In briefDifferentiable structure learning for cyclic linear Gaussian models with hidden confounders
SummaryThis paper studies learning causal structure from observational data in linear Gaussian models that may contain directed cycles and a bounded, unknown number of latent confounders. The authors cast structure learning as minimizing a Gaussian negative log-likelihood with a logarithmic complexity penalty, using Bernoulli gates that are optimized jointly with the coefficients. They prove consistency of global minimizers under stated assumptions and report lower recovery error than previous methods in several experimental settings.
Sadegh Khorasani, Ali Najar, Saber Salehkaleybar et al.
In briefFixing Where, Keeping How: Training-Free Referential Guidance for Compositional Generalization in Vision-Language-Action Models
SummaryVision-Language-Action models often rely on task-irrelevant visual shortcuts, which limits their ability to recombine seen objects, destinations, and backgrounds in new configurations. The authors propose ReGuide, a training-free wrapper that uses object poses from a grounding module and combines semantic and geometric rebinding to steer the end-effector toward demonstration-supported target configurations. In simulation across multiple backbones and on a real robot, ReGuide raises success rates under compositional shifts by up to 56.8 and 75.0 percentage points, while preserving standard-task performance.
In briefExo2EgoHOI: Hand-object interaction-aware exocentric-to-egocentric video generation
SummaryThis paper presents Exo2EgoHOI, a video generation framework that converts third-person manipulation videos into first-person views while preserving hand-object interactions. It combines a unified 4D interaction prior with a dual-branch residual adapter, and uses decomposed gated cross-attention to keep objects consistent across viewpoints. On ARCTIC-HOI, the method improves object mIoU by 32.3% and reduces MPJPE and PA-MPJPE by 34.7% and 50.0% over the best baseline.
In briefStreamDecisionBench: A test of whether AI decisions stay correct while live information keeps changing
SummaryThis paper releases StreamDecisionBench, a dataset of eight streaming scenarios across four application families, with reference decisions derived from public rules. It introduces the in-force accuracy metric, which measures how often the applied decision is correct across update intervals of 0.5 to 8 seconds, and attributes errors to judgment, latency, or both. The best evaluated system keeps a correct decision in force only about two-thirds of the time, a substantial gap for real-time use.
In briefChanges to a Neural Postprocessor for Hourly Precipitation Forecasts
SummaryThis note accompanies an earlier preprint on an Attention Residual U-Net that turns deterministic GRAF forecasts into probabilistic hourly precipitation forecasts. It describes methodological changes: one season-conditioned model using FiLM replaces 192 separate per-month, per-lead models, and lead time extends from 48 to 72 hours. Local solar hour and a monthly precipitation climatology are added as inputs. The updated training shows a modest, consistent improvement over the original.
In briefLearning from Biased Samples: A Complete Condition for Regression Under Selection
SummaryThis paper studies when regression is possible when outcomes are observed only after passing through selection filters that depend on covariates and outcomes. It gives a characterization of when regression remains feasible under sample selection bias, and shows the regression function can sometimes be identified even when the selection filter cannot. Under stronger conditions, the authors provide finite-sample estimation guarantees and oracle-efficient algorithms.
In briefA decade on: agentic training data for real-world shadow removal
SummaryThe paper argues that shadow removal looks nearly solved on standard benchmarks but remains fragile in real-world settings, partly because the paired training data have barely changed in nearly a decade. The authors propose an offline agentic workflow combining physics-motivated generation, failure detection, feedback-driven retry, candidate selection, and deterministic correction to turn existing shadow detection datasets into paired supervision. Using it, they build AgenticShadow, a dataset of 17,138 image-mask-target triplets covering general scenes, faces, and remote sensing.
In briefTesting whether AI agents can still understand users when they communicate poorly or change their minds
SummaryExisting benchmarks assume users always state a fixed goal clearly, which rarely holds in practice. This paper defines Interactive Intent Alignment, where agents must recover and track a user's current intent despite miscommunication and shifting goals, and introduces the Drift-Bench++ benchmark and GRIP evaluation protocol. Across tested models and conditions, more interaction helps but performance stays far below oracle levels.
In briefAperture: A Training-Free Multiscale Concept Model for Satellite Imagery Interpretation
SummaryThis paper presents APERTURE, a training-free concept-bottleneck method for remote sensing images that aims to make earth observation models more interpretable. It uses greedy quadtree routing to locate small concepts and pre-trained multimodal LLMs to produce concept scores. On the new SiFC dataset covering three countries, it outperforms the best training-free baselines by over 10 macro F1 points and also beats supervised concept bottleneck models.
In briefSense and Sensitivity: Testing LLM Clinical Triage Advice Against Physician Experts
SummaryThis paper evaluates how reliably large language models give clinical triage recommendations when the wording of patient text varies while the clinical situation stays the same. The authors built a benchmark of over 6,000 clinical scenarios, 7,000 physician annotations, and 225,000 model responses. They found that LLMs recommend unnecessary care more often than physicians, and that their recommendations are more sensitive to gender and tone changes than human recommendations.
Abinitha Gourabathina, Haoran Zhang, Yuexing Hao et al.
In briefJARQ: A Joint Alternating Refinement Method for Model Weight Quantization
SummaryThis paper proposes JARQ, a plug-in method that refines the output of any group-wise quantizer for large language models. It alternates a joint least-squares refit of all group scales with bounded Babai proposals that move many integer codes together on the current grid. Tested on Llama-2, Llama-3, and Qwen models, JARQ lowered perplexity in 90 of 96 comparisons and cut three-bit RTN perplexity by up to 36%.
In briefReinforcement Learning with Complex-Valued Memories: Using Unitary RNNs for Partially Observable Tasks
SummaryThis paper revisits unitary recurrent networks (uRNNs), which represent recurrence and hidden state in a complex vector space to preserve norm and propagate information over long sequences. The authors propose three uRNN variants that can directly replace recurrent PPO architectures, reporting significant gains over baselines on several memory-dependent tasks, including continuous control. The work also explores preserving the phase of the complex hidden state for a phase-aware policy, drawing an analogy to quantum state measurement.
In briefPixelUMM: An Encoder-Free Model for Unified Image and Video Understanding and Generation
SummaryThis paper presents PixelUMM, a model that handles image and video understanding and generation directly in pixel space without a separate encoder. Images are represented as spatial patches and videos as spatiotemporal tubelets, linked to a shared backbone through single-layer linear projections. The model uses a Mixture-of-Transformers architecture that supports text prediction and pixel-space flow matching, and experiments report competitive results across these tasks.
In briefPrompt2Skill: Optimizing LLM Skills Without Supervised Data
SummaryThis paper presents Prompt2Skill, a framework that builds skills for large language models from a natural-language task description alone. It derives a task specification, discovers or synthesizes datasets, and refines the skill through closed-loop reflective editing. Across four domains (question answering, reading comprehension, spreadsheet manipulation, and mathematical reasoning), the authors report consistent gains over direct prompting, with an average improvement of 10.8 across open-source and frontier models.
In briefStereoGaussians: Feed-Forward 3D Gaussian Splatting from Stereo Images
SummaryThis paper presents StereoGaussians, a method that predicts a metric 3D Gaussian Splatting (3DGS) scene representation directly from a single calibrated stereo pair, without per-scene optimisation. It reuses intermediate features from frozen pretrained stereo networks to predict Gaussian attributes, with calibrated disparity anchoring the geometry. The authors report improvements over strong view-synthesis baselines on unseen real and photorealistic stereo benchmarks.
In briefImage restoration that adapts to new tasks without forgetting old ones
SummaryThis paper addresses catastrophic forgetting when image restoration models learn a stream of new tasks without access to past data. The proposed RwF framework uses parameter-space integrated gradients to locate degradation-critical filters, then generates task-specific filters via low-rank factorized transformations and places them back only at localized positions. Experiments on six restoration tasks report that RwF avoids forgetting and uses about 10 times fewer additional parameters than LoRA-style adaptation.
In briefNeurDuo-EEG: A Long-Sequence EEG Foundation Model with Persistent State and Explicit Memory
SummaryThis paper presents NeurDuo-EEG, a causal EEG foundation model with channel-resolved persistent memory that models continuous EEG over long durations with a fixed-size state. It was pre-trained on 3,955 hours of EEG from 17 public datasets. The model achieves the best results on four of five downstream benchmarks, including seizure detection, where AUC-PR rises from 0.285 to 0.471 over the strongest non-NeurDuo baseline.
WhyProposes channel-resolved persistent memory for EEG foundation models, achieving top results on most benchmarks with large-scale pretraining and comparative experiments.
In briefFlexRouter: Learning complementary model sets for flexible LLM routing
SummaryExisting LLM routing methods score models independently and pick the top-k, which can select redundant models that fail in the same way. FlexRouter explicitly models model complementarity, optimizing answer coverage, the probability that at least one selected model answers correctly, using Determinantal Point Processes. At inference it greedily uses log-determinant gains to pick subset sizes adaptively without a fixed budget. Experiments on the RouterEval benchmark report higher coverage and lower redundancy than strong baselines on in-domain and out-of-domain tasks.
In briefRetargeting Motion to Diverse Skeletons via Learnable Flattening
SummaryThis paper proposes a Transformer autoencoder that transfers motion between skeletons with different topologies. Its core method is a learnable flattening of skeletal graphs, with graph-based positional encodings integrated multiplicatively. The model handles diverse skeletons in one architecture and trains fully unsupervised without paired data. In zero-shot tests it reduces global joint position error by 43–47%, and a user study with 37 participants, including expert animators, ranked it highest for motion alignment and physical plausibility.
Kia-J\"ung Yang, Fabian H. Sinz, Pawe{\l} A. Pierzchlewicz
In briefGenerating creative chess puzzles under constraints using diffusion models
SummaryThe paper proposes masked diffusion models that generate chess puzzles conditioned on specific tactical themes and partial board positions. It adds an auxiliary task of simultaneous best-move prediction, which raises solution uniqueness and theme-conditioning accuracy. A reinforcement learning framework adapted from DDPO further increases the yield of unique, theme-matching positions, and the authors release the first open-weights models for this task.
In briefToward Model as a Library: offline, community-built speech AI for African languages
SummaryThe paper introduces Model as a Library (MaaL), a software architecture that packages small, community-enrolled speech models as versioned on-device dependencies. Vocabulary is enrolled directly by speakers from a few example recordings at deployment, rather than from web-scraped text. Its central mechanism uses keyword spotting to turn closed-vocabulary text forms into voice forms that are filled and submitted on-device. The work is a position and system-design paper, presenting the concept, mechanism, and an analytical feasibility case.
In briefWhy Drawing Fresh Sketches Each Step Speeds Up Ridge Regression (arXiv paper)
SummaryThis paper studies whether iterative ridge regression benefits from drawing a new random sketch at every step instead of reusing one. The authors prove that fresh sketching yields provable advantages and derive residual-aware sampling rules from this view. Experiments on synthetic and real data, including ridge probes on Qwen2.5 representations, show substantially faster convergence.
In briefLongTake: Training AI to keep videos moving over long durations
SummaryThis paper introduces LongTake, a two-stage training pipeline for autoregressive video diffusion models that aims to keep long generated videos dynamic and visually high quality. It uses Long-Horizon Teacher Forcing on curated real long videos, conditioning the model on long ground-truth prefixes to predict later frames. The authors report that, on 30-second rollouts, their initialization yields higher dynamic degree than short-horizon initialization at comparable aesthetic quality, and that Hybrid DMD achieves the highest dynamic degree at 30s and 60s.
In briefParameter-free zeroth-order optimization using ellipsoidal sampling
SummaryThis paper proposes POEM-ES, a parameter-free stochastic zeroth-order algorithm that extends the POEM method by combining subspace preconditioning with ellipsoidal random sampling. It uses a preconditioner encoding low-dimensional problem geometry, with an empirical effective dimension guiding sampling and smoothing schedules. The authors prove a dimension-reduced convergence rate under low-rank assumptions and report experiments on hinge-loss binary classification using LibSVM datasets.
In briefDual-Scale AI Framework for Detecting Mycosis Fungoides Skin Lymphoma
SummaryThis paper proposes an automated framework for detecting mycosis fungoides, a rare skin T-cell lymphoma often misdiagnosed early on. It combines late fusion of convolutional neural networks analyzing histopathology images at 10x and 20x magnification with a random forest classifier trained on 16 clinical features. On a dataset of 6,267 images from 463 patients, the image-based model reached 83.58% accuracy and 89.13% sensitivity, while the clinical model reached 96.6% accuracy and 93.8% sensitivity.
In briefClassifying how humans and AI interact in clinical trials
SummaryThis paper proposes a multidimensional framework for classifying human-AI interactions in clinical trials, based on AI tasks, human-AI relationships, interaction configurations and interacting human groups. The authors sampled 15 trials from a previously reported dataset and had each categorised independently by two human reviewers and six large language model classifiers. The results suggest LLM-assisted categorisation is promising, while human judgement remains important when trial records are incomplete or ambiguous.
Sandra Woolley, Tim Collins, Khalid Khattak et al.
In briefDemographic Pluralism: Modeling Diverse Human Opinions at Inference Time
SummaryThe paper proposes Demographic Pluralism, an inference-time framework that estimates population-level opinion distributions by generating multiple perspectives within demographically grounded groups, without needing opinion-distribution training data or task-specific fine-tuning. Across four backbone models on GlobalOpinionQA and VITAL, it reduces Jensen-Shannon distance by 8.4%-26.4% compared with Modular Pluralism. Among weighted, equal-weighted, and inverse-weighted aggregation, equal weighting performs best overall.
In briefFairness theatre: testing after-the-fact fixes for vendor-controlled student early warning systems
SummaryThis paper evaluates six post-hoc fairness interventions on a research early warning system, using student records from a public college in Ontario, Canada, under simulated procurement constraints. It compares fairness, accuracy, and demographic disparities, and uses error-type profiling to show how interventions shift false positives and false negatives. The interventions redistributed disparities without consistently reducing them, and two implementations favored already-advantaged groups while small, marginalized groups stayed poorly served.
In briefUniversal Wasserstein Barycenters via Flow Matching
SummaryThis paper studies weighted averages of probability distributions, known as Wasserstein barycenters, across all weight choices rather than a single fixed weight. The authors propose BaryFM, a flow matching model that transports marginal distributions to any barycenter in the Wasserstein simplex and generates samples via an ODE. The method is tested on domain adaptation, generalization, Bayesian posterior aggregation, and algorithmic fairness. On 10 domain adaptation benchmarks, BaryFM achieves the best average rank among 15 competing methods.
In briefMedKIT: Testing whether LLMs can actually use newly integrated medical knowledge
SummaryThis paper introduces MedKIT, a benchmark for checking whether language models can apply medical knowledge updates in realistic sequences, not just recall facts. The authors ran a large study of 12 knowledge integration strategies across 5 models, including general and medical LLMs. Most methods improved the original update task and lexical variations, but relational generalization was limited and compositional or operational tasks showed no meaningful gains.
In briefThinkV2V: Using reasoning in multimodal AI to guide video edits from instructions
SummaryExisting instruction-guided video editing methods mostly use multimodal large language models as semantic encoders, which struggles with implicit edits needing causal or semantic reasoning. This paper proposes ThinkV2V, a framework that activates the model's explicit thinking before visual generation, using progressive curriculum training and inference-time thinking scaling. The authors also release the ThinkV2V-150K dataset and the ThinkV2V-Bench benchmark. The reported experiments show their 5B-parameter DiT model outperforms larger 10B-scale baselines on complex and standard editing tasks.
In briefWhen a Flatness Proxy Is Not a Function: Robustness Certificates and Training Interventions
SummaryThe paper examines a last-layer relative-flatness proxy and shows that a valid curvature upper bound does not by itself justify using it as a robustness certificate or as a training intervention. It finds that empirical-risk stationarity does not remove pointwise first-order loss terms and derives a globally valid, gauge-invariant feature-space repair. It also shows that common softmax row shifts make the proxy unbounded, while row centering restores invariance.
Vicente Opazo, Jose Calatayud-Mateu, Cristobal Rojas et al.
In briefReLaG: A scalable method for splitting data with hidden relations
SummaryRandom splitting can produce train and test sets that are not independent when samples are related, as in biochemical studies, which makes generalization estimates overly optimistic. This paper introduces ReLaG, a modality-agnostic framework that models sample relatedness with a hierarchical latent-variable process and uses proximity graphs and community detection to infer related groups for independent splits. On molecular and protein datasets, ReLaG matches existing relation-aware methods while scaling substantially better.
In briefDoes This Action Still Explain the Task? Reverse Scoring for Diffusion LLM Agents
SummaryThe paper finds that agents built on diffusion large language models often get stuck in retry loops, repeatedly reissuing actions that have already failed. The authors attribute this to masked decoding committing to confident positions, which lets the failed action be filled in without facing the failure feedback. They propose Reflect Reverse, a training-free method that scores candidate actions by the likelihood of the task given the state and action. On four multi-turn embodied benchmarks, it improves task success and progression rates over forward-scoring baselines.
Jiacheng Qiu, Christopher E. Mower, Jan Peters et al.
In briefInstance-specific classifier-free guidance scaling for text-to-image diffusion and flow-based image editing
SummaryThis paper addresses prompt-based image editing with text-to-image diffusion and flow models, where most methods use a fixed classifier-free guidance (CFG) scale even though different edits need different amounts of change. The authors show that choosing a CFG scale per instance can substantially improve LPIPS, which measures fidelity to the source image, without hurting CLIP-T, which measures alignment with the target prompt. To avoid the cost of repeatedly running the editing pipeline, they train a lightweight network that predicts these metrics for a given edit and CFG scale, enabling fast selection at inference. Experiments across multiple editing methods and benchmarks show the approach outperforms baselines.
In briefHow positional encoding choice changes the mechanisms models use for in-context retrieval
SummaryThis paper studies how the choice of positional encoding (such as RoPE, NoPE, and hybrids) shapes the internal mechanisms language models use for in-context retrieval. Across 22 open-weight models from eight families, standard RoPE models rely mainly on positional retrieval, while hybrid models shift toward semantic retrieval. A controlled pre-training ablation shows that confining positional encoding to local layers produces this semantic shift and degrades positional representations. The authors also find that the long-context gains of hybrids mask a retrieval trade-off, since SWA NoPE worsens when distinguishing competing keys.
In briefRevisiting scaling laws for reward optimization: how performance depends on preference data and KL budget
SummaryThis paper studies how performance scales when optimizing a policy against a proxy reward model in AI alignment, jointly considering the number of training preference comparisons and the policy's KL-divergence budget. It proposes a scaling law where performance grows roughly as Θ(√min{log M, K}), with M comparisons and budget K, supported by an information-theoretic upper bound and a constructive achievability proof. Experiments using a 70B gold reward model to generate feedback and smaller 0.6B to 4B proxy models report fits with R² between 97% and 99%.
WhyProposes a joint scaling law in comparisons and KL budget, Θ(√min{log M,K}), with an information-theoretic upper bound and a constructive tightness proof, validated with 70B gold and 0.6B–4B proxy rew
In briefGenerative sequence modeling for infinite-memory processes via predictive states, a method for predicting next-step distributions
SummaryThe paper studies estimating the one-step-ahead conditional distribution of a multivariate stochastic process, including processes with possibly infinite-range memory. It introduces an estimator based on the process's predictive states, which avoids restrictive assumptions like finite memory or sparsity. The authors show fast convergence rates when the past can be compressed into a low-dimensional sufficient statistic, with complexity governed by the intrinsic dimension of the predictive state space. Guarantees are given for a deep neural network instantiation, supported by experiments.
In briefMM-FinEval: A Multi-Task Multimodal Benchmark for Real-World Financial Forecasting
SummaryThis paper introduces MM-FinEval, a benchmark for testing multimodal large language models on financial tasks drawn from earnings conference calls. It covers 2,045 S&P 500 earnings calls from 2019 to 2022, each providing a transcript, presentation slides, and audio, with 12 financial task labels. Across 19 baseline models, small Any-to-Any models using all three modalities performed strongly, outperforming larger proprietary models limited to two modalities.
In briefShamAN-Q: A Shampoo-Augmented Quantization Method for Sub-1-Bit LLM Weights
SummaryThis paper presents ShamAN-Q, a sub-1-bit post-training quantization method that extends NanoQuant by replacing its diagonal reconstruction geometry with a dense curvature metric inspired by the Shampoo optimizer. It fits a Kronecker product to the empirical Fisher information of a small calibration set and redistributes ranks across layers at the same bit budget. On Qwen3-Base, ShamAN-Q lowers WikiText-2 perplexity at about 1 bpw (for example, 27.56 to 22.96 for 0.6B), and the 0.6B model at about 0.8 bpw matches NanoQuant's published perplexity at about 1.0 bpw.
Jonathan Mei, Sang Hyub Kim, Oliver Knitter et al.
In briefGazeFlow: Generative Prediction of First-Person Eye Gaze from Human Gaze Behavior
SummaryThis paper presents GazeFlow, a framework for predicting egocentric (first-person) gaze, which is challenging because human gaze is stochastic. It models gaze as a joint distribution of temporal gaze positions conditioned on both bottom-up visual features and top-down task information. Using conditional flow matching, it transforms Gaussian noise into plausible gaze trajectories. The method reportedly achieves state-of-the-art per-frame performance on standard datasets and generates trajectories that better match human gaze dynamics.
In briefDoes Text Steer Neural PDE Surrogates? A Controlled Diagnostic with OperatorCLIP
SummaryThis paper compares an unconditioned FNO, a constant-sentence FiLM control, and a fixed task-description model trained with contrastive alignment, using OperatorCLIP across Darcy2D, ShallowWater2D, and three-dimensional compressible Navier-Stokes over three seeds. Constant conditioning achieves lower mean test error on both 2D tasks. The authors argue that lower error alone does not show the model uses the meaning of the text, and that pathway controls are needed.
In briefFrom Solo to Social Learning: How LLMs Improve by Learning from Each Other
SummaryThe paper studies whether LLM agents, each pursuing its own reward, can learn from one another to improve the whole group. Agents revise skill files and decide whether, when, and whom to copy from, all within a shared token budget. Three LLMs earned less reward per token from peers than solo learners, though copying made learning more efficient. Skill exchanges concentrated the population around fewer independent discoveries.
In briefVAmoS Part Deux: A Harder, More Realistic Voice-Agent Benchmark
SummaryThis paper introduces VAmoS Energy, a benchmark of 100 calls about utility billing and payment help, where each caller makes two to four requests. The agent uses sixteen tools connected to a stateful billing simulation. Across fourteen voice systems, completion ranges from 17.3% to 44.7%, and background television lowers pooled completion from 38.7% to 8.6%.
In briefDEdit: Iterative Draft Editing for Faster LLM Decoding
SummaryThis paper introduces DEdit, a diffusion-based drafter that speeds up large language model generation through speculative decoding. Unlike methods that propose tokens independently, DEdit can iteratively edit its draft so later predictions help repair earlier errors and extend the accepted prefix. Across seven benchmarks on Qwen3-4B and Qwen3-8B, it reports the highest macro-average token acceptance and speedups of 5.72x and 5.97x over autoregressive generation under greedy decoding.
In briefLVMH: A Meta-Optimization Framework for Long-Video Agent Memory
SummaryThis paper presents LVMH, an automated multi-agent framework that treats memory-policy design for long-video agents as multi-objective code search around a fixed answer model, with safeguards against self-certification and overfitting. Using it, the authors discover SEEM, a "text-first, vision-second" policy that answers from compact structured event memory and evidence specialists, consulting frames through a selective visual gate. According to the paper, SEEM with GPT-5 reaches 65.7% average accuracy across three benchmarks, 7.7 points above the prior state of the art, with roughly 32x lower evaluation token use and 91x shorter response time.
In briefAutoregressive Frontier Expansion: Growing Tree Structures with Graph Machine Learning
SummaryThis paper proposes Autoregressive Frontier Expansion, a generative method that builds tree-like structures by iteratively expanding a growth frontier, mimicking how real trees grow. At each step, a flow-matching model using an SO(2)-equivariant graph neural network predicts whether each active branch bifurcates or terminates. The method is tested on cortical neurons and botanical trees in unconditional, class-conditioned, and morphology-guided generation, with outputs closely matching reference distributions and specified targets.
In briefEmpirical study comparing traditional and AI-enabled Simulink controller architectures
SummaryThis paper analyzes 62 real-world Simulink models spanning 8 controller types and 10 application domains, and surveys 13 practitioners. It compares the structure of traditional and AI-enabled controllers using a taxonomy derived from the literature. The study finds that subsystem organization dominates controller structure in both paradigms, while constraint enforcement blocks largely disappear from AI-enabled models. As a result, safety mechanisms move from explicit structure to implicit training-time artifacts, which breaks traceability.
In briefWhat to Attend, What to Keep: Skill-Conditioned Visuotactile Representation with Progress-Guided Event Memory
SummaryThis paper studies how robots should combine vision and touch for manipulation, with each primitive skill needing different sensor information. The authors propose a skill-conditioned representation that fuses short-term visual and tactile observations based on the current skill and uses a sparse event memory of terminal observations from recent skills. On three contact-rich tasks, it reduces slip-detection delay by 87% versus fine-tuned state-of-the-art progress models and twist-completion delay by 67.5% versus a vision-only ablation.
Amir-Hossein Shahidzadeh, Seungjae Lee, Eadom Dessalene et al.
In briefChecking whether a chosen speaker was fully removed from psychiatric speech recordings using audio AI models
SummaryThis paper addresses verifying that a designated speaker, either clinician or patient, has been removed from psychiatric interview recordings. Each role's audio is first redacted, then audio-language and large-language models scan the output for missed speech. Tested on 48 dyadic recordings with four open-weight models, a disjunctive OR ensemble reached a combined F1 of 0.478, with recall of 0.870. The authors report that the models are complementary.]
Joseph T Colonel, Daniel Katzman, Kelsey Kirker et al.
In briefContrastive Covariance Subspaces for Steering LLMs Without Confounding Factors
SummaryThis paper introduces Contrastive Covariance Steering (CCS), a training-free method that controls large language model behavior by projecting steering vectors onto a low-dimensional subspace. The subspace is chosen to keep variation in the target concept while suppressing directions shared with other classes, removing confounds such as topic or style. Across three LLMs and four domains, CCS outperformed all baseline steering methods in all 12 settings and also sustained performance in multi-turn conversations.
Alexandros Xenos, Manos Plitsis, Niki Foteinopoulou et al.
In briefMemory for clinical AI agents that tracks patient state changes while keeping past records
SummaryThis paper introduces STAM, a memory framework for language model agents working with clinical records. It logs state changes as new entries arrive, using semantic retrieval and typed clinical relations to find affected memories, keeping current information in an Active store and superseded or resolved information in a History store. A query-dependent gate selectively serves historical memory at read time. The authors evaluate STAM on four longitudinal clinical benchmarks using question answering, state-maintenance diagnostics, and comparisons at roughly matched context lengths.
In briefGaussian Stippling: sorting-free 3D Gaussian rendering with hybrid sampling and spatiotemporal reconstruction
SummaryThe paper proposes Gaussian Stippling, which replaces the continuous alpha of 3D Gaussian Splatting with discrete stochastic visibility samples, removing the need for depth sorting. It adaptively combines primitive-based and fragment-based stippling and uses a lightweight Gaussian-aware spatiotemporal network to reconstruct images from sparse samples and suppress noise. The authors report that the method works on unmodified 3DGS assets, generalizes to unseen scenes and runs interactively on mobile devices, with higher quality under scene-specific training.
In briefMultidimensional observer model of human image quality perception
SummaryThis paper proposes a framework for understanding the perceptual space humans use when judging image quality. It models images as distributions in a latent perceptual space and treats human judgment as comparing noisy samples, constrained by primate ventral stream neural representations and fit to large-scale behavioral data. The authors find that the perceptual spaces needed to explain human quality judgments are very low-dimensional, and that their structure differs between low-level and high-level judgments.
WhyA multidimensional observer model fit to large-scale behavioral data reveals extremely low-dimensional, task-dependent perceptual spaces for image quality judgments.
In briefAnthropomorphism in the Age of LLMs: Risks and Mitigations Overview
SummaryThis paper surveys recent research on anthropomorphism, the tendency to describe AI systems in human-like terms. It covers theoretical frameworks, how language shapes human-like perceptions of AI, and the risks involved. The authors propose a taxonomy grouping twenty-one concerns into five categories and link them to interventions in design, communication, education, and governance.
Ismael T. Freire, Marceau Nahon, Maud van Lier et al.
In briefDeep learning for tracking medical images over time: a scoping review
SummaryThis scoping review examines 102 studies published between 2018 and 2025 that apply deep learning to longitudinal medical imaging, where scans are analyzed across multiple timepoints. Neurological disorders and ophthalmic conditions were the most common applications, with MRI the dominant modality. The most frequent approach combined CNNs with temporal models such as LSTM/RNN, and classification was the most common task, while only 24% of studies performed external validation. The authors conclude the field is promising but needs newer temporal architectures and larger datasets.
Francesca Mussa, Divyanshu Tak, Atlas H. Avval et al.
In briefPosition-level gradient conflict and position-aware modulation in unified multimodal models
SummaryUnified multimodal models train image understanding and autoregressive image generation on shared parameters, and the two objectives interfere with each other. The paper introduces a position-resolved interference map that attributes understanding-generation gradient conflict to visual-token positions, finding conflict varies with sequence position: the first quarter averages a gradient cosine of -0.18 against understanding, the last quarter -0.02. Based on this, the authors propose position-aware modulation (PAM), which removes the anti-aligned generation gradient component only at high-conflict positions, improving MME and GenEval on Show-o over layer-wise separation.
In briefRetroGEF: Dynamic Graph Edit Flow for One-Step Retrosynthesis
SummaryRetrosynthesis finds practical synthetic routes to target molecules, which matters for drug discovery and materials design. The paper proposes RetroGEF, a flow-based generative model that starts from a target molecule and builds possible reactants by adding atoms and changing bonds. It handles changes in molecular structure and graph size within one generative process and learns directly from product–reactant pairs without a fixed edit order. The authors report state-of-the-art results on representative retrosynthesis benchmarks.
In briefPANDA: A decentralized, fault-tolerant architecture for coordinating many AI agents
SummaryThis paper presents PANDA, a decentralized architecture that lets many independently run AI agents discover each other's skills and form small specialized teams for each task. It separates collective communication from team communication to scale, supports star, chain, and mesh orchestration patterns, and recovers from failures by replanning around broken components. A web-of-trust model governs which agents may interact. On the HotPotQA benchmark, the authors report scaling to thousands of agents, millisecond team assembly, accuracy matching state-of-the-art at up to 8x efficiency, and 100% task completion under faults.
In briefKlinikeBench: A Benchmark Testing Language Models Beyond Diagnostic Accuracy
SummaryThis paper introduces KlinikeBench, a benchmark of 333 clinician-authored tasks that simulate a full clinical encounter with a virtual patient, clinical tools, and success criteria. Language models must take history, request examinations, follow constraints, and record a final diagnosis. The best models reached 90.7% diagnostic accuracy but succeeded on less than 30% of tasks, showing a large gap between diagnosis and interactive clinical assessment.
In briefExplainable pedestrian routing using captions to estimate perceived safety
SummaryThis paper proposes an explainable pedestrian routing method that first generates and stores a vision-language caption for each street-level image, then judges perceived risk solely from structured features of that text, so every segment score can be inspected. Across nine captioning conditions it matches a direct CLIP image-embedding baseline. Field validation in Manchester and Huddersfield against 3,669 human ratings showed a statistically significant but modest correlation (r=0.262).
In briefCurating synthetic data so vision models work on specific tasks
SummaryThis article argues that for specialized vision systems the key question is which synthetic data to generate, not how much. It discusses curated synthetic data designed around a given task and compares three paradigms: procedural rendering, physically-based simulation, and generative AI. Examples cover industrial surface defect detection, restoration of degraded digitized autochrome plates, and 6DoF pose estimation from RGB and event data. The authors conclude that curated synthetic data complement real observations, and that hybrid pipelines combining controllable supervision with learned appearance are the most promising route to reliable sim-to-real transfer.
Saptarshi Neil Sinha, Paul Julius K\"uhn, Michael Weinmann
In briefComparing six RAG search methods for answering science questions
SummaryThis paper compares six retrieval strategies for scientific question answering over 463,971 arXiv papers from 2024 and 2025. The strategies include dense retrieval, query rephrasing, reranking, multi-query fusion, an agentic tool-call pipeline, and ColBERTv2 late-interaction retrieval. All pipelines share the same generator, prompt, and evaluation protocol. The authors release their code and a synthetic dataset of 19,484 questions for reproducible evaluation.
Bhagyesh Rathi, Eshan Chawla, William B. Andreopoulos
In briefDiffusion-2BC: a hybrid diffusion and regression method for offline self-driving imitation learning
SummaryThis paper proposes Diffusion-2BC, which adds an auxiliary deterministic behavior-cloning loss on a shared visual encoder to a diffusion denoising objective. The auxiliary branch is used only during training, while inference remains diffusion-based. The method is tested in the Claw environment and in bird's-eye-view CARLA navigation, including cross-map evaluation from Town01 to Town02. The authors report lower mask-distance error in Claw and longer distances traveled in route-free CARLA than both baselines.
In briefMaking random binary codebooks for quantized scalars match a target similarity law
SummaryHyperdimensional computing often encodes quantized scalar levels as dense binary codebooks, where similarity between levels should follow a prescribed function of their separation. Randomly built codebooks deviate from this target at finite dimension due to sampling noise, initialization imbalance, and update-count fluctuations. The paper proposes a transition-based derandomization framework for two similarity families, exponential and linear decay, and derives exact finite-dimensional expressions for bias, variance, and root-mean-square error. Simulations show that each constraint removes or reduces a specific source of similarity mismatch.
Dmitri Rachkovskij, Evgeny Osipov, Olexander Volkov et al.
In briefBACKDROP: How the surrounding world costs an agent its capability
SummaryThis paper introduces BACKDROP, a benchmark that tests how much of an agent's ability survives in a world where other people also act. It adds four everyday hazards, unauthorized authority, injected text, out-of-scope requests, and ambiguous write failures, individually and together across 3,678 variants and 16 models. With all four present, the average pass rate drops from 69.5% to 31.3%, with the strongest models falling furthest.
In briefPAMI: Generating Human-Object Interaction Motion from Text Using Body-Part Anchors
SummaryThis paper presents PAMI, a framework that generates full-body human-object interaction motion from text descriptions. Inspired by the Hough Transform, body-part anchors vote for object motion, and a learned interaction latent space guides a coarse-to-fine generation and refinement process. On the InterAct dataset, PAMI reports more faithful interactions and 14.5% higher contact recall than the previous state of the art.
In briefComparing System One decision models with trained classifiers and language models for automated decision gates
SummaryThis paper benchmarks eight decision-model checkpoints from six families, including the hosted model Jev, against four open generative models and supervised and zero-shot classifiers on workflow, intent, emotion and social-science tasks. With labels, a fine-tuned DeBERTa-v3-large achieves the highest accuracy on most benchmarks, while without labels Gemma-4-31B matches Jev on some tasks at higher cost and latency. The authors also report sensitivity to swapping yes and no answers and the effect of training-data exposure, and derive condition-dependent design rules for automated decision gates.
In briefDoes gradient conflict predict the understanding-generation trade-off? A controlled audit of conflict-metric validity in unified multimodal models
SummaryThis paper tests the assumption that reducing gradient conflict between understanding and generation objectives improves the trade-off in unified multimodal models. Using GRIDUMM, a controlled testbed where the true trade-off can be computed exactly, the authors examined 63 configurations and 372 checkpoints. No directional conflict metric showed a meaningful correlation with the eventual trade-off. The authors conclude that conflict metrics must be validated before use as diagnostic targets, and they release their audit protocol as a reusable standard.
WhyControlled audit of whether gradient conflict predicts the understanding-generation trade-off, with a dose-response test and 372 checkpoints.
In briefFast surrogate model for jointly matching production and 4-D seismic reservoir history
SummaryThe paper proposes a surrogate-accelerated Bayesian history-matching framework tested on a synthetic reservoir model with about 100,000 cells. A Fourier neural operator replaces the forward simulator and is trained with a black-oil residual penalty, while production and 4-D seismic data are assimilated through ensemble-based inversion. The surrogate reportedly reaches R² above 0.999 for pressure and saturation on held-out data, and a rollout takes 0.579 s, about 2500 times faster.
Clement Etienam, David Blom, Hari Pattathil et al.
In briefTrafficSignBench: A Rule-Focused Closed-Loop Benchmark for Traffic-Sign Compliance in Self-Driving
SummaryThis paper introduces TrafficSignBench, a benchmark that tests whether autonomous driving planners follow traffic rules, rather than relying only on aggregate scores. It combines real-map simulation with procedurally generated scenarios targeting specific rules, covering 34 traffic signs with automatic rule checkers. The authors find that current planners can score well on standard metrics yet comply poorly with traffic rules. They also show how existing planners can be turned into rule-compliant trajectory experts to generate oracle trajectories for fine-tuning.
Victoria Smirnova, Viktoriia Zinkovich, Gregorii Bukhtuev et al.
In briefNAQD-Env: A benchmark testing whether language agents selectively withdraw planned actions
SummaryThis paper introduces NAQD-Env, a synthetic environment for testing whether language agents can suspend only the actions affected by changed evidence, revoked permission, or stop instructions, while preserving unaffected work and resuming after sufficient repair. Evaluating three open-weight models on 350 frozen scenarios, the authors find withdrawal recall of at most 0.06, no valid resumptions at eligible opportunities, and only one episode matching the full reference policy. The authors argue selective withdrawal should be assessed as a distinct part of agent reliability, but note the setting measures policy application on trusted structured inputs and does not establish real-world containment or source-verification ability.
In briefPrivMeSA: A Privacy-Aware Self-Evolving Medical Multi-Agent System Combining Local and Remote LLMs
SummaryThe paper presents PrivMeSA, a multi-agent system for clinical use in which a local agent decides what information to disclose to remote specialist models, trained with reinforcement learning to balance accuracy against privacy risk. Completed consultations are distilled into generalized lessons that later cases can reuse without further remote exchanges. On an emergency-department benchmark built from MIMIC-IV-ED records, mean task accuracy improves by up to 15.8 percentage points over delegation, while personal detail disclosure falls from 98.0% to 0.2% of cases.
In briefGrokking Explained Through Minimum-Norm Interpolation Theory
SummaryThis paper studies grokking, where a model fits training data long before learning the underlying signal, using a statistical theory in high-dimensional regression. It shows how regularization geometry and signal sparsity shape generalization near interpolation, including a zero-one generalization law in noiseless overparameterized settings. Experiments on diagonal linear networks and transformers trained on modular arithmetic support the theoretical predictions.
Gil Kur, Ileana Rugina, Cl\'ementine Carla Juliette Domin\'e et al.
In briefCHOIR: Free-list probing reveals distinct voices among LLM ensembles
SummaryThe paper introduces CHOIR, a framework adapting free-list elicitation from cognitive anthropology to LLM ensembles. It collects ranked lists from models, clusters items into prompt-level concepts, and measures concept salience across models, prompt variants, and personas. On the Infinity-Chat 100 prompt bank, 93 of 100 prompts show above-chance surface agreement, and narrow and broad prompts are separated. Base-model identity is the strongest recoverable signature, while persona prompts shift surfaced concepts within each base model's signature.
In briefA low-input deep learning model for predicting national crop yields from weather, tested on Brazilian soybeans
SummaryThe paper proposes a deep learning framework that forecasts crop yields using only routine weather data plus two lightweight static inputs, aiming to be transferable across crops and regions. It is tested on 20 seasons of Brazilian soybean data (2001/02 to 2020/21) with leave-one-year-out validation, comparing several neural architectures against ridge regression and a five-year moving-average baseline. The Transformer performed best, reducing error by 47.6% relative to the baseline.
Fernando Dupin da Cunha Mello (1), Prashant Kumar (2, 3 et al.
In briefWhat Pretraining and Midtraining Make Learnable from Rewards
SummaryThis paper studies how pretraining and midtraining supply the information and computation that make reward-based adaptation effective. It characterizes sequential state computation and contextual memory mechanisms that fit every training reward yet predict different held-out answers, and shows that task-independent source observations resolve this ambiguity. Experiments on pretrained Qwen2.5 checkpoints across eight worlds report 82.61% success with correct source and first-operation supervision versus 44.15% for a private-random source control.
In briefAIM: An Agentic Framework for Managing Research Ideas in Automated Science
SummaryThis paper introduces the Agentic Idea Manager (AIM), an autonomous framework for organizing and exploring research directions in automated, idea-driven research with LLMs. Inspired by Bayesian optimization, it uses an agentic surrogate and acquisition mechanism to select ideas, plus a solution auditor and resource planner to maintain idea-solution alignment and allocate experiment budget. On 10 AutoLab benchmark tasks, AIM outperforms the strongest baseline by 1.6 points on System Optimization and 4.9 points on Model Development & CUDA, reaching baseline performance up to 3.1x faster.
Hyeong Kyu Choi, Bhavana Dalvi Mishra, Jiefeng Chen et al.
In briefAudible World Models: Spatially Aware Sound Generation for 3D Worlds
SummaryThis paper presents a training-free framework that adds sound to text- or image-generated 3D worlds. It builds a panoramic 3D proxy, separates semantic layers, identifies sound-producing objects and ambient regions, and synthesizes dry audio for each sound label. These sources are anchored to reconstructed geometry, and spatial audio is rendered using geometric acoustic propagation so it changes with listener position and motion. Experiments on 80 generated scenes show better spatial consistency than text-, video-, and panorama-conditioned baselines, with competitive semantic alignment.
In briefBIND: Linking 3D Robot Actions to 2D Image Features
SummaryThis paper proposes BIND, a new action representation for visuomotor robot policies that links candidate 3D end-effector positions to the corresponding 2D image features in each camera view using camera geometry. Actions are selected by scoring each candidate's position and its image-bound features. On a real robot, BIND achieves near-perfect success on tasks with as few as 5 demonstrations and degrades gracefully under large camera-viewpoint shifts and unseen object positions, where coordinate-regression baselines fail completely.
Cameron Smith, Arsh Tangri, Vitor Guizilini et al.
In briefPre-trained autoencoder for fast surrogate models of gyrokinetic plasma turbulence
SummaryThe paper presents PreVAE-Turb, a framework that uses a pre-trained Stable Diffusion variational autoencoder to compress turbulence fields and ConvLSTM networks to model their evolution in latent space. It is validated on 2D Hasegawa-Wakatani drift-wave turbulence and extended to gyrokinetic turbulence from the GENE code, predicting four fields simultaneously. Once trained, the model generates thousands of time steps in seconds on a single GPU.
Minglei Yang, Marshall Nicholson, Diego Del-Castillo-Negrete et al.
In briefTACIT: Optimization Models That Learn from Their Mistakes
SummaryReal-world optimization models often miss constraints or objectives because experts' tacit knowledge is hard to formalize, leading to solutions that do not match operations. This paper proposes TACIT, which automatically repairs misspecified formulations using historical past solutions and later user overrides. It combines LLM reasoning with optimization: the LLM proposes structural repairs that optimization calibrates and validates, while optimization infers cuts from observed decisions that the LLM turns into interpretable constraints. On 38 misspecification scenarios across nine problem classes, TACIT repaired 78.9%, versus 60.5% for the best baseline.
In briefAI-Experiment: an augmented reality art project made with AI
SummaryThis paper describes a physical artwork created through collaboration between an artist, a computer scientist, and generative AI tools. A latent diffusion model was finetuned on the artist's photographs and generated synthetic works in a similar style. Selected outputs were repainted in acrylic on canvas by the original artist, and an augmented reality layer uses generative video models to animate the piece. The authors release part of the training data, the generation code, and the AR experience.
In briefMOBA-VL: Reinforcement-trained AI for real-time MOBA match commentary
SummaryThe paper introduces MOBA-VL, a 9B-parameter vision-language model that narrates live MOBA esports matches second by second. It uses game telemetry marking exact event times as a supervision signal and trains with event-localized multi-turn reinforcement learning that rewards turns describing each event. The authors also release MOBACast, a dataset of 860 professional matches with timestamped commentary, and MOBACast-Bench, a held-out benchmark on which MOBA-VL outperforms baseline models and improves event recall.
In briefPolicy-Conditioned AI-Use Detection: A Framework for Checking Rule Compliance in Publishing
SummaryThe paper argues that common AI text detectors estimate whether a model wrote a text, which does not match what conferences and journals need when they check whether human-AI workflows follow role-specific rules. It proposes policy-conditioned AI-use detection, which takes the governing rule as explicit input and reports hypotheses, evidence, calibration, and uncertainty instead of verdicts. Evaluation uses true positive rate at a false positive rate set in advance, and the peer-review case shows that venues must also support structured disclosure, confidentiality-respecting tool routing, and a way to contest findings.
In briefLoopVL: A recurrent approach to vision-language models that reuses shared parameters across loops
SummaryLoopVL is a vision-language model that studies whether Loop Transformers can be extended to multimodal settings. It iteratively updates a unified vision-language state through shared modules and is trained from scratch in language pre-training, multimodal training, and post-training. The authors report that it outperforms similarly sized and larger non-recurrent models on multimodal understanding and visual reasoning benchmarks, and observe shifts in visual attention across loops.
In briefGraph anomaly detection as finite-horizon control with training-free empirical Bayes scoring
SummaryThis paper proposes EB-GAD, a training-free framework for detecting anomalous nodes in graphs by modeling normality as a graph-aware generalized Ornstein-Uhlenbeck relaxation. Empirical Bayes fits the graph precision, and anomaly scores are computed as closed-form finite-horizon control energies. Across 11 benchmarks and without labels, EB-GAD achieves the best or tied-best AUROC on 9 of them.
WhyReformulates graph anomaly detection as training-free finite-horizon control energy with empirical Bayes priors; strong novelty, best results on 9 of 11 benchmarks, notable subfield impact.
In briefWhat Was Said, Not What Was 'Thought': A Logic for Checking LLM Chain-of-Thought Reasoning
SummaryThe paper introduces Type-6 logic, a dynamic epistemic logic with uncertainty and recurrence operators, to model the reasoning flow of large language model chain-of-thought outputs. It builds a verifier that converts a reasoning trace into a graph and checks it against the logic's axioms and inference rules, tested on formal and informal reasoning corpora. The authors report that the verifier catches structural flaws missed by surface heuristics, that derived contradiction is the most common hard failure, and that agreement between different verification methods is low.
In briefAutomated Colorectal Cancer Tumor Segmentation Using Adaptive Augmentation and Model Ensembles
SummaryThis paper presents an automated pipeline that segments whole-slide histopathology images to label colorectal tumor grades 1 to 3 and normal mucosa. It combines dense prediction transformers with different encoders, overlapping patches, and test-time augmentation, and uses an adaptive augmentation policy guided by large language models. Top models are combined by soft voting and refined with post-processing. On a colorectal cancer grade dataset, the F1 score improved from 62.92 to 69.84.
In briefSynthetic Data Without Training: A Inference-Only Pipeline for Tables, Time Series, and Relational Databases
SummaryThis paper proposes GENSCRIPT, a pipeline that generates synthetic data without training any generative model. It computes a statistical profile of the source data and uses a language model to infer field semantics and constraints, then a coding agent builds an auditable sampler. The authors report that the approach works across single-table, temporal, and relational data, with benchmark results on generation speed and fidelity.
In briefVidHarness: Automatically Evolved Agent Harnesses for Cheaper Long Video Understanding
SummaryVision-language models can answer questions about hour-long videos, but processing every frame is very costly. VidHarness automates the design of harness programs that wrap a frozen VLM and selectively observe the video, evolving them with Monte Carlo tree search and uncertainty-aware multi-fidelity validation. A mixture-of-harness routes each question to a harness suited to its frame budget. The paper reports new state-of-the-art results on LongVideoBench, Video-MME, and Video-Holmes, and generalization to Video-MMMU and MMVU with fewer than half the frames of uniform sampling.
In briefA large dataset of Brazilian Portuguese financial news for sorting articles into categories
SummaryThis paper presents a corpus of about 160,000 Brazilian Portuguese financial news articles, addressing the shortage of domain-specific datasets for this language. The authors benchmark three similarity-based methods (cosine similarity, KNN, and FAISS) for weakly supervised multi-class classification, using subsets labeled manually and by a large language model. On the manually labeled subset, KNN achieved 47.3% accuracy and 31.9% F1-score. On the LLM-labeled subset, accuracy reached about 50% but F1-scores fell to nearly 20%, showing sensitivity to label quality and class imbalance.]
Leomax Bandeira, Gabriel Aguiar, Dalton C\'ezane Gomes Valadares et al.
In briefArgGYM: A benchmark that tests whether AI can reason through incomplete, revisable arguments, with answers checked by a logic engine.
SummaryThe paper introduces ArgGYM, a benchmark and training environment for defeasible reasoning, where conclusions can be provisionally supported, defeated, or reinstated. It splits the reasoning into twelve tasks and uses a symbolic argumentation engine to verify model outputs. The frozen benchmark contains 1,440 verified instances, and models show different reasoning profiles, with performance declining in configurations with longer dependencies. The benchmark, generators, and verifiers are released.
\.Ibrahim Ethem Deveci, Funda Tan \c{C}al{\i}k, Bar{\i}\c{s} Deniz Sa\u{g}lam et al.
In briefDisagreement-Regularized Imitation Learning: Image-Based Continuous Control with Gaussian and Beta Policies
SummaryThis paper studies whether Disagreement-Regularized Imitation Learning (DRIL), which turns disagreement among cloned policies into a reinforcement-learning reward, improves image-based continuous control. Experiments on CarRacing compare Gaussian and Beta policies with demonstrations of one or 20 trajectories. DRIL gives its largest gains when demonstrations are few, while bounded Beta behavior cloning stays strong with 20 trajectories.
Irving Giovani Bronzatti Petrazzini, Eric Aislan Antonelo
In briefTesting whether AI language models can reliably fill gaps in stories
SummaryThis paper introduces a benchmark of about 9.2K instances for narrative infilling, created by masking one to three sentences across encyclopedic text, commonsense stories, news articles, and visual narratives. It evaluates 20 instruction-tuned open-source LLMs ranging from 1.5B to 70B parameters. The results show that model size does not reliably predict infilling quality, with the smaller Gemma-2-2B scoring highest, and explicit reasoning offers only limited benefit.
In briefC2C: Narrowing Bug Search from Codebase to Lines Using Semantic Retrieval and Hierarchical Reinforcement Learning
SummaryThis paper presents C2C, a framework that localizes software bugs by narrowing the search across files, functions, and lines of code. It first retrieves candidate buggy code via semantic similarity search over CodeBERT embeddings fine-tuned with contrastive learning, using bug report text and stack traces. A hierarchical reinforcement learning stage then localizes the bug progressively from files to functions to individual lines. According to the authors, experiments on real-world Java and Python datasets show improved retrieval precision and localization accuracy.
Ankur Garg, Corey Yang-Smith, Rishav Rishav et al.
In briefGestAdapt: Generating Speech Gestures for Humanoid Robots That Fit the Available Workspace
SummaryThis paper presents GestAdapt, a framework that generates co-speech gestures for humanoid robots while conditioning them on the available wrist workspace, such as when a robot stands near a wall. Instead of correcting unconstrained motions afterward, it learns from six co-speech corpora through a shared motion representation and can be retargeted to different robot bodies. In a user study, its constrained gestures scored 3.24/5, above the no-workspace variant at 2.43/5 but below reference motions at 3.68/5. On the Reachy2 robot, its motions ranked first in 69.7% of comparisons.
In briefSimTrace: Generating realistic synthetic user click trajectories for online user modeling
SummarySimTrace is a framework that generates synthetic, fine-grained multimodal clickstreams using a computer-use agent grounded in anonymized real user interactions and a simulated web environment. It was applied to e-commerce and outperformed baselines on 7 of 8 fidelity metrics. Models trained on the synthetic data performed comparably to those trained on real data for purchase prediction and recommendation, and augmenting real data with synthetic data improved next action prediction accuracy by 11.0% relative. The authors release SimTrace as an open-source package.
In briefUnified optimality conditions linking rough path and Itô approaches to stochastic optimal control
SummaryThe paper studies stochastic optimal control using both Itô calculus and rough path theory, each of which yields a different Pontryagin Maximum Principle. The authors show that the adjoint equations of the two principles are linked by a conditional expectation and derive a unified principle connecting them. Two applications are presented: a rederivation of the adjoint matching method for fine-tuning generative models, and an indirect shooting method for a class of feedback problems.
In briefWhich tasks remain after self-supervised learning? A paper on which downstream task information survives in SSL representations.
SummaryThis paper studies which downstream task information survives same-instance self-supervised learning (SSL) and introduces semantic recoverability, the fraction of a task's posterior score captured by the represented function space. For centered and whitened representations, the authors show it determines directional class-distance-normalized variance (CDNV) and governs few-shot nearest-centroid classification. They also give a spectral characterization of a canonical two-view SSL objective and test the predicted relationships on synthetic and real datasets.
