements are kept exactly. Experiments on grayscale and RGB video report strong PSNR, SSIM, and temporal-gradient results.
In briefRLEC: Reinforcement Learning for Explicit Confidence Targets
SummaryThis paper addresses how language models give plausible but wrong answers when uncertain, and why existing abstention training with a fixed error penalty cannot adapt to different confidence requirements. The authors propose RLEC, which lets a user set a target accuracy p in the prompt so a single model decides whether to answer or abstain. Experiments across benchmarks and model scales report that RLEC largely preserves general capabilities and achieves competitive calibration.
In briefFrom Questions to Queries to Insights: A Local, Schema-Grounded AI Workflow for Crash Data Analysis
SummaryThis paper presents CrashDB-QA, a fully local agentic workflow that lets users analyze crash databases using natural language. Specialized components interpret questions, generate and validate read-only SQL, run queries, and explain results, grounded in the agency's data dictionary and safety-analysis conventions. The authors evaluate it on 4.4 million Wisconsin crash records with 300 questions and nine local open-weight models.
In briefLOCUS: A visual grounding method that first locates an object, then verifies its box coordinates until stable
SummaryThis paper proposes LOCUS, which combines a regression box head with a vision-language model's own coordinate tokens. The head first drafts a bounding box, then the same backbone checks the coordinate tokens through Jacobi iteration until they settle. On RefCOCO-family benchmarks, verification raises Locus-2B's Acc@0.9 from 54.87 to 59.13, and exact answers are produced 1.5 times faster than vLLM greedy decoding of the same tokens.
In briefRisk-aware output control should be the priority of machine unlearning
SummaryThis paper argues that most machine unlearning research aims for parameter-space similarity to models retrained from scratch, which does not match practical user concerns. The authors contend that unlearning should prioritize statistically guaranteed output-space control instead. They support this position with empirical evidence and recent theoretical results.
In briefCA-HNM: Constraint-aware hard negative mining for dense retrieval using structured domain models
SummaryThis paper proposes CA-HNM, a framework that selects hard negatives from retrieved candidates using deterministic incompatibility rules derived from a structured domain model, and uses them to fine-tune a dual-encoder dense retriever. Structured information is needed only during offline negative construction. On MOOCCubeX and Course-Skill Atlas, CA-HNM-Pure achieves higher mean NDCG@10 than a retrieval-rank-matched control in six of eight dataset-encoder-loss configurations. The results depend on the training objective and the composition of the negative pool.
Ngoc Luyen Le, Marie-H\'el\`ene Abel, Bertrand Laforge
In briefDriveUniTok: A Unified Tokenizer Linking Vision, Depth and Motion for Driving World Models
SummaryThe paper introduces DriveUniTok, a tokenizer that encodes driving appearance, projected LiDAR depth, and ego motion into one decodable continuous latent space. A frozen tokenizer is paired with a pretrained diffusion model, DriveUniWAM, which predicts future unified latents that are decoded into future appearance, depth, and trajectories. On NAVSIM, the authors report better future prediction and planning than separate-stream methods at lower downstream computational cost. After reinforcement learning post-training, DriveUniWAM reaches a PDMS of 91.54, which they describe as state-of-the-art among evaluated WAM-based planners.
In briefSimPOTrans: Reinforcement learning that lets offline speech translation models adapt to simultaneous translation
SummarySimultaneous speech translation aims to translate speech as it happens with low delay and good quality. SimPOTrans is a reinforcement learning framework that uses a two-stage post-training pipeline to adapt offline speech translation models to dialogue-style simultaneous translation, without synthesized read/write trajectories or explicit latency optimization. At inference it needs no external decision policy, and quality-latency trade-offs are controlled by adjusting the speech chunk size. Experiments on ACL 60/60 and MCIF across three directions show it outperforms several strong baselines.
In briefInfrared thermal wake detection for underwater vehicles without cooling-water discharge
SummarySmall uncrewed underwater vehicles are hard to detect with sonar and magnetic sensors. This study builds AVP-IR-Wake, a real-world infrared benchmark of 161,485 frames collected over three years. It finds that individual wake vortices look statistically identical to wind-induced background vortices, but the whole wake is organized by vehicle trajectory while background clutter follows wind-driven thermal fronts. The authors conclude that reliable detection requires trajectory-constrained global physical modeling rather than local feature extraction.
In briefMapping which grokking tasks generalise on Latin squares
SummaryGrokking is a delayed shift where a network first memorises training examples, stays near chance on unseen inputs, and then suddenly generalises. The paper models pairwise input tasks as edge colourings of K_{q,q}, using Latin colourings, and sorts tasks into all-good, mixed, or no-good by their minimum-norm endpoint sets. The authors prove that finite-abelian-group families are asymptotically all-good from partial observations, while under a uniform random Latin square all-good tasks become asymptotically rare. Experiments also show that relabelled addition still generalises.</summary>
In briefMacro Mondrian Conformal Risk Control: a method that controls risk per data group
SummaryThis paper proposes a new conformal risk control method that controls risk on average over user-defined groups of the data, rather than only marginally over the whole distribution. Each data point is assigned a group by a taxonomy, and its risk is reweighted by its group probability so that rare points count more. The approach is illustrated on two semantic segmentation datasets.
Julien Combes (SVH, Michelin), Jean-Francois Coeurjolly (SVH) et al.
In briefAutoPRL: Learning Vehicle Route Representations with Automatically Generated Views
SummaryThe paper proposes AutoPRL, a self-supervised framework for learning representations of paths traveled by vehicles on road networks. It uses variational autoencoders to generate diverse path views, decoupled cross-view contrastive learning, and adaptive co-training to balance training objectives. The authors report consistent effectiveness and efficiency across several downstream tasks on three real-world datasets.
In briefPatentAgentBench: A Benchmark Testing AI Agents on Professional Patent Tasks
SummaryThis paper introduces PatentAgentBench, a benchmark of 510 instances across eight patent task suites, including novelty search, freedom-to-operate search, and patent drafting. The tasks are built from real patent records and professional materials, with expert-reviewed annotations and rubrics. Seven LLMs were evaluated across four agent harnesses. The authors report that current systems still struggle to identify novelty, assess risk, and understand practitioner intent.
In briefComputer-generated parent-child conversations do not match how real children and adults talk
SummaryThe paper compares the synthetic TinyDialogues dataset with natural dialogue across many linguistic measures. Synthetic data matches natural data at the lexical level but greatly underestimates the messy, fragmented character of spoken interaction. The authors propose instruction-tuning with data built from real dialogues, and a pilot study shows a small model approximates natural data's statistical tendencies after moderate parameter-efficient fine-tuning.
In briefBellowsKV: A Training-Free Method Sharing One Latent Across Layers to Compress the KV Cache
SummaryLarge language models store attention keys and values for past tokens in a KV cache, whose memory use often limits GPU serving. BellowsKV is a training-free method that maps the hidden states of adjacent layers to one shared latent per token, with each layer reading its own keys and values through a layer-specific projection. The latent is compressed by low-precision storage (BellowsKV-Quant) or by making one shared eviction decision per block (BellowsKV-Evict). On Llama-3.1-8B, Mistral-7B and Qwen2.5-7B, BellowsKV-Quant keeps about 99% of full-cache accuracy at 3.6x compression.
WhyNovel cross-layer latent design with precision-based compression retains near full accuracy on three models, substantially aiding LLM serving memory bottlenecks.
In briefAttention with Bounded Key-Query Similarity: Space Complexity and Transformer Anisotropy
SummaryThis paper studies the space needed to approximate attention when key-query cosine similarities are assumed to lie in a small interval around zero. The authors give a nearly space-optimal algorithm and a matching space lower bound. The similarity distribution in their lower bound shows the anisotropy previously observed empirically, and uniform sampling is shown to be nearly optimal over typical parameter ranges.
WhyGives near-optimal space algorithm and matching lower bound under bounded cosine similarity, explaining anisotropy and uniform sampling effectiveness, with moderate practical reach.
In briefZero-shot depth estimation for prosthetic vision
SummaryThis paper examines whether monocular depth estimation foundation models can support visual perception in lower-limb prostheses. The authors introduce the LP dataset, a benchmark for depth perception from an emulated knee-level prosthetic viewpoint. Four state-of-the-art monocular depth estimation models are evaluated on it, providing the first systematic benchmark from a prosthetic viewpoint.
Ishi Vijayvergiya, Marc H\"olle, Youssef Dawoud et al.
In briefWhen Pointing Becomes Profiling: Bias and Refusal in Text-Steerable Vision Encoders
SummaryThe paper audits SteerViT, a text-steerable visual encoder, and finds its language-driven pointing can amount to demographic profiling, such as crime-related prompts selecting Black people more often. It shows the model cannot refuse problematic requests and introduces SafeSteer, which adds a learned abstention output and counterfactual-invariance training. SafeSteer largely removes race routing while keeping localization and retrieval within about one point of the released model.
WhyShows text-steerable encoders turn language pointing into demographic profiling and proposes SafeSteer with abstention; counterfactual mosaics and multi-model comparisons give solid evidence.
Seyedarmin Azizi, Parsa Razmara, Sajjad Ghiasvand et al.
In briefMultiscale Policy Optimization: A Training Method for Long-Horizon LLM Agents
SummaryLong-horizon LLM agents interact with environments through complete assistant turns, generating each action token by token. The authors observe that critic values vary mainly across turns and propose Multiscale Policy Optimization (MPO), which combines Multiscale Advantage Estimation (MAE) with turn-level importance sampling to align credit assignment and policy updates at the turn scale. Experiments on search, tool-use, and automation tasks with 4B and 30B-A3B models show consistent gains over strong RL baselines.
In briefCritic's Eye: Graph-based reasoning for finding recurring motifs in long works
SummaryThis paper studies how recurring literary motifs, such as blood in Macbeth, gain symbolic meaning through repetition with variation across a whole work. It defines Computational Motif Discovery, which predicts missing line-motif links using the full narrative structure of a play. The proposed Critic's Eye models the play as a heterogeneous Narrative Topology Graph and reports 84.8% Hit@5, far above Gemini-3-pro-preview at about 22.6%. The authors conclude that explicit structural priors are more effective than scaling parameters for this task.
In briefWhen Persona Clashes with the Line: Finding Where Speech Synthesis Loses Intent
SummaryThis paper examines where chain-of-thought text-to-speech systems lose the intended emotional direction when a requested persona conflicts with a line's lexical sentiment. The authors build a bilingual Japanese-English benchmark of congruent and conflicting persona-line pairs and trace the expected acoustic direction through planner outputs, emotion vectors, and rendered waveforms. Planner outputs are usually direction-correct, but downstream agreement is considerably lower. Supervising QwenEmotion raises direction agreement to 75% in Japanese and 84% in English.
