In briefGeoPrior-Mamba: Using Language-Model Process Priors to Reconstruct Fine-Resolution CO2 Maps
SummaryThis paper presents GeoPrior-Mamba, a multi-directional Mamba model that reconstructs fine-resolution XCO2 fields from sparse satellite observations. A language model is used offline, before training, to organize relative knowledge about biospheric uptake, respiration, and emissions into prior tables, rather than predicting XCO2 directly. Using OCO-2 data from 2018 to 2020, the model reports an RMSE of 0.81 ppm and R2 of 0.93 on held-out observations, with a 48.2% RMSE reduction versus CAMS background interpolation and 3.1% versus Trans-XCO2.
In briefRLHND: Video Foundation Models as Physically Grounded Hand Trackers for Robot Learning
SummaryThis paper introduces RLHND, a model that jointly estimates hand pose and tactile information such as contact and force from monocular egocentric videos. It builds on the pre-trained Cosmos 3 video diffusion model and adds a separate tactile stream for dense surface contact and force. The authors report state-of-the-art results on pose, contact, and force estimation benchmarks. They also show its use for robot learning through retargeting and real-world robot experiments.
In briefIris-3B: A pixel-space diffusion study testing whether skipping latent compression helps
SummaryThis paper trains Iris-3B, a 3B-parameter pixel-space text-to-image transformer, and converts the latent FLUX.2 Klein 4B model to pixel space. Both are fine-tuned for monocular depth estimation and image restoration/super-resolution. The authors find no significant gain from the pixel-space generative prior: Iris-3B matches the latent model on depth, and neither pixel model beats the latent one on 4x DIV2K restoration. Iris-3B does show that pixel-space pretraining scales to 3B parameters with text-to-image quality matching Qwen-Image on OneIG at 1024², and the weights and code are released.
Hanqiu Li Cai (SperidLabs), Chema Garabito (SperidLabs)
In briefZero-shot Chinese character recognition using radical-level verification after global matching
SummaryThis paper addresses recognizing Chinese characters never seen during training. The proposed two-stage method first retrieves candidates with STG-CLIP, which adds tree-position and radical geometry priors to the character decomposition, then uses a Radical Verification Module to check radicals against matching regions in the input glyph. On the ICDAR2013 benchmark under the character-level zero-shot setting, it reports 83.06% top-1 accuracy with 2,755 seen classes.
In briefArctic Questions, Missing Answers: A Benchmark for LLM Abstention in Arctic Science
SummaryThe paper introduces ArcticQA, a dataset of 194 multiple-choice questions drawn from primary Arctic research, with automated checks of answer support and distractor contradiction. It also presents ArcticAbstain, a paired benchmark comparing conditions where the correct answer is present versus replaced by a distractor, with an explicit abstention option. Evaluating eight models, the authors find answer-present abstention rates from 0.0% to 63.0%, and abstention rising by 5.05 percentage points on average when the correct answer is removed. They conclude that abstention frequency and responsiveness to answer availability should be assessed jointly.
Benjamin Wilcox, Dawei Gao, Pradeeban Kathiravelu et al.
In briefLighTROcc: Lightweight 4D Occupancy Forecasting Using Instance-Centric 3D Gaussians
SummaryThis paper addresses forecasting future 3D occupancy from surround-view cameras for autonomous driving, noting that dense voxel and bird's-eye-view methods become costly and struggle to keep object identities consistent. The authors propose LighTROcc, which represents movable objects with compact learned queries, estimates their 3D centers via attention-guided lifting, and models each instance as anisotropic 3D Gaussians propagated with predicted displacements. Experiments on nuScenes and nuScenes-Occupancy show higher instance-level forecasting accuracy than the evaluated baselines while keeping strong voxel-level occupancy quality.
In briefTIRA: Zero-Shot Prediction of MSI and TMB Across Cancer Types Using Immune Patterns
SummaryThis paper proposes TIRA, a framework that uses spatial immune topology to refine frozen pathology foundation-model representations for predicting MSI-H and TMB-H biomarkers from histology. The model is trained only on TCGA-COAD+READ and evaluated zero-shot on CPTAC-COAD, TCGA-STAD, TCGA-UCEC, and CPTAC-UCEC. With UNI2, AUROC on TCGA-STAD improved from 0.633 to 0.766 for MSI and from 0.651 to 0.772 for TMB.
In briefMixture of Layers (MoL): A Method for Routing Visual Features Across Encoder Layers for Visual Reasoning
SummaryThe paper notes that most multimodal large language models use only the final or near-final layers of a vision encoder, which can miss small objects and fine details. It proposes Mixture of Layers (MoL), which predicts routing weights over intermediate encoder layers based on the text query and sparsely aggregates their features. Across 7 fine-grained visual reasoning tasks, MoL improves over baseline models, including +18.9% overall accuracy on V*.
In briefInscriptionOCR: A Dataset and Method for Reading Ancient Inscriptions
SummaryThis paper presents an end-to-end AI pipeline for reading ancient Brahmi-script inscriptions, covering image enhancement, character recognition (OCR), transliteration into Roman script, and neural machine translation of the Prakrit text into English. The pipeline is designed for low-quality photos taken directly from stone inscriptions. The authors also release two datasets: an OCR dataset of over 200,000 Brahmi character images across about 600 classes, and a Prakrit-English parallel corpus of over 2,000 sentence pairs.
Jaidev Sanjay Khalane, Akbar Ali, V. N. Prabhakar et al.
In briefControllable crowd generation via world-model planning
SummaryThis paper presents Ctrl-CWM, a multi-agent controllable crowd world model that combines crowd generation with run-time control for applications such as robot navigation and autonomous driving. It learns human motion dynamics from real-world pedestrian videos, then uses an actor to generate imagined crowd trajectories and a planner that combines a critic's scores with user costs to select actions. The authors report that Ctrl-CWM outperforms the state-of-the-art method on most crowd realism and collision metrics and adapts to user objectives introduced during simulation.
In briefAn extended deep energy method for simulating crack growth under combined heat and mechanical loading
SummaryThe paper presents an extended deep energy method that simulates crack growth when heat conduction and thermal stresses act together, keeping the crack as a sharp polyline. Two neural networks represent temperature and displacement, receiving the crack through a discontinuous embedding function, with displacement enriched near the tip by Williams expansion. Stress intensity factors are extracted by an interaction integral and compared with published and independent finite element results, including a 0.11% agreement on a thermal edge crack.
Han Zhang, Mehrisadat Makki Alamdari, Babak Shahbodagh et al.
In briefSkillCycle: Co-Evolving Agent Policies and Skill Banks
SummaryThis paper presents SkillCycle, a framework in which an agent's policy and its skill bank co-evolve through a loop of skill internalization and rule revision. The method alternates between policy learning with a fixed skill bank and rule revision with a frozen policy. On WebShop, a 3B model reaches a 74.74% success rate without inference-time skill inputs, a relative gain over the prior state of the art. Ablations on ALFWorld and WebShop show that continual skill revision outperforms a static skill bank.
In briefEvent-Aligned Visual Action Reasoning for World Action Models
SummaryWorld-Action Models use predicted future visual frames as an intermediate step to guide robot action generation. This paper proposes a framework that organizes visual and action prediction around task-critical interaction events, giving detailed reasoning to key state changes and coarser treatment to connecting transitions. It also adds an execution validity head that identifies the valid part of each predicted action sequence. Experiments report a 10.26 percentage point gain in DOMINO success rate over the baseline and competitive results on RoboTwin 2.0.
In briefRSI-Forge: Turning Research Papers into Environments for Self-Improving AI Agents
SummaryRSI-Forge is a multi-agent pipeline that converts published papers into executable tasks with automated evaluators for training and testing self-improving agents. The authors built 210 environments across 18 fields, 90 of which were reviewed by independent human domain experts, and reimplemented each paper's method to set a baseline score. In tests of four models over three successive attempts on 120 environments, at least one model improved after the first attempt in 84% of environments, and models outperformed the reproduced paper methods in 68 of them.
In briefCoMoE: Making MoE inference affordable on consumer GPUs
SummaryThis paper presents CoMoE, a communication-efficient inference system for Mixture-of-Experts models on consumer GPUs such as the RTX 5090. These GPUs lack P2P interconnects and have limited PCIe bandwidth, which creates communication bottlenecks. CoMoE uses host-centric routing, including host-backed token multicast for dispatch and fine-grained token-level aggregation for combine. The authors report up to 1.46x higher inference throughput on RTX 5090, approaching NVLink-capable A800 performance at 23.4% of the hardware cost.
In briefSpatial Latent Reasoning for Pointing-Gesture Object Grounding
SummaryThis paper proposes Spatial Latent Reasoning (SLR), a framework that guides pointing-gesture visual grounding through an ordered sequence of geometric and visual intermediate states. The first spatial ray state is supervised by fingertip position and pointing direction, followed by four states aligned with target-region features, using a parity pooling method to build visual targets. On EgoPoint-Ground, it improves mIoU over same-backbone supervised fine-tuning across three models, and on YouRefIt it reaches 77.6% precision at IoU 0.5, a numerical margin of 5.2 points over the reported state of the art under differing evaluation protocols.
In brieftrACT: A Real-Time Airborne Camera Trap for Autonomous Drones
SummaryThe paper presents trACT, a lightweight real-time aerial robotics framework for consumer drones that detects and verifies subtle ground motion in cluttered environments such as forests. It combines Temporal Max Pooling for encoding motion over a rolling window, self-supervised motion anomaly detection to separate target motion from wind and drone drift, and motion prediction with gimbal-stabilized optical zoom verification. The authors report field experiments in dense wildlife habitats and surveillance scenarios showing it bridges wide-area monitoring and precise target verification.
Oliver Bimber, Rakesh John Amala Arokia Nathan, Mohamed Youssef et al.
In briefEfficient Reasoning with Flow Language Models: How Continuous States Help Models Reach Answers in Fewer Steps
SummaryThis paper compares the reasoning efficiency of Flow Language Models (FLMs) with discrete diffusion language models under matched denoising steps. FLMs keep a continuous sequence representation throughout denoising and decode to discrete tokens only at the end, and the authors argue and show through intermediate-state interventions that this lets evidence for multiple candidates persist. On maze planning and Sudoku tasks, FLMs achieve higher sequence accuracy than discrete diffusion baselines with few denoising steps. On Maze15, FLM reaches 95% accuracy at 64 steps with 36.5% fewer parameters than MDLM.
In briefGenerative models whose training loop is shaped by evolving user tastes
SummaryThis paper studies generative models trained in self-consuming loops, where users select preferred outputs that then feed back into training, so model outputs and user preferences co-evolve. When training uses only user-curated synthetic data, initial biases are amplified and the system converges to one of several singleton equilibria, with the initially advantaged instance eventually dominating. Injecting reference data at a sufficiently large rate yields a unique globally attracting equilibrium. The authors also propose an efficient algorithm that selects a reference distribution and its mixing weight to steer outcomes toward desired attributes at lower data collection cost.
In briefBalancing Relevance and Diversity in LLM Retrieval
SummaryThis paper examines whether diversifying retrieved results helps retrieval-augmented generation, finding its effect depends mainly on redundancy in the candidate pool. The authors propose a query-adaptive rule that diversifies only when nearest-neighbor top-k results contain fewer distinct documents than the query requires, computed from existing embeddings. They also introduce RNG-Score, a geometric reranker.
In briefShared Geometry as a Rosetta Stone: Aligning Separately Trained Models Without Paired Data
SummaryThe paper argues that independently trained models for different modalities often share similar representation geometry. It proposes a Wasserstein Procrustes method with coarse geometric initialization that estimates a single orthogonal map to align two embedding sets without any paired examples. The method aligns consistently across datasets, modalities, and unimodal models, and standard geometric metrics predict when alignment is possible. With very few pairs it substantially outperforms existing methods, and the alignments enable text-to-image generation without paired data.
WhyShows unpaired cross-modal alignment via Wasserstein Procrustes, validated across datasets, modalities, and text-to-image generation, with solid empirical support.
Dominik Schnaus, Thomas Dag\`es, Daniel Cremers et al.
In briefA paper on the stability and diversity of networked self-consuming generative AI systems
SummaryThis paper studies how multiple generative AI models that consume each other's synthetic data form a networked self-consuming training loop. The authors model the models as nodes in a directed, weighted graph and analyze long-term behavior under retraining, establishing convergence conditions and characterizing fixed points. They also examine how each model's access to real data, cross-model data consumption, and the interaction graph's structure shape long-term stability and diversity.
In briefTiTok: An Audio-Visual LLM That Finds Multiple Event Segments in Videos
SummaryThis paper addresses audio-visual multi-segment grounding, where a model must locate several event segments in untrimmed videos for a query using both audio and visual evidence. The authors propose TiTok, which uses Time Token Interleaving and decoupled multi-segment rewards optimized with GDPO. They also introduce CountF1 to measure event-count miscalibration, and report state-of-the-art results of 65.7 mIoU and 0.58 CountF1.
In briefNoise, Denoise, Correct: Three-Step MCMC Posterior Sampling with Diffusion Priors
SummaryThis paper introduces diffusion waltz, an MCMC method for posterior sampling in inverse problems with nonlinear, non-differentiable forward models. It uses SDEdit-style noising and denoising as a proposal, corrected by Metropolis-Hastings to achieve exact posterior sampling. The authors also propose a gradient-free ensemble Kalman update that injects observations into the proposal. On a non-differentiable Navier-Stokes initial condition recovery task, the method outperforms existing baselines across different noise and nonlinearity levels.
So Takao, Gregory David Bellchambers, Luke Ye et al.
In briefLetting the Library Describe Itself: Applicability-Aware Method Selection for Formal Proving
SummaryThis paper addresses LLM-based formal provers, arguing that relevance alone does not show whether a mathematical method can be applied to a given theorem. It proposes self-advertisement, where a model first generates a problem-specific proposal for each candidate method describing its target, action, and required conditions, before ranking. The authors organize 82 reusable Putnam 2000–2014 methods as Method Contracts. Self-advertisement reaches 95.0% hit@5 on Putnam 2015–2025, versus 84.2% for the strongest reranker, and 91.7% on IMO ProofBench versus 88.3%.
In briefTutorLoop: Sensor-Driven Generative Feedback That Adjusts Student Learning Behavior
SummaryThis paper presents TutorLoop, a system that uses webcam-derived sensor signals to estimate students' real-time cognitive states and provide adaptive feedback. A deep reinforcement learning agent optimizes the feedback type across the whole learning process, and a large language model rewrites it into context-aware messages. In a user study with 187 participants, a model trained offline was applied to a new learning task without retraining. The results report less frequent but more effective interventions, with improved attention, reduced workload, higher engagement, and better learning outcomes.
In briefOne Frame, Full Heartbeat: ECG-free cardiac MRI synthesis that generates a full cardiac cycle from one frame
SummaryThe paper presents PhaseFlow, a generative framework that synthesizes full-cycle cardiac cine MRI without ECG signals. It estimates cardiac phase from a left-ventricular area curve derived from segmentation, and uses a rectified flow model to generate motion, decoded into a diffeomorphic displacement field that warps end-diastole images. On the ACDC benchmark, it reports the best LV volume curve R², SSIM and FID among the compared baselines.
In briefARCS: Precise Text-to-SQL Through Structured Disambiguation
SummaryThis paper addresses how ambiguous user questions cause errors in text-to-SQL systems. It proposes structured disambiguation, which resolves ambiguity through explicit, constrained interactions instead of free-form dialogue. The authors introduce ARCS, a benchmark of naturally occurring, unconstrained ambiguities over real databases, fully annotated with ambiguity points, interpretations, and SQL queries. Experiments show that gpt-6-sol reaches only 51% execution accuracy, and no open-source model exceeds 27%.
In briefQuantifying Volumetric Risk: Class-Aware Asymmetric Weighted Conformal Prediction for 3D Segmentation
SummaryThis paper addresses the lack of calibrated uncertainty in deterministic medical segmentation models such as MedSAM under distribution shift. It proposes CA-WCP, which combines density-ratio weighting for covariate shift with directional quantiles and class-specific asymmetry factors for volume bounds. Evaluated on 3D brain tumor segmentation (BraTS 2020) and a synthetic multi-organ CT benchmark, the method reports coverage consistent with the nominal 90% level and 8–14% narrower intervals than symmetric weighted conformal prediction.
In briefPersona Hierarchy Model: How Fine-Tuned LLMs Generalize Beyond Their Training Context
SummaryThis paper proposes the Persona Hierarchy Model, which holds that a shared default persona shapes model behavior across different contexts. Across 120 fine-tuned models, generalization narrowness correlated positively with the similarity between the training context's persona and the default persona. The authors also introduce persona-preserving regularization (PPR), which in reinforcement learning cut reward hacking from 42-55% to at most 0.2% while keeping accuracy gains.
In briefPPCAR-Net: a network that rebuilds 3D coronary arteries from a few X-ray angiogram views
SummaryPPCAR-Net directly reconstructs 3D coronary arteries from sparse X-ray angiographic views. It predicts a branch-structured representation of centrelines and vessel radii without explicit point matching, triangulation, or an intermediate volume. A coarse predictor is followed by projection-guided refiners that correct the estimates using the input views. On simulated angiographic data, it gives better connected arteries and strong centreline accuracy, especially for RCA, with coarse-to-fine inference taking 121 ms.
In briefScribbleEdit: A Benchmark for Editing Images with Scribbles Alone
SummaryThis paper introduces ScribbleEdit, a benchmark that tests whether image editing models can perform edits using only user scribbles. It finds that existing VLM- and LLM-based editing models struggle to capture the intent behind scribbles. The authors also provide an automated data pipeline, an evaluation protocol for intention alignment, and a soft-token baseline that outperforms standard image editing models on the benchmark.
In briefJudging Generated Customer Feedback Taxonomies as Whole Trees, Not Just Nodes
SummaryThe paper argues that AI-generated category trees for customer feedback can pass generic naming and structure checks yet be unusable in production. The authors build six taxonomies from two proprietary feedback corpora and find that every generated tree has at least 97.7% of leaf names restating an ancestor, with some records falling under multiple top-level categories. They propose whole-tree metrics, structural discriminators and team partitionability, and conclude that evaluation must measure the entire tree.
In briefA unified model that synthesizes missing 3D MRI contrasts using efficient plane-by-plane generation.
SummaryAcquiring all MRI contrasts takes time and is uncomfortable for patients, so this paper studies synthesizing missing contrasts from existing ones. The authors propose MPAD, a latent diffusion framework that uses a 3D autoencoder and 2D diffusion operations generated autoregressively across planes to keep volumes coherent. Compared with 3D latent diffusion baselines, it reduces training and inference FLOPs by 7x and 3x, and the authors report superior performance on multiple datasets.
In briefFrom Retrieval to Customer Context: Evaluating Frontier-Model Systems for Voice-of-Customer Analysis
SummaryThis paper proposes a customer context graph that links customer feedback, business operations, and analytical objects through typed relationships, so an agent can trace why customers complain, who is affected, what action followed, and whether it was resolved. The authors compare three systems on 9,432 public Cursor feedback records using 30 questions, scoring answer quality, analytical depth, and evidence quality. The graph-backed agent scored 0.961 overall versus 0.710 for the Deep Research Agent and 0.651 for Agentic RAG, with its largest gain in analytical depth.
In briefDUDA-Bench: Testing AI Agents on Multimodal Data-Driven Urban Diagnosis
SummaryThis paper introduces DUDA-Bench, a hierarchical, interactive benchmark that frames data-driven urban diagnosis as a multi-stage AI agent workflow. It includes 86 atomic tasks and 22 workflow tasks, built on multimodal data from 12 cities covering five urban problem types. Evaluations of seven backbone models and five agent systems show a substantial gap between isolated analytical skill and end-to-end diagnosis, with unresolved evidence gaps propagating across stages.
In briefExpert Coupling in MoE Pretraining: Cutting Communication Overhead by Placing and Shuffling Correlated Experts
SummaryThis paper studies the costly all-to-all communication in Mixture-of-Experts model training under expert parallelism. The authors find that routers learn correlated expert selections early in pretraining, and they use this to place often co-selected experts on the same GPU and to shuffle tokens toward the GPU holding their next-layer experts. In Megatron-LM experiments, the two methods reduce all-to-all time by 1.16 to 2.63 times and end-to-end step time by up to 1.41 times, without changing routing decisions or expert parameters.
In briefThe Confidence Game: How AI Agents May Strategically Misreport Confidence When Users Delegate Tasks
SummaryThis paper studies how an AI agent that seeks user engagement or revenue may strategically distort its reported confidence when a user decides whether to delegate a task. The authors model this as a repeated signaling game and place an LLM in the agent role, finding it claimed high confidence on 56% of tasks it was told it would probably fail. The distortion destroys 68% of the gains from delegation, according to the paper.
WhyFormalizes confidence reporting as a repeated signaling game and validates strategic distortion with LLM experiments; novel framing, but experiments are limited in scope.
Raghu Arghal, Saswati Sarkar, and Shirin Saeedi Bidokhti
In briefImmiscible Diffusion Policy: Label-free noise assignment to keep multimodal robot actions
SummaryDiffusion policies were expected to learn multimodal action distributions, but the authors find they often collapse to a single mode even with balanced data. They propose Immiscible Diffusion Policy, a label-free training-time add-on that assigns action-noise pairs to keep denoising paths relatively distinct, without changing the policy architecture or inference. Across five simulated and two real-world humanoid manipulation tasks, the method increases the share of non-dominant modalities by 6.0x to 14.6x on two-modality tasks and recovers modalities missing from vanilla rollouts on four-modality tasks.
In briefA camera-native stereo VR180 dataset of firsthand recordings with calibration and tools
SummaryThe authors present a stereo VR180 video and still dataset they captured themselves with two Blackmagic URSA Cine Immersive cameras, containing 1,211 samples. The data is released in camera-native RAW, native fisheye HEVC and half-equirectangular HEVC formats, with factory lens calibration, conversion tools and AI-generated annotations. They report that native fisheye coding needed a median of 38% more bitrate than half-equirectangular coding for equal viewport quality across 24 clips.
In briefVerifying contrail avoidance with satellite imagery
SummaryContrails are thin ice clouds left by aircraft that contribute substantially to aviation warming. The paper presents a small 8.4M-parameter diffusion model that detects contrails in satellite images, with a controlled study of which components matter. The model reaches 0.476 PR-AUC, above a DeepLabV3+ baseline at 0.414 and an adapted MedSegDiff at 0.119.
In briefFrom Chunks to Functional Evidence: Function-Aware Retrieval for EDA Documentation QA
SummaryThis paper addresses question answering over electronic design automation (EDA) documentation, where retrieval mismatches are a key bottleneck. It redesigns the RAG retrieval unit by grouping scattered artifacts into functional units linked to their source chunks, training an encoder to match queries to these units, and combining this with direct chunk retrieval before a unified reranker selects evidence. On the new EDADocEval-QA dataset it reports ROUGE-L gains of 37.1% over Chunk RAG and 55.6% over the strongest graph baseline, and a 30.0% gain over the strongest baseline on ORD-MMBench.
In briefTopoGraphRAG-Bench: Testing How Well Multimodal GraphRAG Reasons Over Document Layouts
SummaryThis paper introduces TopoGraphRAG-Bench, a benchmark of 2,024 questions over 201 long, visually rich documents, built from text, figure, and table evidence under three controlled reasoning structures. The authors evaluate text-only GraphRAG, page-level visual retrieval, and multimodal GraphRAG systems. Multimodal GraphRAG performs best overall but still fails when visual-textual evidence alignment or multi-unit composition is incomplete.
In briefGlobal exponential convergence of two-layer linear network training
SummaryThe paper proves global exponential convergence with an explicit rate for wide two-layer linear networks trained with smooth Polyak-Lojasiewicz predictor losses. For initial covariance Σ0 = σ²Id, the loss reaches the global minimum at a linear rate of at least 4σ²κ, where κ is the PL constant. It also covers stability under finite-width sampling, extension to deep linear ResNets, and convergence conditions for heavy-ball momentum.
In briefMultimodal LLMs Learn to Read Brain Signals: A Vision-Language Model for Multi-Task EEG Decoding
SummaryThe paper introduces BraVista, a framework that encodes multichannel EEG signals as structured images and performs multi-task learning using instruction-conditioned vision-language models. It continues post-training a general-domain VLM without a separate large-scale EEG pretraining stage. The model is evaluated on four datasets covering sleep staging, emotion recognition, cognitive workload classification, and abnormal EEG detection, showing strong performance. The authors report that the EEG-to-image representation choice is critical and that performance degrades gradually as noise increases.
Parastoo Azizeddin, Omid Sharafi, Maryam M. Shanechi
In briefLearning unknown constraints from demonstrations without unsafe data, using optimality and counterfactual regularization
SummaryThis paper studies how to infer unknown constraints from expert demonstrations, such as in robotic control. The authors propose Counterfactual KKT (CF-KKT), which uses a learned differentiable dynamics model to impose KKT-inspired optimality conditions on demonstrations and generates reward-improving counterfactual behaviors as synthetic infeasible data. In high-dimensional robotic control tasks, the approach learns neural constraint representations with better safety and data efficiency than state-of-the-art offline inverse constrained reinforcement learning baselines.
In briefRelevance Is Not Sufficiency: Closing Evidence Gaps in Long-Term Memory QA
SummaryLLM agents that span many sessions store past interactions in external memory and answer questions from a few retrieved records. The paper argues that relevance-ranked memories can jointly miss a needed fact, and recasts retrieval as building a sufficient memory set. It proposes Budgeted Flat Reconstruction (BFR), which uses Formal Concept Analysis to select a compact candidate set and then iteratively acquires more records within a budget. On LongMemEval-S, judged accuracy rises from 72.4% to 82.2%, with Turn Hit at 91.4%.
In briefOnlineQAT: On-policy distillation for compressing language models to ultra-low bit widths
SummaryThis paper introduces OnlineQAT, a two-stage method for recovering accuracy in large language models quantized below four bits. It first applies block-wise quantization-aware training to obtain a usable low-bit initialization, then performs on-policy distillation on responses generated by the quantized student, guided by a frozen full-precision teacher. On Qwen3-1.7B, it achieves average scores of 57.28 at W3A16 and 32.52 at W2A16, outperforming ReasoningQAT by 2.90 and 0.44 points respectively.
In briefDo image editors blur like real lenses? A rendered-ground-truth pilot audit
SummaryThis paper checks whether two general image editors, Gemini 3.1 Flash Image and GPT-image-2.5, add blur that follows thin-lens optics when asked to mimic a shallow aperture such as f/1.4. It compares their outputs with Blender Cycles renders that have true depth of field, using the same blur-width estimator for both. The editors' blur ratios across depth and aperture are far smaller than the rendered reference, and their aperture response stays attenuated.
In briefTileSkipper: Adaptive Tile Pruning for Faster 3D Gaussian Splatting
SummaryThis paper addresses tiled 3D Gaussian Splatting renderers that apply one scene-wide contribution cutoff, although different content tolerates truncation differently. TileSkipper selects a static per-Gaussian cutoff policy for a frozen checkpoint using calibration renders that measure pair savings and a distortion proxy, with no parameter updates. Across 13 scenes, a fixed policy gives dataset-macro speedups of 1.088x at standard resolution and 1.238x at 3840 pixels wide, with mean PSNR changes of -0.007 and -0.023 dB.
In briefData-free MoE compression via shared low-rank basis factorization
SummaryThis paper studies compressing Mixture-of-Experts language models, which have many parameters despite sparse routing. The authors analyze expert pruning, expert merging, and weight reconstruction, showing that pruning and merging can incur structural errors tied to routing and expert heterogeneity. They propose Shared Low-rank Basis Factorization (SLBF), a data-free reconstruction method with rank-k bases shared across experts, which outperforms all three method families on five architectures from 16B to 122B parameters.
In briefBenign Overfitting When Fusing Heterogeneous Inputs
SummaryThis paper studies benign overfitting in minimum-norm linear interpolation under a heterogeneous Gaussian design, comparing two statistically dependent input blocks with their fusion. For regression, the authors identify a full-spectrum covariance certificate that is preserved by every positive-semidefinite joint covariance consistent with the two marginals, though this does not hold for all benign regression problems. For one-sparse Gaussian classification, benignity depends on the balance between surviving predictive signal and nuisance contamination, which fusion can shift in opposite directions.
In briefSearchWorld: World-Model Imagination with Spatial Value Guidance for UAV Object Search
SummaryThis paper presents SearchWorld, a recurrent state-space world model for UAV object search in urban environments. It combines explicit bird's-eye-view spatial memory with value-guided imagined rollouts, decoding a task-aware spatial value layer to guide search without a separate scalar critic. On the UAV-ON benchmark, it reaches a 23.8% success rate (versus 19.5% for the strongest published agent) and 35.5% oracle success, with 19.9% success on unseen scenes.
In briefVM-ArrayDPS: Adding Virtual Microphones to Improve Unsupervised Blind Speech Separation
SummaryThis paper addresses blind source separation, which aims to split mixed audio from several speakers into individual signals without prior knowledge of the sources or mixing process. It extends ArrayDPS, a diffusion-based method using a pretrained speech model, by adding higher-SNR virtual microphones that provide extra multi-channel consistency constraints. Experiments on 2-speaker and 3-speaker datasets show VM-ArrayDPS significantly outperforms ArrayDPS. Ablation studies examine the effect of the number of virtual microphones and the weight of the consistency objective.
In briefA new framework for open-task infrared-visible image fusion, CRT-HMAR
SummaryThis paper proposes CRT-HMAR, a framework for fusing infrared and visible images so that the result suits many downstream vision tasks. Existing methods are usually designed for a fixed set of tasks and generalize poorly to unseen ones. CRT-HMAR traces task-specific requirements to image-level maps and uses hierarchical multi-agent regulation to balance tasks and mitigate gradient conflicts. Experiments on open-task scenarios with five downstream tasks report improved generalization to unseen tasks while maintaining performance on seen tasks.
In briefGRC-Net: A Network for Unsupervised Multimodal Anomaly Detection
SummaryThis paper proposes GRC-Net, a method for detecting anomalies in manufactured products using both images and 3D information. It combines a global-attention MLP, which enforces consistency across patch embeddings, with a stable reconstruction module that reduces reconstruction noise. Experiments on MVTec 3D-AD and Eyecandies report that GRC-Net outperforms existing methods at both image and pixel levels.
In briefVisual Jev Rewards: Reference-Bound Checking for Multi-Subject Image Generation
SummaryThis paper proposes reference-bound Visual Jev rewards, which check whether requested attributes, actions, and relations hold for the specified reference subjects in multi-subject image generation. A Qwen3.5-4B verifier is trained with binary supervision on fixed questions, and the mean of its Yes probabilities serves as a GRPO reward. On a 897-task MICo-Bench subset, the GPT-5.4 composite score rises from 41.78 to 52.50, though each reward was tested in only one GRPO run.
In briefVIS-Ground: Interactive Video Storytelling with Context-Based Grounding
SummaryThis paper addresses interactive video generation, where viewer requests depend on both the source material and the video already rendered. It proposes VIS-Ground, which converts heterogeneous inputs into executable generation constraints through structured context abstraction, constraint induction, and constrained video generation. Across three video generation backbones, VIS-Ground achieves the highest composite score, with an average absolute improvement of 10.3 points over the strongest baselines.
In briefConstrained diffusion to expand scarce orbital data for constellation tasking simulations
SummaryThis paper studies how to augment scarce reference orbit trajectories with a constrained diffusion model, so that Monte Carlo evaluations of constellation tasking policies have more geometric diversity. The model learns a 13-dimensional orbital prior and propagates samples with Basilisk under five force-model tiers. Diffusion matches baseline distributional fidelity but samples farther from training priors and fails to reproduce the shifted population. The authors conclude it supports only local, in-support augmentation and cannot replace orbital dynamics.
Omar Ramadan, Sam Siavoshian, Amir Kashif Saeed et al.
In briefSpeech language models robust to dialects using synthetic pseudo-dialect data
SummarySpeech language models often perform worse on dialects because dialect speech data is scarce. The authors generate dialect text with an LLM and convert it to speech with a standard-language TTS model, so no real dialect speech is needed. They also add intermediate standard-text prediction during training to normalize meaning for downstream tasks. Evaluated on dialect-to-English speech translation for Japanese, German, and Chinese, the pseudo-dialect augmentation improves Japanese and German scores, and the intermediate prediction further boosts Japanese and Chinese results.
In briefSAREO-FM: a foundation model that separates semantic and reconstruction supervision for SAR and optical imagery
SummaryThis paper presents SAREO-FM, a foundation model for combining synthetic aperture radar (SAR) and electro-optical (EO) imagery. It uses separate token streams: modality tokens are trained by masked reconstruction to preserve how each sensor observes the scene, while learnable semantic queries learn scene content under guidance from a pretrained vision foundation model. The model is pretrained on the million-scale SAR-1M corpus. According to the paper, it transfers well to SAR-only and EO-only inputs and gains substantially from joint SAR-EO observations on tasks that benefit from complementary sensing.
In briefBeyond Type-Checking: A Holistic Evaluation Framework for Formal Specification Generation
SummaryThis paper studies specification generation (SpecGen), a step in pipelines that produce machine-checked code from natural language requirements. It notes that theorem provers can verify proofs but offer no definitive check that a generated specification matches user intent. The authors build a unified dataset of 350 Lean tasks and a framework covering formal validity, reference similarity, and behavioural adequacy. They report that metric coverage changes configuration rankings, and that a specification with perfect postcondition scores can still accept 0% of required inputs, motivating separate feedback on input and output constraints.
In briefWhy vision-language models miss small objects, and when zooming in is safe
SummaryThe paper studies why vision-language models often miss small objects in large images, using a theory based on the visual tokens across an object's side and the content each call must cover. It derives limits on whole-image processing and a recall-cost trade-off for any search strategy, including zoom agents. Tested on about 177,000 requests over 797 images, the predicted orderings held on controlled images, and decomposition rules did not significantly lower recall on construction drawings.
WhyProposes a provable theory of small-object limits and decomposition, with bounds validated across about 177,000 requests on 797 images.
In briefDenoising Blocks Instead of Tokens: Faster Long-Text Generation via Compressed Continuous Diffusion
SummaryThis paper introduces Branching Latent Diffusion (BLD), which compresses a 1024-token sequence into 64 block latents and decodes each latent with parallel local autoregressive branches. Latents are generated in groups, each conditioned on previously generated ones. Relative to the similarly sized ELF-L baseline, BLD reports over 80x fewer generation FLOPs and over 6x higher throughput. It also reports over 6x higher throughput and over 4x lower latency than an autoregressive baseline, while keeping competitive local fluency and diversity, though long-range coherence remains challenging.
In briefMovieSTAGE: A Framework for ADHD Classification from Movie-Watching fMRI Scenes and Transitions
SummaryThe paper introduces MovieSTAGE, a framework that classifies ADHD from naturalistic movie-fMRI data. It combines hypergraph-structured functional connectivity within scenes, unsigned connectivity differences between adjacent scenes, and whole-movie connectivity. Tested on 260 participants from the CMI-HBN Despicable Me cohort, it achieved AUROCs of 0.69, 0.73, and 0.75 across three classification tasks, with the highest mean point estimates among compared methods.
In briefKuration SDK: Narrowing the Virtual2Real Gap through Data Curation
SummaryThe paper finds that diffusion world models trained on CounterStrike gameplay data can feel very different to play despite similar FVD, LPIPS, and JEDi scores, a mismatch the authors call the Virtual2Real gap. They argue that curating raw gameplay data and measuring diagnostic properties before training gives a more robust signal, and they release Kuration SDK, a general-purpose data curation toolkit for physical AI. The SDK helped identify the root cause of a case where two models trained on identical data behaved very differently.
In briefemg2face: Facial Animation Driven by High-Density Surface EMG Signals
SummaryThis paper proposes high-density surface electromyography (HD-sEMG) as a non-optical way to capture facial movements, useful when the face is hidden by devices such as VR headsets. The researchers recorded 64 EMG channels from two textile grids and synchronized them with video using a novel audio-burst method. A deep neural network was trained to predict facial expression blendshapes from EMG signals alone, enabling real-time animation. Recordings from 25 participants were used to demonstrate the approach.
Ganidhu Abey, Wendy Greening, Ashika Kamboj et al.
In briefNode-Level Graph Neural Architecture Search (N-GNAS) Framework
SummaryThis paper proposes N-GNAS, a node-level graph neural architecture search method that automatically selects a suitable network architecture for each subset of nodes when updating node features. It addresses limitations of traditional GNNs, which apply uniform convolutions to all nodes and can suffer from over-smoothing as layers increase. The method also adds a contrastive learning loss to separate features of different categories. Experiments on eight node and graph classification datasets reportedly outperform leading GNN architecture search methods and human-designed GNNs, reaching 78.26% accuracy on CiteSeer.
In briefThe AI Evaluation Ecosystem: How Actor Dynamics Shape Benchmark Validity
SummaryThe paper argues that valid AI benchmarks must be designed with attention to how model providers, users, funders, and regulators interact within the evaluation ecosystem. The authors build a simulation combining rule-based market dynamics with LLM-driven strategic actors, modeling benchmarks and capabilities across six dimensions. In a case study, moving from public to private holdout benchmarks narrowed the gap between benchmark scores and user satisfaction on most benchmarks but widened it on a few.
In briefRT-Safe: Testing AI Agent Safety in Real-Time Physical Settings
SummaryThis paper introduces RT-SAFE, a simulated urban benchmark that tests how safely embodied AI agents perform navigation tasks when the world keeps changing during reasoning and action. Across eight vision-language models, agents achieved high task completion but rarely finished safely, with only 0.7% of episodes free of safety events in the hardest setting. Real-time execution increased collisions by 12.3 times compared with static evaluation. The authors also show that RT-SAFE can support offline reinforcement learning that substantially reduces collision rates.
In briefminiTE: A Framework That Closes the Execution Gap for Low-Bit NVFP4 Reinforcement Learning in LLMs
SummaryThe paper finds that matching model weights and numerical formats between separate training and sampling engines does not ensure consistent policy evaluation, so cross-engine numerical execution matters. It introduces miniTE, a lightweight framework combining execution diagnostics, cross-engine numerical alignment, and efficient native W4A4 GEMM computation. Experiments on Qwen3-30B-A3B show stable long-horizon RL training with benchmark performance close to BF16, and higher throughput than BF16 and TE NVFP4 on a single NVIDIA B300 node.
In briefFaceKit: an interpretable toolkit for quantifying facial traits, generating synthetic images, and testing privacy in rare diseases
SummaryThe paper presents FaceKit, a framework that measures facial features quantitatively from frontal photographs. It extracts landmarks and 120 morphological features, reporting z-scores relative to reference distributions built from the FairFace dataset. The authors evaluated it on 50 rare-disease cohorts from the GestaltMatcher Database and also assessed whether synthetic face images could reveal identifiable patient information.
Hongzhuo Chen, Zhanliang Wang, Florent Pollet et al.
In briefLeCuration: A Tiny World Model for Organizing Robot and Game Data
SummaryThis paper proposes a data curation approach for physical AI, training LeCuration, a small world model that helps organize data for a larger downstream model. It combines the LeWorldModel encoder and predictor with a diffusion transformer decoder to generate visuals from autoregressive gameplay rollouts. In a qualitative proof-of-concept on CS:GO gameplay data, the embeddings were found useful for anomaly detection and content-based clustering, and for checking action-state consistency. The authors do not yet report quantitative curation metrics or downstream training results.
In briefSelf-attention summary networks for velocity-model building from common-image gathers
SummaryThis paper proposes a multiscale self-attention network that compresses 3D common-image gather volumes into compact embeddings for probabilistic subsurface velocity inversion. A flow-matching model, conditioned on these embeddings, learns to transport a Gaussian distribution to the posterior of plausible velocity fields. Numerical experiments show improved posterior velocity inference compared with conditioning on raw gathers, with greater robustness to background-model mismatch.
Shiqin Zeng, Yunlin Zeng, Abhinav Prakash Gahlot et al.
In briefTwist Flow for Inverse Problems: A Generative Method for Better Posterior Sampling
SummaryThis paper addresses posterior sampling in Bayesian inverse problems and proposes joint twist-flow, an augmented flow-matching method that learns continuous transport from an augmented source state to an augmented terminal state. An auxiliary coordinate tied to observation consistency is added to reduce likelihood-inconsistent variation while preserving variability in weakly constrained directions. The method is validated on low-dimensional problems, image restoration, and seismic velocity-model inversion, where it better preserves multimodal posterior support than a direct conditional-flow baseline and increases posterior variability while maintaining observation consistency.
In briefCATune: Bayesian tuning of database settings that respects built-in rules between settings
SummaryModern database systems have hundreds of configuration settings, making automatic tuning expensive. CATune treats the ordering rules between settings, as stated in the database documentation, as part of the search space and optimizes only within rule-consistent options. It uses an LLM-based pipeline to extract these rules. Experiments on PostgreSQL and MySQL with TPC-C and TPC-H workloads report better sample efficiency and final tuning quality.
In briefEM-SNN: An Efficiently Modulated Spiking Network for Remote Sensing Image Dehazing
SummaryThis paper proposes EM-SNN, a spiking neural network designed for dehazing remote sensing images. It addresses the problem that haze-induced high-frequency attenuation interacts with discrete spike thresholding, which weakens recovery of edges and textures. The model combines a statistics-driven Threshold-Modulated LIF neuron with a Spike Sobel Modulation module. Experiments on HRSD, RICE, RRSHID, and SateHaze1K show competitive dehazing performance using one quarter of the energy of the ANN baseline SFRDP-Net.
In briefHardware-aware calibrated clustered attention to speed up 3D reconstruction transformers
SummaryThis paper addresses the latency bottleneck of global attention layers in VGGT, a transformer that jointly infers camera poses, depths, and dense geometry for 3D scene reconstruction. The authors propose blockwise clustered attention (BC attention), which limits clustering to hardware-friendly neighborhood blocks, and add hashing hyperplane calibration and threshold-based error compensation to reduce clustering errors. On GPUs, the calibrated method reportedly speeds up global attention layers by 2.10-2.63x and the whole backbone by 1.77-2.35x, with about 1% loss for large scenes.
Weitian Wang, Shubham Rai, Cecilia De La Parra et al.
In briefTesting whether trajectory features improve routing of final-layer attention
SummaryThe paper asks whether hidden-state extrapolation error, curvature, and error change can predict the value of attention better than simpler signals for deciding which queries to execute. Experiments use frozen SmolLM3-3B-Base and Qwen3.5-4B-Base checkpoints on 100 held-out PG-19 books under a 20 percent invocation quota. After familywise correction, none of six prespecified comparisons shows a positive gain, and a parameter-matched fixed-projection control performs better than the trajectory router.
In briefThe Symbol of the Surrogate: Checking Whether Neural PDE Solvers Copy Numerical Errors
SummaryNeural PDE surrogates are trained on numerical solver outputs that mix true physical evolution with solver-specific discretization errors. The paper proposes a Fourier-symbol diagnostic that probes a trained model's linearized one-step operator with individual Fourier modes and compares it with exact-evolution and training-scheme references. In linear advection, the learned surrogates reproduce the training schemes' amplitude and phase errors. The authors conclude that agreement with solver-generated test data does not establish fidelity to the exact evolution, and that Fourier-symbol measurements provide a direct diagnostic of numerical provenance.
In briefPhyDiCT: Training-free CT reconstruction from sparse X-rays using physics and diffusion priors
SummaryThe paper proposes PhyDiCT, a training-free framework that reconstructs 3D lung CT images from a few X-ray projections. It combines a differentiable physics-based forward model based on the Beer-Lambert law with a text-conditioned diffusion model used as a prior, guided by Split Gibbs sampling. A test-time refinement step is also added. The authors report a 7.5% SSIM improvement over fully trained methods.
In briefAdaptive visual token reduction to speed up image understanding in vision-language models
SummaryLarge vision-language models need heavy computation to process high-resolution images. This paper proposes ReFIT, a framework that reduces visual tokens based on the instruction, using relevance-guided window reshaping and instruction-guided token refinement. Experiments on four VQA benchmarks report higher answer accuracy with lower computational cost.
In briefEfficient evaluation of Best-of-N sampling policies for aligning model outputs at inference time
SummaryThis paper studies Best-of-N (BoN), an inference-time method that picks the highest-scoring answer among N samples from a reference model. It proposes a framework that evaluates and selects BoN policies using only samples, without needing response likelihoods, and introduces a doubly robust estimator called BoN-DR. The authors also give two selection rules and test them on synthetic data and GSM8K.
Jonas Schweisthal, Yuxin Wang, Athiya Deviyani et al.
In briefAn Informational Curse of Horizon in Goal-Conditioned Robot Policy Learning
SummaryThis paper identifies an informational "curse of horizon" in goal-conditioned policy learning, in which longer goal relabeling horizons reduce policy generalization and performance beyond known bias from temporal-difference backups. Controlled experiments with oracle planners show that behavioral cloning policies degrade with training horizon even when tested on nearby subgoals, and that reinforcement learning objectives mitigate this effect. The authors attribute the phenomenon to lower conditional mutual information between actions and hindsight-relabeled goals, and report that distilling input Jacobians from short-horizon policies into long-horizon ones yields significant gains, especially on combinatorial manipulation tasks.
In briefRisk-averse multi-population mean-field games beyond nominal equilibria
SummaryThe paper introduces risk-averse multi-population mean-field games, in which each population maximizes its worst-case expected reward over dynamically feasible ambiguity sets of other populations' mean-field flows. Using occupation measures and set-valued analysis, it establishes geometric properties of the ambiguity sets and the existence of a novel risk-averse equilibrium. Under entropy regularization, the fixed-point operator is shown to be contractive, and a risk-averse fictitious-play scheme is proposed whose exploitability decays to zero. Numerical experiments illustrate convergence and risk-averse behavior.
In briefWe Query, Therefore We Compute: An abstract machine that treats an LLM as an oracle, unifying agents and workflows
SummaryThe paper builds an abstract machine that treats a large language model as an Oracle, extended with an instruction that passes a whole stack as a query and appends the answer back to it. It uses this model to describe both agents and workflows within one framework and to analyse caching and scheduling under stated conditions. The authors also propose ArchNights, an extended RISC-V system that runs on gem5 and with an LLM as the Oracle, which they say will be open source.
In briefPooling Representation Autoencoders for Efficient Diffusion: Merging Neighboring Visual Tokens to Cut Image Generation Cost
SummaryThe paper introduces PoolDINO, a learned affine pooling operator that merges neighboring tokens to reduce the cost of Representation Autoencoders (RAEs), which generate images from pre-trained visual features. The operator is trained jointly with the RGB decoder, keeping the standard two-stage RAE procedure without a separate feature autoencoder. On ImageNet-256, 4x compression keeps generation quality comparable under internal guidance, while 16x compression trades some quality for efficiency, raising latent-sampling throughput by 3.7x and 9.0x at 100 sampling steps.
In briefWinner's curse in LLM self-improvement: selection noise and evaluation reuse
SummaryThe paper studies self-improving LLM loops that keep self-proposed changes scoring better on a small evaluation set. The authors treat this as selection under measurement noise, and find that reusing the selection set overstates gains, with greedy loops exceeding held-out accuracy by large margins on small sets. They conclude that self-improvement studies should report held-out gains with their uncertainty.
In briefA method for studying language agent behavior through trajectory abstraction
SummaryThis paper proposes a research procedure for studying language agent behavior by recursively learning and testing a hierarchy of trajectory abstractions. It first measures role- and phase-indexed events, proposes temporally constrained relations and checks their stability, then builds episode-level motif variables from selected relations. The authors give a finite-depth bound, a finite-sample test for intervention consistency, and constructed examples illustrating the approach.
In briefUsing Data Symmetries to Tighten Causal Effect Bounds
SummaryThe paper proposes using known data symmetries, meaning invariance of a causal effect under certain data transformations, as new constraints for partial identification of causal effects. These are implemented as a shape constraint on the causal function and as a change of measure via simple data pre-processing. The authors show the approaches sharpen bounds under two canonical partial identification models, with theoretical results for the population case and experiments for finite samples.}
Uzair Akbar, Zulfiqar Zaidi, Niki Kilbertus et al.
In briefDiagnosing LLM judges under different deployment conditions
SummaryThe paper proposes Conditional Accuracy Profiling (CAP), a post-hoc framework that breaks pairwise LLM-judge accuracy into eight conditions grouped under content sensitivity, robustness, and rationale quality. The authors apply CAP to seven LLM judges across six benchmarks, including a controlled testbed they created, judgerEva-Standard. They report that aggregate accuracy hides differences between judges, such as sensitivity to omitted qualifications.
In briefQuantizing by Drift: Label-Free Mixed-Precision Compression for Text Embedding Models
SummaryThis paper proposes a label-free method for mixed-precision post-training quantization of text embedders, using quantization-induced representation drift as the sensitivity signal. Each module is quantized, the corpus is re-encoded, and the drift of output embeddings from full-precision positions is measured, with precision allocated under a hard packed-byte budget. Across five development embedders, configuration-level drift reached a macro Spearman of 0.911 against held-out retrieval quality. On three held-out embedders, drift-based allocation had numerically lower drift than same-budget uniform precision at the main budget, and the authors conclude it is a robust coarse sensitivity signal rather than a universally optimal allocation objective.
In briefFLoRa: Drone-assisted data collection from battery-saving LoRa sensors
SummaryThis paper proposes FLoRa, a system where a drone collects data from LoRa IoT devices that sleep on a duty cycle to save battery. It combines simulated annealing for route planning, CMA-ES for choosing hover positions, and partially observable Markov decision processes to decide how to probe each device. The authors introduce a Value of Information metric that rewards fresh data and penalizes failed probes. The paper reports that FLoRa outperforms metaheuristic, greedy, and deep reinforcement learning baselines in expected VIP, node coverage, and successful collections.
In briefPVSync: A unified model that checks lip-sync timing and mouth shapes in talking-head videos
SummaryPVSync is a model that estimates audio-video timing offsets and scores how well mouth movements match individual speech sounds. It combines window-level contrastive learning for synchronisation with an objective that aligns audio and video of the same viseme class. Viseme labels are generated automatically from forced-aligned transcripts. On videos from 13 generation models, its rankings correlate with human judgments more closely than LSE-C does.
In briefData-free diffusion distillation via consistent distribution matching for faster generation
SummaryFlow and diffusion models generate images slowly because they rely on costly numerical integration. This paper proposes Consistent Distribution Matching, a simulation-free, data-free distillation method that uses one student network for both sample generation and score estimation, with only a frozen teacher and a trainable student. The authors prove that minimizing their objective implies Wasserstein convergence of the student flow-map pushforwards to the teacher marginals. On ImageNet 256×256, they report an FID of 2.04 with one function evaluation and 1.37 with four, within 40 training epochs.
WhyProposes data-free distillation with one student and a frozen teacher, proves Wasserstein convergence, and reports strong ImageNet few-step FIDs.
In briefCM-DPO: Constraint-Margin Direct Preference Optimization for Training LLM Planners
SummaryThis paper proposes CM-DPO, a method for training large language models to plan. It replaces DPO's binary preference signal with a continuous margin from a deterministic symbolic verifier, scaled by violation severity, and separates hard and soft constraints via a lexicographic objective. On TravelPlanner, NaturalPlan, and out-of-distribution PlanBench, an 8B model fine-tuned with CM-DPO reports 89.2% pass rate and 93.4% solve rate.
Rabimba Karanjai, Qun Gu, Hemanth Hegadehalli Madhavarao et al.
In briefRLDiscover: LLM-Driven Co-Evolution of Reinforcement Learning Algorithms
SummaryThe paper presents RLDiscover, a framework that uses LLM-guided program evolution to self-evolve model-free deep RL algorithms such as SAC, PPO, and DQN. It combines Progressive Co-Evolution, moving from targeted component edits to joint evolution, with Progressive Probabilistic Evaluation using staged training and repeated evaluation. Experiments on four benchmark suites report per-family median mean-return gains of 32%-84%, and independent searches repeatedly find interpretable combinations of adaptive robust losses, progress-dependent value targets, and running statistics.
In briefA deterministic evidence layer for vision-language autism screening from home video
SummaryThe paper addresses unstable verdicts from vision-language models when screening for autism from home videos of preschool children. The VLM is kept frozen and produces a timestamped event table, while a deterministic weight-of-evidence scorer computes the final risk category, making decisions traceable and reproducible. On 43 clips, the pipeline reaches AUC 0.851 ± 0.012 and 86.0% accuracy, and labels 74.4% of clips correctly in every run.
In briefAGAR: a reinforcement learning framework for evolving programs with LLMs
SummaryThe paper formalizes LLM-driven program evolution, where a model rewrites candidate programs and a search loop selects survivors, as a Markov decision process whose action is the modular prefix the model is conditioned on. It presents AGAR, a substrate in which credit assignment, value estimation, exploration and experience memory are separate components that can be replaced or disabled without changing the controller, with no gradient training of the backend model. Across 19 tasks, two backends and three seeds, AGAR improves on the stronger of two published baselines on most tasks, with gains concentrated in competitive programming.
In briefCurveTQ: Rotation-free trellis quantization of LLM weights using curvature-weighted search
SummaryThis paper proposes CurveTQ, a two-bit trellis quantization method for large language model weights that skips random orthogonal rotation. It adds the Hessian's diagonal weights from the LDL factorization into the Viterbi branch metric so the search follows curvature within each coding block. The authors report that on three 4-8B Instruct models, CurveTQ is 1-3 points higher in mean downstream accuracy than QTIP and Proteus, and also show a result on a 35B mixture-of-experts model.
