In briefSEPAL: A Method That Keeps Separate AI Teams Apart Until Their Final Answers Are Voted On
SummaryThis paper introduces SEPAL, a method for making collaboration among large language models more reliable when answering questions. It assigns three private Actor-Critic teams to direct reasoning, evidence grounding, and verification, with each team revising answers internally. Only the final answers are combined by majority voting. Across five open-weight backbones and five question-answering benchmarks, the authors report a 1.81 percentage point average accuracy gain over a matched single Actor-Critic pair.
In briefRiboUnmix: Separating Shared Ribosome Translation Patterns from Experimental Bias in Ribo-seq Data
SummaryRibosome profiling (Ribo-seq) measures where ribosomes sit along mRNAs, but the measurements include experiment-specific distortions and random noise. The authors propose RiboUnmix, a probabilistic framework that models each measured profile as a shared sequence-dependent signal scaled by a dataset-specific factor, with a negative-binomial model for replicate variability. On a controlled synthetic benchmark where the true components are known, both recovered components correlate strongly with their targets. RiboUnmix also outperforms sequence-to-profile baselines on four organism-specific real datasets, and shared profiles learned from subsets of 114 HEK-derived datasets remain concordant.
Gabriele Martino, Denis Skibinski, Ivo L. Hofacker et al.
In briefPredicting Cross-Lingual Transfer Without Training Using Typological Features
SummaryThis paper examines whether cross-lingual transfer can be predicted from freely available typological features, avoiding costly multilingual pre-training. A random forest trained only on typology reaches a leave-one-language-out Spearman correlation of 0.705 on a 24-language transfer matrix, outperforming a non-typological control. The signal holds under leave-one-script-out and leave-one-family-out evaluation, and the authors propose it as a zero-compute screening tool.
Dalton Raphael Harmsen, Swier Garst, Thomas van Osch et al.
In briefMARS: Zero-label tabular learning via LLM-elicited feature response surfaces
SummaryThis paper proposes Marginal Response Surface Elicitation (MARS), a method that uses large language models to supply domain priors from task descriptions and feature semantics, building a reusable zero-shot tabular classifier without labeled data. MARS selects representative feature values from unlabeled data, obtains class support scores and feature weights from the LLM, aggregates responses by median, and predicts via weighted sums without further queries. Across eight tabular benchmarks, MARS achieves the highest average AUC and AP, outperforming direct prompting by 1.97 and 6.21 percentage points while substantially reducing end-to-end costs.
In briefA study comparing Italian and Romanian "brain rot" TikTok videos across text, sound, and visuals to see what makes them popular.
SummaryThis paper compares Italian and Romanian "brain rot" meme videos to examine what makes them appealing and how the two languages differ. The authors built CRIB, a dataset of 240 manually curated TikTok videos, and analyzed their text, audio, and visual features. They found that popularity was not significantly correlated with textual or acoustic features in general, but in Romanian, faster cutting and overall pace predicted success. Italian videos were found to be more negative in text, higher in perplexity, and more rhyme-heavy.
Anca Dinu, Andra-Maria Florescu, Marius Micluta-Campeanu et al.
In briefPrediction Beats Detection: Using Drones to Control Traffic Jams
SummaryThis paper studies a team of drones that detect and predict traffic jams in a simulated road network, with their reports driving an adaptive traffic-signal controller in a closed loop. Performance plateaus when fleet size approximates the number of monitored junctions, and the authors offer a fleet-provisioning rule for persistent monitoring. Adapting signals to predicted rather than detected jams roughly doubles the reduction in jam duration, suggesting onboard prediction can be more valuable than adding robots.
In briefSAGE: Finding and Generating Salient Factors Using Visual Foundation Representations
SummaryThis paper introduces SAGE, a method that separates factors specific to a target dataset from content shared with a background dataset, using a frozen representation autoencoder's spatial latent space. The learned salient representations reveal subtypes without subtype labels and condition a diffusion transformer to generate new examples of them. On Digits-ImageNet and FFHQ eyeglasses, the authors report high reconstruction quality and better subtype discovery than baselines. On retinal OCT, the salient space separates three diseases using only normal/disease labels.
In briefA Dropped Correction in PPO Restored Exactly on Trees, Improving Sample Reuse
SummaryThe paper examines why PPO and related methods omit the state-visitation ratio and shows that under history-injective dynamics, such as autoregressive generation, this ratio equals a product of per-step policy ratios that can be computed from existing log-probabilities. It proposes a one-parameter family interpolating between PPO and the full correction, trading bias against variance. On hard credit-assignment scheduling tasks, a short corrected warmup with aggressive early reuse learns faster than PPO.
In briefImproving exploration in sequential test-time scaling
SummaryThis paper studies two ways language models spend extra compute at inference to improve reasoning. The authors find that sequential methods often get trapped in an attractor, a set of answers that blocks exploring other answers, with 53.8% of trajectories entering one within four iterations. A simple model-mixing intervention reduces the attractor hit rate by 21.2 percentage points on average and widens solution coverage.
In briefPEG-Tab: Repairing and Controlling Release of Synthetic Table Records at Sampling Time
SummaryPretrained tabular generators can reproduce training records even when overall utility is high, and retraining may be impractical. PEG-Tab is a post-training framework that works on frozen generators: for each generated row it creates two alternatives and uses a shared calibrated score to pick the lowest-risk candidate, followed by a final release check. Tested on GReaT, CTGAN, TVAE, and TabDDPM across five datasets, it cut mean Near Copy from 0.078 to 0.027 and reduced aggregate Exact Copy to zero.
In briefCertification-Based Differentially Private Learning: Using Formal Methods for Tighter Privacy Guarantees
SummaryThis paper uses formal methods, namely abstract interpretation, to obtain tighter differential privacy analysis for machine learning. It extends the abstract gradient training framework to private prediction in continuous, unbounded regression and introduces Abstract Gradient Sampling for private learning. On regression benchmarks, the proposed bounds are tighter than global-sensitivity baselines and give the first finite privacy guarantees where global prediction sensitivity is unbounded.
In briefMIND: Marginal-Invariant Neural Dependency Diffusion for Generating Mixed-Type Tabular Data
SummaryThis paper introduces MIND, a model that generates tabular data containing variables of different types. It first maps each column's variables into a unified latent dependency space using column-wise marginal transport, then uses a conditional diffusion model to learn relationships across columns. Experiments on nine tabular benchmarks show MIND improves marginal fidelity and dependency preservation over existing unified approaches.
In briefIntroduction to Computer Vision: A Code-First Textbook
SummaryThis book is a code-first introduction to computer vision, organized into 44 short chapters across three parts covering classical 2D image processing, classical 3D vision, and deep learning. Each technique is implemented in Python with NumPy or PyTorch and checked numerically against the corresponding OpenCV or PyTorch library functions. The material was distilled with AI assistance from freely available online course notes. It is intended as a self-contained reference for students and practitioners.
In briefUsing StyleGAN's intermediate latent space to parameterize reservoir properties for ensemble-based data assimilation
SummaryThis paper studies how to parameterize reservoir properties for ensemble smoother history matching, where complex non-Gaussian geology is hard to handle. The authors compare VAE-GAN, Latent Diffusion, and StyleGAN models, and apply StyleGAN2 data assimilation in two 2D cases, using both the latent z-space and the intermediate w-space. All three models performed well, with StyleGAN2 producing geologically realistic samples. Assimilation in the w-space gave better results than in the z-space.
Marcio A. Sampaio, Paulo H. Ranazzi, Martin J. Blunt
In briefD-Scope: Steering Diffusion Transformers via Sparse Autoencoder Features
SummaryThe paper introduces D-Scope, a framework linking sparse autoencoder (SAE) features in diffusion transformers to controllable image generation. It aggregates SigLIP 2 embeddings of highly activating patches into visual centroids, which are matched to text descriptions for feature retrieval. The authors evaluate 150 SAEs across two model families and five layers, plus a 100-concept benchmark. Results show high reconstruction fidelity can coexist with low dictionary utilization, and contrastive retrieval gives larger regional gains without consistently preserving outside-region content.
In briefExpandDiff: Diffusion-Based Single-Image HDR Reconstruction with Dynamic Range Expansion
SummaryThis paper presents ExpandDiff, a conditional diffusion pipeline that reconstructs lost shadow and highlight detail from a single LDR image. It introduces Dynamic Clipping Synthesis, which randomly samples shadow and highlight clipping percentiles when building training inputs from HDR targets. A pixel-space diffusion model with spatially-adaptive normalization predicts perceptually encoded HDR in one sampling trajectory. On the SI-HDR benchmark, it improves PU21-PSNR by 3.43 dB over the strongest evaluated competitor, and by 7.34 dB under two-sided clipping.
Mehmet Emre And{\i}ran, Zhuoqian Yang, Liying Lu et al.
In briefSemantic watermarking to detect malicious image manipulation and its direction
SummaryThis paper proposes a robust semantic watermarking framework that treats the watermark as a recoverable semantic reference rather than an opaque identifier. It combines a β-VAE-based binary watermark (CLIP-VAE) with channel-aware training, and a lightweight module, SDA-Net, uses the recovered embedding to indicate not only whether but in which semantic direction an image was altered. Compared with binary hashing baselines under realistic InstructPix2Pix bit-error rates, CLIP-VAE achieves the highest reconstruction cosine similarity and uniquely supports direction-of-drift detection.
In briefHybrid Imputation Methods for Robust Tabular Data
SummaryThis paper proposes two hybrid imputation methods, NuclearForest and SoftForest, which combine low-rank initialization via Singular Value Thresholding or SoftImpute with a single non-iterative Random Forest refinement step. The authors also introduce an adaptive step-size rule for the SVT component and establish its convergence. Benchmarks across diverse datasets and missingness mechanisms show accuracy matching or exceeding MissForest, with speedups of about 5.81 and 9.52 times.
In briefIs This Evidence Decision-Critical? Learning to Verify Rule-Governed Decisions
SummaryThis paper proposes InterPact, an intervention-based learning framework for checking whether evidence is critical to decisions made under explicit rules, such as eligibility checks and contract reviews. It edits case facts with a frozen language model to create counterfactual training pairs and trains a propagation verifier to link evidence changes to condition and decision changes. On single-case evidence criticality verification, InterPact reaches 68.28% accuracy, outperforming six baselines.
In briefFOMO: Selectively Unlearning Video Concepts While Keeping the Scene Intact
SummaryThis paper introduces FOMO, a training-based method for selectively unlearning concepts in generative video models, so they stop producing unwanted content. It aims to remove targeted concepts while preserving the rest of the scene, and can redirect generation toward a specified safe alternative. The method also extends to motion unlearning, where concepts are defined by temporal behavior, and the authors report the best trade-off between concept removal and scene preservation.
{\L}ukasz Rudnik, Agnieszka Polowczyk, Alicja Polowczyk et al.
In briefWhy conventional world models fail to learn cellular automata
SummaryThis paper uses cellular automata as a simple testbed to show that conventional transformer and convolutional world models can predict most pixels correctly yet rarely reproduce full rollouts. The authors attribute the failures to three issues: spatial locality, temporal locality, and temporal stability. Simple modifications addressing each, such as two-dimensional rotary positions, neighbourhood-aware tokens, and causal freezing, substantially raise rollout completion rates.
In briefGroundingPI: A grounding model that locates target objects for robots and self-driving cars
SummaryThe paper introduces GroundingPI, a 4-billion-parameter model that outputs points and boxes as quantized coordinates to identify and locate objects. It is trained with multimodal and spatial pretraining, supervised fine-tuning, and GRPO reinforcement learning. The model averages 73.68% across 34 grounding benchmarks and is reported to improve robotic manipulation and autonomous driving when used as a visual backbone.
In briefGroundAnything: Fast Parallel Decoding for Precise Visual Grounding
SummaryThe paper presents GroundAnything, a 4B-parameter visual grounding foundation model that pairs fast parallel decoding with precise localization through blockwise bidirectional diffusion. Across 30 grounding benchmarks, the autoregressive variant GroundAnything-VLM scores 72.42%, and the authors report that a self-speculative mode gives a 4.51× speedup over it with a 0.74-point drop in COCO F1mIoU.
In briefText-to-3D Policy: a paper on aligning fine-grained language with robot behavior for unseen instructions
SummaryThis paper presents T3DP, a framework that helps point-cloud-based 3D diffusion policies follow fine-grained language instructions, such as target position, displacement, or articulated state, that were not covered in demonstrations. Instead of compressing instructions and demonstrations into one global embedding, it builds token-level correspondences between language elements and behavior segments. Across Meta-World, ManiSkill, and RoboTwin, T3DP improves held-out-specification success over global alignment, and on real-robot tasks average success rises from 47.5% to 65.0%.
In briefConvergence analysis of practical Muon optimizer with few Newton-Schulz steps and Nesterov momentum
SummaryThis paper analyzes practical Muon, an optimizer that keeps momentum per weight matrix and applies a small fixed number of Newton-Schulz iterations for orthogonalized updates. It studies these layer-wise finite-step updates jointly on a coupled nonconvex objective under gradient-dependent smoothness and bounded layer-wise variance assumptions. The authors prove an O(T^{-1/4}) rate for the expected average Frobenius gradient norm, and add a nuclear-norm result that quantifies rank dependence under a stronger spectral condition.
In briefStructural Limits of Splitting Uncertainty into Aleatoric and Epistemic Parts
SummaryThis paper examines the standard information-theoretic method that splits machine learning uncertainty into aleatoric (AU) and epistemic (EU) components. It shows that, in finite settings, parts of the assumed AU and EU range cannot be reached, with the boundary given by AU ≤ log(2)/N, where N is the number of Monte Carlo samples such as ensemble members. The authors argue this limit helps explain epistemic collapse, and caution against treating AU and EU as independent when AU is low.
Jakob L{\o}nborg Christensen, Christian F. Baumgartner, Morten Rieger Hannemose et al.
In briefSPOON: Coherent multi-view 3D scene generation from uncalibrated images
SummarySPOON is a framework that generates 3D scenes from uncalibrated multi-view images by treating object placement as scene-level, geometry-based pose reasoning. It coordinates per-view object pose hypotheses using multi-view geometry derived from reconstruction, through a Guide-Route-Reconcile process. Experiments on ARSG-110K and MIDI-3D-Front show improved object placement and scene composition across different numbers of input views. On ARSG-110K, it reduces scene-level and object-level Chamfer distances by 12.7% and 17.7% over a strong baseline.
In briefKilometerVision: A Benchmark for Large-Scale Spatial Understanding in Vision-Language Models
SummaryThis paper introduces KilometerVision, the first benchmark that tests vision-language models on geographical layout understanding using real-world videos covering distances up to 1 km. It evaluates models against three stages of human spatial awareness drawn from cognitive science: anchoring with landmarks, connecting them via routes, and integrating them into global mental maps. The experiments find that current models rely almost entirely on 2D visual recognition and text matching rather than true path integration or geometric survey knowledge.
Aravindh Mahendran, Michael King, Matthew Koichi Grimes et al.
In briefA generalizable, explainable method for detecting AI-generated videos using gradient statistics
SummaryThis paper proposes a method for detecting AI-generated videos by applying the first-digit law to Sobel gradient values. Linear discriminant analysis visualises the discriminatory signal, and a multi-layer perceptron performs classification, with no generator- or scene-specific features used. The model is trained and tested on GenBuster-200K, GenBusterBench, GenVA, FaceForensics++ C23, and CelebDF. The authors also report that zero-shot detection fails, even though the feature set carries a discriminatory signal.
In briefMADGRAV: A Multilevel Anomaly-Detection Pipeline for Gravitational-Wave Searches on LIGO Data
SummaryThe paper presents MADGRAV, a deep-learning pipeline for searching high-mass compact binary mergers in LIGO data from the third and part of the fourth observing runs. The pipeline uses sequential convolutional neural networks for anomaly detection, glitch classification, coherence testing, and signal ranking. It reports 47 gravitational-wave detections with a false alarm rate below one per year, 44 of which are shared with the coherent WaveBurst search. The authors suggest anomaly detection can serve as an independent detection channel complementary to matched filtering in the high-mass, high-SNR regime.
Gianluca Inguglia, Huw Haigh, Ulyana Dupletsa et al.
In briefRobust transfer learning for recognizing paper ECGs, using a contrastive method that handles degraded images.
SummaryThis paper introduces RobECG-CL, a rank-aware contrastive learning framework for learning robust representations of paper ECGs. It builds progressively degraded paper ECG views with varied layouts from standard 12-lead recordings and balances same-recording invariance with degradation-aware ordering. On synthetic stress tests using CODE-II and EchoNext, it performs better under severe degradation and few-shot transfer, and surpasses the waveform-based ECG-FM in the 1% labeled setting. On 312 hospital samples with 37 labels, it achieves the best macro AUROC.
In briefAVERT-VLN: Abstention-Aware Recovery and Training for Robot Navigation Following Language Instructions
SummaryThis paper presents AVERT-VLN, a framework that helps vision-and-language navigation agents recover when they stray from instructions in unfamiliar environments. A separate vision-language monitor detects deviations, pauses autonomous movement, and requests human guidance. The corrected interactions are then used in preference learning to improve later navigation decisions. Under human-assisted evaluation, the full system reports success rates of 76.2% on R2R-CE and 66.3% on RxR-CE val-unseen splits.
In briefThinking Outside the Box: Can Language Models Selectively Rely on External Guidance?
SummaryThis paper examines whether language models can use helpful external workflows while rejecting unreliable ones, a skill the authors call thinking outside the box. They introduce Box²-Bench, which keeps the model and task fixed while varying workflow reliability. Training two open-weight models with counterfactual supervised fine-tuning and outcome-based reinforcement learning improves robustness, and the behavior also extends to peer correction and corrupted memory.
In briefA framework for metacognitive reasoning that sets compute effort from uncertainty
SummaryThe paper notes that current LLM-based AI struggles to estimate its uncertainty before answering and to allocate compute dynamically. The authors propose MERITED, which combines Instance-Based Learning Theory with Energy-Based Models to control reasoning effort based on uncertainty. They also train and openly share a 191M-parameter reasoning EBM.
Tailia Malloy, Prateek Kumar Rajput, Serge Lionel Nikiema et al.
In briefECHO-G: A framework for generating speech-synchronized full-body motion for humanoid robots
SummaryThe paper presents ECHO-G, which generates full-body humanoid robot motion that accompanies speech, conditioned on both speech audio and timed transcripts. Its Speech-Grounded Diffusion Transformer combines frame-aligned acoustic features with token-level linguistic context and is trained with rectified flow matching to generate motion directly in robot space. The authors also release a BEAT2-derived audio-text-robot dataset and a benchmark, and demonstrate deployment on a physical humanoid robot.
In briefBeyond Risk Prediction: Evidence and Factor Verification for Explainable Suicide Risk Assessment
SummaryThis paper proposes a framework for identifying suicide risk in social networking posts, comprising Risk Assessment, Evidence Grounding, and Factor Identification. Risk Assessment routes posts by length, while Evidence Grounding uses a constraint to keep evidence consistent with risk predictions, and Factor Identification combines two verifiers' probabilities. The framework achieved a Weighted F1 of 0.8088 for risk assessment, a test Macro F1 of 0.7605 for evidence grounding, and a Macro F1 of 0.5562 for factor identification.
In briefSteering Fields: adaptive vector fields for controlling image generation
SummaryThis paper addresses controlling text-to-image flow models, such as suppressing harmful content. It proposes Steering Fields, which re-estimate the steering direction at every step of generation instead of applying one fixed global vector, reducing unintended global changes. The method offers a continuous trade-off between steering strength and content preservation and can induce and inhibit concepts simultaneously. The authors report state-of-the-art results on safety steering benchmarks and show it can also serve as a structure-preserving image-editing technique.
Simone Facchiano, Jan Eric Lenssen, Bernt Schiele et al.
In briefHow Ambiguity and Vagueness in Words Affect the Training and Internal Workings of Language Models
SummaryThis paper uses artificial homonyms and hypernyms as pseudowords to study how lexical ambiguity and underspecification affect language model training. Both increase model performance in proportion to their effect on the language's type-token ratio, but the accuracy of generating sentences with ambiguous words or their synonyms drops. The authors also show that internal representations reflect disambiguation of pseudo-homonyms, while underspecification of pseudo-hypernyms persists during generation.
In briefSelf-Spec Verifiable Code Generation: LLMs Write Their Own Specs, Code, and Proofs
SummaryThe paper notes that existing benchmarks evaluate specification and code generation separately, often using an oracle specification, so they miss end-to-end success and cover few practical software tasks. The authors introduce VeriCodeBench, with 400 language-native problems in C, Java, Rust, and Python, where models rely only on their own specifications and code. They also propose CodeNova, which uses constraint-guided specification and verifier feedback to repair implementations. Results show self-generated specifications remain a major bottleneck, and CodeNova substantially improves all metrics, with Claude Sonnet 5 achieving the best results under this protocol.
In briefInvariant shape analysis of closed 3D surfaces that tells mirror images apart
SummaryThe paper proposes a descriptor for closed 3D shapes that removes dependence on parameterization, pose, and scale, which standard spherical harmonic invariants fail to do. It uses a conformal parameterization normalized by its conformal barycenter, followed by polynomial invariants of the rotation group, so mirror images can be distinguished. Benchmarks confirm the guarantees, and on bilateral anatomical structures the descriptor separates mirror-image pairs from asymmetric pairs.
In briefFrom Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning
SummaryThis paper presents CASE, a set of clinical agents that learn to actively seek evidence across patient records and longitudinal MRI rather than reasoning over preselected information. The authors build a benchmark on UK Biobank with 50,401 questions from 4,739 participants, where agents choose visits, organs, modalities, slices, and tools. The authors report that their Qwen3-VL-8B-Instruct based agent achieves over 16% and 10% relative accuracy gains over GPT-5.4 and Claude Opus 4.8 under matched evaluation conditions.
WhyProposes evidence-seeking longitudinal medical agents with a large benchmark and combined SFT, RL, and self-distillation; reports over 16% and 10% relative accuracy gains over frontier models.
In briefA2Z GameSpec-Bench: How Faithfully Can Coding Agents Build Games from Design Documents?
SummaryThis paper introduces A2Z GameSpec-Bench, a benchmark of 100 long-form game design documents for testing whether coding agents can build complete games that follow them. Each document is converted into a dependency-aware contract, and games are checked through source-code inspection and scenario-based replay and adaptive playtesting. The authors find that current agents struggle to satisfy interdependent requirements across code and actual play. Requirement-specific feedback raised GDD Fidelity by 10.9% relative to self-revision after two rounds.
Seonho Lee, Wonryeol Jeong, Alberto Cereser et al.
In briefRESUME: Efficient Video Language Modeling via Recurrent State Updates from Motion and Residual Signals
SummaryExisting video language models encode sampled RGB frames independently, which either exhausts the token budget on long videos or loses changes between frames. RESUME is a stateful codec representation that initializes a compact latent state from an anchor I-frame and treats each predictive frame as an update to that state, with a shared readout producing tokens for the language model. Across ten benchmarks, gains concentrate on temporal reasoning, with improvements over LLaVA-Video-7B and CoPE-7B on TempCompass, TOMATO, and MVBench, while remaining competitive on general and long-form QA.
WhyProposes stateful codec representation updating a latent state per predictive frame; gains on three temporal benchmarks over RGB and codec baselines, with frozen-transition analyses.
In briefCandidate Retention for Abductive Learning: Choosing Which Candidate Explanations to Keep
SummaryThe paper addresses abductive learning, where multiple valid explanations can assign conflicting labels to the same inputs. It proposes Abductive Candidate Retention (ACR), which greedily selects a retained subset of candidates to balance supervision sharpness and model-mass coverage. Experiments report that ACR improves concept accuracy over single-candidate baselines and A3BL in most evaluated aggregated mod-addition settings.
In briefSelf-Repulsion: diversifying sampled answers in diffusion language models
SummaryThis paper introduces Self-Repulsion (SR), a sampler for masked diffusion language models that lowers a token's logit in proportion to how many peer paths have already chosen that token at the same position, so the sampled paths become more diverse. On LLaDA-8B-Instruct with ten paths and 128 denoising steps, deterministic SR reaches 80.38% plurality accuracy on GSM8K, versus 70.17% for unpenalized greedy decoding. The authors report that voting gains come mainly from higher coverage of correct answers, and that the gains vary by benchmark and decoding regime.
In briefWIPSNet: Deep Learning to Detect Wheeze in Children from Overnight Breathing Signals
SummaryThis paper proposes WIPSNet, a 3D ResNet model that analyses scalograms of overnight impedance pneumography signals to detect wheezing in children. On 60 nights from 15 patients, it reaches an AUC of 0.783 ± 0.026, higher than the clinical Expiratory Variability Index and other baseline models. The authors conclude that structured time-frequency representations with 3D convolution are effective for long, irregular physiological time series.
In briefDivide and Collapse: exact decomposition of multi-agent path post-optimization into independent sub-instances
SummaryThis paper studies MAPFC, a post-optimization step that removes avoidable moves from a feasible Multi-Agent Path Finding plan while keeping it feasible. The authors observe that most sub-problems involve a single agent and can be split off, and they use a very lightweight solver for these cases. Coordination-heavy cases fall back to Judgelight or a CBS-like solver. The framework is exact, does not use a commercial ILP solver, and achieves a median 10.5x per-instance speedup over Judgelight.
In briefPerformance-guaranteed, task-agnostic exoskeleton control using human torque estimation
SummaryThis paper studies how to achieve task-agnostic control of robotic exoskeletons using human torque estimation. The authors define matched assistance as cases where the robot positively contributes to human movement, and design a desired interaction torque that guarantees a lower bound on the matched assistance probability. The guarantee holds across the entire torque distribution, including data beyond the training tasks. The method was implemented on the ABLE upper-limb exoskeleton and tested in a multi-task setup, where experiments validated the theory and showed smoother movement with reduced human physical effort.
In briefNEUROTOKEN: Decoding Which Speaker a Listener Attends To from EEG
SummaryThis paper addresses the cocktail-party problem, identifying which speaker a listener attends to from EEG signals. It proposes NEUROTOKEN, a single network whose conditional flow-matching head (ATTUNEFLOW) scores candidate speakers by the likelihood of the attended envelope given EEG. Reported results on KU Leuven, DTU, and NJU datasets at 5-second windows show improved per-segment source decoding over the strongest non-generative baseline.
WhyUnifies directional AAD, source AAD, and envelope reconstruction under a conditional likelihood using flow matching; multi-dataset experiments and strict-protocol reproductions give fairly strong evid
In briefEffGS: A faster Gaussian Splatting method that keeps high image quality
Summary3D Gaussian Splatting renders new viewpoints in real time, but existing general speed-up methods lose quality on complex, large scenes. This paper proposes EffGS, which combines frequency-aware guidance, localized density control, and adaptive primitive scale modulation. It aims to speed up training and rendering while matching or exceeding vanilla 3DGS quality on bounded and large-scale scenes. Experiments report a favorable balance between reconstruction quality, training time, and primitive count.
In briefGhost in the Encoder: Hidden Artist Identity in Lyrics-to-Song Models
SummaryThis paper examines the internal representations of a text-to-song generation model and finds that artist identity can be linearly decoded from its activations using song lyrics alone. The authors conducted a controlled case study of ACE-Step 1.5 using 2,000 songs across 100 artists. They report that this conditioning signal propagates from the lyric encoder to the diffusion backbone during inference, suggesting lyrics form an artist-level channel that prompt-side safeguards do not address.
Arhan Vohra, Choenden Kyirong, Laura Ib\'a\~nez-Mart\'inez et al.
In briefRankEvolve: A multi-agent framework that automates improving ranking models
SummaryRankEvolve is an automated research framework in which LLM agents propose, implement, train, and evaluate changes to generative ranking models. A runtime-enforced protocol controls the workflow, and several coding-agent products review and repair each other's work. In a budget-matched evaluation, heterogeneous composition raised execution accuracy from 45.8 percent to 62.5 percent. On the HSTU recommender, the framework reported NDCG@10 gains over the published anchor.
In briefHyperbolic Prototype Routing: A Rehearsal-Free Method for Class-Incremental Learning
SummaryThis paper addresses catastrophic forgetting in class-incremental learning and proposes HyPro, a rehearsal-free framework. It assigns a dedicated LoRA-Expert module to each incremental task and projects routing features onto a Poincaré ball for geodesic nearest-prototype matching. The authors report that HyPro consistently improves average and final-stage accuracy over strong baselines on standard and few-shot class-incremental benchmarks.
HongWei Zhao (Beihang University), Rui Liu (Beihang University), Yong Chen (Beijing University of Posts et al.
In briefTeaching GUI agents to judge unreliable app knowledge before acting
SummaryGUI agents built on large language and vision-language models still struggle with unseen apps and complex multi-step tasks. The paper proposes a framework that automatically explores interfaces, builds a UI state-transition graph, and synthesizes task trajectories with a VLM, then trains the agent on trajectories injected with five types of realistic errors. This teaches the agent to assess whether retrieved prior knowledge is reliable before acting. Experiments on physical devices and online emulators report that the method finds more unique screens, covers more benchmark tasks, and better rejects erroneous priors.
In briefSteering diffusion models away from unwanted samples with variance-reduced SMC
SummaryThis paper addresses inference-time steering of pretrained diffusion models so they satisfy new constraints while suppressing an undesired reference distribution. The authors frame specificity-aware generation as a target-design problem, derive a target from an overlap-based objective, and sample it with a variance-minimized Sequential Monte Carlo sampler. Experiments on synthetic tasks, class-contrastive generation, text-to-image, and peptide-MHC binders report stronger suppression of undesired regions, less mode shift, and reduced SMC weight collapse compared with negative-guidance baselines.
In briefEvolving Agent-Prover Tools to Cut Theorem-Proving Costs in Rocq and Lean
SummaryThe paper proposes an evolutionary method in which a frontier model proposes new features incrementally and keeps only those that improve smaller models' overall performance. Using this method, the authors developed ROCQ-MCP-EVOLVE, a new MCP server for the Rocq prover. On the held-out test split of miniF2F-Rocq, agents using it outperformed a compiler-only baseline and an established MCP server across four models in success rate, cost per solve, and time per solve. The server was also ported to Lean, where it improved cost and time per solve on a PutnamBench subset.
In briefComparing methods for adapting pretrained models to generate captions of driving behavior in videos
SummaryThis report compares several existing fine-tuning and prompting methods applied to autonomous driving. It adapts a large language model on a video dataset to test how well it understands driving situations. On the BDD-X dataset, full fine-tuning performs well on some automatic metrics and even surpasses the baseline on some. The authors also test LoRA and prompt engineering on the VideoLLaVA model and discuss their limitations.
In briefForking: Sudden overfitting when training data is replayed
SummaryThis paper studies "forking," a generalization failure found while running NanoGPT autoresearch. When training data is replayed, models with an over-encoding n-gram memory branch show training and validation loss that sharply separate at epoch boundaries, resembling a fork. The authors reproduce the effect in a controlled vanilla NanoGPT setup and in a DeepSeek-style model with Engram. They also report it in short-budget, heavily repeated SFT and RL-like training.
In briefA reusable semantic web framework for gathering evidence for EU AI Act fundamental rights impact assessments
SummaryThe paper addresses the EU AI Act's requirement that deployers of high-risk AI systems carry out fundamental rights impact assessments, noting that the needed evidence is scattered across incompatible sources. It presents a semantic web framework covering employment and access to essential public services, offering a SPARQL-queryable knowledge graph of 1,351 RDF triples built from 150 annotated records. Evaluation against a 69-record gold standard found that LLM-assisted classification in the employment domain reached only κ = 0.045, suggesting caution with automated fairness-related evidence retrieval.
In briefCATCH: A controllable testbed for studying reward hacking in coding reinforcement learning
SummaryThis paper introduces CATCH, a controllable testbed for studying reward hacking, where language models exploit environment loopholes to earn rewards without real capability gains, in coding reinforcement learning. It provides execution-based gold labels by comparing success under a vulnerable evaluator with correctness under an independent audit, and lets researchers adjust initial hacking tendency and reward difficulty. Experiments find that a chain-of-thought monitor initially suppresses hacking, but this protection erodes as the policy learns to mislead the monitor with code comments.
WhyProposes a controllable testbed with execution-based gold labels, and experiments show chain-of-thought monitors get misled over training, useful for the subfield.
In briefWEIRDO: Aligning Diffusion Models via Weak Residual-Regularized Doob's h-Transform Guidance
SummaryThis paper studies how to estimate the guidance that steers a diffusion model's output toward a reweighted target distribution at inference time. Using stochastic optimal control, the authors show the exact drift correction is the gradient of the log of Doob's h-function and propose a penalized least-squares estimator whose penalty is the residual of the space-time harmonicity equation. They derive high-probability error bounds for the guidance estimate, transfer them to total variation distance between samplers, and support the approach with numerical experiments.
In briefLens Flare Removal and Reconstruction: A Dataset and Diffusion Model for Removing Large Flares, Plus a Multi-View Consistent Flare Representation
SummaryThe paper notes that lens flares degrade downstream tasks such as 3D scene reconstruction, and that existing methods struggle with large flares covering the whole image. The authors build a large-flare removal dataset combining public real-world data with procedural generation, and fine-tune a diffusion model on it. They also propose a flare representation exploiting the flare's symmetry about the camera principal point, jointly optimized with 3D Gaussian splatting to decompose a scene into flare and scene. The explicit, re-renderable flare can be edited and transferred to new images and scenes.
Tarun Yenamandra, Jonathon Luiten, Daniel Cremers et al.
In briefAdding auxiliary supervision to close representation gaps in amortized Bayesian inference
SummaryThis paper treats Bayesian inference as training a neural network to learn an amortized posterior, noting that under finite compute and time budgets, internal representations often retain avoidable gaps. The authors propose auxiliary guidance losses on internal representations to improve training, which speeds convergence when data is abundant and improves performance when data is scarce. They also offer a generic diagnostic to separate summary failures from inference failures, and report better convergence and accuracy on several challenging real-world inference problems.
Hans Olischl\"ager, Svenja Jedhoff, \v{S}imon Kucharsk\'y et al.
In briefCIDER-FM: A foundation model for causal inference across diverse experiments
SummaryThis paper introduces CIDER-FM, a causal foundation model that combines observational data with surrogate interventional datasets to predict a target conditional interventional distribution. It uses an intervention-aware representation and hierarchical three-axis attention to share information across variables, samples, and experimental regimes. The model is evaluated against many baselines on synthetic graph and mechanism families and on simulated and real Causal Chambers data. The authors report strong prediction performance and that experimental context can improve predictions over observational data alone.
In briefReferential Uncertainty in Human-AI Collaboration: When AI Should Say It Is Unsure
SummaryThis paper studies referential uncertainty, where it is unclear which candidate object a description refers to, in a puzzle task where a human instructs an AI to place pieces. The authors find that a separately elicited belief distribution is better calibrated than raw action probabilities, though models rarely ask for clarification. In a human study, precise descriptions and well-targeted hedges reduced acceptance of wrong placements, but hedges poorly targeted could do more harm than good.
Christian Poelitz, Finale Doshi-Velez, Si\^an Lindley
In briefSpike-driven Vision-Language-Action Model: A Low-Power Neuromorphic Approach to Robot Control
SummaryThis paper proposes the first spike-driven vision-language-action (VLA) framework that can be trained end-to-end for robotic manipulation. It combines spiking visual and instruction encoders, a Multi-Winner Spike Fusion module using bidirectional top-k winner-take-all routing, and a Spike Action Chunking Transformer that generates continuous action chunks. Experiments on LIBERO and Meta-World show competitive performance with fewer parameters and lower estimated inference energy than conventional VLA models.
In briefInterpretable synthetic medical tabular data generation for clinical decision support using fuzzy cognitive maps
SummaryThis paper proposes generating synthetic medical tabular data with Fuzzy Cognitive Maps, which describe clinical features with interpretable fuzzy sets and encode dependencies between features as edge weights. Synthetic patient records are produced by propagating activation vectors until convergence and then defuzzifying them. On UCI medical benchmarks under a Train-on-Synthetic-Test-on-Real protocol, the method reaches up to 0.81 accuracy and 0.90 AUROC on the Heart Disease dataset, comparable to TVAE and Gaussian Copula baselines while running on CPU.
Michael Vasilakakis (1), Dimitris K. Iakovidis (1) ((1) Department of Computer Science, Biomedical Informatics et al.
In briefCan Domain Generalization Be Guaranteed in Small-Sample Learning?
SummaryThis paper studies theoretical guarantees for domain generalization when training data are limited. Using the concept of stability, the authors establish the first theoretical guarantees for Structural Risk Minimization (SRM) in domain generalization, deriving learning consistency and generalization error bounds. Under a specific hypothesis space assumption, they further establish stability, learning, and generalization bounds for SRM and discuss their applicability to deep learning.
In briefPartiCam: Reward-Guided Camera-Controlled Video Generation Without Retraining
SummaryThis paper presents PartiCam, a training-free method based on particle filtering that steers pretrained video diffusion models to follow a specified camera trajectory. It extends Sequential Monte-Carlo (SMC) guidance with a local refinement stage based on particle filtered resampling, forming a global-local refinement framework. According to the experiments, the method improves camera trajectory adherence, reduces drift, and improves visual quality without requiring model retraining.
In briefGeometry of randomized smoothing when predictions are filtered to feasible sets
SummaryRandomized smoothing gives probability guarantees for a classifier's output under Gaussian noise. When a predictor keeps only proposals passing a feasibility or confidence filter, the reported conditional probability can certify a radius that crosses a decision boundary. The paper separates geometric conditions under which conditioning preserves Gaussian comparisons from the certification methods, and gives finite-sample bounds. Experiments on a learned image classifier and a released confidence filter are reported.
WhyIdentifies a geometric flaw in conditional randomized smoothing, separates convex and nonconvex retained sets, and validates bounds with image classifier experiments.
In briefWhen the Right Answer Is Missing: An Arithmetic Rejection Bottleneck in Jev
SummaryThis paper examines Jev, a typed decision model that chooses from predefined options, and finds it often picks incorrect numerical answers when none of the candidates is correct, even with a rejection option. The authors report 99% accuracy when a correct answer is present but only 7% correct rejection when it is absent. A simple decision threshold tuned on separate development problems raises rejection accuracy to 79% while keeping 97% accuracy on answer-present cases, without retraining.
In briefDisentangling Self-Distillation: Measuring Knowledge Acquisition and Retention
SummaryThis paper studies self-distillation with privileged context, where a language model conditioned on a reference response teaches its context-free copy. The authors propose a unifying framework covering three entangled axes: rollout source, teacher coupling, and KL direction. They ran 1,200 adaptation experiments on Qwen2.5-7B and Ministral-3-3B across ordinary and contradictory tasks. Findings show rollout source matters mainly for contradictory tasks, teacher coupling rate most affects acquisition, and KL direction's effect on retention depends on the model.
In briefFront-to-Back: Testing AI Models on Matching the Same Car Across Front and Rear Cameras
SummaryThis paper introduces Front2Back-ReID, a benchmark of 500 manually verified vehicle handovers from 20 recording sequences in South Africa. Models must match a vehicle seen in a front-camera image to the same vehicle among at least three candidates in a rear-camera image. The authors compare seven zero-shot vision-language models, four image-retrieval baselines, and 25 human participants. The best model reached 76.6% Rank-1 accuracy on target crops, versus 74.0% for a frozen SigLIP2 baseline, a difference that was not statistically clear, while humans scored about 92–94%.
In briefOmniReasoning: Pushing the Limits of Audio-Visual Joint Reasoning
SummaryThis paper notes that existing benchmarks, training data, and methods mostly treat audio and vision separately, so joint audio-visual reasoning is poorly evaluated. The authors introduce OmniReasoningBench with 1,150 questions, a data engine OmniQA, and a self-distillation method MFSD. Their model OmniReasoning-30B-A3B scores 50.0% on OmniVideoBench and 42.5% on OmniReasoningBench, improving its base model by 12.8 and 9.3 points.
In briefRobust time series learning via sample-adaptive capacity modulation
SummaryThis paper observes that samples in deep time series learning differ in reliability, and uniform regularization can under-regularize corrupted samples while over-restricting clean ones. The authors propose Capacity-Centric Modulation and SACM, a task-agnostic framework that uses spectral sparsity to assign sample-wise dropout probabilities along internal activation paths. Across 301 dataset-backbone pairs, SACM reduces forecasting MSE by 6.7% on average and improves classification accuracy and point-adjusted F1 by 3.04% and 17.05% over unmodified backbones, with zero test-time overhead.
In briefCorrecting CondOT: exact finite-step sampling for Gaussian flow matching
SummaryThis paper studies how the choice of sampling schedule affects numerical error when flow matching uses a finite number of steps, focusing on Gaussian targets and the explicit midpoint sampler. Error is measured by the squared Wasserstein distance between the target and the sampler's final distribution. The authors show that the standard CondOT schedule cancels the leading midpoint error and improves the general convergence bound. They also prove that scalar noise schedules approaching CondOT exist that yield exact Gaussian sampling for sufficiently large step counts, supported by controlled Gaussian experiments.
In briefFrom Wrecks to Wisdom: Recovering Crash Mechanics from Real-World Multi-View Photos
SummaryThis paper studies whether crash mechanics can be recovered from post-crash vehicle photographs alone when structured crash metadata is missing. The authors frame it as supervised prediction from multi-view photo sets, estimating six Collision Deformation Classification descriptors and the longitudinal and lateral components of change in velocity. They train on 15.2k cases from the Crash Investigation Sampling System and evaluate on 1.15k validation and 1.15k test cases. Results show that photos provide usable signal for several descriptors, while weakly observable and long-tailed targets remain difficult.
Ond\v{r}ej Valach, V\'aclav Divi\v{s}, Ivan Gruber
In briefCAMOS: A coupled oscillator state-space model for multimodal clinical time series
SummaryThe paper shows that linear state-space models handle a missing modality by masking inputs while leaving the transition operator unchanged, so they cannot represent interactions between modalities that are jointly present or jointly absent. The authors propose CAMOS, which uses per-modality oscillators coupled through an availability-gated matrix so the transition itself depends on which measurements are observed. On ADNI, CAMOS outperforms uncoupled oscillatory state-space models and clinical fusion models on same-visit staging, landmark prediction, and longitudinal forecasting. Under zero-shot transfer to OASIS-3, it is the only model that avoids collapsing to the majority class.
WhyProves linear state-space layers cannot model cross-modality interactions and proposes availability-gated coupled oscillators, a novel fix; validation is mainly on ADNI.
Maxx Richard Rahman, Mostafa Hammouda, Wolfgang Maass
In briefWho Owns That? Testing How LLMs Judge Ownership Among Competing Claimants
SummaryThe paper introduces the Competing Ownership Attribution Task (COAT), with 42 scenarios, to compare how 24 LLM configurations and 108 human participants allocate ownership. Overall human-model similarity is close to human-human similarity, but models are more homogeneous and divide ownership more evenly within answers. Models also respond differently than humans to changes in material value and public recognition across scenarios.
In briefHow workflow graph structure affects graph-learning predictions of task resource intensity
SummaryThe paper studies predicting task-level resource intensity, such as CPU and memory use, for large workflows on cloud systems represented as directed acyclic graphs. The authors build a benchmark to systematically measure how graph topology affects prediction accuracy and compare a range of modeling approaches. They find that topology is a critical feature, that even simple handcrafted topological features outperform baselines, and that graph-native models achieve the lowest errors, improving further when combined with simple topological features.
In briefMatrixReward: Building Rewards for Open-Ended Generation from Pairwise Rubric Comparisons
SummaryThis paper addresses reward design for open-ended generation, where no standard answers exist. MatrixReward builds a win-rate matrix by pairwise comparing sampled responses under each rubric, then derives data-dependent rubric weights and positive and negative ideal profiles. Each response's reward is based on its relative closeness to these ideals. Evaluated with Qwen3-8B on four open-ended query-answering benchmarks, it achieves an average score of 63.02, about 2.0% above the strongest baseline.
In briefSmall Language Models Can Still Give Reasonable Self-Confidence Estimates
SummaryThis paper systematically tests how well language models of different sizes estimate their own uncertainty by judging their answers on question-answering tasks, from general to specialized knowledge domains. As models get smaller and domains more specialized, accuracy drops, but the reliability of the self-evaluated confidence stays largely stable. The authors conclude that the most capable model is not necessarily the best at judging its own predictions, and that smaller models can provide reasonable self-assessed confidence for resource-constrained deployments.
In briefBeyond Plato's Cave: Testing AI Agents for False Memories Using Counterfactual Reasoning
SummaryAutonomous agents rely on memory to generalize, but memories can introduce biases into their beliefs, which the paper calls false memory. The authors propose FAME, a training-free framework that detects false memory by measuring how agent beliefs drift under counterfactual scenarios, using hidden states before answer generation. The paper reports that monitoring answers alone often fails to detect false memory, while FAME achieves AUROCs of 76.2% to 96.7% and outperforms the best baseline on benchmarks including GSM-Symbolic, GitChameleon, and BigBench-Hard.
In briefT2SPO: Trajectory-to-Step Policy Optimization for Agentic Reinforcement Learning
SummaryThis paper proposes T2SPO, a method that uses past successful interaction trajectories to give LLM agents step-level feedback when rewards only reflect final outcomes. It derives remaining-distance targets from successful trajectories and uses a pretrained TabPFN regressor to estimate distance to success at each state of a new rollout. Changes in this estimate across consecutive states provide auxiliary credit for agent steps. Experiments with 1.5B and 7B models on ALFWorld and WebShop show T2SPO consistently improves overall task success over GRPO.
In briefDensePed-Lite: Quality-Aware Adaptive Pedestrian Detection Under Occlusion in Dense Scenes
SummaryThis paper presents DensePed-Lite, a lightweight framework for detecting pedestrians in crowded scenes with heavy occlusion, varied scales, and real-time demands. Its core principle is that the network should adapt to the quality of what it observes rather than assume complete information. The method targets three points where occlusion hurts most, and the authors report experiments on CityPersons and CrowdHuman showing a better accuracy-efficiency trade-off than recent lightweight methods.
In briefT-ARC: A clustering method that adds topology-aware correction to K-means
SummaryThis paper proposes T-ARC, a clustering method that embeds topological information into the objective to correct K-means' geometric bias. It couples the K-means data term with a graph-cut penalty, modeling the latent graph with a Stochastic Block Model and tuning its parameter via distributionally robust optimization. The topological prior comes from zero-dimensional persistent homology. Experiments on synthetic non-convex data and random Fashion-MNIST subsets report higher accuracy than K-means on curved and interleaved clusters, and greater stability on real data.
Serena Grazia De Benedictis, Andersen Ang, Nicoletta Del Buono et al.
In briefHead geometry in federated regression: a natural selection rule and unconstrained feature model analysis
SummaryThe paper studies federated averaging when local objectives have multiple optimal shared linear heads, so the aggregate depends on which head each client returns. Using an unconstrained feature model, it proposes a rule where each client returns the optimal head closest to the broadcast head, and shows that global minimization with a vanishing proximal penalty realizes this rule. The shared Gram matrix converges to the Bures-Wasserstein barycenter of the clients' optimal Gram matrices, which generally differs from the centralized optimum, and a one-time exchange of target means and covariances recovers the centralized result in one round under exact local optimization.[]
In briefRight-Wing Rock or Just Rock? A Computational Analysis of Frei.Wild
SummaryThis paper examines whether the German band Frei.Wild should be classed as politically right-leaning or as part of mainstream German rock. The authors built a right-wing rock reference corpus and ran lexical analyses and classifiers. They find the band keeps its political affiliation ambiguous but leans right, with two classifiers labeling more than half of its songs as right-wing extremist.
Carlotta Schneeberger (1), Kevin Tang (1, 2) ((1) Heinrich Heine University D\"usseldorf et al.
In briefMutual Equilibrium: a paper on multimodal learning via two-way feedback between inputs
SummaryThis paper proposes MEQ, a mutual feedback architecture that refines two inputs, possibly from different modalities, into a pair of coupled embeddings, each reflecting the other's information. The model's final output is defined as the fixed point of this interaction. The authors provide theoretical analysis and test on classification and visual grounding tasks across several datasets, reporting competitive or better results than concatenation-based baselines.
In briefFrom Speech to Editable Concepts: Probing Emotion Recognition with Concept Bottlenecks
SummaryThis paper adapts concept bottleneck models to speech emotion recognition to examine how individual predictions depend on transcripts, acoustic descriptions and speaker attributes. Three large language models are tested on CREMA-D, IEMOCAP and MELD. In the zero-shot setting the models are strongly biased toward the transcript, and fine-tuning removes this bias. The authors conclude that aggregate performance changes alone do not reveal how removing concepts affects individual predictions.
In briefResARC: residual-aware autoregressive coding for ultra-low bitrate image compression
SummaryThis paper presents ResARC, an image codec that compensates for two residuals in progressive autoregressive compression at the decoder. A diffusion transformer conditioned on the autoregressive context generates the quantization residual without extra side information, while the generation residual is computed at the encoder and sent through a learned codec. Experiments report competitive perceptual similarity and improved distributional fidelity over leading generative codecs at ultra-low bitrates.
In briefRobust linear regression under distribution uncertainty, viewed through adversarial training
SummaryThis paper studies linear regression using Wasserstein distributionally robust optimization, which guards against uncertainty in the data distribution. It shows that this general method includes square-root Lasso and adversarial linear regression as special cases. The authors prove that key properties of those special cases, such as non-asymptotic in-sample error bounds, insensitivity to noise level, and solution equivalences, carry over to the general method. The method can be solved efficiently, and numerical simulations validate the findings.
Elis Stefansson, David V\"avinggren, Ant\^onio H. Ribeiro
In briefAnalyzing synthetic data quality through how easily models learn it
SummaryThis paper examines how easily machine learning models learn from synthetic data generated by large language models, comparing different model families and sizes with human-written data as a reference. The authors generate synthetic datasets for classification, labeling, and tree prediction tasks, then derive sample-level data distributions from encoder training dynamics and test how robust these distributions are across encoders. They also evaluate how data selection strategies based on these learnability signals affect synthetic and human-written data differently.
In briefDuplexAct-Bench: Evaluating Full-Duplex Speech Systems on Proactive Real-Time Interaction
SummaryThis paper introduces DuplexAct-Bench, a bilingual English and Chinese benchmark covering six real-time interaction behaviors, including interruption, yielding, proactive initiation, active silence, and backchanneling. It tests these behaviors under Pre-session, In-session, and No-explicit conditions across 1,290 streaming trials. The authors evaluate 12 full-duplex speech systems on timing and content, finding large variation across behaviors, conditions, and systems, and frequent mismatches between semantic quality and timing. The results suggest current systems are far from reliably managing when, whether, and how to participate in real-time conversation.
In briefRouting on Raw Inputs for Mixture-of-Adapters in Time Series Foundation Models
SummaryTime series foundation models usually adapt to new data with a single trainable head on a frozen backbone, which underfits heterogeneous regimes. On instance-normalized backbones, replacing this head with a mixture of experts suffers routing collapse, where one expert absorbs every input. The authors propose Raw-Routed Mixture of Adapters (RR-MoA), which routes on the pre-normalization raw input, and report 54/54 wins against the strongest fixed adapter under a strictly frozen backbone, with effects reported across six backbones and an imputation task.
WhyIdentifies normalization-induced routing collapse and fixes it via raw-input routing, backed by eight causal controls and a mutual-information decomposition; scope is mainly time series models.
In briefCAST: Reinforcement fine-tuning for image diffusion models with spatially weighted, step-by-step rewards
SummaryThe paper proposes CAST, a reinforcement learning method for fine-tuning pretrained diffusion models. It sets the exploration window from each model's denoising trajectory, splits prompts into independently checkable semantic atoms via Causal Scene Graphs, and projects atom-level advantages onto pixels to weight the training objective. Applied to FLUX.2-dev and Qwen-Image-2512 and evaluated on GenEval 2 and Qwen-Image-Bench, CAST reports up to 3.07x the improvement of Flow-GRPO on the hardest prompts, with overall quality also improving.
In briefStudy showing principal component regression beats other monotone filters in linear regression
SummaryThis paper compares finite-sample, instance-wise risks of monotone spectral filters for linear regression, including principal component regression, gradient descent, and ridge regression. It shows that optimally tuned principal component regression has risk at most a constant factor larger than any such filter on every problem. For step-function-like filters such as ridge and gradient descent, it identifies problem instances where principal component regression's risk is smaller by a polynomial factor in sample size.
WhyProves PCR dominates all monotone spectral filters, substantially generalizing prior GD-over-ridge results with instance-wise sharp bounds.
In briefFaithfulness benchmark testing whether open LLMs reason from supplied geographic context or their own recall
SummaryThe paper presents an open pipeline that defines a point's surroundings by its pedestrian street network, retrieves OpenStreetMap and GHS-POP population data, and injects computed indicators into prompts as a spatial brief. It builds a faithfulness benchmark that labels each model claim by source and correctness and tests it with planted false premises that the brief refutes. Sixteen open-weight configurations were evaluated across three cities. Resistance to the false premise varies more by model family and generation than by scale.
In briefFrom Imitation to Reward Discovery: On-Policy Warmup for RL Agents
SummaryThis paper addresses slow early training in reinforcement learning with verifiable rewards (RLVR) for language-model agents, where sparse outcome rewards give little signal. The authors propose On-Policy Warmup (OPW), in which a student trains with teacher supervision on its own trajectories before switching to RLVR. In the main comparisons, RLVR initialized this way reaches high performance earlier and ends with higher final performance. The paper also offers a theoretical link between on-policy reverse-KL distillation and trajectory-level distribution matching to explain the effect.
In briefEvaluating how reliably AI signals its uncertainty for brain tumor (glioma) diagnosis
SummaryThis paper evaluates uncertainty quantification in a multi-task deep learning model that analyzes MRI scans for glioma tumor segmentation and predicts IDH mutation, 1p/19q co-deletion, and tumor grade. It uses Monte Carlo Dropout and compares it with Deep Ensembles and Monte Carlo Deep Ensembles across calibration, error detection, and selective prediction. Uncertainty estimates supported error detection, with calibration depending on dropout rate, and a composite trust score did not consistently beat classification uncertainty.
Gonzalo Esteban Mosquera Rojas, Sebastian R. van der Voort, Carolin M. Pirkl et al.
