In briefMoonGS: High-Quality Lunar Surface Representation via Gaussian Splatting from Image Pairs
SummaryThis paper addresses the difficulty of reconstructing high-quality 3D lunar terrain from sparse rover images, which have limited overlap and weak textures. It proposes MoonGS, the first feed-forward 3D Gaussian Splatting framework designed for lunar scenes, which predicts Gaussian primitives from only two input images in a single pass without per-scene optimization. The method outperforms state-of-the-art feed-forward baselines on the LuSNAR and synthetic MoonBlender benchmarks and shows the best visual quality on Chang'e mission imagery.
In briefSignature Features for Human Activity Recognition: A Reproducible Study
SummaryThis paper examines whether signature transforms, which encode temporal order and cross-channel interactions, improve human activity recognition from sensor data. Using the UCI HAR dataset with a leakage-aware, reproducible pipeline, the authors compare raw signals, time-augmented paths, and lead-lag paths across truncation depths and classifiers. With Random Forest, time-augmented six-channel signatures at depth 6 reached 0.858 accuracy, beating the strongest raw baseline at 0.816. MLP models did not exceed the raw baselines.
Kamal Jarrar (LMAP), Jacky Cresson (LMAP), Christian Paroissin (LMAP)
In briefJudgeMoE: Distribution-Based Aggregation for LLM Judges
SummaryThis paper introduces JudgeMoE, a lightweight aggregator that assigns example-specific weights to cached score distributions from LLM judges and fuses them before computing a final score, preserving uncertainty and disagreement information. On the original 10-cell benchmark, it improves mean Spearman correlation over uniform log pooling by +0.079. Extended to 16 cells, it gains +0.0393 over the strongest local single judge, with positive differences in 12 of 16 cells and a one-sided Wilcoxon signed-rank p-value of 0.0091. The authors also find that the preferred aggregation method depends on the task and judge pool.
In briefPatient-specific clinical explanations traced to proofs, built on knowledge-graph reasoning
SummaryThis paper presents the CKG Clinical Explanation Engine, a downstream layer for a frozen, training-free clinical knowledge-graph reasoner. It converts patient data and disease knowledge into typed facts, rule-application traces, provenance-linked conclusions, and policy-licensed recommendations. Across five usable workbooks covering 6,720 patients, structural and provenance auditability metrics reached 100% or near 100%. The authors state that these results show structural and implementation auditability only, not clinical correctness or utility.
In briefEmpirical Variational Autoencoder: a generative framework for continuous-valued sequences
SummaryThis paper proposes Empirical Variational Autoencoder (EVA), a general generative framework for continuous-valued, non-vector-quantized sequences. EVA builds on the VAE evidence lower bound but learns autoregressive latent priors empirically from training data, requiring only one additional linear layer on top of a VAE. Replacing the standard Gaussian constraint with self-predicted priors reduces the gap between prior and posterior latent distributions. Experiments on image and sound synthesis report generation quality competitive with autoregressive diffusion baselines, with much faster inference.
In briefA detailed analysis of the LoRA fine-tuning optimization landscape and its implications for data selection
SummaryThis paper studies how to choose the adapter rank in LoRA, a parameter-efficient fine-tuning method. The authors introduce LoRA-RIP, a data-dependent restricted-isometry metric that characterizes the conditioning of the cross-entropy objective along low-rank directions, and prove that sufficient rank over-parameterization eliminates spurious local minima. Experiments on language and vision tasks suggest that rank and data quality should be considered jointly.
In briefWaveGSSM: A graph wave model that tracks how spatio-temporal patterns move
SummaryThis paper introduces WaveGSSM, a second-order graph state-space model that keeps two linked states per node: one for the current pattern and one for its rate of change. A graph-wave transition uses these to advance the pattern while capturing spatial propagation and temporal evolution together. The authors evaluate it on four temporal-graph benchmarks and global weather forecasting, reporting the best mean performance on the benchmarks and a 20.2% average reduction in geopotential RMSE for 1- to 5-day forecasts versus a backbone-matched snapshot model.
In briefAnatomy of LLM Sycophancy: What a Flip Rate Hides
SummaryThe paper introduces SycoLens, a modular replay protocol that measures how user pushback and evaluation settings affect answer flip rates in language models. Across eleven frontier models and about 760,000 controlled replays, it finds that which models appear sycophantic depends on how the user pushes back. The authors argue that a single flip rate cannot separate correction from capitulation, and propose a multi-factor reporting profile instead.
In briefGenerative AI model reproduces varied neuron voltage responses to the same stimulus
SummaryThis paper introduces a conditional generative model, built with flow matching, that learns the range of voltage responses a single neuron can produce to one input current. Tested on biophysically detailed models of two human cortical interneuron types, the generated responses reproduce electrophysiological feature distributions, spike timing, and excitability, and match experimental recordings from human cortical neurons. The model recovers both firing and non-firing responses near threshold and both low- and high-firing modes near depolarization block, whereas a neural operator baseline suppresses spiking near threshold and blurs those modes.
Cameron Schofield, Luca Ghafourpour, Philip H. Wong et al.
In briefTest-time adaptation of reasoning strategies using Bayesian memory
SummaryLarge language models reason through chains of thought, which can be costly, and insights gained during inference are hard to reuse. This paper introduces a Bayesian Cheatsheet, a structured memory that clusters behavior embeddings with a Hierarchical Dirichlet Process Gaussian Mixture Model and retrieves relevant behaviors at inference time. It updates online during test-time training and creates new components when behaviors are sufficiently novel. The authors report performance gains over existing memory modules on AIME'25, Omni-MATH, and PhysReason, including in a cold-start setting.
In briefANT: A network traffic dataset and benchmark for auditing LLM agent behavior
SummaryThis paper introduces ANT, a dataset of network traffic from LLM agent executions, annotated with risk, scenario, and behavior primitive labels. It contains 3,114 episodes, 276,417 bidirectional flows, and 40,049 behavior primitive segments. Benchmarking 13 traffic analysis baselines shows that existing methods recover useful but uneven behavioral signals, struggling with malicious workflows resembling benign tasks.
In briefSOL: Measuring Text Distribution Gaps with Double Sliced Wasserstein Distance
SummaryThe paper proposes SOL, a distance between text distributions that represents each sequence by the empirical measure of its hidden states from a fixed transformer, then compares these measures with the double sliced Wasserstein distance. The authors prove SOL is a metric when the transformer is injective. Experiments show it detects distributional failures, recovers expected model trends, and gives stable sample-based estimates, and it is used to re-evaluate several models trained on OpenWebText.
Gregor Kornhardt, Moritz Piening, Jannis Chemseddine et al.
In briefArtifactArena: Judging models by the physical artifacts they build
SummaryThis paper introduces ArtifactArena, an open-ended platform where models must co-design hardware and software to engineer functional robots that compete in a simulated arena. The authors assess frontier models through three harnesses that refine robot designs using text descriptions, physics simulator feedback, and gameplay data. Performance is benchmarked with an Elo ranking derived from head-to-head tournaments between the models' robots. The authors release the framework and tournament infrastructure for ongoing community submissions.
In briefXaurora: a generative weather forecasting model built from a pretrained deterministic foundation model
SummaryMost atmospheric foundation models for weather forecasting are deterministic, which limits their ability to produce large ensembles for uncertainty and extreme-risk assessment. This paper converts the pretrained Aurora model into a generative ensemble model, Xaurora, using Denoising Stochastic Interpolants and a replay buffer for SDE rollouts. Xaurora is finetuned from the small Aurora version and approaches state-of-the-art global ensemble metrics, competing with the large Aurora. It generates 15-day forecasts in 13 minutes.
Eliot Walt, Miltiadis Kofinas, Nikolaj M\"ucke et al.
In briefA general pipeline that creates dense lighting color maps from physically rendered synthetic scenes for training illuminant estimation models
SummaryThis paper presents a reusable pipeline that derives dense illuminant chromaticity maps from physically based 3D-rendered scenes, aiming to ease data collection for illuminant estimation in computational photography. By repurposing an existing 3D scene collection, it generates pixel-wise lighting annotations under controlled conditions. The authors produced 74,321 synthetic images and used them to pre-train single- and multi-illuminant estimation models, reporting consistent gains, especially when real data is scarce.
In briefHierarchical procedural understanding for better proactive AI assistance
SummaryProactive AI assistants watch a user's activity and decide when to offer guidance or stay silent. The paper introduces ProactiveCoach, a suite with a training dataset (ProactiveCoach-Instruct), an evaluation benchmark (ProactiveCoachBench), and fine-tuned vision-language models. Hierarchical supervision improves overall performance over fixed-granularity supervision by up to 9.6 percentage points across backbones. An adaptive guidance system built with a lightweight router outperforms an in-context adaptation baseline by 57.1 percentage points across four guidance-level transitions.
In briefStudy on how text-to-image models can generate harmful content and where moderation fails
SummaryThe paper systematically tests five open text-to-image models by turning news captions into unsafe prompts across several harm categories. A human evaluation of 1,500 images found high harmful-content rates, especially for gore (89.2%). The authors also report detection gaps in automated moderation systems and show that synthetic image detectors trained only on benign data perform worse on explicit content.
In briefNeuroCBIR: A Fast, Accurate Brain MRI Search System for Whole-Brain and Regional Scans
SummaryNeuroCBIR is a framework that searches for similar 3D T1-weighted brain MRI scans, either for the whole brain or for specific regions. It extracts 103 cortical and subcortical regions and uses a variational autoencoder with contrastive learning to create scan embeddings. The embeddings achieved mean average precision at top-5 of at least 98.4% for subject re-identification and carry information useful for zero-shot age and pathology tasks. The software and more than 26,000 precomputed embeddings are publicly available.
Felix Nieto-del-Amor, Jingru Fu, J.-Sebastian Muehlboeck et al.
In briefYou Changed Your Mind, The Model Didn't: Intent Handling in Multi-Turn Dialogue
SummaryThis paper studies how large language models behave in multi-turn tasks when a user proposes a change and then rejects it. The authors introduce Intent-Eval, a benchmark spanning tool actions, code, databases and mathematics, and find that merely mentioning a rejected or superseded change can degrade task accuracy. To address this, they propose Intent-OPSD, a self-distillation framework in which a frozen teacher model, initialized from the same model, supervises a student to follow only the requirements that remain in effect.
In briefTopology-informed, prompt-conditioned model segmenting uterine structures in ultrasound and MRI
SummaryThis paper proposes TPUS, a framework that segments multiple uterine structures across ultrasound and MRI images. It combines modality-specific input stems, a shared graph-based encoder-decoder, task-aware class prompts, a dynamic convolution module, and a topology-informed loss to handle differing label spaces across datasets. On a uterine ultrasound dataset and a T2-weighted uterine myoma MRI dataset, it reports Dice scores of 0.898 and 0.693 on the two held-out test sets, outperforming several generic and universal segmentation baselines.
In briefpolyview: A Python package for multi-view machine learning
SummaryThis paper presents polyview, a scikit-learn-compatible Python package for multi-view embedding, clustering, fusion, view augmentation, and handling incomplete views. The authors demonstrate it on five real multi-view datasets and compare its canonical correlation analysis-based components with those of two established libraries. The package is intended as a practical toolkit for benchmarking and prototyping multi-view methods, and as a foundation for future research.
Gwendal Debaussart-Joniec (ENS Paris Saclay, CB), Argyris Kalogeratos (CB et al.
In briefGPlaceRL: An Open-Source Graph Reinforcement Learning Framework for Detailed Placement
SummaryThis paper presents GPlaceRL, an open-source framework that uses graph reinforcement learning to refine detailed placement in circuit design. It represents legalized placements as graphs and offers a modular environment for testing graph encoders, policies, rewards, and local placement actions. Using PPO policies with graph attention network encoders on five benchmarks, the best greedy results achieved HPWL improvements of 3.27% to 32.87%.
Pavlos Stoikos, Foteini Oikonomou, Christos Poulos et al.
In briefFairProp: Fair Node Representation Learning via Differentiable Propagation Layers
SummaryThis paper studies group fairness in graph neural networks for node classification, link prediction, and node regression, bounding the demographic parity gap for any number of sensitive groups. It identifies two sources of bias: differences between group means and within-group covariance of final representations. The authors propose FairProp, which embeds fairness constraints into the APPNP propagation process and provably converges linearly to a unique fair optimum. Experiments on three tasks report strong fairness-utility trade-offs compared with baselines.
In briefAECP: A protocol where AI coding agents coordinate only through structured artifacts
SummaryThis paper introduces AECP (Artifact-Exclusive Communication Protocol), which requires multiple AI agents working on code generation to communicate only through structured artifacts. The execution harness supplies relevant findings, checks implementations against recorded interface commitments, and requires affected agents to revisit revised agreements. On Doc2Repo, NL2Repo, and CodeProjectEval, AECP improved average test pass rate by 28.2% and reduced average wall time by 16.5% compared with free-form inter-agent messaging. It also blocked the relay of malicious instructions between agents.
In briefBehavior-Preserving KV Cache Compression: A Training-Free Method
SummaryKey-value (KV) caches consume large memory during long-context inference in language models. This paper proposes a training-free framework that decides which cache entries to evict by estimating how much removing each one would change the full-cache model's next-token distribution, measured by KL divergence. The authors report that, at matched cache budgets across several architectures, the method improves downstream task quality over attention-based heuristics, with the largest gains under aggressive compression, while still giving an end-to-end speedup over full-cache inference in their tested settings.
In briefBeyond Successor Accuracy: Keeping Model States in Recursive Recommender Self-Improvement
SummaryThe paper studies recursive self-improvement in recommenders, where outputs feed later training. It argues that pre- and post-update models can hold complementary ranking decisions, which it calls distributed progress, and measures this with cross-generation advantage. Experiments across four datasets and three sequential recommendation encoders show the best retention strategy varies by architecture, and a label-free rank-separation statistic can select the stronger model pairing.
WhyIntroduces distributed progress and cross-generation advantage, showing state retention is a new Rec-RSI problem. Validated across four datasets and three encoders, with the selection rule working in
In briefMaRO-GS: Robust 3D Object Reconstruction from Inconsistent Multi-view Masks
SummaryThis paper presents MaRO-GS, a Gaussian Splatting method that reconstructs only the target object from multi-view images instead of the whole scene. It addresses inconsistent 2D segmentation masks across views by filtering unreliable views, suppressing irrelevant Gaussians, and using a silhouette-aligned object loss. The authors report improved PSNR, segmentation accuracy, and efficiency across diverse datasets, with a maximum PSNR gain of 2.05 dB on the small-object LERF-Mask dataset.
In briefOdyssey: A Long-Horizon Closed-Loop Driving Benchmark with Explicit Navigation Routes
SummaryThis paper introduces Odyssey, a closed-loop benchmark for evaluating end-to-end driving over long horizons, with 100 scenarios reconstructed from 100-second nuPlan logs. It replaces ambiguous directional commands with explicit standard-definition map routes and adds metrics including SD Route Compliance, a Pre-Lane Change Score, and RouteDS, which penalizes route deviations and failed lane preparation. The authors adapt state-of-the-art planners, including vision-language-action models, and evaluate their navigation performance with these metrics. Benchmark code and adapted baselines are said to be released publicly.
In briefLatent Flow Matching for Molecular Graph Generation: generating whole molecular graphs in a compressed latent space
SummaryThe paper generates entire molecular graphs in the latent space of a pretrained Variational Autoencoder with high reconstruction fidelity, using flow matching and decoding only at the final step. Across molecular benchmarks of increasing size, it achieves strong validity and FCD, with a favorable quality-efficiency trade-off against state-of-the-art explicit graph generators. The latent representation is learned once and reused across multiple generative objectives.
Mathis Goupillon, Roman Bresson, Konstantinos Divriotis et al.
