In briefIs Sycophancy a Form of In-Context Updating?
SummaryThis paper examines factual sycophancy, where language models defer to users' false beliefs. The authors argue it is better understood as a failure of in-context updating, the process by which models revise answers based on information provided during an interaction, rather than only a bid for user approval. Behavioral and mechanistic evidence supports this view, showing the same components drive updating more broadly. The authors note that reducing sycophancy may also reduce useful updating, so evaluations and mitigations should measure and separate both behaviors.
In briefPACE: A Training-Free Method for Response Timing in Streaming Video Understanding
SummaryThis paper addresses when a model should answer in streaming video understanding, noting that premature responses are a common failure. The authors propose PACE (Plan, Attend, Check, Emit), a training-free framework that drafts a plan, then checks each answer for stability, grounding in cited frames, and event completion before emitting it. Experiments on ProactiveVideoQA and OVO-Bench Forward show gains over official protocols across proprietary and open-source backbones, reaching state-of-the-art performance.
In briefManifold-Informed Neural Networks for Inferring Electrolytes from Scarcely Labeled ECGs
SummaryThis paper proposes the Manifold-Informed Neural Network (MINN), which solves a discrete Poisson equation with a learned source term on a k-nearest-neighbor graph in latent space, replacing the known operator residual used in physics-informed networks. On three-class classification of potassium, calcium, and magnesium from 12-lead ECGs in MIMIC-IV-ECG, MINN reaches a potassium Macro-F1 of 0.61, above the best baseline's 0.55. It also shows a 31% relative improvement with 1% labeled data.
In briefStabilizing Filters for Graph Signal Generation: Designing Filters That Stay Robust to Graph Errors
SummaryThis paper studies how stable graph-based generative models are when the underlying graph is noisy or imperfectly estimated. The models combine a graph filter with a learned graph neural network to generate graph signals. The authors derive a Wasserstein stability bound that quantifies how graph perturbations affect the generated signal distributions. They then propose a filter design framework and an explicit family of polynomial filters with provable stability guarantees, and experiments on synthetic graphs and real fMRI signals support these findings.
In briefABACUS: A Benchmark for Testing Whether Image Editors Add, Remove, or Change the Right Number of Objects
SummaryThis paper introduces ABACUS, a benchmark, evaluation protocol, and scoring metric for measuring how accurately instruction-based image editing models handle object counts. Instead of relying on vision-language model judges, it checks each edit at the instance level against panoptic ground truth confirmed by detector consensus, verifying the requested count and that the rest of the scene is preserved. Evaluating leading image editors, the authors find substantial numerical editing failures that grow worse as scenes become more complex and object counts increase.
In briefBias-corrected neural networks for amortized Bayesian inference on semi-parametric targets
SummaryThis paper studies prior-data fitted networks (PFNs), an end-to-end approach to amortized Bayesian inference, for semi-parametric targets such as means under missingness or causal effects. The authors show that even an idealized, Bayes-optimal PFN can concentrate posteriors slowly, depending on nuisance functions and prior diversity. They propose two debiasing strategies, label correction and prior correction, based on efficient influence functions, which achieve concentration rates matching frequentist bias-corrected estimators. Under additional assumptions, the debiased PFNs attain the Bernstein-von Mises property and minimax optimality.
WhyProposes two influence-function-based debiasing strategies, linking amortized Bayesian inference with semi-parametric efficiency theory and proving asymptotic optimality guarantees.
Valentyn Melnychuk, Emil Javurek, Stefan Feuerriegel
In briefTraining one geospatial encoder to align with several foundation models
SummaryGeospatial foundation models differ in pretraining objective, modality and resolution, and none leads across all benchmarks. The authors introduce Pangaea-Latents, which provides frozen embeddings from 5 such models over the 11 datasets of the Pangaea benchmark, and find the embeddings are neither redundant nor disjoint. They then propose GeoMixture, which aligns a ViT encoder to several models through a multi-representation alignment objective. Across nine datasets, the encoder ties the best model in average rank, trading peak performance for cross-task consistency.
In briefFrom Abstraction to Action: Domain-Specific Languages for Interpreting Abstract Instructions
SummaryNatural-language instructions such as "stack blocks in a pyramid" rely on abstractions that language models struggle to unroll into concrete actions, especially at higher abstraction levels. The paper proposes domain-specific languages (DSLs) as an explicit representation layer whose primitives expand abstractions into actions, tested on the Hexagons and Text2CAD benchmarks. Human-curated DSLs yield large gains, up to 10 F1 points, while model-induced DSLs recover only part of the benefit and are inconsistent across runs.
In briefRecommending bacterial growth conditions from 16S rRNA sequences
SummaryThis paper introduces DSNI, a benchmark of 65,023 bacterial strains that pairs 16S rRNA sequences with four cultivation conditions: media, temperature, pH, and oxygen, using phylogeny-aware data splits. It proposes a Variability-Gated Encoder that learns position-dependent importance weights based on region type, position, and local complexity. Experiments report a 4.7% gain in joint accuracy over flat Transformers, 44.2% exact-match joint accuracy when combined with RNABERT, and 90.2% of ground-truth protocols within the top 25 recommendations.
In briefReward Misspecification Lowers Chain-of-Thought Monitorability
SummaryThis paper studies how misspecified rewards in reinforcement learning affect whether a language model's chain-of-thought (CoT) can be monitored. The authors build a testbed pairing true task-aligned rewards with misspecified rewards, and find that CoT monitorability stays low as hack rates rise, since reasoning still tracks task intent while outputs violate it. They propose training models on inferred reward descriptions so the CoT explicitly reasons about the misspecified reward, which improves monitorability by an average of 24% or 42% absolute, depending on the metric.
In briefLearning normal coordinates on Riemannian manifolds using invertible neural networks
SummaryThe paper trains a conditional invertible neural network, called a flow chart, to approximate normal coordinates around points on a Riemannian manifold. In these learned coordinates, the exponential and logarithmic maps reduce to fast Euclidean operations. The authors also introduce a nonnegative loss function that is zero exactly when the chart is normal, under suitable assumptions.
In briefAdaptive Replay Refinement: Shifting Test-Time Compute from Revising Answers to Resampling
SummaryThis paper proposes Adaptive Replay Refinement (ARR), a training-free method that detects when repeated revision of a solution reaches a fixed point and then reallocates that compute to a new independent sample. It treats the current solution as a draft and verifies unchanged tokens in parallel to detect the fixed point efficiently. Across seven models and seven benchmarks, ARR matches majority voting on three of five math reasoning benchmarks at equal compute while cutting critical-path serial passes by up to 2.4 times.
Ahsan Bilal, Muhammad Ahmed Mohsin, Muhammad Umer et al.
In briefAGGR-Mamba: a Mamba-based network that refines undersampled MRI reconstructions with attention and gated residuals
SummaryUndersampled MRI reconstruction recovers high-quality brain images from partial k-space data to shorten scan times. The paper proposes AGGR-Mamba, a lightweight output-stage refinement method built on a Mamba backbone, using an attention refine head and a GDFN-style gated residual branch to correct residual local errors and high-frequency artifacts. Experiments on the IXI brain MRI dataset with T1- and T2-weighted images at R=4 and R=8 show competitive PSNR and SSIM results.
In briefCPVB-RG: Prior-Aware Chest X-Ray Report Generation Using Cross-Prior Visual Bridging
SummaryThe paper proposes CPVB-RG, a framework that generates chest X-ray reports by combining current and prior radiographs with available clinical text, so that new, resolved or progressing findings can be identified. Its Cross-Prior Visual Bridge uses progression-guided attention, contrasting progression-difference attention with shared-spatial attention, plus two cross-attention units to link longitudinal visual evidence with clinical semantics. Experiments on MIMIC-CXR and IU X-Ray report consistent improvements across language and clinical metrics.
In briefLocal Work, Global Progress: A Proof That AI Agent Teams Converge
SummaryThis paper treats collaboration among large language model agents as block optimization, where each agent improves one part of a shared solution and a coordinator checks proposals against the global objective. The authors prove linear convergence up to an error floor that vanishes when corrections are exact. In robot routing experiments, a three-agent team found collision-free plans in every run, and merging the best checked combination of proposals cut search time by about 30%.
WhyFormalizes LLM agent collaboration as block optimization with a linear convergence proof; experiments on robot routing and SQLite query planning against a single-agent baseline support it.
In briefHRoPE: Homography-based relative positional encoding that helps multi-view transformers respect scene surface geometry
SummaryThis paper introduces HRoPE, a positional encoding for multi-view transformers that models each image patch as a slanted 3D plane. It injects the plane-induced homographies into attention as relative transformations, helping the model correct perspective distortion and align patch features across views. The method needs no ground-truth geometry, since plane parameters can be learned end-to-end from a view-synthesis objective or taken from foundation models. The authors report that their best variant outperforms the compared geometric positional encodings for novel view synthesis on RealEstate10K, CO3Dv2, and DL3DV.
In briefChartEvo: Self-evolving agents that improve reasoning on rare chart types
SummaryMultimodal large language models perform well on common charts but lose accuracy on long-tail charts with diverse visual structures. ChartEvo is a recursive self-improvement framework that lets chart agents evolve reusable chart-specific Skills and executable Tools from experience, and learn reusable tool-call chains to control inference overhead. On long-tail charts from ChartBench and ChartX, it improves accuracy over its base model, Qwen3.7-Plus.
In briefForty Questions: Identifying a Language Model from One-Word Answers
SummaryThe paper proposes identifying the model behind an API from its one-word answers to everyday questions such as "Name a sound" or "Name a poet". On a panel of 94 models, 40 answers identified the exact model 92.5% of the time, against 1.1% by chance. Applied to 96 endpoints of 13 open models on OpenRouter, 92 matched the model they claim to be.
WhyNovel one-word-answer fingerprinting reaches 92.5% accuracy on 94 models and exposes real-world model substitution and billing anomalies on endpoints.
In briefGALA: A Geometry-Aware Language Model for Controllable Object Placement in 3D Scenes
SummaryThis paper introduces GALA, a language model that explicitly perceives point cloud geometry to place objects in 3D scenes under fine-grained constraints such as support and containment. It predicts the 6D pose of one target object at a time, conditioned on scene context and a placement instruction, using two-stage training. The authors also release a multi-source scene dataset and report that GALA outperforms baselines lacking explicit geometric reasoning on their benchmark.
In briefHow Language Models Learn Taste: Brand Preference in LLMs
SummaryThe paper compares 102 models across 42 brand categories and finds that the brand a model names differs from the brand it recommends in 26 of them. When recommending, models tend to converge on shared niche or specialist brands, while named brands are mostly market leaders. The authors report that preference optimization (DPO) training shifts these recommendations most, though they state they do not know exactly how the preferences arise.
In briefVGL-stGen: A shared model linking lung tissue images and gene activity
SummaryVGL-stGen is a multimodal large language model adapted from a pretrained medical vision-language model. It was trained on paired lung H&E images, spatial transcriptomics data and single-cell references. One checkpoint supports thirteen evaluated tasks, including tissue and cell-type classification, gene prediction and conditional expression completion.
In briefPI-MLLM: A Psychophysics-Inspired Multimodal Model for Analyzing Emotions in "Oddly Satisfying" Videos
SummaryThis paper studies Oddly Satisfying Videos, which provoke strong viewer reactions through repetitive motion, visual texture, and ASMR-like sounds rather than explicit semantic content. The authors build the OSVD dataset of 12,000 annotated 5-second clips and propose PI-MLLM, a multimodal large language model that models visual motion and acoustic dynamics. Experiments report improved viewer emotion prediction and sensory-grounded explanations.
In briefReliability-aware reconstruction of V-shaped surfaces using a low-cost multi-zone SPAD sensor
SummaryThis paper addresses unreliable depth estimates near edges when using low-cost multi-zone SPAD sensors for V-shaped surfaces. It learns point-wise reliability from sensor confidence, peak amplitude, and peak sharpness, and uses it in a graph-based depth optimization. Tests on 60, 90, and 120-degree targets reduced the mean V-angle error from 11.21 to 9.66 degrees (13.8%).
In briefSplatFlux: feed-forward 4D Gaussian scenes from video latents for reconstruction and generation
SummarySplatFlux is a feed-forward framework that decodes video latent representations into 4D Gaussian splatting scenes, supporting both video-based 4D reconstruction and conditional 4D generation. It uses a static-dynamic decoupled 4D Gaussian representation that predicts dynamic probabilities for Gaussian primitives, separating static structure from moving regions. An auxiliary distillation strategy supplies pseudo dynamic masks and optical-flow cues so the model can learn without costly annotations. The authors report state-of-the-art performance across 3D and 4D reconstruction and generation experiments.
In briefGeometry-aware block selection for sparse variational Gaussian processes: choosing better data partitions to tighten the variational bound and improve prediction.
SummarySparse variational Gaussian processes (SVGPs) are a scalable method for Gaussian process modelling that give a lower bound on the full GP marginal likelihood. Recent work shows that a block-diagonal structured scaling matrix yields a tighter bound. This paper studies whether selecting blocks that tighten the variational objective improves test RMSE and NLL at a fixed inducing-point budget. The authors propose a simple, scalable, geometry-aware algorithm that recovers much of the log-determinant contribution by choosing a data partition, and a doubly stochastic training framework. They also explore using alternative conditional posterior structures to learn hyperparameters and inducing points.
In briefLoRA-adapted vision models reach similar brain MRI performance via different internal changes
SummaryThis paper studies how four vision foundation models, pre-trained differently, adapt to brain tumour segmentation using LoRA. Although their segmentation performance is similar, their internal representations change in different ways, and their effects on Alzheimer's disease classification differ. The authors find that adapting only early-to-mid layers can match or approach full-layer segmentation while reducing classification degradation.
Cecilia Diana-Albelda, Fiona Kekwick, Raghav Mehta et al.
In briefSafe multi-robot collaboration via trust-based link exclusion
SummaryRobot swarms communicate over directed, time-varying links, some of which may be compromised to send malicious data. The paper proposes a communication wrapper that uses Safe Anytime-Valid Inference to test trust values from an anomaly detector on each robot and select safe links, requiring only that the detector perform better than chance. Hardware experiments on 3 LoCoBots and 8 GoPiGos cover neural implicit mapping and heading consensus, and simulations using hardware-sampled trust values suggest scaling to swarms of hundreds of robots.
Thomas Kaminsky, \'Aron V\'ek\'assy, Hammad Izhar et al.
In briefTPO: Three-level robust learning for preference optimization under distribution shift
SummaryThe paper observes that Direct Preference Optimization (DPO) assumes training and deployment preference distributions match, which often fails across group, prompt, and preference levels. It proposes TPO, a tri-level distributionally robust framework using a hierarchical rectangular ambiguity set and Wasserstein distance. The authors derive a tractable objective with finite-sample convergence guarantees and report experiments showing it outperforms existing robust DPO methods.
In briefNorm-optimal model order reduction in kernel Hilbert spaces
SummaryThis paper studies model order reduction in reproducing kernel Hilbert spaces (RKHSs). The authors generalize known interpolatory conditions for norm-optimal reduced models and propose an alternative to the IRKA algorithm for RKHSs with differentiable kernels. Using generalized Malmquist-Takenaka functions, they extend the framework to non-differentiable kernels. Numerical experiments cover machine learning models such as support vector machines and deep ReLU networks, as well as polynomial fitting and a finite-element-like method for PDEs.
Tam\'as D\'ozsa, Andrea Angino, Rolf Krause et al.
In briefLocal inference of long-context LLMs: a data systems perspective on single-GPU bottlenecks
SummaryThis paper studies why serving long-context large language models on a single user's GPU is inefficient, beyond simple memory capacity limits. The authors compare several models across multiple GPU generations rather than one model on one GPU. They find that self-attention dominates long-context inference runtime and leaves key GPU resources underutilized, with the newest GPU often least well served. The authors frame many of these inefficiencies as data management problems, such as KV cache paging and cache utilization, and present them as experimental observations with recommendations for future work.
In briefCan Language Models Develop Aphasia-Like Failures From Neuron Removal?
SummaryThe authors systematically ablate neurons in LSTM and Transformer language models and apply limited fine-tuning as a proxy for compensatory plasticity to study how neuron loss shapes capability decline. They find the architectures fail differently: LSTMs degrade gradually, while Transformers show late threshold-like collapse under feedforward ablation. The paper suggests these failure profiles may offer computational perspectives on aphasic syndromes and neurodegenerative diseases.
In briefExeBench-FHD: Fixing Fidelity Problems in Decompiler Benchmarks
SummaryThis paper examines ExeBench, a popular benchmark for LLM-based decompilers, and identifies nine issues that reduce how faithfully it measures decompiler performance. The authors release ExeBench-FHD, a backward-compatible revision that adds 16 new fields and five warning flags. In an -O2 evaluation across four architectures, the new fields raise re-execution rates on some function categories by up to 26.5%.
In briefImage-based screening of food packages against draft R3337 with abstention and selective VLM review
SummaryThis study presents a workflow that extracts nutrition, ingredient and product-form evidence from food-package photos to apply South Africa's draft R3337 warning criteria. A detector localises relevant regions, OCR and structured extraction recover the evidence, and deterministic rules produce outcomes, abstaining when evidence is incomplete or conflicting. Selective review with Qwen3-VL 8B reassesses some unresolved cases. On a manual validation set of 65 products drawn from 442 products, selective accuracy was 97.4% and coverage rose from 51.7% to 65.0%.
Mayimunah Nagayi, Alice Scaria Khan, Tamryn Frank et al.
In briefLearn to Modify: A Self-evolving Multi-agent Framework for Refining Conversational Text-to-Motion Generation
SummaryExisting text-to-motion methods usually generate motion in a single shot, so users who dislike part of the result must regenerate the whole sequence. This paper proposes a conversational paradigm in which the system handles instruction understanding, generation, localized editing, and quality assessment. The authors present SMART, a multi-agent framework with a semantic-guided masked motion editing mechanism that maps natural-language feedback to spatiotemporal regions, and a memory-based self-evolving strategy that reuses past successful edits. According to the paper, experiments show that SMART supports multi-turn conversational generation and efficient refinement.
In briefJAXbounce: A Pong-like continuous-control benchmark for testing active inference beyond simple environments
SummaryThis paper introduces JAXbounce, a pong-like continuous-control benchmark for evaluating active inference agents beyond discrete, low-dimensional settings. It supports state-vector or pixel-grid observations with composable noise models, and is differentiable and parallelisable in JAX. Over 10,000 simulations compare a deterministic baseline, a state-based agent using Unscented Kalman Filter perception with expected free energy planning, and a pixel-based agent with a pretrained neural encoder. The results characterise how performance varies with planning, observation modality, and noise knowledge.
Chaitanya Kaul, Markus Klar, Sebastian Stein et al.
In briefMaD-RSI: Treating Models as Data to Recursively Self-Improve Agent Harnesses
SummaryThe paper proposes MaD-RSI, a framework that optimizes a single shared agent harness across several heterogeneous LLMs by using their diverse failure trajectories as training evidence. Instead of checking model differences only after optimization, it uses them to guide the harness's self-improvement. The authors also introduce two complementary metrics for evaluating harness optimization across models. Experiments on multiple datasets and LLMs reportedly show better generalization than existing approaches with lower training costs.
In briefParameter-efficient adaptation of time series foundation models for electrical load forecasting
SummaryThis paper compares three ways of using time series foundation models for electrical load forecasting: pretrained inference, full fine-tuning, and low-rank adaptation (LoRA), across four load aggregation levels. The primary backbone is MOIRAI 1.0 Large, and Chronos-2 is used to check whether the findings transfer across model families. Rank-16 LoRA updates about 1% of parameters and reduces normalized mean absolute error by 7.5% on average, retaining most of the gain from full fine-tuning.
In briefRethinking object-centric reinforcement learning by switching from absolute to relative position representations
SummaryThe paper argues that object-centric reinforcement learning agents usually receive object positions in a fixed global frame, even though decisions often depend on relative positions, which makes performance collapse when a scene is shifted. The authors propose Object-Relative Coordinates fixed to Anchor (ORCA) and a transformer architecture, ORBiT, that processes anchor-relative inputs. They report that relative representations improve both MLP and transformer agents, and that ORBiT achieves an aggregate HNS IQM above 1.0 across 13 Atari games.
Cedric Derstroff, Jannis Bl\"uml, Janosch Moos et al.
In briefAttractive Faces Skew Text-to-Image Reward Models and Leaderboards
SummaryThis paper studies how human preference judgments in text-to-image evaluation are biased by facial attractiveness. In a controlled user study, annotators rated more attractive faces as higher quality, even when comparing real photographs of similar fidelity. The authors show this bias can distort Elo leaderboards and appears in several human preference models, where attractiveness is the strongest correlate of score. They also report that preference-aligned generation shifts faces toward younger, more feminine looks.
WhyShows facial attractiveness overrides quality in preferences, reproduced across several reward models and leaderboards with broad impact.
Harvey Mannering, Zhiwu Huang, Adam Pr\"ugel-Bennett
In briefAdaptive Pullback: Reweighting training signals at sudden penalty spikes in on-policy distillation
SummaryThis paper studies on-policy distillation, where a student model learns from a teacher using per-token log-ratio penalties on its own generated responses. It finds that abrupt increases in average penalty, called penalty jumps, coincide with decreasing teacher confidence, a pattern the authors call penalty-confidence tension. The proposed Adaptive Pullback OPD method rescales supervision coefficients after each detected jump so the following block's weighted penalty matches the preceding block's mean. The authors report improved performance on mathematical reasoning and code generation.
In briefOffline Calibration Improves Online Video Cache Compression (CALM)
SummaryThis paper addresses KV-cache memory use in autoregressive video generation by proposing CALM, which allocates token-merging ratios across layers and heads rather than applying one uniform ratio. It uses offline calibration, with attention-output error as a lightweight proxy, to build a fixed layer-head risk map. At inference, dynamic programming picks the allocation for a given cache budget. Across three autoregressive video models, CALM outperforms uniform merging under the same memory budget, reducing LPIPS to original videos by about 21%.
In briefUnified RAW image exposure correction and restoration via linear illumination and frequency-adaptive enhancement
SummaryThis paper proposes RIECR, a unified framework that corrects both under- and over-exposed RAW images and restores their detail. It first estimates exposure-aware linear gains from channel-specific RAW cues, then uses a dual-branch frequency network with dynamic convolution to restore remaining degradation. The authors also build a real paired RAW benchmark of 223 scenes and 1,784 pairs, and report state-of-the-art results against sixteen methods.
In briefCrossFit-GS: Out-of-fold pose estimation for COLMAP-free 3D Gaussian Splatting
SummaryThis paper addresses 3D Gaussian Splatting (3DGS) built directly from unposed image sequences, avoiding traditional Structure-from-Motion (SfM) preprocessing. The authors propose CrossFit-GS, which splits the input into interleaved subsets with two complementary submaps, refining each candidate pose against a frozen submap that never saw the target frame. Checkerboard validation and spatial coverage checks reject unreliable updates before the final global model is trained from scratch. The authors report reduced cumulative trajectory drift and improved pose estimation and novel-view synthesis across multiple benchmarks.
In briefFinding Shared Audio-Text Neurons in Large Audio-Language Models
SummaryThis paper examines how large audio-language models (LALMs) combine audio and text representations before generating text. The authors hypothesize a shared subspace where both are aligned and identify cross-modal neurons that represent both audio and text. Across three LALMs and four acoustic attribute types, audio and text neurons show alignment to varying degrees. Causal interventions that steer and deactivate these neurons indicate they are necessary for prediction and can be manipulated.
In briefForeword: Using the Jacobian Lens to Predict and Characterise LLM Test Failures Before Generation
SummaryThe paper presents Foreword, a framework that predicts which inputs will cause LLM test failures before any output is generated, based on the model's internal state. It learns from both passing and failing reference inputs and represents inputs through an interpretability technique called the Jacobian lens. On ten tasks and three open-weight LLMs, Foreword outperforms the prior method Clotho in failure prediction and test prioritisation. A case study also shows it can point to ambiguities in labels or task specifications.
In briefSees the Hazard but Fails to Brake: A Perception-Action Binding Gap in Driving VLAs
SummaryThe paper finds that reasoning-augmented vision-language-action driving models encode in-path hazards in their visual representations, yet longitudinal plans are dominated by ego-speed history and respond only weakly to hazard proximity. Without retraining or changing weights, the authors inject a hazard-conditioned energy into the frozen flow-matching sampler of Alpamayo-1.5, reducing footprint intrusion from 58.1% to 14.5%. In their braking-relevant closed-loop tests, this eliminated collisions entirely.
WhyProposes a new perception-action binding gap diagnosis and a training-free energy-guidance fix, cutting footprint intrusion from 58.1% to 14.5% with closed-loop evidence.
Matteo Cercola, Katie Z Luo, Simone Formentin et al.
In briefCAPS: Anatomy-Prioritized MRI Synthesis for Filling Missing Scans in Brain Tumor Segmentation
SummaryThis paper proposes CAPS, a framework that synthesizes missing MRI modalities so brain tumor segmentation can proceed when a scan is absent. It uses a hybrid loss combining L1, an anatomical feature term, and a DualDice topology term to preserve deep anatomy and boundaries. On the BraTS Adult Glioma dataset it reaches 82.38% mean Dice for downstream segmentation and generalizes zero-shot to Meningioma, Pediatric, and Sub-Saharan African cohorts. A blinded radiological assessment reported improved perceived image quality and clinical utility.
Cecilia Diana-Albelda, Arthur Longuefosse, Mario Rub\'en Paz Campos et al.
In briefClustering Customers to Maximize Peak-Demand Coincidence Using Smart Meter Data
SummaryThis paper proposes a way to group electricity customers so that their peak demands coincide more strongly, which matters for distribution planning. The authors formulate the grouping as maximizing a weighted average of group coincidence factors and prove that jointly optimizing memberships and peak times is strongly NP-hard. They introduce a refinement procedure that alternates between identifying group peak times and reassigning customers, and test it on four public smart-meter panels with 27,654 customers. On held-out days, the method outperforms random grouping and five benchmarks.
Ramiro H. G\'alvez, Gabriel Martos Venturini, Iv\'an Mondrzak
In briefTEPCA: Aligning LLM reasoning with time-evolving knowledge graphs
SummaryThis paper addresses why large language models struggle with temporal reasoning, since their static memory does not track how knowledge changes over time. The authors propose TEPCA, which aligns evidence and parameters: time-constrained path pruning and mutual-information pruning select temporally coherent facts, and DG-LoRA adapts low-rank updates based on timestamps. The paper reports state-of-the-art results on two temporal knowledge graph benchmarks and further evaluates DG-LoRA on eight commonsense QA datasets.
In briefAfrica-Papers: A 10-billion-token multi-domain academic corpus for training African-focused language models
SummaryThis paper introduces Africa-Papers, a multi-domain pre-training corpus of over 844,807 full-text academic papers totaling about 10.15 billion tokens. It draws from African journals indexed in DOAJ, the CGIAR agricultural repository, African-related arXiv preprints, and OpenAlex publications by African-affiliated scholars. The papers are converted to layout-aware Markdown that keeps section structure, equations, and tables, and are organized under a standardized 8-domain taxonomy. The authors released the corpus openly on the Hugging Face Hub to support continual pre-training and domain evaluation of foundation models.
In briefDocLayout-YOLO: A Real-Time Detector for Analyzing Document Page Layouts
SummaryThe paper proposes DocLayout-YOLO, a real-time detector specialized for document layout analysis. It combines MeshLayout-1M, a large synthetic pretraining corpus built with a Mesh-candidate BestFit algorithm, and Page-Axis Layout Attention (PALA), which models long-range dependencies along page axes. The authors report state-of-the-art mAP50:95 scores of 72.4%, 75.5%, 96.0%, and 73.1% on D4LA, M6Doc, HJDataset, and PRIMA-LAD, with a better accuracy-efficiency trade-off than existing real-time detectors.
WhyCombines mesh-guided synthetic pretraining with page-axis attention for document-specific YOLO; strong results across four benchmarks.
In briefFrom Code Generator to Engineering Companion: A Plant-aware Agent for PLC Programming
SummaryThe paper presents a conversational copilot for PLC programming that is aware of a specific plant. It extends retrieval-augmented code generation with plant-specific context ingestion, a chat-first model for explanation and debugging, and graduated tool access from chat to deployment. The system targets real plant infrastructure and keeps the engineer in control of every step with real-world effects.
In briefPredicting silent failures of learned inference under restricted access, and what physics and priors explain
SummaryThis paper studies silent failures of learned priors, where reconstructed uncertainty intervals exclude the truth in poorly measured regions while all ground-truth-free diagnostics look normal. The authors propose a score based on a linearised VAE prior to predict such errors before target data are seen, and test it prospectively on real X-ray CT data against two learned inference systems. Under bounded measurement reach, the score gave no detectable improvement over a physics-based baseline, and the authors conclude that silent failure structure mainly follows from measurement access and the decision quantity rather than the learned prior.
In briefScalable stress-constrained topology optimization using machine learning, optimizing networks simultaneously
SummaryThis paper proposes SNTO, a framework that represents state, adjoint, and design variables with neural networks optimized jointly in one loop. To handle large problems, the design domain is split across multiple GPUs and sliced to limit peak memory. On 2D benchmarks, SNTO gives sharper boundaries and fewer gray areas than a finite element approach. For 3D problems with up to eight million collocation points on eight GPUs, training time falls nearly in proportion to GPU count.
In briefCurriculum Learning for Generative Flexible Docking
SummaryThis paper addresses flexible protein-ligand docking, predicting the joint complex structure from a protein apo structure and a ligand SMILES string. The authors propose curriculum learning that gradually incorporates large apo-to-holo conformational change examples during training, using a linear pacing function. On the PDBBind ESMFold apo benchmark, the resulting model GCLDock reports a 46.52% top-1 success rate (ligand RMSD under 2Å), and 40.26% when also requiring PoseBusters validity, outperforming existing baselines.
In briefProof that computing approximate correlated equilibria is as hard as finding Nash equilibria
SummaryThe paper shows that computing an approximate correlated equilibrium of a polytope game is PPAD-complete, making it polynomial-time equivalent to finding a Nash equilibrium. This challenges the view that correlated equilibria are easier to compute. The authors link correlated equilibria to a generalization of fixed points called fixed sets, and prove that computing a fixed set is PPAD-hard even at inverse polynomial precision.
WhyProves approximate correlated equilibria are PPAD-complete via a new fixed-set generalization, overturning the view that they are easier than Nash.
Ioannis Anagnostides, Constantinos Daskalakis, Gabriele Farina et al.
In briefNear-optimal regret bounds for UCB-style Gaussian process bandits using data thinning
SummaryThis paper studies regret upper bounds for Gaussian process upper confidence bound (GP-UCB) algorithms under a Gaussian noise assumption. It uses data thinning, a technique that creates independent datasets from a single dataset, to obtain tighter confidence intervals. A direct extension of GP-UCB then achieves a near-optimal simple regret bound, and combining it with a phased elimination algorithm yields a near-optimal cumulative regret bound. Numerical experiments are presented to demonstrate the approach.
In briefCausal discovery in linear models with hidden confounders using multi-domain identifiable representation learning
SummaryThis paper studies linear causal models with latent confounding, represented as acyclic directed mixed graphs (ADMGs). The proposed LinADMG framework uses multi-domain settings to identify noise variables, without relying on non-Gaussianity or bow-freeness assumptions. The authors show that identifying the noise variables is sufficient to recover all direct causal relations, and that under an additional latent-confounding no-cancellation condition, the whole ADMG can be identified. The methods are validated on simulated and semisynthetic datasets.
In briefSIGNN++: Robust Graph Neural Star Identification for Real Terrestrial Night Skies
SummaryThis paper presents SIGNN++, a learning-based star identification framework for celestial navigation in nighttime environments where GNSS is unavailable. It extends the earlier graph neural network method SIGNN with a more realistic training noise model, Fourier Transform of angular-distance edge features, and RANSAC-based Perspective Star Alignment. The authors report that SIGNN++ is more robust to terrestrial observation disturbances and achieves higher identification accuracy and solved rates than SIGNN and public baselines, with comparable attitude accuracy and practical computation time.
In briefOV-Foam: How 3D Representations Affect Lifting Semantic Features
SummaryExisting open-vocabulary segmentation methods assign features to 3D primitives through simple backprojection, but in 3D Gaussian splatting the primitives overlap heavily, which leads to suboptimal reconstructions. The paper proposes OV-Foam, which uses PowerFoam, a power-diagram representation that partitions the scene so primitives do not overlap and each ray passes through far fewer of them. This simplifies the lifting problem and allows near-optimal feature assignment. Experiments report improved open-vocabulary segmentation compared with Gaussian baselines.
Yousef Rajeh (1), Alejandra Esparza Pelcastre (2), Ivan Viola (1) et al.
In briefGAMF-Net: A Smartphone Skin Lesion Classifier That Combines Images with Patient Clinical Data
SummaryThis paper proposes GAMF-Net, a network that classifies skin lesions from smartphone photos combined with ten clinical variables grouped into demographics, history-related risk, and lesion presentation. Only the clinical groups with available information are fused with equal weights, and the model falls back to image-only prediction when no clinical data is present. Evaluated on PAD-UFES-20 with a patient-disjoint five-fold protocol, it outperforms the image-only and multimodal baselines reimplemented under the same setup, with the strongest baseline being MetaBlock.
In briefDe-Diffusing WAMs: Single-Step Robot Control That Predicts Only What Matters for Actions
SummaryThis paper introduces DirectWAM, which converts diffusion-pretrained cascade world-action models into single-step predictors without stochastic sampling. The future predictor is trained by decoding actions from its own predicted futures and backpropagating demonstration action error through the action expert. DirectWAM reports 97.1% average success on four LIBERO suites and 85.4% on RoboTwin 2.0, and reduces inference latency by 28.0× compared with the full iterative teacher.
In briefInterruption-aware recurrent reinforcement learning for drone navigation after temporary overrides
SummaryThis paper studies how recurrent reinforcement learning navigation policies for drones behave after a temporary override by a remote pilot or onboard safety system returns control to the agent. The authors find that interruptions make task information harder to decode from the recurrent hidden state and raise collision risk after handback. They propose three interruption-aware architectures, and in simulations the strongest one cuts post-interruption collisions by 74% and interruption-induced mission failures by 69% compared with the baseline, without hurting interruption-free performance.
Christopher Conrad, Deepak Panda, Weisi Guo et al.
In briefPointForward V2: Point-aligned interaction for feedforward 3D driving scene reconstruction
SummaryThis paper presents PointForward V2, a method for feedforward reconstruction of dynamic driving scenes into a 3D representation. It lets lifted 3D point queries exchange information through compact static and dynamic scene registers, combined with a 4D positional encoding that uses aligned spatial coordinates and timestamps. The approach avoids dense global attention while enabling spatio-temporal reasoning. The authors report state-of-the-art results on Waymo, nuScenes, and Argoverse 2.
In briefText-guided vision-language method for more reliable medical image segmentation
SummaryThis paper presents RELiF, a vision-language framework that uses medical text to help segment medical images where targets are visually ambiguous. Its Reliability-guided Fusion Gate adaptively controls cross-modal updates, and its Text-guided Skip Refinement Adapter refines high-resolution features using text. The method is evaluated on radiographic, CT, and endoscopic segmentation tasks, with additional analyses of efficiency and stability.
Sivan Talaei, Fatemeh Daneshfar, Abdulhady Abas Abdullah et al.
In briefOvisOCR3: A Compact 0.8B Model for Many OCR Tasks
SummaryThis paper introduces OvisOCR3, a 0.8B-parameter model that handles document parsing, general OCR, key information extraction, figure captioning, and cross-page table reconstruction in one model. It is trained with continued pretraining, domain-balanced mid-training, multi-teacher distillation from five 4B teachers, and mixed-task reinforcement learning. The authors compare it with general vision-language models and specialized OCR models on public and in-house benchmarks, reporting top results on several of them. The model is released on Hugging Face.
In briefNoise-Robust Spiking Neural Networks via Conditional Information Bottleneck
SummaryThis paper addresses the vulnerability of Spiking Neural Networks (SNNs) to input perturbations, hardware noise, and adversarial attacks. It proposes SCIB, a robust learning objective based on the Conditional Information Bottleneck, which replaces the standard compression term with conditional mutual information and uses a learnable class-conditional prior. Experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 report improved robustness under most evaluated noise conditions and higher adversarial accuracy without adversarial training, along with a 24.9% reduction in SynOps-based single-sample inference energy relative to an unconstrained SNN baseline.
In briefGradient clipping guarantees under weak heavy-tailed gradient noise
SummaryThis paper studies gradient clipping for stochastic optimization when gradient noise has heavy tails, using a weaker weak-L^α tail condition than standard moment bounds. It derives high-probability convergence guarantees for clipped SGD and an accelerated variant on smooth convex problems. For 1<α<2 the additive stochastic rate is K^{-(α-1)/α}, and matching lower bounds plus Pareto-noise experiments support the results.
Alyona Nekrasova, Makar Sidorov, Savelii Chezhegov et al.
In briefReading Too Much Into It? Interpretive Preferences and Commitment in LLM Mentalisation
SummaryThis paper introduces MentalLoom, an evaluation of how large language models interpret ambiguous social situations, using 604 episodes, matched depression and social anxiety disclosures, and four-turn dialogues with unchanged event evidence. It tests eight LLMs and finds that disclosures initially produced more negative explanations overall but more positive narrator-directed interpretations. Under scripted repeated questioning, models showed more emotional support, stronger positive endorsement of the narrator, and higher certainty, even though no new event evidence was added.
In briefA Semi-Analytical Framework for Inference in Structural Causal Models
SummaryThe paper proposes a method for probabilistic inference that runs message passing directly on Structural Causal Models (SCMs). Structural mechanisms and probability densities are represented in a shared finite-dimensional function space, with forward propagation based on the Law of the Unconscious Statistician and densities recovered by spline projection. Experiments on randomly generated continuous SCMs show close agreement with Monte Carlo reference solutions, improving as basis resolution increases.
In briefPolyscene: Automating Scenario Planning as Diversity Planning
SummaryThe paper automates scenario planning, a manual foresight practice, by treating scenarios as different narratives on an analyst's drivers. Its generic engine, Polyscene, produces a set of scenarios from a topic within a budget. Case studies cover geopolitical conflicts built from GDELT data and shipping-route disruptions.
Mustafa F. Abdelwahed, Joan Espasa, Felipe Meneguzzi
In briefBAIT: Can LLMs Spot Cognitive Bias Triggers in News? A Benchmark
SummaryThis paper introduces BAIT, a benchmark for detecting cognitive bias triggers, which are linguistic features that exploit readers' mental shortcuts, in news articles. It includes a nine-trigger schema and 50 English news articles annotated by trained annotators at article and span levels. Six frontier LLMs were evaluated, approaching human agreement on which triggers appear in an article but rarely recovering all of them and reaching under half of human agreement in locating trigger passages. The authors conclude the models respond mainly to surface cues rather than the mechanisms of manipulation.
In briefLatticeRank: A training method for 3D medical image segmentation that handles varying voxel spacing
SummaryThe paper addresses how 3D medical image segmentation networks cope with changes in voxel spacing when working directly on native grids. It proposes LatticeRank, a training-only framework that aligns candidate grid predictions in physical space and weights consistency supervision toward the coarsest grid, while inference uses a single forward pass. On MSD Task04 across five seeds, it improves coarse-grid Extra Dice by 0.03436 over mean consistency and meets the prespecified reference-grid noninferiority criterion.
In briefFrom Complex Histories to Simple Predictors: Predictive Dimension in Dynamic Graph Learning
SummaryThis paper argues that continuous-time graph learning need not reconstruct full interaction histories, since future predictions can be captured by a low-dimensional predictive state. It defines an intrinsic predictive dimension that measures the effective predictive degrees of freedom a task requires. Based on this, the authors propose PreDyG, a lightweight model whose capacity is matched to that dimension, and report superior predictive performance.
In briefWildAIGC-Bench: Testing AI Image Detectors on Real Online Fake and Edited Pictures
SummaryThis paper presents WildAIGC-Bench, a benchmark built by continually collecting AI-generated or edited images circulating online from news sites, fact-checking sources, and social media. It contains 407 human-verified images, each with a source link and evidence summary, and keeps real-world conditions such as compression, screenshots, and recapture. Tests of dedicated detectors, multimodal large language models, and a simple Perception Encoder baseline show substantial recall gaps compared with existing benchmarks.
WhyBuilds a benchmark of wild AI-generated and edited images, revealing substantial recall drops for existing detectors; methodological novelty is limited.
In briefMoDDiT: Token Routing to Cut Compute in Diffusion Transformers
SummaryThis paper proposes MoDDiT, a simple modification to Diffusion Transformers that learns to route image tokens through both MLP and self-attention blocks. This enables dynamic compute allocation across tokens, images, and denoising timesteps. The authors report that processing as few as 12.5% of tokens significantly reduces training and inference costs while improving generation performance over unrouted baselines. They also apply MoDDiT to video generation and to distilling a dense video model into a sparse autoregressive student.
Jarek Liesen, Matthew T. Jackson, Benjamin Ellis et al.
In briefInsightRAG: Combining Scholars' Judgments to Generate Scientific Surveys
SummaryAutomatic survey generation struggles with incomplete scope, weak taxonomies, and unreliable citations. InsightRAG uses citation relations and explanations from authors' Related Work sections to guide retrieval, reconstruct the research scope, and organize relational evidence. It aggregates these judgments under a minimum-conflict objective into a taxonomy that guides generation. Results on ten topics show gains in scope reconstruction, taxonomy quality, citation support, and reference reliability.
In briefIECG: A Weakly Supervised, Interpretable ECG Foundation Model
SummaryThis paper presents IECG, an electrocardiogram foundation model that performs diagnostic classification and abnormality localization together. It removes the need for manual abnormal-region annotations by generating heatmap pseudo-labels from clinical ECG knowledge and using them to guide attention via a KL-divergence loss. Pretraining combines this attention supervision with ECG-text contrastive alignment and report captioning. Experiments reportedly show state-of-the-art abnormal region localization with strong classification performance.
Huynh Dang Nguyen, Trong-Thang Pham, Anh-Thuan Do et al.
In briefPULSE: Opportunistic Learning for Personalized Wearable Models
SummaryWearable devices need personalized models, but data after deployment often differs from training data, and a single device has limited local data. PULSE is an event-driven framework that treats each encounter with a nearby device as an online choice: either no interaction or collaborating with a promising peer. It uses peer gradients only when they align with the local learning direction, and past interactions guide later peer selection. It reduces peer-assisted exchanges by 80-94% compared with opportunistic model averaging while preserving or improving personalized accuracy, and it runs on wearable hardware.
Mattia Giovanni Campana, Dariush Salami, H\"useyin Yi\u{g}itler et al.
In briefSlot-JEPA: Object-Structured Prediction Interfaces for Latent World Models
SummaryThis paper proposes Slot-JEPA, which splits dense image representations into stable semantic slots and time-varying spatial masks to form a reusable object-structured interface. Building on it, the Transition Memory Predictor stores visual transitions as compact masks and retrieves context using slot statistics, with Unbalanced Optimal Transport supervising mask prediction. The authors report that the features transfer across seven visual-understanding benchmarks and improve goal-conditioned planning on six planar environments and four robotic arm tasks.
In briefMuPACE: Learning to Set Muon's Step Size for LoRA Fine-Tuning
SummaryThis paper proposes MuPACE, a small recurrent policy that meta-learns only the global step size for the Muon optimizer during LoRA fine-tuning, leaving its update direction unchanged. The policy watches the training trajectory and makes bounded adjustments to an analytic schedule. It was meta-trained on models with up to 16 layers and reportedly transfers one-shot to models up to 64 layers deep.
In briefSame Judgment, Different Yes: Multilingual Judgment Profiles in Language Models
SummaryThe paper collects 1.1 million replies from eight models in six languages to up to 105 moral scenarios, comparing how different question forms and answer readings (token or sentence) affect the judgments obtained. Judgment profiles largely agree across tested readouts, except for the yes/no token after a negative question read under a preset mapping. The authors find models' profiles are often closest to their own profiles in other languages, and argue that scoring a yes/no token requires first establishing which proposition it affirms.
In briefQuantifying the Threat: Bayesian Probabilistic Heavy Metal Health Risk Model for Children in Ghana's Densu Basin
SummaryThis study assesses non-carcinogenic health risks from six metals (Fe, Mn, Ni, Pb, Cd, As) in 94 groundwater samples from Ghana's Densu River Basin. It uses a Bayesian probabilistic framework with uncertainty propagation and spatial interpolation to produce continuous risk maps. Approximately 31% of sampling sites have a probability above 0.5 of exceeding a hazard index of 1, with Pb, As, and Cd dominating cumulative risk. The authors suggest the approach can support groundwater monitoring and public health planning in data-limited regions.
In briefTrajEvolve: Reinforcement learning for LLMs using trajectory evolution and hindsight credit assignment
SummaryThis paper addresses difficulties in applying reinforcement learning to long-horizon tasks for large language models, where successful trajectories are scarce and sparse final rewards cause redundant steps. The authors propose TrajEvolve, which refines low-quality steps in historical trajectories to create evolutionary trajectories and recasts step-level credit assignment as a binary classification of whether each step is necessary. Step labels are obtained through counterfactual verification, and the authors report superior results on two datasets.
In briefAuditing Where Concept Bottleneck Model Accuracy on CUB Comes From
SummaryConcept Bottleneck Models aim to make predictions that humans can inspect and correct through concepts like "curved beak," but benchmark evaluation on CUB-200-2011 relies only on task and concept accuracy. This paper audits where CBM and Concept Embedding Model accuracy comes from, finding that CUB annotations are noisy and that much class information enters concept scores through non-local features. Experiments with local patches and occlusion show that locality makes grounding measurable but does not guarantee it, so CUB accuracy gaps should not be overinterpreted.
Sascha Saralajew, Eleonora Poeta, Eliana Pastor et al.
In briefUniversal Online Convex Optimization with Delayed Feedback: An Algorithm That Adapts Without Knowing Loss Curvature
SummaryThis paper studies online convex optimization when feedback arrives with delays. It presents a single algorithm that adapts simultaneously to general convex, strongly convex, and exp-concave losses without prior knowledge of their curvature class or parameters. Its regret bounds match the best known dependence on the time horizon and delays, up to doubly logarithmic factors. A new exp-concave base learner further improves the delay dependence by replacing the maximum delay with the maximum number of simultaneously missing observations.
Hao Qiu, Andrew Jacobsen, Emmanuel Esposito et al.
In briefSlots Are Not Speakers: A Training-Free Linker for Frozen Streaming Diarizers
SummaryThe paper proposes a causal, training-free linker that lets a frozen streaming speaker diarizer track more speakers than its fixed number of output slots. It uses only the host's slot posteriors and audio, splitting slots at speaker changes and matching pieces to previously seen speakers, so it changes only speaker confusion. On 1418 sessions from eight corpora, the 4-slot host's DER above its cap fell from 36.8% to 22.6%.
In briefEvaluating How Language Pretraining Shapes Visual Representations in Fine-Grained Classification
SummaryThe paper proposes a compact explainability-based framework for comparing vision encoders trained with classification-only versus caption-generation supervision on CUB-200. Saliency maps from attention rollout and gradient-weighted rollout are assessed with five faithfulness and localization metrics. Caption supervision yields representations more sensitive to removing salient regions, better aligned with annotated object parts, and less reliant on background, while classification supervision is slightly more robust to removing low-saliency regions.
In briefA digital twin that simulates carbon-fiber SMC strip deposition to generate labeled images for camera-based inline inspection.
SummaryThe paper addresses the shortage of large, accurately labeled image datasets for inline quality inspection of Sheet Molding Compound manufacturing. It builds a photorealistic digital twin that combines rigid-body deposition simulation with physically based polarization rendering, producing images with synchronized oriented boxes, point annotations, density maps, and polarization renderings. The authors report that detectors trained only on synthetic data reach mAP50 up to 0.91, and that the synthetic-to-real gap is similar in size to the gap between two real acquisition series.
In briefOn-Policy Transition Distillation: Using Layer-to-Layer Prediction Changes to Improve Teacher-Student Learning
SummaryThe paper examines whether changes in predictions between layers can supply extra supervision in on-policy distillation (OPD). A diagnostic study found that closer teacher-student agreement on intermediate transitions goes with higher answer correctness. The authors propose On-Policy Transition Distillation (OPTD), which aligns teacher and student layer-to-layer log-probability increments on student-generated rollouts. Across three backbones and eight benchmarks, OPTD achieves the best or tied-best overall performance in both dual- and three-domain distillation. In the three-domain setting it recovers a larger share of the teacher-student capability gap than OPD.}
In briefPoisoning-based backdoors can be reconstructed, unlike cryptographic authentication.
SummaryThe paper studies backdoors installed by poisoning training data labels. It argues that when a verifier enters training only through poisoned labels, a defender can reconstruct it from observed triggers, so no cryptographic-style asymmetry arises. Experiments on instruction-tuned large language models are reported to support this dichotomy.
WhyProves poisoning-installed backdoor verifiers are reconstructible by defenders, refuting cryptographic asymmetry; experiments are limited to instruction-tuned LLMs.
In briefUsing letter positional probabilities to measure how complex French words are
SummaryThis paper uses letter positional probabilities (LPPs) to compare the spelling of simple and complex French words. It draws on about 38,000 French words scored by reaction time in a lexical decision task, using the 5,000 slowest and 5,000 fastest as high- and low-complexity groups. The authors report that many LPPs are linked to word complexity and that word length alone gives 77.27% classification accuracy, rising to 83.95% with binary variables from the first six letter positions.
In briefWhen does training on synthetic data provably transfer to real data?
SummaryThis paper studies when a binary classifier trained on synthetic data transfers to a real data distribution, deriving finite-sample excess-error bounds. It also analyzes the case where real data lies outside the synthetic distribution's coverage. Experiments on 32,200 generated ImageNet counterparts show an adversarial classifier with 0.5% synthetic error but 99.25% real error, so the authors recommend validating models on real data.
Leroy Odunlami, Tobias Riedlinger, Hanno Gottschalk
In briefGeometry-aware scene graph generation for 360-degree panoramas
SummaryThis paper addresses panoptic scene graph generation for equirectangular panoramas, which suffer from distortion and a seam. The authors propose a two-stage spherical framework that fuses overlapping perspective predictions and adds spherical-area-aware relation reasoning. They also build a dataset of 340 high-resolution panoramas. Reported results include a 36.03% PQ gain from spherical fusion and a 52.99% mR@50 gain under SGDet.
In briefFrom RL Papers to Framework Code: How Seven Open-Source Training Frameworks Include and Review Methods
SummaryThe study traces commit histories of seven open-source reinforcement learning post-training frameworks and links them to 6,939 arXiv papers from January 2025 to June 2026. Of 1,598 papers proposing general RL training methods, 37 were included in at least one framework within 180 days, a number that stayed roughly steady across quarterly cohorts. Included papers more often trained policies of 30B parameters or more and modified existing loss components than comparable papers. Most methods no framework included received no public discussion in the frameworks' issues and pull requests, and the authors release a tracker linking these records.
WhyLinks seven RL frameworks' commit histories to 6,939 papers, quantifying inclusion patterns with large-scale, same-quarter comparisons; the analysis is novel and affects widely used training tools.
In briefA survey of reinforcement learning for accelerating and optimizing FPGA and chip hardware systems
SummaryThis survey reviews more than 100 studies that implement reinforcement learning on FPGA and integrated-circuit platforms. It organizes the work by RL algorithm family, abstraction level, hardware realization strategy, and application domain. The paper compares resource use, latency, throughput, and power consumption across studies. It also identifies recurring architectural patterns, practical limitations, and open research directions.
In briefScaling Continued Pre-Training for Embodied Agents Across Diverse Virtual Worlds
SummaryThis paper studies how continued pre-training scales for embodied agents in virtual worlds with different visuals, dynamics, controls, and goals. Across 120 configurations varying data volume, diversity, and model size, the authors find more data helps in-corpus performance, diversity aids transfer to related worlds, and larger models are more data-efficient but generalize worse to distant worlds. They also fit an empirical scaling law linking validation loss to these factors. The authors conclude that building generalist embodied agents requires training stages beyond continued pre-training.
In briefP³-VLM: A point-based 3D vision-language model that reduces reliance on bounding-box shortcuts
SummaryThe paper finds that current 3D vision-language grounding benchmarks rely heavily on bounding boxes as prompts, which leaks category information so a box-only "blind" language model performs well. Switching to single 3D point prompts drops that blind model's accuracy on ScanNet Nr3D from 36.4% to 18.0%. The authors propose P³-VLM, a detector-free model with three complementary pathways that works without bounding boxes and reports 52.1% accuracy on point-grounded captioning, along with gains in robustness and generalization.
Anna-Maria Halacheva, Jan-Nico Zaech, Sombit Dey et al.
In briefRANGER: A knowledge-graph retrieval agent trained with semantic rewards from an LLM
SummaryThis paper proposes a hybrid method in which a large language model acts as a mentor during training, giving a reinforcement learning agent dense, semantically informed rewards. The agent learns to navigate a knowledge graph to retrieve information relevant to a query. Experiments on open question answering datasets report that it outperforms a baseline using classical graph search and an RL agent trained with heuristic rewards.
Artem Mikhaylov, Pavel Konoplev, Nikolay Efanov et al.
In briefAdaptMIL: Domain adaptation for multiple instance learning in digital pathology
SummaryPathology foundation models use multiple instance learning (MIL) on whole-slide images, but their performance often drops when applied to data from a different hospital. The paper proposes AdaptMIL, a training strategy that integrates domain adaptation into MIL training. It uses a two-stage procedure that separates representation-level domain alignment from MIL aggregation learning. The authors report improved cross-hospital performance and better results than alternative adaptation methods across several datasets, tasks, models and objectives.
Ben Vardi, Dana Schonberger, Yuval Friedmann et al.
