In briefAtoms, Not Knowledge: What Large-Scale Addressed-Memory Writes Store
SummaryThis paper tests whether deterministic hash-addressed conditional memory can inject knowledge after training, using models from 124M to 7B parameters and a large FP8 model. The writes behave as independently retrievable, precisely addressable, reversible atoms that do not compose into reasoning, and two-hop composition does not exceed the single-hop baseline. The authors propose a layered evaluation protocol and recommend reporting coverage, synergy ratio, and distortion rate before deployment claims.
In briefSENRI: Studying how AI links events to places in real-life egocentric videos
SummaryThis paper introduces SENRI, a task that examines how events and spatial locations connect in long, continuous first-person videos of real life. It defines six task types across two directions, Event→Space and Space→Event, and builds SENRI-Bench with 500 manually annotated multiple-choice questions over about 25.3 hours of video. The authors also propose SENRI-Agent, which uses Target-Guided Visual Check to improve existing memory-based baselines without changing their stored memories.
In briefShared-private cross-modal learning for detecting audio-visual deepfakes from imbalanced multiple sources
SummaryThe paper studies detecting audio-visual deepfakes when training data come from several sources with uneven distributions, using FakeAVCeleb and LAV-DF with known source identity. It adds shared and source-specific feature streams before the first block of the Multi-Scale Cross-Modal Transformer Encoder, and a gated residual module injects private evidence into the shared representation. Under two controlled source-composition protocols, the method reports higher macro and worst-source AUC than the MRDF and MSCT baselines.
In briefLearning When to Abort in Transformer Planning: A Mechanistic Study
SummaryThis paper studies whether Transformers can detect their own planning errors, using path-finding in graphs as a test setting. The model is trained to output an abort token after an invalid transition, and the authors analyze the internal mechanism behind this detection. Experiments report that combining error detection with retries improves planning reliability on graphs, Countdown, and mazes.
In briefAligning Language Models with Indigenous Community Worldviews via Constitutional AI
SummaryLarge language models are mostly trained on Western, English-language data, which can clash with Indigenous communities' assumptions about territory, authority, and nature. The authors built a 12-principle cultural constitution from interviews with five Indigenous leaders in Latin America and applied a two-phase Constitutional AI pipeline (supervised fine-tuning and direct preference optimization) to Qwen2.5 models. Using only 24 preference pairs, they report modest but detectable gains in intergenerational transmission, linguistic caution, and consultation and consent.
Silvia Fern\'andez-Sabido, Jorge Reyes-Maga\~na, Jes\'us Adri\'an Chay Tun et al.
In briefWorld Model Interpretability: Understanding the Latent Worlds Behind Predictive AI
SummaryThis Perspective paper examines Joint-Embedding Predictive Architectures (JEPAs), which are evolving from representation learning into latent world models for prediction, planning, and control. The authors argue that these latent worlds lack a human-interpretable interface, so interpretability should shift from what latents encode to how they interact, evolve, and causally drive prediction. They propose a four-stage agenda, Ground, Structure, Verify, and Act, noting that existing JEPA interpretability work offers early evidence only for the Ground stage.
In briefMotionShift: Motion transfer that fits the target scene's physics
SummaryExisting motion transfer methods copy movements from reference videos accurately but often ignore whether the motion fits the physics of the target scene. MotionShift uses a large language model as a planner to generate motion trajectories and a vision-language model as an evaluator to iteratively correct them, without training. The adapted motion then conditions an image-to-video model to render the target scene. The authors report high-fidelity, physically compatible transfer across diverse scenes and show it can be integrated into existing trajectory-conditioned I2V models.
In briefThe Explainability Illusion: Must ViTs Show Patch-Decoding Interpretability?
SummaryThis paper examines whether decoding individual patch tokens in vision transformer classifiers yields interpretable outputs. Across diverse architectures, scales, and training objectives, the authors find some models show aligned explanations, while others show inverted maps or no correlation with image semantics. They conclude this explainability is not an inherent property of trained transformers but depends on the problem setting, so it should not be assumed a priori.
In briefVision as Proprioception: Estimating Humanoid Robot Pose from Video
SummaryThis paper presents HumanoidPose, a framework that recovers a humanoid robot's full-body joint configuration in world space from video alone. It combines pretrained human-pose features with robot kinematics and retargets human motions to sixteen humanoid embodiments for learning from synthetic data. Unlabeled online videos are used for self-training a 2D robot-pose detector, and the authors report improved joint-angle and joint-position accuracy over the evaluated baselines.
Jiaxi Jiang, Matthias Heyrman, Ren\'e Zurbr\"ugg et al.
In briefA stochastic model of ADHD symptoms and methylphenidate treatment using brain imaging data
SummaryThis paper proposes a mathematical and computational framework that treats methylphenidate as an optimal control policy acting on ADHD's nonlinear dynamics. ADHD neural processes are modeled as stochastic processes whose trajectories tend toward stable but dysfunctional patterns linked to clinical symptoms. The drug is formalized as an external control signal that changes the system's drift and steers trajectories toward neurotypical regimes. Latent representations are inferred from preprocessed functional MRI data using a deterministic hybrid architecture combining a one-dimensional CNN and a Vision Transformer.
In briefLearning Lane Fields to Guide Online HD Map Building
SummaryThis paper introduces the Lane Field, a dense BEV-aligned representation that encodes local lane structure at each location, predicted from onboard observations or built from external map priors. Building on it, the authors propose a plug-and-play framework that improves online HD mapping through BEV feature modulation and deformable attention guidance. Evaluated primarily on OpenLane-V2, the method consistently improves existing approaches and reports state-of-the-art lane segment perception performance.
Andrea Bertogalli, Anthony Dell'Eva, Marcello Ceresini et al.
In briefData-driven forecasting of tropical cloud forest flowering and fruiting cycles using dynamical systems and machine learning
SummaryThis study proposes a hybrid method combining dynamical systems theory with machine learning to forecast flowering and fruiting cycles without relying on predefined functional forms. Using 35 years of data from a Venezuelan tropical cloud forest, the authors reconstruct the system's state space with delay vectors inspired by Takens' theorem and include temperature and precipitation as climate drivers. Gaussian Process Regression provides point predictions with uncertainty estimates, and the authors report higher predictive accuracy for flowering and fruiting.
In briefBridging Concepts: Multi-Agent Reasoning on Concept Graphs to Predict High-Impact Research
SummaryThis paper builds scientific concept graphs (SCGs), where nodes are fine-grained research concepts extracted from papers by LLMs and edges show which concepts co-occur in papers. It aims to forecast high-impact concept pairs that are not yet connected but would have high impact if linked. The authors release benchmark datasets from CIKM, KDD, and SIGIR papers and propose MiroGraph, a multi-agent framework, which reportedly outperforms various impact forecasting and temporal link prediction methods.
In briefHolo4D: Using Holistic 4D Reconstruction to Guide Video Diffusion for Camera Control
SummaryThis paper presents Holo4D, a framework that re-renders a monocular video along a new camera trajectory. It reconstructs 3D trajectories from the source video with a feed-forward 4D reconstructor, projects them into the target views, and uses the encoded geometry and appearance to condition a pretrained video diffusion model. Experiments show that this trajectory-based conditioning supports camera control and novel-view synthesis, and the framework works with different 4D reconstructors without retraining the generator.
In briefGeoART: Geometry-aware test-time refinement for camera-controlled video generation
SummaryCamera-conditioned video diffusion models often have inaccurate camera control and geometric inconsistencies because camera constraints are only weakly enforced during denoising. GeoART is a training-free self-refinement framework that generates multiple candidates at selected denoising steps and evaluates them with a geometry prior, without directly modifying the denoising trajectory. Experiments report improved camera controllability, multi-view consistency, and visual quality over existing methods, without training or modifying the pretrained model.
In briefEG-GRPO: Evidence-guided step-level credit assignment for deep search agents
SummaryThis paper proposes EG-GRPO, a reinforcement learning framework for LLM-based deep search agents that derives step-level supervision from evidence annotations in synthetic search data. It identifies the first supporting citation for each required evidence claim to build step-level process rewards, and combines them with the standard trajectory-level outcome advantage. Experiments on five deep search benchmarks reportedly show that EG-GRPO generally outperforms standard GRPO and other baselines across model scales.
In briefSPARC: A Benchmark Testing Whether Transcribed Speech Reveals Who Is Speaking
SummaryThis paper examines whether transcribed words alone reveal the speaker's identity and demographic attributes such as sex, age, and accent. The authors built SPARC, a benchmark from two podcast corpora with over 12,500 speakers and 1.5M filtered utterances. Experiments show that LLMs can tell same-speaker pairs apart from text alone and verify sex and accent labels above chance. Releasing only short 50-word probe texts partly reduces observable linkage.
Efthymios Tsaprazlis, Bowen Yi, Tiantian Feng et al.
In briefEasyResearch: Honest Regression, Forecasting and Signal Classification Tools for Non-Programmers
SummaryThe paper presents EasyResearch, open-source tools for regression, time-series forecasting and signal classification, plus EasyResearch Desktop, a code-free Windows application. All learning uses training data only, and the selected model receives a separate final score. Using held-out data, the authors report that common practices overstate R² and forecasting accuracy, while EasyResearch's final scores are closer to true performance.
In briefHow many variables remain at coarser scales in AlphaEarth embeddings
SummaryThis paper examines whether the 64-component vectors that AlphaEarth assigns to 10 m land pixels can be reorganized into ranked global nonlinear coordinates. It asks how many of these coordinates are needed for a given reconstruction accuracy and whether that number falls at coarser spatial scales. At 3% validation tolerance, the count drops from 24 at 10 m to 7 at 78.9 km, while the PCA rank drops only from 46 to 38.
In briefPanoDent-KD: a panoramic dental X-ray detection pipeline from general distillation to specialist and embedded deployment
SummaryThis paper presents PanoDent-KD, a pipeline for detecting dental structures in panoramic radiographs. It trains a 12-class source model under three arms (baseline, single-teacher and multi-teacher distillation), transfers shared classifier rows to a five-class specialist for the OralXrays-9 dataset, and deploys the model on a Jetson Orin Nano Super. All three arms reach macro box F1 near 0.82, with the supervised baseline scoring highest on average, and the deployed engine runs at about 77 ms median latency. The authors state that the results show feasibility of an auditable, device-specific workflow but do not establish clinical validity or accuracy gains from distillation.
In briefBatchParScale: A sparse attention method letting parallel reasoning streams share information
SummaryThis paper introduces BatchParScale, an attention mechanism that lets parallel reasoning streams exchange information through a small set of communication positions, unlike earlier parallel scaling methods where streams stay independent. The authors evaluate it on a new Parallel Needle-in-a-Haystack retrieval benchmark and in pretraining, fine-tuning, and cooperative multi-agent reinforcement learning. It outperforms ParScale on language-model loss, but fact retrieval performance for pretrained and fine-tuned models remains unclear.
Hans Peter Lyngs{\o}e Raaschou-jensen (University of Copenhagen), M. Onurcan Kaya (Technical University of Denmark), Desmond Elliott (University of Copenhagen)
In briefPersistent Representational Misalignment Causes Vision-Language Model Failures
SummaryThis paper examines why vision-language models (VLMs) fail at some tasks that are easy for humans. The authors argue that a mismatch between vision and language representations is a major cause. They show that existing vision and language representations cannot be linearly aligned, and that this misalignment is inherent to the data, so it would persist even with greater scale. More representationally aligned VLM backbones showed fewer failures, which the authors present as evidence against a strong version of the Platonic Representation Hypothesis.
Yonatan Gideoni, Yoav Gelberg, Tim G. J. Rudner et al.
In briefMulti-hop retrieval to boost hybrid sequence model expressivity
SummaryThis paper examines how giving efficient layers in hybrid language models access to a full-attention layer's KV cache affects expressive power. The authors prove that conventional hybrid blocks and those with top-κ KV-cache access need Ω(√n) capacity for two-function composition. They introduce the Multi-hop Retrieval Attention (MuRA) block, which retrieves from a shared KV cache over multiple hops, and show that an m-hop block can compute m-function composition, establishing a depth-based expressivity hierarchy. Synthetic experiments support this hierarchy.
In briefVisual Scaffolding: Toward Native Multimodal Reasoning
SummaryThis paper proposes visual scaffolding, a mechanism that lets a multimodal LLM autonomously retrieve native visual features at any reasoning step, with a lightweight retriever inserting them into the chain of thought for reuse. The authors also introduce NMR-Bench, covering Burden Counting and Path Tracing. Adding about 88M parameters to Qwen3-VL-4B-Thinking, the model reaches 66.43% and 72.50% accuracy, which the authors report surpasses the evaluated GPT-5.5 and Gemini 3.1 Pro baselines.
In briefEstimating causal bounds from observational data using neural generative models
SummaryThe paper introduces a framework that uses neural generative models to estimate causal bounds from discrete observational data, along with a new extrapolation procedure to improve accuracy. The authors report that across multiple causal inference settings, the bounds are consistently closer to the true bounds than those from existing general neural approaches. The framework is validated on benchmark problems where exact bounds are available.
In briefGeoV2V: A geometry-guided video model that generates driving scenes along new paths, such as lane changes
SummaryGeoV2V is a video-to-video diffusion model that generates temporally consistent driving videos along novel trajectories, such as lane shifts. It combines LiDAR and depth-derived geometric priors with full-video conditioning on the Wan 2.1 backbone. The authors also introduce Para4D, a synthetic dataset of synchronized multi-lane driving videos that supplies supervision for lane-shift generation. The paper reports state-of-the-art results on Waymo, nuScenes, and Para4D in visual quality, geometric consistency, and robustness to lighting changes under large trajectory shifts.
In briefEvidence on Loss of Control from Agentic General-Purpose AI
SummaryThis paper reviews 19 recent empirical studies against eight properties drawn from threat models, such as strategic deception, situational awareness, and autonomous replication. It rates evidence quality on construct validity, content validity, and external validity. The authors conclude that current evidence supports a judgment of loss of control as "weakly plausible," meaning the evidence points in a direction but is not conclusive. They also note that for the most dangerous properties, high external validity may be structurally hard to reach under responsible research practices.
Risto Uuk, Santeri Koivula, Lorenzo Pacchiardi et al.
In briefComparing Cloud and Self-Hosted OCR for Arabic, English, and French Documents
SummaryThis study compares cloud and self-hosted OCR systems on bounding box detection, text recognition, and mathematical expression recognition across Arabic, English, and French documents. Text recognition was scored by character error rate against a Gemini-generated reference transcription. The authors found that near-perfect page scores could hide errors such as replaced Arabic variables or converted digits, so they argue reliable evaluation needs uncertainty estimates and structural checks.
Jumana Alsubhi, Ahmed Alhusayni, Abdulrahman Gharawi et al.
In briefMap2Vid: Generating navigation videos from top-down maps for easy human control
SummaryThis paper introduces Map2Vid, a task that generates embodied navigation videos from bird's-eye-view (BEV) maps, which show scene layout and intended routes. The authors build a benchmark dataset of over 100K navigation sequences across 61 indoor scenes, pairing egocentric videos with BEV maps, camera trajectories, and language instructions. They propose a unified framework with curriculum training that gradually reduces non-map inputs, and report experiments across multiple foundation generation models.
In briefDROD: Dual-Ratio Ordered Dropout for federated continual learning across devices of different sizes
SummaryThis paper studies federated learning where clients hold models of different sizes and their data streams change over time. The authors train a ResNet-18 fleet from scratch on Split-CIFAR-100 and propose DROD, which extends FjORD's Ordered Dropout with two independent reduction axes, a shortcut ratio and a branch ratio, to extract sub-models. Under a shared parameter and FLOP budget, DROD reports the highest mean accuracy among the heterogeneous-model federated strategies compared, and this advantage holds across replay-buffer sizes and under ER, DER++, and BiC.
In briefHidden Cartographer: Hidden-state-guided data selection for budgeted instruction tuning
SummaryThis paper addresses how to choose which unlabeled examples to annotate and train on when expert labeling and fine-tuning budgets are tight. The proposed Hidden Cartographer (HC) method has two stages: it uses hidden states to progressively pick candidates for annotation, then ranks annotated candidates by exact training-dynamics scores to form a compact training subset. According to the paper, HC achieves the highest average accuracy among budgeted baselines across medical, legal, and financial domains, annotating 20–30% of the pool and fine-tuning on 10–15%.
In briefDo LLMs Really Lack Difference Awareness? Heuristic Substitution in Fairness
SummaryThis paper examines whether large language models truly lack difference awareness, the ability to recognize meaningful differences between social groups in fairness decisions. The authors propose Fairness Heuristic Substitution, arguing that models may fall back on simpler formal-equality reasoning when identifying relevant differences and judging their fairness implications is demanding. They introduce TDD, a two-stage distinguish-then-decide framework, which outperforms one-stage baselines across four benchmarks and ten models, and neuron-level probing offers mechanistic evidence for the hypothesis.**
WhyProposes fairness heuristic substitution, a two-stage framework, and neuron-level probing evidence across four benchmarks and ten models.
In briefCalibration-free early stopping for chain-of-thought reasoning using E-processes
SummaryThis paper proposes an adaptive early-stopping framework for reasoning large language models that treats stopping as sequential evidence accumulation along the reasoning trajectory rather than threshold calibration on held-out data. It runs two E-processes on a shared per-chunk signal, one tracking the signal's position relative to its slow trend and one tracking its signed velocity. Under signal-level null assumptions, Ville's inequality gives an anytime-valid bound on false-alarm probability at user-specified significance levels. The authors report that across several model and dataset combinations, the method matches or exceeds calibration-based stopping on the token-accuracy Pareto front while using no calibration data.
In briefNeuroAct: An AI Research Agent That Improves EEG Brain-Signal Analysis Through Self-Directed Experiments
SummaryThis paper introduces NeuroAct, an autonomous research agent for analyzing and decoding EEG brain signals. An inner loop tests hypotheses and refines analysis choices using evidence about signal structure and model behavior, while an outer loop turns results into conditional, testable recommendations that are validated on later tasks. Across four experimental settings, the authors report gains over a baseline agent and improved balanced accuracy on several EEG tasks. They also report that reusing past experience raised cross-task performance while cutting LLM token usage by 64.02%.
In briefBrandEval: Two-Track Multi-Agent Benchmark for Risk-Sensitive LLM Crisis Communication
SummaryThis paper introduces BrandEval, a benchmark for testing how LLMs handle brand crisis communication over time, covering product safety, quality, service, and trust incidents. It has a static track for stakeholder-aware strategic analysis and a dynamic track that stress-tests responses in BrandPolis, a multi-agent sandbox with consumer feedback, competitor reactions, and trust-risk trajectories. The authors also release BrandSRD, a Chinese crisis-response dataset, and report that static strategic quality does not reliably predict volatility or delayed recovery, while dynamic tests reveal tail-risk failures.
In briefA Thousand Faces: A Persuasion Benchmark Built on Personal Experience and Psychology
SummaryThis paper introduces A Thousand Faces (ATF-Persu), a persuasion benchmark with 1,000 personas across 200 topics and nine domains. Based on personality and attitude theories and conversational implicature, the authors build five personas per factual context that link experiences to judgments, expression, and reasons for revision. Evaluations show lower persuasion success than three comparison benchmarks under their own protocols.
In briefContinuous-depth language model for scientific tasks, inspired by dynamical systems
SummaryThis paper introduces Continuum-LM, which reuses pretrained residual transformations as dynamical operators in a continuous-depth language model, with hidden states evolving through an ODE solved by higher-order numerical integration. Across eight scientific tasks, it improves mean exact-match accuracy by 7.77 and 10.96 percentage points over discrete counterparts at 0.6B and 1.7B scales. The authors argue that continuous hidden-state evolution is a practical design dimension for scientific language models beyond parameter scaling.
In briefViCoR: Selecting Medical Image-Text Training Data by Estimating Visual Necessity
SummaryThe paper proposes ViCoR, a method for selecting fine-tuning data for medical vision-language models. It estimates how visually necessary each sample is using image-dependent optimization residuals, calibrates these scores with medical supervision compatibility, and builds a diversity-aware subset. Across two backbones and five benchmarks, subsets of 20% selected by ViCoR achieved the strongest average performance, and gradient and perturbation analyses indicate stronger image-dependent training signals.
WhyProposes counterfactual-residual scoring to pick visually necessary samples, showing gains over full-data fine-tuning and baselines across several external medical benchmarks.
In briefRAIL-5G: A Real-World 5G Dataset for Testing AI Indoor Positioning
SummaryThis paper releases RAIL-5G, a public benchmark of real 5G NR channel impulse response measurements recorded in an operational railway maintenance facility. The 811 m² indoor site uses eight distributed radio units, robot-mounted user equipment trajectories, and high-accuracy iGPS reference positions, with heavy metal-induced multipath. The authors evaluate geometric TDoA, supervised fingerprinting, unsupervised channel charting, and self-supervised wireless pretraining. On the full dataset, TDoA averages 7.02 m error, the best unsupervised method 2.46 m, and supervised fingerprinting 0.52 m. Self-supervised pretraining reduces labeling effort by up to an order of magnitude but still lags supervised learning.
Jonas Pirkl, Andreas Eidloth, Maximilian Kasparek et al.
In briefChronoChart: A label-free, setup-agnostic channel charting method for indoor localization
SummaryThis paper proposes ChronoChart, a channel charting framework that learns low-dimensional representations of channel state information using only CSI and timestamps, without position labels or prior distance metrics. It jointly optimizes a triplet loss built from temporally derived samples, a temporal-proximity loss for smooth trajectories, and a joint-embedding loss for consistency across augmentations. Evaluated on real-world 5G, MIMO, and passive UWB data, the method matches or exceeds prior channel charting baselines that rely on tailored distance metrics and is applicable across heterogeneous radio systems.
Majid Sharghi Foroushani, Jonas Pirkl, Jonathan Ott et al.
In briefShallow probes for near-optimal personalization depth in federated learning
SummaryThis paper does not propose a new federated optimizer. It tests whether a small client-specific measurement budget can guide cost-aware personalization-depth choices under a utility that trades off Macro-F1 against deployment cost. Three shallow probes (P0–P2) are evaluated against a non-deployable offline oracle, and a training-free greedy probing method achieves regret 0.0071 and 98.3 percent oracle recovery on CIFAR-10/100 and patient-disjoint PTB-XL, outperforming metadata-only policies and FedPBA. However, agreement drops to 0.578 on deep-oracle rows, and PTB-XL deep-oracle cases remain high-regret without a validated escalation gate.
Beyazit Bestami Yuksel, Emrah Dikbiyik, and Ayse Yilmazer Metin
In briefAcharya 2.0: A Hindi-language AI tutor for self-study subjects
SummaryThis paper presents Acharya 2.0, a RAG-powered AI tutor web application that provides academic assistance to Hindi-speaking students. It uses a client-server architecture and includes several features that help students access and interact with educational content. The authors note that retrieval-augmented generation has mostly been applied to English and has seen limited use in Indic languages.
Shantipriya Parida, Debasish Dhal, Sonal Khosla et al.
In briefSampling and curvature in first-exit comparisons for momentum SGD
SummaryThis paper studies first-exit prediction for momentum SGD, showing that pointwise gradient laws omit sampling information that affects when training exits a region. For persistent sampling in one dimension it gives an explicit O(h) mean-exit comparison, and retaining realized minibatches and sampled curvature yields finite-window exit brackets. The authors also report exact-real certificates and replication experiments across several trained networks.
In briefSCOPE: a multi-agent system for profiling news outlets on credibility and bias
SummaryThis paper presents SCOPE, a multi-agent system that profiles news media outlets for bias, factuality, and credibility from a given outlet URL. Specialized agents handle article acquisition, content analysis, external evidence search, scoring, and report generation, following an MBFC-style methodology. On a benchmark of 200 outlets, the authors report that SCOPE outperforms LLM-only, article-only, search-only, and media-background-check baselines on most metrics.
In briefUsing knowledge graphs to rerank large language model answers
SummaryThis paper addresses the difficulty of answering natural language questions over knowledge graphs by extracting subgraphs linked to candidate answers, which are obtained from a large language model, using shortest paths between entities. It proposes various features that can be taken from these subgraphs together with reranking models to select the most probable answer. Experiments on Wikidata compare feature types and reranking method combinations, and the authors also released a public web application for visualizing subgraphs and generating natural language descriptions of them.
In briefA framework to measure how people's concept meanings vary, using vector representations of their word associations
SummaryThe paper proposes a measurement framework that makes individuals' concept representations empirically accessible through text elicited under standardized conditions. Using transformer-based sentence embeddings, standardized responses for concepts are mapped into a shared high-dimensional vector space to analyze interindividual variation. Three analytical avenues are illustrated with a proof-of-concept on a large-scale free-association dataset, and a ready-to-use GitHub implementation is provided.
In briefGenerating conversational recommendation dialogues from user search and click data using knowledge graphs
SummaryThe paper introduces SearCl, which automatically creates conversational recommendation dialogues from user search and click behavior without manual annotation. A probabilistic state machine controls the dialogue flow, and two LLM agents play the seeker and recommender roles, with the recommender using knowledge graphs built from item metadata. The authors generated 3,538 dialogues covering 1,731 unique target items. Under an LLM-as-a-judge evaluation, these dialogues scored higher in quality than those from existing benchmarks.
In briefMemory, Know Your PLACE: governing lifecycle state and query-time use of persistent agent memory
SummaryThe paper argues that persistent memory turns past agent interactions into durable state, yet existing systems offer little control over whether that state stays valid or should affect later queries. It introduces PLACE, a governance layer that adds write-side state certification and read-side authorization to a memory backend. A factorial evaluation on Mem0 with Qwen2.5-3B and 7B shows write-side governance substantially improves lifecycle state correctness, while read-side governance eliminates measured causal leakage from ineligible memory. Combining both mechanisms gives the strongest joint performance while preserving query-relevant support.
In briefAn information-geometric framework for designing reconfigurable wiring in large AI clusters
SummaryThis paper presents a unified information-geometric framework for designing inter-pod wiring in AI clusters built with Optical Circuit Switches (OCS), which make wiring programmable. Wirings are treated as points on a probability simplex, with KL divergence measuring departure from a symmetric Fat-Tree prior and a smoothed DAG critical-path length as the cost. The authors provide theoretical guarantees, convergence results, and a traffic-change detection method, and validate the approach against integer programming on pipeline, data, and MoE workloads.
In briefComparing World Models, Trees and Foundation Models for Retail Forecasting on Four Public Datasets
SummaryThis paper compares forecasting models at several grains of retail decision-making, such as item-by-store and category-by-store, using four public datasets: M5, Dominick's, Breakfast at the Frat, and Corporación Favorita. The best model depends on the grain, with an equal-weight average often having the lowest error, while category-level results were frequently not statistically separated. The findings apply to these datasets, and a retailer's own data may yield different results.
In briefLAMB: Latent multi-hypothesis belief memory for long-horizon language agents
SummaryThis paper introduces LAMB, a recurrent framework that stores interaction evidence and competing hypotheses about unobserved states as latent representations instead of natural-language text. It combines a shared Common state with weighted Delta components, and trains the belief modules and policy jointly. Across ALFWorld, BabyAI, ScienceWorld, and Crafter, LAMB matches or outperforms PABU, with reported gains on several metrics.
In briefM-JEPA: Predictive self-supervised molecular graph learning, evaluated with scaffold splits on Tox21
SummaryThe paper adapts Joint Embedding Predictive Architectures to molecular graphs by masking connected subgraphs and training an encoder to predict the masked target embeddings from an EMA teacher. It introduces a compute-matched, leakage-resistant three-phase evaluation protocol. Under matched compute, M-JEPA attains lower ESOL proxy RMSE than an InfoNCE baseline (2.12 vs 3.75 mean). On Tox21, gains in ROC-AUC and PR-AUC vary by assay and do not uniformly improve calibration.
In briefIndistinguishable Agents: Training Agents to Keep Private Facts Hidden from Observers
SummaryLLM agents act for users in search, shopping, and messaging, and each action can be seen by some observer. Existing privacy training rewards agents for not mentioning protected items, but an observer may still infer a private fact from the agent's behavior. The paper proposes training on twin worlds that differ only in that private fact, scoring observer views with a frozen distinguisher's log-likelihood ratio, and optimizing task success minus this ratio. On shopping and messaging tasks, the method removes leakage missed by disclosure judges and query-level rewards while the agent still completes its tasks.}
In briefReal-world chats: how users ask LLMs for legal help
SummaryThe study analyzes 2,122 naturally occurring WildChat-1M conversations using a multi-axis taxonomy that captures whether legal guidance is sought, how requests are phrased, and which legal domains they concern. A 200-conversation human-reviewed gold set is used to compare ModernBERT models with a TF-IDF cascade. In this sample, 71.1% of legal-content conversations seek personalized or practical guidance.
In briefCan LLMs Reproduce the Age Gap in Loneliness? Testing Against European Survey Data
SummaryThis paper tests whether large language models can reproduce the loneliness gap between younger (18-34) and older (65+) adults within France, Italy, Poland, and Sweden, using the 2022 EU Loneliness Survey as the benchmark. Some models reverse the age ordering in several countries, mainly by overestimating older-adult means, while another keeps the ordering but underestimates both group means. The authors conclude that group means and their contrast must both be checked against human survey data before LLM outputs are used as age-group proxies.
In briefLabel-efficient medical image segmentation via bilevel feature-optimized auxiliary learning
SummaryThis paper studies auxiliary learning, which improves a main segmentation task using auxiliary tasks whose labels are automatically derived from existing segmentation annotations, requiring no extra manual labeling. The authors formulate their framework as a bilevel optimization problem and use a feature communication module to transfer knowledge between tasks. Evaluated on three medical image segmentation benchmarks with limited labels, the method outperforms state-of-the-art approaches. The code is publicly available.
In briefMorpheus: a local de-identification tool for Greek clinical notes, benchmarked
SummaryThis paper presents Morpheus, a fully local pipeline that removes patient identifiers from Greek clinical documents, and evaluates it on 500 synthetic discharge summaries. It removed 98.49% of direct identifiers and 96.59% overall, while over-redacting 2.82% of protected clinical content. The authors describe it as the first benchmark evaluation of clinical-text de-identification for Greek.
Filippos-Paschalis Rorris (1, 2), Nikoleta Stanitsa (3) et al.
In briefLearning Without Overwriting: Structural Knowledge Fusion for Continual Learning
SummaryThis paper introduces the Knowledge Fusion Network (KFN), a continual learning method that reduces catastrophic forgetting. Each new task is learned in a separate lightweight specialist network, then merged into the global model by a Structural Fusion Module that allocates dedicated channel capacity based on a novelty score. The authors report that on Split CIFAR-100 and EuroSAT, KFN retains more accuracy and forgets less than the compared baselines while scaling parameters more efficiently.
In briefFHKAN: Fractional Hermite Kolmogorov-Arnold Networks for Learning from Data and Physics
SummaryThis paper introduces FHKAN, a Kolmogorov-Arnold Network that uses a trainable fractional-order generalized Hermite basis instead of bounded Chebyshev polynomials. The basis is orthonormal on the whole real line and bounded under a Gaussian envelope, so it handles unbounded inputs without squashing. The authors derive closed-form derivatives and analyze differentiability at the origin. Experiments cover synthetic regression, image classification and denoising, sentiment analysis, and physics-informed problems such as the Lane-Emden and Burgers equations.
Arsha Farnoush, Amirhossein Sadr, Aida Pakniyat et al.
In briefPLANE: Plug-in visual experts that reduce false details in vision-language AI models
SummaryLarge vision-language models often hallucinate, inventing objects or misstating attributes and spatial relations in images. The authors propose PLANE, which adds three LoRA visual experts on top of a CLIP encoder, each trained on captions about objects, attributes, or relations, plus an instruction-aware gate that fuses their features. They also introduce GIFT, a manually annotated hallucination benchmark that does not rely on GPT-4 for evaluation. Experiments report reduced hallucinations across models of different types and sizes with minimal added inference latency.
In briefA Multi-granularity Interaction Network for Extracting Entity Relations from Raw Sentences
SummaryThis paper studies relational triple extraction, which identifies relations between entities in raw sentences without relying on entity type information. Traditional methods perform poorly on "unseen" entities because they fail to build relevant contextual dependencies, so the authors propose a multi-granularity interaction network that models word-word, entity-relation, and head-tail interactions. Through multiple iterations, the network refines head and tail entity representations and recovers triples missed earlier. On two rearranged datasets it performs consistently well, improving Unseen-metric F1 by 3.1% on NYT-R and 20.5% on WebNLG-R.
In briefImage Dehazing Reframed as In-Context Tabular Regression
SummaryThis paper asks whether single-image dehazing can be framed as an in-context tabular regression problem. It introduces DehazeTab, which turns overlapping hazy image patches into tabular samples using handcrafted haze features, pretrained visual embeddings, or both, and uses a pretrained tabular foundation model to predict transmission without fine-tuning. The predicted transmission map is refined and used in the atmospheric scattering model to recover haze-free images. Experiments on benchmark datasets show competitive restoration, with handcrafted physically motivated features outperforming pretrained embeddings.
In briefA Spectral Adaptive Graph Transformer for Efficient, Scalable Traffic Forecasting
SummaryThis paper proposes a spectral adaptive graph transformer for traffic forecasting. Its spatial operator is learned rather than derived from the graph Laplacian, is directed, and depends on the current traffic state, with cost linear in the number of sensors. It adds a sparse road-graph diffusion branch and a zero-initialized gate. Across six public benchmarks with 170 to 1843 sensors, it achieves the lowest error of any single end-to-end model on one benchmark and stays competitive with leading attention models on the others, using about a quarter of their memory and training about twice as fast.
In briefReflection-Guided Dynamics Rectification for Text-to-Image Generation
SummaryThis paper addresses fine-grained correction of existing image generations under a limited extra budget during test-time optimization. The authors propose Reflection-Guided Dynamics Rectification (RGDR), which turns reflections on completed outputs into corrective interventions applied to the sampling trajectory at every step. Under matched extra generation budgets, RGDR's cross-model average final checklist score gain over the reference is reported as 1.51 times that of TIR prompt refinement and 1.81 times that of Dynamic CFG.
In briefHullProbe: Measuring How Much Maritime Vessel Classifiers Depend on Hull Parts
SummaryThe paper introduces HullProbe, a method that swaps registered anatomical parts (hull and superstructure) of vessels with same-class instances to directly measure how much each part drives a classifier's decision, instead of relying on worst-group accuracy. Tested on planted shortcut data and 26 vessel types in MARVEL, the classifier places about 46% of its corrected sensitivity on the generic hull. The group-label-free de-biasing methods evaluated did not detectably reduce this hull reliance.
WhyProposes a direct interventional measure of part reliance using exhaustive 64 swaps and Walsh decomposition, validated against known planted shortcuts and worst-group accuracy.
In briefTIRE: Generating Physics Training Problems from Formulas with Verified Answers
SummaryThe paper presents TIRE, a framework that builds physics problems by instantiating reusable physical formulas and their applicability conditions, and computes reference answers by execution before problem generation. Using TIRE, the authors built TIRE-19k, covering nine physics subfields. Across 3 backbone LLMs and 4 benchmarks, it outperforms 5 representative datasets with equal SFT data by 2.86 percentage points on average. GRPO further raises Qwen3-8B's average accuracy by 1.52 percentage points over its TIRE-19k SFT checkpoint.
In briefDecision-Target Reversal: Informative Label Missingness Can Help Under One Loss and Hurt Under Another
SummaryThis paper studies whether exploiting the pattern of label availability can improve classification when labels are missing informatively. With the data-generating model and statistical experiment held fixed, the authors show that informative partial classification can be favourable under one loss and unfavourable under another, a phenomenon they call decision-target reversal. For binary Gaussian discrimination, they derive a local cost-sensitive excess-risk expansion and compute explicit critical cost ratios in one dimension where the risk ordering reverses.
In briefMemBubble: A benchmark measuring how agent memory narrows LLM response diversity
SummaryThis paper examines how memory in LLM-based agents, which personalizes responses using user preferences from past conversations, can reduce response diversity. The authors introduce MemBubble, a benchmark of 1,000 instances built from real user-assistant conversations, using test queries that preserve topical relevance. Across multiple LLMs and memory mechanisms, entropy fell by 34.7% on average compared with providing the full interaction history directly. Simple user-side interventions and changes in information availability did not effectively close this diversity gap.
In briefManaging skill libraries for AI agents by tracking each skill's real contribution
SummaryThis paper introduces TrackRL, a reinforcement learning framework for LLM agents that manages a skill library by tracking how much each skill actually contributes. Within each GRPO group, it runs online randomized controlled experiments, injecting skills at random and comparing outcomes within the same task, then accumulates reliability-weighted evidence. The authors report that TrackRL outperforms recent skill-augmented RL baselines both when skills are available and when they are not.
WhyTurns GRPO group sampling into online randomized experiments to isolate each skill's causal contribution, with evidence-gated lifecycle management. Outperforms recent baselines with supporting analyse
In briefThe Bottleneck Is Retrieval, Not Reading: Diagnosing Evidence-Based RAG for Long-Context QA
SummaryThis paper diagnoses structural limits of retrieval-augmented generation (RAG) for long-context question answering. It introduces evidence-based generation (EBG), which requires line-level citations, and a joint Ans+Evid metric scoring correctness and grounding together. Tested across ten proprietary and three open-weight LLMs, the study finds that similarity retrieval finds some relevant chunks but rarely all needed chunks for multi-fact queries, and chunking removes document meta-information needed for identity-bearing queries.
WhyProposes evidence-based generation and a joint metric, diagnosing RAG's structural flaws across many LLMs with useful subfield implications.
Debopriyo Banerjee, Owais Aijaz, Ahmer Tabassum et al.
In briefCapability-Aware Multi-Teacher On-Policy Distillation: Adjusting teacher supervision by a student's remaining capability gap
SummaryThe paper proposes CA-MOPD, a method for merging several domain specialist models into one student model. It first merges reinforcement-learning specialists into a strong initialization, then uses verifier-estimated capability gaps to set how strongly each domain is distilled, reducing supervision as gaps close. Across six benchmarks covering mathematics, code generation, and instruction following, it reports an average score of 32.22, above naive MOPD, Open-MOPD, and the merged initialization.
In briefMemories That Outlive Their Reasons: Tool-Using LLM Agent Memory Tied to Dependencies
SummaryThe paper observes that conclusions stored in an AI agent's memory rest on premises, and when a premise is invalidated, the related conclusions often cannot be retrieved together, so outdated verdicts get published. The authors propose Tether, which records at write time the exact versions each verdict depends on and, at read time, withdraws any verdict whose support has failed anywhere in the store. In simulations, Tether publishes far fewer unsafe upstream-failing objects than baseline methods, though its benefit drops when dependencies are not captured correctly.
In briefBuilding semantic maps like experts by fusing data with semantic knowledge
SummaryThis paper studies semantic map models (SMMs), which build graphs showing conceptual similarity by comparing cross-lingual form-meaning mappings. Current automatic methods ignore the prior semantic knowledge experts use, enlarging the search space of candidate graphs and hurting accuracy and efficiency. The authors fuse distributional semantic knowledge from language model representations into a top-down, data-driven algorithm that re-sorts candidates by conceptual semantic similarity. Experiments on a self-collected benchmark of 10 datasets are reported to show the method builds SMMs effectively and efficiently.
In briefClinSDT: LLM-Guided Clinical Semantics for Offline Sepsis Treatment with Decision Transformers
SummaryThis paper proposes ClinSDT, a framework for learning sepsis treatment policies offline from retrospective ICU trajectories. A frozen large language model serves as a semantic bridge, encoding Profile, Context, and Advisory clinical views as independent tokens for a Decision Transformer. A treatment-response forecaster screens low-uncertainty counterfactual candidates for policy refinement, and inference uses closed-loop rollout with refreshed patient state. Experiments on MIMIC-III and MIMIC-IV report consistent relative improvements over imitation learning, offline reinforcement learning, and Decision Transformer baselines.
In briefLC-TSRNet: A Land-Cover-Guided Two-Stage Network for Reconstructing Annual MODIS Surface Reflectance Time Series
SummaryMODIS surface reflectance products often have data gaps caused by clouds and sensor faults. This paper proposes LC-TSRNet, a two-stage network that uses an annual land cover map as prior knowledge to seamlessly reconstruct annual MODIS surface reflectance data. The second stage integrates multi-scale spatiotemporal context and joint spatio-temporal-spectral constraints to handle large, persistent gaps. The authors also built the AIRMOD-TSR dataset. The method outperforms comparison methods, improving PSNR by 0.56 dB and reducing SAM by 0.13 degrees in spatially adjacent scenarios, and a vegetation monitoring application confirms the practical value of the results.
In briefMemory as Belief: Bayesian Evolution of Memory for Self-Evolving LLM Agents
SummaryThis paper studies how LLM agents should judge the value of stored memories before reusing them. It proposes BEAM, which treats memory value as a context-dependent belief that is updated using execution feedback. Experiments on reasoning and code generation benchmarks report better memory utilization and continual self-evolution.
In briefPolynomial-time learning of positive ReLU networks under Gaussian inputs
SummaryThe paper studies positive, bias-free, one-hidden-layer ReLU networks with standard Gaussian inputs. It gives algorithms whose sample and arithmetic complexity are polynomial, without separation or conditioning assumptions on hidden directions. For conditional label means that are width-k networks with bounded output weight, the learner reaches a target squared prediction error. For arbitrary bounded-second-moment labels, it constructs a positive ReLU network of width O(k log(2/ε)) with a stated risk bound.
In briefCompile once, run cheaply: frozen reusable robot skills and the role of recursive self-improvement
SummaryThe paper studies a route to robot manipulation skills in which a strong general model acts only as an offline author, writing and repairing a skill in simulation until it passes an exam on layouts drawn after freezing. The frozen skill then runs as code alone or under a cheap caller that reads only its description. On 7 RoboTwin 2.0 tasks, the author compiled 4 tasks into 7 frozen skills, at author cost of USD 4.29 to 8.22 each, while the other three tasks did not compile in 18 campaigns. The frozen skills kept 148 of 150 successes on new layouts with no model, and 147 of 150 with a cheap caller.
WhyProposes a frozen, portable skill compilation protocol with hidden layouts and controls testing inheritance effects; evidence is solid but lacks a direct-control baseline.
In briefA learner-free swarm reaches the explore-exploit equilibrium
SummaryThe paper frames the explore-exploit trade-off in multi-agent resource sharing as the ideal-free distribution, the Nash equilibrium of a congestion game. It shows that a minimal foraging swarm using Levy-flight exploration, threshold-cascade recruitment, and dispersal from crowded patches approaches this allocation with no learner, reward, or trained parameters. Compared with PPO learners trained in the same environment, the swarm matches the equilibrium gradient more faithfully at zero training, while learners collect more where resources are scarce.
In briefExplore or Confirm? LLM Agents Misallocate Lab Measurements Under Noise
SummaryThe paper introduces Wellocate, a simulated assay benchmark built on a generated landscape and two scrambled real protein datasets (GB1 and PhoQ) with known true values. Five language models tested as minimal agents misallocated their fixed measurement budgets: three over-confirmed by re-testing too often, while one trusted single noisy readings and ended with worse final picks than a basic scripted strategy. The authors release Wellocate as a benchmark and reinforcement-learning environment with frozen splits and scripted baselines for measuring and training budget allocation.
In briefGeneration-Heavy Hybrid Machine Translation: A Method for Language Pairs with Scarce Source Resources
SummaryThis dissertation proposes Generation-Heavy Hybrid Machine Translation (GHMT), an approach for language pairs where source-language resources are scarce and target-language resources are rich. It needs only a source syntactic parser and a simple one-to-many translation dictionary, with no transfer rules or complex interlingual representations. Rich target-language symbolic resources generate multiple structural variants from a syntactic dependency representation of the source sentence, and statistical target-language resources then select among them. The author argues that this source-target asymmetry makes the systems easier to retarget to new source languages.
In briefWhy the heavy-ball method does not speed up on one-variable functions
SummaryThis paper studies whether Polyak's heavy-ball method accelerates beyond its optimal rate on convex quadratics, a long-standing open question. Using a Fourier-based analysis, the authors prove that the method does not accelerate over univariate smooth strongly convex functions. They also enlarge the known root-of-unity cycling region and give a higher-dimensional cycling example not covered by root-of-unity cycling.
Konstantinos Fotopoulos, Michael Helcig, Karl Deck et al.
In briefDeep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models
SummaryThis paper introduces Deep Fusion Memory (DFM), a parametric memory architecture that helps large language models use external knowledge without online retrieval. A causal predictor generates memory from preceding context, and this memory is fused into intermediate hidden states through gated cross-attention while the backbone stays frozen. Compared with existing parametric memory methods, DFM requires far fewer retrieval queries and retrieved items to build its training data. The authors report performance gains on six QA tasks and on domain adaptation.
In briefPre-query retention with bounded bit and word reads from an archive
SummaryThe paper studies an agent that stores fixed-length information before a future question arrives and later retrieves evidence from an immutable binary archive under a worst-case limit on raw bit or word reads. It characterizes feasible one-bit-read memory cells, builds encoders that partition the full record space, and combines explicit encoders with decision-tree counting and an entropy converse for multi-bit reads. In a matched dialogue-agent study, a categorical sufficient statistic preserves source decisions with less storage and lower decoding latency than a universal encoder.
In briefAgentic retrieval-augmented generation for retail workflows that span heterogeneous data sources
SummaryThe paper argues that enterprise retail agents must chain evidence across structured product and order records and unstructured inquiries and reviews, a challenge existing agentic RAG benchmarks rarely capture. The authors introduce RetailBridge, a compositional benchmark built from persona-driven simulations, and RetailRAG, a framework with auto-schema briefing, source-routed plan chaining, and diagnostic plan revision. Evaluating ten LLM backbones, they find that multi-source chaining is the main bottleneck, suggesting reliable retail agents need explicit cross-source orchestration rather than scaling alone.
In briefAdaptive Sequence-Level Clipping for Training Reasoning Models with RL
SummaryThis paper proposes CSPO, which uses conformal prediction to adaptively calibrate sequence-level clipping bounds from recent training statistics, replacing the fixed bands used in GSPO. The authors bound the probability that a new score falls outside the band, accounting for calibration and drift terms. Experiments on Qwen3-1.7B models show gains over GSPO on difficult math reasoning benchmarks and ALFWorld.
In briefVisual reasoning over space and time for grounded video segmentation
SummaryThis paper addresses Video Reasoning Segmentation, where a model must identify an implicitly described target and predict its pixel-level mask trajectory across frames. The authors propose Temporal Chain of Video Primitives (TCoVP), a chain-of-thought format that links each reasoning step to entity references, temporal support, and timestamped spatial evidence. They build the TCoVP-SFT-15k and TCoVP-RL-13k datasets and train models with supervised fine-tuning and chain-aligned reinforcement learning. Experiments on five benchmarks and three backbones are reported as showing strong performance.
In briefDAGLoop: Aligning Method Design with Experimental Feedback in Automated AI Research
SummaryThis paper proposes DAGLoop, a framework for automated AI research agents that tests each design decision of a candidate method through targeted ablation-style experiments. Decisions and their dependencies are organized as a directed acyclic graph, and outcomes are stored as structured research memory. The memory guides the agent in keeping supported decisions, revising unsupported ones, and planning further experiments. The authors report substantial performance improvements across dozens of research tasks.
Dengyuxing Wang, Yonggang Zhang, Qianshu Cai et al.
In briefBeyond Factual Accuracy: A Benchmark for Judging Whether LLMs Stay in the Requested Viewpoint
SummaryLarge language models often give answers that are factually correct but mix in viewpoints the user did not ask for, which factual accuracy metrics cannot detect. The authors introduce PerspectiveBench, a benchmark with scenarios where several valid reasoning paths coexist, and a Perspective Alignment Score (PAS) to measure adherence to a specified perspective. Experiments show that current models include the target content but fail to exclude non-requested perspectives. Adding retrieved documents via RAG improves factual grounding but does not fix this perspective drift.
In briefDual-Slot: Turning Serial Loops into Parallel Slots in Language Models
SummaryThis paper proposes Dual-Slot, an architecture that reorganizes the sequential loop computation of looped language models into parallel latent and prediction slots to reduce autoregressive decoding latency. On a 1B dense model, Pile validation perplexity drops from 7.25 to 7.00 and average five-shot accuracy across nine task configurations rises by 2.26 points. After supervised fine-tuning, zero-shot GSM8K accuracy improves by 5.83 and 3.49 points for the dense and A0.6B MoE models, while retaining 93.1-99.6% of vanilla decoding throughput.
In briefFinding the key frames a video model relies on
SummaryThis paper addresses explaining video-understanding models, noting that existing attribution methods spread evidence across near-duplicate frames. The authors introduce Sufficient Frame Subsets: compact frame sets where masking all other frames still preserves the model's prediction. They evaluate on UCF101 and Something-Something v2 with CNN and transformer models, finding compact, effective subsets and that models relying on more such subsets tend to perform better.
In briefCoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?
SummaryThis paper proposes CoEvo-RSI, a self-improving agent framework in which execution harnesses and model capabilities co-evolve through bidirectional inheritance from a shared evolutionary archive. The approach aims to reduce the mismatch between a harness's demands and the model's learned abilities. Using only supervised fine-tuning on the model's own rejection-sampled trajectories, Qwen3.5-9B is reported to be competitive across several tasks. It is said to surpass DeepSeek-V4-Pro on TraceElephant and GLM-5.3 on HomeBench VM.
In briefP2N: Reusing Deep Representations to Increase Effective Depth
SummaryThis paper proposes P2N, which adds the previous token's deep hidden state to the current token's shallow representation so later layers can build on earlier deep computation. The method adds only a single vector addition per decoding step, aiming to raise effective computational depth with negligible overhead. In experiments with Qwen-style models, the authors report that a 1.5B-parameter model improves GSM8K accuracy by over 10 percentage points with nearly unchanged decoding latency. Gains are also reported for language modeling and average general downstream benchmarks.
In briefMoCKEFA: Model-Conditioned Keyframe Extraction and Fair Attribution for Explaining Video Models
SummaryThis paper presents MoCKEFA, a method that represents a video with just a few keyframes while preserving the model's prediction about the video. The keyframes are extracted conditioned on the model's prediction, and the video's importance to the model is fairly distributed among the keyframes using game-theoretic Shapley values. Qualitative and quantitative experiments show that keyframe-level attribution outperforms traditional frame-level attribution, and the authors preliminarily explore highlighting important changes by comparing adjacent keyframes.
In briefMoWE: Spatially Adaptive Wavelet-Expert Pooling for Ultrasound Lesion Segmentation
SummaryThis paper proposes MoWE (Mixture-of-Wavelet-Experts), a spatially adaptive routing method that replaces max-pooling and strided convolutions in U-Net for ultrasound lesion segmentation. A context-aware router assigns local image regions to fixed wavelet families based on local texture and global features, and the decoder reuses the encoder's routing maps during inverse transforms. Evaluated on ultrasound and cardiac benchmarks including BUSI, the authors report improved Dice and boundary-distance scores and fewer trainable parameters for U-Net.
In briefDimension-free inequalities linking f-divergences for Gaussian mixtures
SummaryThis paper proves comparison inequalities between KL divergence, chi-squared divergence, and squared Hellinger distance for Gaussian location mixtures with compactly supported mixing distributions. Both divergences are bounded by the squared Hellinger distance, with constants depending only on the support radius, not on the ambient dimension or the number of components. The proof uses a Gaussian interpolation argument, and the result answers an open question raised in jia2023entropic.
WhyResolves an open question with dimension-free KL and chi-square bounds via Gaussian interpolation, giving a complete proof with broad theoretical relevance.
In briefAuto-MD-bench: Testing AI Agents on Real Medical Modeling Tasks
SummaryThis paper introduces Auto-MD-bench, a benchmark of 15 real-world medical tasks from 10 clinical datasets, covering survival analysis, classification and sequential prediction, with external validation across cohorts, hospitals and time periods. It evaluates four autonomous agents: AIDE, AI Scientist v2, Codex and Autoresearch. The strongest agents achieve competitive primary scores, but their predicted risks are less well calibrated than simple baselines, and models with similar primary scores can differ markedly in calibration. The authors conclude that primary task performance alone is insufficient for assessing autonomous clinical modeling.
In briefEdge FPGA Scene Memory for Robot Localization and Neural Splatting
SummaryThis paper presents LENS, an edge scene-memory acceleration system on the AMD KR260 low-power FPGA/ARM board that lets robots localize cameras and render nearby views without a desktop GPU. Rather than introducing a new scene model, it co-designs an INT16 scene-coordinate regression accelerator and a bit-exact fixed-point ordered 3D Gaussian splatting renderer with the board data path. According to the reported board measurements, LENS achieves better performance per watt than a Jetson Orin NX on every deployment task and is faster in absolute latency for localization and rendering. On-device training reaches ACEsplat-level quality but does not match the GPU in time or energy.
In briefOne Skill Takes Over: Conflicting AI Agent Skills and a Single-File Fix
SummaryThis paper studies what happens when two agent skill files for the same job are both loaded into a coding model's context: does it combine their compatible rules or follow only one? The authors built SkillClash, 16 same-job skill pairs, and tested seven models. Rule-following fell from 93–99% when a skill was loaded alone to 61–62% when a same-job skill was present, usually with one skill taking over. Adding one worked example that follows both skills to each file reduced takeover to 1–15% on six tested models.
WhyIdentifies takeover when same-job skills coexist and shows a one-example fix; checked across seven models with verifiable reference outputs.
In briefLightCRS: A Lightweight Multi-Domain Conversational Recommender
SummaryThis paper proposes LightCRS, a lightweight framework for unified multi-domain, multimodal conversational recommendation. Product representations are precomputed offline with frozen encoders, and a compact 5.3M-parameter module fuses them per dialogue using query-conditioned cross-attention shared across domains. The authors also introduce MM-R2A-25, a 25-domain benchmark pairing Reddit recommendation conversations with Amazon listings. The reported results include state-of-the-art average metrics, a 35.4% HR@1 gain over the strongest baseline, and 266 times lower per-query latency.
