In briefA categorical map of parameter-efficient fine-tuning methods, comparing them geometrically
SummaryThe paper treats parameter-efficient fine-tuning methods as smooth maps starting from a pretrained model's weight point, and compares them using slice categories. It identifies invariants governing expressivity and training dynamics, and proves results about methods such as zero-initialised LoRA and LoRA+. The authors classify 382 methods from 138 papers by seven coordinates read from their formulas.
WhyUnifies hundreds of parameter-efficient fine-tuning methods categorically and proves several exact results on LoRA initialisation, rank bounds and optimizer symmetries.
In briefZebra3D: Faster hybrid AI models for reconstructing 3D scenes from many photos
SummaryThis paper presents Zebra3D, a hybrid attention framework that makes pretrained 3D foundation models more efficient without training them from scratch. It assigns linear attention to earlier, more tolerant network stages and sparse tile attention to sensitive later stages. The authors report speedups up to 7.1× at 1,024 views, and after post-conversion training, Zebra3D-VGGT reaches 80.5% RealEstate10K AUC@30 at 94% mean attention sparsity.
Utkarsh Saxena, Ruijing Yang, Mehdi Rezagholizadeh et al.
In briefSpatial Window or Past History? Input Sizing for Neural Surrogates of Partially Observed Fields
SummaryThe paper studies neural surrogates that forecast a physical field from sparse, noisy sensors, trading off a spatial window of nearby sensors against a history of past snapshots. It uses an exact Gaussian analysis as a Bayes-optimal reference and validates on a linear field, viscous Burgers flow, and the Kuramoto-Sivashinsky equation. The authors conclude that the window should be sized in correlation lengths of the field, and that a fixed input budget should be split according to the lead time the model will serve.
WhyExact Gaussian analysis shows nearby sensors suffice, and a lead-time-dependent window-history trade-off inversion is validated across multiple fields with known optima.
In briefUnsupervised Multimodal Reward Modeling: Vision-language reward models without human preference labels
SummaryThe paper builds a vision-language reward model without human preference labels by training a linear head on an intermediate layer of a frozen Qwen3-VL backbone. The head is trained to score a web caption above a perturbed version of the same caption, with labeled data used only to select the layer. The authors report that this recipe matches or beats GPT-4o on three benchmarks at 32B scale and analyze how the perturbation sampler shapes the learned reward.
WhyTrains a linear probe on a frozen backbone's intermediate layer using perturbed captions as signal, with layer ablations, GPT-4o comparisons, and hallucination evaluations supporting the claims.
Yueying Li, Jingxuan Fan, Kiant\'e Brantley et al.
In briefRePrior: Measurement-Consistent Prior Correction for Sparse-View Cone-Beam CT Reconstruction
SummaryThis paper addresses sparse-view cone-beam CT (CBCT), where too few X-ray projections leave reconstruction under-constrained. It proposes RePrior, a multi-stage framework that selects actionable local corrections to a learned prior, then reconciles them with patient-specific X-ray measurements. The authors report about 2.6 dB higher 3D PSNR than R²-Gaussian on LUNA16, PANORAMA, and PENGWIN, without needing test-time ground truth or case-specific fine-tuning.
In briefAI-driven, timing-aware chip power optimization after routing
SummaryThe paper proposes an AI-driven framework for reducing chip power after routing, where a learned model ranks power-recovery edits and repair actions while deterministic EDA analyses approve every change. An ensemble of gradient-boosted models predicts power change, damage reduction and constraint-regression risk, with a deterministic fallback when uncertain. The paper specifies the contract, learned policy, transactional flow and a public evaluation protocol, but reports no trained-model results.
In briefComparing Model Attention with Human Gaze: Explainability of CNN, Mamba, and ViT via Eye-Tracking
SummaryThis paper systematically evaluates the explainability of Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Vision Mamba (ViM), using eye-tracking data as a reference for human visual attention. Model attribution maps and gaze heatmaps are normalized into spatial probability distributions and compared with three similarity metrics. CNNs show more global focus, ViTs attend to fine-grained patch-level features, and Mamba performs well on medical images but diverges in natural images.
Marco Stefani, Edoardo Parpaiola, Matteo Bino et al.
In briefProduct Monte Carlo: a method for estimating expectations under a product of densities
SummaryThis paper proposes Product Monte Carlo (PMC), an algorithm that computes arbitrary expectations under a product of densities using only Monte Carlo draws from the individual densities. It reformulates the intractable product expectation as a regression problem and provides theoretical guarantees on consistency and convergence rates. Experiments report higher estimation accuracy, better uncertainty coverage, and lower computational cost than existing methods. Since no density evaluation is required, the method applies to simulation-based inference.
In briefActivating Annotators: Using Lived Experience in Data Labelling
SummaryThis paper introduces AXLE, a pre-annotation stage that uses a scenario vignette to guide a short conversation, which is then turned into a first-person micro-narrative shown to the annotator before labelling. In a 2x2 between-subjects study with 104 annotators and 1,040 annotations on ambiguous hate speech, the stage was associated with higher identity resonance and sensitivity to harm. It also raised label entropy and lowered inter-annotator agreement, which the authors interpret as surfacing plurality rather than manufacturing consensus.
Sheza Munir, Ramaravind Kommiya Mothilal, Syed Ishtiaque Ahmed
In briefPrimal-Dual Flow Matching for Generating Data Under Per-Sample Constraints
SummaryThe paper studies how to learn a generative model whose samples satisfy a prescribed per-sample constraint with a user-specified probability. It builds a primal-dual algorithm, Primal-Dual Flow Matching (PDFM), on a flow-matching parameterization and derives guarantees on the constraint satisfaction probability under regularity assumptions. PDFM needs only binary membership feedback, without differentiable distances, projections, or convexity of the constraint set. The authors report that it improves constraint satisfaction over vanilla flow matching while remaining competitive with constraint-aware baselines and preserving distributional fidelity.
In briefA density-dependent two-stage method for maximizing algebraic connectivity in network design
SummaryThis paper addresses an open network design problem: building graph topologies with maximum algebraic connectivity for fixed numbers of vertices and edges. It proposes a two-stage approach that first builds a structural backbone chosen by graph density, then completes the remaining edges with reinforcement learning. Numerical results report high algebraic connectivity and favorable runtime scalability, outperforming existing methods.
In briefCritical Mass: How a Splitting Rule for Molecular ML Drifts as Libraries Grow
SummaryThe paper argues that the structure-disjoint splitting rule widely used in molecular machine learning was fixed at a smaller library scale and breaks down as libraries grow. The authors propose an optimized version that makes the criterion computable on the 227,690-molecule public libraries. At that scale the graph percolates, with one connected component holding 95.5% of the library, and held-out compounds are systematically larger than training compounds.
Tomo Oga, Devesh Shah, Cailum M. K. Stienstra et al.
In briefFin_Agent at FinEval 2026: Task-aware methods for multilingual financial NLP tasks
SummaryThis paper presents Fin_Agent's submission to the FinEval 2026 shared task, covering Greek financial named-entity recognition, Japanese implicit communication recognition, and multilingual financial question answering. Greek NER combines annotation-aware numeric rules with zero-shot Qwen2.5-32B classification, reaching 0.9521 public and 0.8884 private F1. Japanese recognition uses an interpretable cue-based ordinal scorer with 98.0% test accuracy. The QA system pairs BGE-M3 retrieval with Qwen3-32B generation, scoring 0.3023 and 0.3419 ROUGE-1 F1 on the public and private splits.
Mohotarema Rashid, Tanzina Hossain, Xiaoying Song et al.
In briefAI detection of low-voltage intervals in neonatal EEG as a brain injury biomarker
SummaryThis paper proposes low-voltage intervals (LVIs), defined only by amplitude and duration, as an alternative to inter-burst intervals in neonatal EEG, and develops a ConvNeXt-based AI detector. Using 654 hours of EEG from 123 neonates, the model's agreement with readers reached expert-level equivalence. In a combined dataset of 444 neonates, the percentage of LVIs was associated with EEG grades, HIE grades, and seizure burden, and showed seizure prediction capacity.
John M. O'Toole, Sean R. Mathieson, Aurel Luca et al.
In briefOne-step amortised sampling for Bayesian posteriors
SummaryThis paper proposes OSIS, a one-step implicit sampler for drawing samples from high-dimensional Bayesian posterior distributions. It uses short-time MCMC dynamics to define a local target for an implicitly modelled distribution. The authors analyse its convergence properties and relate it to other generative modelling and sampling methods. Reported experiments show it performs well against multi-step amortised samplers and non-amortised MCMC baselines while being much more efficient at sampling time.
Vincent Pauline, Kirill Tamogashev, Arran Carter et al.
In briefWoodBR-Cross: public wood-image dataset and tree-disjoint species benchmark
SummaryThis paper introduces WoodBR-Cross, a public dataset of 4,875 transverse wood-surface images across 38 taxonomic classes, drawn from 303 trees in a Brazilian Forest Service collection. It proposes a benchmark that keeps all images from each tree in a single data split, avoiding optimistic results from shared trees. Several models were compared, and frozen DINOv2 features achieved the highest reported tree-level macro-F1 score.
Daniel Casanova, Alexandre Bahia Gontijo, Arnaldo Candido Junior et al.
In briefPlanning drone and ground robot teams for aerial monitoring with energy and visit-order limits
SummaryThe paper studies routing for teams of one UAV and one UGV that must visit monitoring points in a required order. The ground vehicles act as mobile chargers and carry the drones between aerial tours. The authors propose a polynomial-time algorithm that jointly plans routes and timing to minimize mission completion time under bounded travel-time disturbances. They provide feasibility guarantees, complexity analysis, and numerical comparisons.
In briefGeoFitBench-LLM: A Benchmark Testing Whether AI Models Can Identify the Geospatial Data Needed for Location-Based Analysis
SummaryThis paper introduces GeoFitBench-LLM, a benchmark that tests whether large language models can infer, from natural-language requests, the datasets, resources, and constraints needed for geo-analytic tasks. It is organized around two reasoning dimensions, Spatio-Temporal Relational and Semantic Grounding, and encodes ground-truth requirements in a fixed, provider-independent schema. In experiments on seven models, identifying necessary resources was often achievable, with GPT-5.4 Thinking recovering the complete set in 66.67% of requests. The results also indicate that accurate fitness-for-use specification becomes harder as inferential demands increase.
In briefA recurrent harness that revises past reflections for continual learning in language agents
SummaryThe paper proposes ReHarness, a recurrent harness framework for language model agents that learns continually from a strictly online stream. Beyond forward reading and writing, it revises retrieved reflections backward using each instance's resulting trajectory as feedback. Across three continual learning benchmarks and three model backbones, the authors report consistent gains over raw-context, memory/reflection, and prompt optimization baselines.
In briefperturber: A stateless tool that makes meaning-preserving text changes
SummaryThis paper presents perturber, a stateless service and Python library that applies exactly one semantics-preserving text perturbation per call. It offers a catalog of 68 perturbations across 11 families, and callers chain calls to compose them. The authors evaluate the tool's correctness, reproducibility, and cost rather than any model's robustness.
In briefIncremental RAG for human-in-the-loop decision support, where past human decisions become retrievable precedents
SummaryThe paper introduces iRAG, a decision-support system that combines Retrieval-Augmented Generation with Hybrid Decision-Making for settings where the knowledge base starts empty and grows through human decisions. Past decisions are stored as retrievable precedents, and the system moves from observation to contestation to autonomous handling with abstention. Compared with an independent baseline, iRAG's error results differ by drift level, and retrieval decay helps on changed knowledge but hurts on stable knowledge, a stability-plasticity trade-off. The authors note that human oversight is an architectural feature whose usability with real users remains to be evaluated.|
In briefLearning graph correspondence to transfer facade patterns from prototype to building massing
SummaryThis paper proposes a workflow in which a designer-made prototype's facade articulation logic is transferred to a larger target building massing. Both are represented as attributed graphs, and a Siamese graph attention network learns correspondences between their segments. On the original 85/15 holdout, prototype-to-massing top-1 accuracy reached 96.5%, versus a reported 76.5% raw-feature baseline. A Rhino demonstrator and qualitative applications on four built projects are also presented.
In briefMagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis
SummaryThe paper finds that about 88% of problems in a large pooled math corpus for RLVR are either solved or failed in every rollout, giving little learning signal. It introduces MagicBeanStalk, which recursively synthesizes new problems from seeds, preserving reasoning structure and calibrating difficulty to the policy and rollout budget. Across 8 math benchmarks, continual RLVR on these problems improves gpt-oss-120b by 8.6 points and DeepSeek v4Flash by 5.4 points.
WhyProposes policy- and budget-calibrated recursive data synthesis and characterizes the zero-advantage trap; gains on 8 benchmarks, mostly math-focused.
In briefData-driven optimal sensor placement for UK aerosol monitoring
SummaryThis paper presents a data-driven framework for placing sensors to monitor aerosols in the UK, using a trained convolutional Gaussian neural process within the DeepSensor optimisation framework. Genetic optimisation with lazy evaluation and caching cuts mean computation time by 60% while matching the greedy solver's sequential placement performance. Against a retrospective oracle benchmark built from the CAMS reference field, Joint-MI-driven placement improves Joint NLL by about 22% after five station locations are identified, and seasonal and wind-regime analyses show agreement with oracle benchmarks varies with transport conditions.
Talha Ansar (1), Waqar Muhammad Ashraf (2), Alejandro Coca-Castro (2) et al.
In briefA physics-informed machine learning framework that reconstructs sparse industrial dye recipes in both directions
SummaryThe paper proposes a hybrid framework for designing high-dimensional sparse recipes in process industries, combining a forward quality model, a sparse inverse reconstructor, and a feasibility-constrained ranking layer. It was validated in an industrial textile dyeing setting, where HGB and GBDT models performed best in the forward model and inverse reconstruction reached a recipe-vector MAE of 0.0295. The authors note that the link between alkali adjustment and lower error is only an observed association. The tool is intended as a human-in-the-loop aid to reduce costly lab trials.
Md. Muqtadir Fuad, Md. Istiak Adnan, Ridwan Al Aziz
In briefKronecker Product Attention: a paper on shrinking the KV cache via structured factorization
SummaryThis paper addresses the growing memory cost of the key-value (KV) cache in autoregressive generation. It proposes Kronecker Product Attention (KPA), which represents keys and values as sums of Kronecker products of compact, token-dependent matrix factors, and is said to generalize Tensor Product Attention (TPA). The authors also describe a matched structured rotary embedding. According to the paper, under identical training conditions KPA matches or exceeds several baselines across four model scales and nine benchmarks, with up to 10x KV cache compression.
Bahareh Nikpour, Hossein Rajabzadeh, Maryam Dialameh et al.
In briefCLEAT: Teaching Interactive AI Agents When to Ask Clarifying Questions
SummaryThis paper introduces CLEAT, a framework that co-trains a controller predicting the value of asking clarifying questions alongside an agent's task policy. The value is estimated by comparing paired rollouts from the same history, with and without asking. With Qwen3-4B, CLEAT achieves the highest mean score in all ten evaluation settings across four benchmarks.
In briefSynAct: Active feedback learning that targets key decisions for LLM agents
SummaryThis paper presents SynAct, a framework for LLM agents that allocates scarce expert feedback to individual decisions rather than whole tasks. It ranks decisions by the mutual information between chosen actions and final outcomes across repeated rollouts, queries experts at the highest-value points, and distills corrections into rules the agent can reuse. The authors report held-out accuracy gains in all 16 evaluated settings across four benchmarks and four models.
In briefFeature Competition in Vision-Language Models: A Stroop-Based Study of Text Versus Color
SummaryThe paper adapts the classic psychological Stroop task into a benchmark with images where text and color conflict, and tests contrastive and generative vision-language models on it. Across models, responses consistently favor the text over the color when the two disagree. Analysis of the embedding space shows textual cues are represented more strongly and separably than color cues, and steering toward text-related directions succeeds more easily than steering toward color-related ones.
In briefComparing Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building
SummaryThis paper compares statistical, machine-learning, recurrent neural-network, and pretrained TimesFM models for predicting next-day electricity use in a university building. All configurations were tested on the same 57 days. TimesFM 3.0 with weather covariates had the lowest overall errors, about 43% lower mean absolute error than weekly seasonal-naive forecasting. XGBoost was also competitive, though paired daily comparisons showed no clear mean-absolute-error advantage for TimesFM.
Amey Vasulkar, Samuel Kernan Freire, Arie-Willem de Leeuw et al.
In briefDistributed feedback control for closing holes in planar robot swarms
SummaryThe paper studies closing holes in a planar multi-agent system whose sensing disks form a time-varying Čech complex. It proposes a harmonic-covariance feedback that uses harmonic edge signals to locate holes and turns them into agent forces, evaluated distributedly via a fixed-degree polynomial of the regularized Hodge Laplacian. Under assumptions such as measurable, bounded hole area and continuous evolution, offline computations yield rational lower bounds on cumulative hole area. In a 15-second numerical demo, all holes close by 9.9 seconds and stay closed.
In briefForcingBooster: Adaptive Per-Patch Denoising to Speed Up Autoregressive Video Generation
SummaryThis paper addresses a limitation of autoregressive video diffusion models, which apply the same number of denoising steps to every spatial patch when generating each video chunk, wasting computation on simple regions. The proposed ForcingBooster framework dynamically assigns denoising steps per patch, letting easy patches finish early and giving difficult patches more refinement. It works with most existing few-step autoregressive diffusion models and needs only a lightweight routing module with modest training. Experiments reportedly cut the total denoising steps across video patches by about 50% while keeping visual quality comparable to the four-step baseline.
In briefMemReg: Multi-frame outdoor LiDAR point cloud alignment using memory buffers
SummaryThis paper presents MemReg, a network that aligns streaming outdoor LiDAR scans by using temporal information from earlier frames. It uses an implicit feature buffer and an explicit pose buffer, plus a dynamic history weighting module, to combine current and past pose-related features. The authors report state-of-the-art results on KITTI, nuScenes, and Apollo-Southbay, outperforming prior pairwise methods and multi-frame SLAM systems. They also report competitive results on indoor 3DMatch, 3DLoMatch, and ScanNet.
WhyIntroduces pose-aware hybrid memory buffers and dynamic history weighting beyond pairwise registration; broad comparisons across KITTI, nuScenes, Apollo, and indoor benchmarks give fairly strong evide
In briefMesh-native tokenisation: adapting grid-trained physics models to irregular meshes
SummaryThis paper introduces Mesh-Native Convolution (MNC), a locality-aware tokenisation layer that encodes unstructured mesh data in a way analogous to CNNs. It allows physics foundation models trained on grid data to be efficiently retrofitted to unstructured meshes. The authors report that the approach often reaches acceptable accuracy in zero-shot deployment and outperforms existing baselines after a few epochs of fine-tuning.
Georg Maierhofer, Miles Cranmer, Payel Mukhopadhyay
In briefGeopoliticalQA: Benchmark for Question Answering over Geopolitical Intelligence Reports
SummaryThis paper introduces GeopoliticalQA, a benchmark for grounded question answering over 415 open-source geopolitical and strategic intelligence reports, with 5,525 question-answer pairs. It evaluates 15 retrieval systems and six open-weight LLMs for retrieval-augmented generation. At chunk-level retrieval, the mid-size bge-m3 encoder matches much larger models and outperforms commercial embedding APIs. Answer generation remains difficult even with gold evidence, especially for inferential and risk-oriented questions.
Veronika Solopova, Ivan Havlytskyi, Viktoria Skorik et al.
In briefSoftwareGym: Scaling Computer-Use Agent Training with Synthesized Software Environments
SummaryThis paper presents SoftwareGym, an automated pipeline that synthesizes GUI software environments with diverse interfaces, executable functions, and runtime-verified behavior. The authors synthesized 182 applications across twelve categories, reaching a mean GUI task pass rate of 91.5%, and built SoftwareGym-Bench, a 120-task evaluation suite. Training a computer-use agent on trajectories from these environments raised Pass@3 on the out-of-domain OSWorld benchmark from 41.39% to 56.97%.
In briefKnowing Is Not Telling: How LLMs Fix Cross-Border Ethnic Identities to One Nation
SummaryThis paper studies how large language models describe nine ethnic communities that span the China and mainland Southeast Asia border. It defines "identity ossification" as models turning a fluid cross-border identity into a single nation-state identity. The authors introduce the TIODF framework, comparing factual questions with open-ended narrative questions, and test GPT-5.1 and DeepSeek-V3.2 in English and Chinese. Ossification rates differ widely (10.1% vs. 62.6% of answers), and models often tell a single-nation story even when answering the factual question correctly.
In briefVeracity labels shift LLM emotional tone but not racial disparities
SummaryThis paper audits four large language models rating simulated emotions on false and true headlines, with and without veracity labels. The labels strongly changed emotional tone, but Black personas were still assigned more negative affect than White personas, even on race-unrelated headlines. The authors conclude that factual correction and demographic consistency are distinct properties, so persona-based systems should be audited for both before use in sensitive contexts.
In briefSelf-Preservation in Language Models: Models Generate Their Own Replay to Limit Forgetting During Continual Learning
SummaryThe paper studies forgetting in continual learning, where a model adapting to new tasks loses prior capabilities that preservation task sets only partly cover. The authors organize task descriptions and examples into Task Banks, and a frozen model uses them to generate replay data. Experiments with Qwen models of different sizes show that generated replay reduces forgetting on tasks outside the bank, and combining partial banks improves retention. At 4B, selecting replay targets informed by the new task beat random selection, while smaller models showed no such advantage.
In briefReflective prompt optimization for a deep-research agent, evaluated on unseen topics across multiple seeds to test whether gains transfer
SummaryThe paper applies reflective prompt evolution (GEPA) to the two system prompts of a multi-agent deep-research agent and evaluates it with train, validation, and test sets that share no topics. Held-out citation reward rises from 0.702 to 0.781, and the gain replicates across three seeds. Ablations indicate that decomposed per-dimension feedback is the causal ingredient, while collapsing it to a single scalar reward makes held-out transfer negative.
Elias Lumer, Sahil Sen, Vamse Kumar Subbiah et al.
In briefSTAPLE: Expanding Pretrained LLMs with Static Sparse Memory Instead of Retraining
SummaryThis paper introduces STAPLE, a method that adds static sparse memory to pretrained language models. It keeps the original feed-forward up-projections and adds token-indexed embeddings as a residual pathway, expanding capacity without sparse pretraining from scratch. On OLMo-2 1B, it improves over continued dense training by 4.27 points on HumanEval, 2.80 on MBPP, 2.27 on GSM8K, and 1.65 on DROP.
In briefBeyond Binary Rewards: Self-Judging Debugging Behavior for SWE Agents
SummaryThis paper addresses reinforcement learning for software engineering agents, where binary executable-test rewards cannot tell apart trajectories with very different debugging quality. The authors propose Behavioral Self-Rewarding (BSR), which has the policy itself compare observable debugging behavior among trajectories from the same task with the same verifier outcome, turning pairwise preferences into bounded auxiliary rewards. Experiments on SWE-bench Verified and SWE-bench-Live show consistent improvements for Qwen3-14B and Qwen3-32B.
In briefSpireBench: Testing AI Models' Long-Term Strategy with the Game Slay the Spire
SummaryThis paper introduces SpireBench, which has large language models play the full commercial deck-building game Slay the Spire at its highest difficulty. Among 20 models tested, frontier models beat simple heuristic baselines and match an average human player but remain far behind experts, with the gap widening over longer runs. The authors also find that models misjudge risk across a run and do not improve from their own cross-run memories.
In briefKnow What You Don't Know: Agentic data synthesis that evolves its generation space from learner feedback
SummaryThis paper introduces EvoSyn, a learner-feedback-driven multi-agent framework that uses a revisable state to decide which interaction factors to generate and where to focus synthesis effort. It reports that on τ²-Bench and CAR-Bench, EvoSyn improves average task success by 13.3 points over the strongest synthesis baseline. The authors also state that it produces higher-quality, structurally broader synthetic experience and finds controllable factors missed without learner feedback.
In briefScene-based visual cues can make multimodal AI models bypass their safety refusals.
SummaryThe paper shows that task-relevant scene semantics alone, without added attack instructions, can shift multimodal large language models from refusing to complying with harmful requests. The authors introduce SceneSafetyBench to test this across eight risk categories and nine white- and black-box models. They also propose Scene-Grounded Attacks (SGA), which reach an average attack success rate of 86.17% across nine target models.
WhyShows task-relevant scene semantics alone flip MLLM refusals; SGA reaches 86.17% average ASR across nine models, a notable and fairly well-supported finding.
In briefHuman pose estimation using mixtures of Gaussian and uniform probability distributions
SummaryStandard human pose estimation models extract keypoints deterministically from discrete heatmaps, which can fail under dense occlusion and out-of-distribution conditions. The paper proposes a post-hoc probabilistic framework that models raw heatmaps as a Gaussian mixture plus a uniform component, requiring no architectural changes or retraining. The authors report that the spatial expectation of the fitted model matches leading sub-pixel decoders, and that it improves out-of-distribution anomaly detection and some calibration measures, with an extension to multi-view 3D models.
In briefRealism vs. Accuracy: Forecasting Event Sequences as Generative Modeling
SummaryThis paper treats multi-horizon event sequence forecasting as a conditional generative modeling problem, assessing both prediction accuracy and distributional realism. It presents a benchmark with metrics for both qualities, strong statistical baselines, adapted recent models, and curated real-world datasets. The evaluation finds that no method excels at both accuracy and realism, revealing a trade-off between them. The authors also introduce Wasserstein BaryBooster, which aggregates samples from a realism-strong diffusion model into a state-of-the-art predictor, showing realism can be turned into accuracy.
WhyFrames event forecasting as conditional generation, proposes a benchmark assessing accuracy and realism, and a diffusion-aggregation method evaluated across multiple baselines and datasets.
In briefOne Model for Variable Loop Counts: Elastic Latent Recurrence in Language Models
SummaryThis paper proposes Elastic Latent Recurrence (ELR), a training recipe for language models that reuse a shared block of layers, letting one checkpoint trade inference compute for accuracy by changing the loop count. The authors report that at one loop an ELR model costs the same to serve as a non-recurrent model of the same size and reaches similar accuracy on the same data, and that additional loops improve accuracy at higher inference cost. Experiments on 4B dense and 30B-A3B MoE hybrid models are described, along with a resulting checkpoint, Koi-4B.
Faisal Ladhak, Shantanu Acharya, Santiago Akle Serano et al.
In briefCLUE: Contrastive Learning for Automatically Inducing and Annotating Event Types
SummaryThis paper proposes CLUE, a framework that induces, names, and annotates event types without manually creating an event ontology in advance. It combines a joint embedding that captures both verb-centric and semantic views of events, clusters events using contrastive-loss guidance, and then names and refines the resulting types. The authors report experiments demonstrating the method's effectiveness, and the code is publicly available on GitHub.
In briefMind the Temporal Gap: Evaluating whether video-language models understand motion in driving videos
SummaryThis paper tests whether video-language models can reliably identify ego-motion actions from driving videos. Standard, visual, few-shot, and chain-of-thought prompting proved unreliable and biased, though newer models gained on specific few-shot tasks. Lightweight LoRA finetuning and simple probes on intermediate visual features recovered strong motion discrimination, suggesting the cues exist inside models but are not used well. The authors conclude these models act more like biased frame-level recognizers than dependable temporal reasoners for driving.
In briefScaling Multi-Teacher Distillation for Digital Pathology
SummaryThis paper studies multi-teacher distillation, which transfers knowledge from several teacher networks into one student network. The authors find that the standard approach performs poorly as the number of teachers grows, and propose design choices such as learnable teacher tokens, a tailored attention scheme, mixture-of-experts layers, and a contrastive loss to learn teacher-specific representations. Evaluated on 39 tile-level and slide-level tasks, the resulting compact student encoders are up to 10x more efficient than teacher foundation models while matching or outperforming them. The authors also release OMNI, a new family of foundation models trained from 10 state-of-the-art teachers.
Sofi\`ene Boutaj, Pierre Marza, Varun Belagali et al.
In briefAn all-transistor physical circuit trained to perform in-context learning
SummaryThe paper introduces an attention mechanism built from transistors that exploits the exponential current behavior of sub-threshold MOSFETs to perform the multiplication needed for dot-product attention. In simulations, the circuit could be trained on a simple in-context learning task despite using realistic reverse components that transmit only approximate gradients. The authors report that cross-entropy loss, rather than mean squared error, was required for training to succeed.
In briefContext-Aware Correction of Parallel Drafts for Faster LLM Decoding
SummaryThis paper addresses a drawback of parallel drafting in speculative decoding, where tokens in a draft block are predicted without seeing each other. It introduces DCOCO, a lightweight corrector that keeps multiple candidate tokens and scores their compatibility across neighboring draft positions in parallel. Across four verifier backbones and ten benchmarks, the authors report the highest average throughput in every configuration, with speedups of up to 6.6x over vanilla decoding. Under 4 to 32 concurrent requests, they report DCOCO outperforms the strongest competing method by 6-23%.
In briefSWE-Duel: A Self-Scaling Arena Where Coding Agents Write and Solve Bugs for Each Other
SummaryThe paper argues that static coding benchmarks leak into training data and get saturated by stronger agents over time. It introduces SWE-Duel, an arena where two coding agents take turns writing a feature with a hidden bug and reviewing each other's pull requests, so tasks are created at evaluation time. In a tournament of 11 model-harness participants over 1,500 turns on 15 repositories, GPT-5.5 with Codex CLI ranked first as the strongest solver, and Claude Opus 4.8 with mini-swe-agent ranked second as the strongest task author. Preliminary results suggest tasks from newer models are harder for a fixed solver.
In briefMIRAGE: a method that fuses LiDAR-inertial metric motion with feed-forward vision for accurate 3D reconstruction and pose estimation
SummaryFeed-forward 3D vision methods can reconstruct scenes from image streams but often lack reliable metric scale and drift over long trajectories. MIRAGE anchors these visual geometry estimates to LiDAR-IMU metric motion and adjusts online pose corrections based on correspondence support and cross-modal agreement. Experiments on Oxford Spires, M2DGR and INS show more accurate metric trajectories than COLMAP. The authors also build a multimodal platform to create 6-DoF reference trajectories for a new vision-only indoor benchmark and evaluate recent monocular large-scale reconstruction methods on it.
In briefCASTER: Turning Synthetic Videos into Zero-Shot Robot Manipulation via Constrained Optimization
SummaryThis paper examines how robots can carry out tasks using trajectories derived from videos produced by generative video models, without real human demonstrations. The authors frame the conversion as a constrained optimization problem, where object motion extracted from the synthetic video serves as an initial guess that the robot refines in a digital twin while maintaining physical feasibility and semantic consistency. By removing simplifications used in prior methods, they report a 34.1% gain in average task success over the best-performing baseline.
Yinlong Dai, Homanga Bharadhwaj, and Dylan P. Losey
In briefSample-adaptive zeroth-order post-training: faster LLM fine-tuning with fewer model queries
SummaryThis paper proposes SAZO, a sample-adaptive zeroth-order optimization method for fine-tuning large language models. It learns each training example's importance online without true gradients and allocates queries accordingly, adding no extra model evaluations. The authors prove a strictly faster convergence rate than standard methods, and experiments on four model families show better average downstream performance and up to 6× fewer queries to reach the same training loss.
In briefFly-by-Code: Coding agents that actively look and test to make drones grasp objects reliably
SummaryThis paper presents Fly-by-Code, a method in which an agent writes robot programs from language instructions and improves them using execution feedback, targeting aerial manipulation. Because payloads limit cameras and dynamic coupling makes physical contact fragile, the authors add Active Feedback, which runs diagnostic programs in simulation to view unseen outcomes or probe failure causes. Across six simulated tasks with 60 trials per condition, the approach cut false completion declarations from 25 to 5 and raised success from 31.7% to 75%. The refined programs also ran on a physical aerial manipulator without source-code changes.
WhyProposes active View-and-Probe diagnostics that fill execution feedback gaps; simulated trials show large gains, transferring to a physical aerial manipulator.
In briefRubric-CEPR: Self-Evolving Image Editing via Reward-Verified Self-Distillation
SummaryThis paper proposes Rubric-CEPR, a framework that improves an image editor using only its own generated samples, without human-edited targets or an external reward model. Candidate edits are verified with a rubric-augmented Contrastive Edit-Preservation Reward computed from the editor's internal features, and the best verified candidate is distilled back via lightweight adapter training. The authors report that on Qwen-Image-Edit, ImgEdit scores rise from 4.36 to 4.60, with transfer to other benchmarks and gains on Step1X-Edit.
Ritesh Thawkar, Shubham Patle, Shravan Venkatraman et al.
In briefDex-One2Many: Learning Dexterous Manipulation from One Human Demo
SummaryThis paper presents Dex-One2Many, a framework that learns a generalizable dexterous manipulation policy from a single human video using a real-to-sim-to-real approach. The video is abstracted into sequential scene graphs that guide reinforcement learning, sampling diverse reset states and providing dense per-stage rewards. The policy is trained entirely in simulation and transfers zero-shot to a real multi-fingered hand. On five tool-use and manipulation tasks, it outperforms baselines by 6.5% on seen configurations and by 71% on unseen scenarios.
In briefAutoIndexer: a causal model trained on chains of edits for flexible-order decoding
SummaryStandard autoregressive decoding struggles with data that has non-causal dependencies, because already generated outputs cannot be revised afterward. AutoIndexer introduces a training and decoding mechanism that simulates a Chain of Edits, letting the model insert, substitute, and delete tokens in hindsight. The authors report that the method keeps performance on standard LLM benchmarks and offers post-generation context editing that most autoregressive and diffusion models lack.
WhyNovel chain-of-edits training lets autoregressive models revise prior tokens post hoc, with benchmark performance preserved; validation scope appears moderate.
In briefDreamTrue: Action-Faithful Robot World Model via Counterfactual Post-Training
SummaryThis paper presents DreamTrue, a multi-view robot world model that predicts videos which follow given actions and remain physically plausible across different robot bodies. It renders action trajectories as image-space conditions with offline geometric calibration, and uses counterfactual post-training to cover more interaction scenarios. A video reward model trained on human-annotated defect data guides reinforcement-learning post-training. On AgiBot, the reported human-assessed interaction defect rate drops from 48.12% to 6.25%.
In briefStatistical review of the METR plot for AI time horizons
SummaryThis paper examines METR's 50% time horizon, the human completion time of software tasks an AI solves with 50% probability. Using 228 tasks and 26 AIs, the authors recompute it with splines and item-response theory instead of assuming a linear link to log human time. The fitted curve is nearly flat between 2 and 30 minutes, so a jump from 3 to 30 minutes is easier than from 30 minutes to 5 hours. The paper also offers point estimates and diagnostic plots for assessing construct validity.
In briefCSF: Context-Aware Safety Filtering for Motion Generators
SummaryThis paper introduces contextual safety filtering (CSF), a training-free filter for text-conditioned whole-body motion generators. It links natural-language safety rules to safe and unsafe reference trajectories from the generator, and enforces them with a safe-reference-tracking CBF-QP. Tested on four pretrained generators, CSF cut the danger-event rate by up to 90% while keeping 88-100% of benign motions. The full system was also demonstrated on a real Unitree G1 robot.}
In briefA Balanced Data Diet: Fixing Wasted Training in Large-Scale Robot Reinforcement Learning
SummaryThis paper addresses a problem in large-scale reinforcement learning for robot control, where simply adding more parallel simulation environments wastes much of the training data on task configurations the policy has already mastered or cannot yet attempt. The authors introduce Success Guided Sampling (SGS), an adaptive sampler that focuses training on configurations near the edge of the policy's capabilities. Experiments with up to over one million parallel environments show SGS lets RL solve multi-terrain quadruped locomotion and contact-rich assembly tasks that prior methods could not.
Octi Zhang, Mateo Guaman Castro, Patrick Yin et al.
In briefWhat 30,000 Hours of Egocentric Video Doesn't Teach: Scaling Data Doesn't Fix World Models' Object Modeling
SummaryThis paper studies how far scaling world models with 30,000 hours of egocentric human video, spanning over 1,000 scene types and 14,000 contributors, can go. The authors directly measure agent and object-interaction fidelity on an out-of-distribution benchmark. Increasing data 100-fold improves both, but agent modeling becomes strong while object fidelity stays far lower and improves slowly. They also propose a supervision scheme that shifts capacity toward object dynamics, which helps but leaves a substantial gap.
In briefOuroWorld: Turning Any 3D Scene into Diverse, Seamlessly Looping 3D Cinemagraphs
SummaryOuroWorld is a mask-free framework that converts any static 3D Gaussian Splatting scene into a dynamic 3D cinemagraph whose motion loops seamlessly from any viewpoint. A vision-language model infers plausible dynamics and guides a video model to produce a reference video, which is lifted into multi-view videos and learned via Inconsistency-Robust Periodic 4DGS. On 39 scenes, the method outperforms all baselines and wins 70.8%-99.0% of user-study comparisons.
In briefWorldGuide: A Goal-Directed Video World Model for Step-by-Step Task Execution
SummaryVideo world models can generate plausible visual sequences, but long procedural tasks require adapting to what has already been generated. WorldGuide takes only an initial image and a task goal, predicts the next atomic action, generates its video clip, and uses the result to choose the next action or stop. Its planner and executor are trained on the same step-level procedural demonstrations, with hierarchical visual memory to track long-horizon state. It reports 33.33% task success on WorldGuide-Bench versus 29.90% for MiniMax-H3, and 47.69% on VideoCraft-Bench versus 32.73% for MiniMax-H3.
In briefOmniCapBench: A Fine-Grained Benchmark for Evaluating Audio-Visual Video Captioning
SummaryThis paper introduces OmniCapBench, a benchmark for testing how well multimodal large language models describe audio-visual content. It breaks evaluation into verifiable units across entity references, visual shots, and audio events, scored with deterministic checks and localized LLM comparisons. Using 786 densely annotated videos, it distinguishes errors such as temporal grounding failures, identity drift, cross-modal misalignment, and hallucination. Results show frontier models perceive local content well but reason weakly over long audio-visual horizons.
In briefHybrid Cinematography: Previewing Hallucination Risk in Generative Video Reshooting
SummaryThe paper proposes Hybrid Cinematography, a workflow that combines physical capture with generative video reshooting. Using an editable 3D shot plan and a proxy of the take, it evaluates hallucination risk in real time so filmmakers can adjust moves, shoot guided pickups, or knowingly accept hallucinated content. The workflow is demonstrated with a mobile augmented reality app and an offline pipeline for existing video. A study with experienced filmmakers examines how previsualizing risk informs camera decisions.
In briefMental-Models for Multi-Agent Systems: agents that infer a partner's hidden beliefs and intentions
SummaryThis paper introduces mental-model-enabled agents, which learn a latent model of their counterpart to infer hidden beliefs, intentions, and likely reactions and use these to guide action. The method jointly learns an amortized recursive Theory-of-Mind representation with first- and second-order mental-state structure and a belief-conditioned reward model. The authors report that explicit mental-state modeling improves interaction quality and Theory-of-Mind performance over base agentic systems on both language-only and multimodal benchmarks.
In briefBrickBench: A Benchmark for Agents Designing LEGO-Style Builds
SummaryThe paper introduces BrickBench, a benchmark for evaluating agents that design LEGO-style assemblies from text prompts. Agents must pick parts from a discrete library and satisfy both semantic requirements and local and global physical constraints so the design can actually be built. Performance is scored on validity, alignment, and design across three settings varying in scale and part availability. The authors also release BrickAgent, an environment for coding agents to construct, inspect, and validate designs. They report that leading agents largely meet verifiable physical and semantic requirements but fall short of human designs.
In briefVersaCamVLA: Camera-Configurable Robot Manipulation Policies
SummaryThis paper proposes VersaCamVLA, a framework that lets robot manipulation models work with varying numbers and poses of cameras instead of a fixed setup. It maps an arbitrary set of posed RGB views into fixed-size scene tokens, using wrist-camera motion to gain pose diversity. Experiments on RoboTwin, LIBERO, and a real robot report better performance than prior VLA methods and multi-view baselines.
In briefPristine Models: A text-to-image design that separates unsafe concept knowledge into a branch that can be switched off at inference
SummaryThe paper notes that training image generators on broad, unfiltered data yields high quality but can teach unsafe concepts, while training only on safe data lowers quality. The authors propose Pristine Models, an architecture whose attention mechanism separates safe and unsafe knowledge into independently parameterized branches, with the unsafe branch detachable at inference. Experiments on ImageNet and CC12M report safer generation with higher image quality and better text alignment than safe-only training or post-hoc concept removal.
Anirudh Sundara Rajan, Krishna Kumar Singh, Richard Zhang et al.
In briefBi-FORK: a generative model for high-dimensional systems with multiple possible solutions
SummaryThis paper introduces Bi-FORK, a generative framework for learning one-to-many solution maps in high-dimensional bifurcating physical systems, where one input can yield several valid outputs. It generates complete space-time trajectories using latent flow matching and uses repulsion-guided sampling to recover distinct solution branches in a single amortized pass. The method is evaluated on buckling beams, mechanical metamaterials, and Allen-Cahn phase separation, with discretizations up to 260,000 points.
Anna Zimmel, Fleur Hendriks, Markus Holzleitner et al.
In briefOne Block, Multiple Depths: Recurrent Vision Transformers with Depth-Programmed Experts
SummaryThis paper proposes reViT, which applies a single Transformer block recurrently in place of a full-depth vision encoder, without intermediate feature distillation. Each recurrent depth's feed-forward network is a convex mix of a small shared expert bank, controlled by a continuous normalized-depth coordinate. Trained from scratch, reViT-B/16 reaches DeiT III accuracy with about 70% fewer stored parameters. The authors also report elastic-depth operation and a dense materialized form for fixed-depth deployment.
Adrian Bulat, Yassine Ouali, Georgios Tzimiropoulos
In briefTelling FIBS: Probes Detect Sabotage and Hidden Deception in Language Models
SummaryThe paper presents a white-box probe method for detecting deception in large language models, along with FIBS, the largest deception dataset to date, released for training probes. The new probe architecture aggregates information across many layers and tokens. On the SHADE-Arena and BashArena agentic coding benchmarks, the probes reach 98.3% and 98.0% AUROC, outperforming action-only text monitors. The authors also report detecting introspective deception that cannot be judged from context alone, and lying about politically sensitive topics in open-weight models under pressure.
WhyIntroduces a multi-layer, multi-token probe architecture and largest deception dataset, outperforming black-box monitors on agentic benchmarks.
Oskar J. Hollinsworth, Alex F. Spies, Tigist Diriba et al.
In briefRounding in Preconditioner Space: Redesigning 4-bit AdamW State Quantization
SummaryThis paper studies how quantizing AdamW's optimizer states introduces errors that affect later adaptive updates, and redesigns 4-bit quantization by focusing on the space where rounding decisions are made. It proposes ZIP-SR, which keeps zero in the second-moment codebook and performs stochastic rounding in preconditioner space, and ZE-EDEN, which uses a zero-excluding codebook with block rescaling. In GPT- and Llama-style pretraining from 130M to 2.7B parameters, both methods narrow the validation-loss gap to 32-bit AdamW at every model size, with the largest gap reduction reaching 70%.
Hanyang Li, Shao Tang, Daniel Thomas Braithwaite et al.
In briefLEGO: A method that generates first-person video from third-person footage without lifting to 3D
SummaryThis paper addresses generating an egocentric (first-person) video from a single exocentric (third-person) recording, a hard novel view synthesis task because the two views overlap little. The authors propose skipping depth estimation, point clouds, and reprojection, instead using an LVSM-style transformer to render the egocentric view directly and a probabilistic correspondence distribution to condition a video diffusion model. The paper reports that this approach outperforms the state-of-the-art explicit pipeline and generalizes to other datasets without retraining.
In briefEstimating density ratios by modeling how one distribution moves onto another
SummaryThe paper estimates the density ratio between a target and a base distribution by parametrizing it with a displacement field acting on the base, modeling the log-ratio as minus the Stein operator applied to the field up to a constant. A single convex optimization problem yields both statistical and dynamical descriptions of the distribution shift. Iterating this estimate-and-move step gives two algorithms, push-forward and pull-back. Applications to distribution shift in simulation-based inference and nonlinear independent component analysis are presented.
In briefAI Agent Populations: How Collaboration Creates a Takeoff Threshold
SummaryThe paper develops an ecological theory of AI-agent populations, using a growth equation in which fitness depends on cybersecurity capability. It shows that without collaboration a population takes off only when individual capability exceeds a critical level, whereas with collaboration collective capability grows with population size, creating a critical population threshold (a strong Allee effect). The authors propose ecological red teaming and gradually scaling populations in controlled settings to estimate this threshold.
In briefVioLA: Training a General Humanoid Robot Controller on Human Motion Data
SummaryThis paper introduces VioLA, a generalist humanoid policy that predicts body and hand motion latents instead of joint commands, which pretrained controllers then execute on the robot. Motion encoders map human and robot motion into shared latent spaces, so human recordings can be used directly in training; the pool contains 140.6 million frames, 93.2% from humans. The authors report that VioLA follows locomotion instructions zero-shot on a real robot with 100% success and reaches 88.6% manipulation success without task-specific fine-tuning.
Mert Albaba, Jens Bei{\ss}wenger, Anna Manasyan et al.
In briefFAITH: Safety-filtered reinforcement learning for high-dimensional robots that stays feasible
SummaryFAITH is a model-free reinforcement learning framework that learns a safety value and uses a feedforward network to filter actions with minimal intervention, so the task policy can optimize return without a competing safety term. When no safe action exists, the filter picks the action with the lowest predicted peak harm. Tested on a double integrator, Safety Gym, a 29-DoF humanoid, and a real Unitree G1 humanoid, it reports high safety rates while retaining much of the unfiltered return.
Songyuan Zhang, Baljeet Singh, Sarthak Ranjeet Kaingade et al.
In briefFastBench: Testing whether streaming video models can perceive fast-changing real-world footage
SummaryThis paper introduces FastBench, a benchmark for evaluating streaming video language models on high-dynamic real-world video streams. It contains 306 question-answer pairs across eight domains, six capabilities, and three temporal scopes. The strongest model, Gemini-3.5-Flash, scores only 50.7%, and the training-free ProactiveFrame baseline remains well below oracle-guided focusing.
In briefRoboRSI: a robot self-evolution system for stable, reusable skill learning in complex real environments
SummaryThis paper presents RoboRSI, a robot self-improvement system built on Top-Down Skill Refinement (TSR), which splits tasks into compound, atomic, and base skills and attributes each execution outcome to the responsible skill branch for revision. A Manager, Planner, Engineer, and Reviewer coordinate planning, execution, diagnosis, and validated release of new skills, and stable skill sequences are consolidated into reusable compound skills. The authors report developing multi-object household cleanup on a mobile manipulator over 104 rounds. In simulation on LIBERO, LIBERO-PRO, LIBERO-Plus, and RoboTwin, the system reportedly achieves the highest success rate, exceeding the strongest baseline by 2.7 to 11.0 percentage points.
In briefPumpire: A Benchmark for Measuring Real-World Distance Estimation in 3D Models
SummaryThe paper introduces Pumpire, a benchmark that tests whether image- and video-based 3D foundation models can estimate distances between point pairs in reconstructed scenes. It uses pumpire-6k, a dataset of 100 real-world scenes totaling 6,400 frames, with physically measured point-pair distances. Experiments cover 29 baseline configurations, aiming to assess how well models perceive physical scale in 3D space.
In briefProof that transformer-based masked diffusion models can learn distributions with latent structure
SummaryMasked diffusion models generate discrete data by iteratively unmasking multiple tokens in parallel, which introduces parallelization and estimation errors. The paper gives end-to-end statistical guarantees for transformer-based models when data follow mixtures of hidden Markov trees. Under this assumption, the model achieves epsilon-accurate sampling in KL divergence with sample complexity governed by latent structure, without knowing the tree topologies or component parameters.
In briefBeyond Spatio-Temporal Priors: A Generalizable Dense Correspondence Method
SummaryDense correspondence matching has relied on assumptions like smooth motion and rigid geometry, which fail in image editing and reference-guided generation. The paper introduces FreeMatching, a framework combining generative and semantic foundation representations with diverse supervision from classical datasets, tracked videos, and synthetic scenes. Teacher-guided iterative refinement improves correspondence in these settings without dense annotations. A single model improves matching on challenging editing image pairs while remaining competitive on classical benchmarks, and its scores correlate with human judgment for measuring identity preservation.
In briefA unified Bellman operator that balances task reward and safety in reinforcement learning
SummaryThe paper proposes a new Bellman operator that merges task performance and safety objectives into one joint value function. The authors show that temporal difference learning with this operator converges under a two-timescale stochastic approximation framework. They report that the resulting policy maximizes task return while keeping safety at all times, and that continuous control experiments with neural approximation show stable convergence and near-zero safety violations at test time.
Nishanth Arun Rao, Royina Karegoudra Jayanth, Benjamin Eysenbach et al.
In briefOneSearch-VL: A Unified Multimodal Research Agent for Images and Videos
SummaryOneSearch-VL is a unified agent that handles deep research on single images, multiple images, and videos by linking visual grounding, external retrieval, and fact composition. It uses a Visually Grounded Evidence Graph to build training data, supervise reinforcement learning, and evaluate operations. The authors also release two new benchmarks, OneSearch-MI-Bench and OneSearch-Video-Bench. The 8B model improves over tool-enabled Qwen3-VL-8B by 20.2 and 17.6 percentage points on these benchmarks, with further gains on seven image benchmarks and VideoDR.
In briefWOVEN: Adding visual world modeling to multimodal LLMs via transition reasoning data
SummaryThe paper argues that multimodal LLMs struggle with spatial, physical, and temporal reasoning because of a shared weakness in visual transition reasoning. The authors introduce WOVEN, a benchmark and training source of 36,076 examples organized by scene, action, and reasoning type. Training on WOVEN subsets improved 22 of 26 external benchmarks, by up to 27.3 percentage points.
In briefMAMHOI: Scene-aware human-object interaction via affordance-based factorization
SummaryThis paper presents MAMHOI, a method for generating realistic human-object interactions in complex 3D indoor scenes. It uses an explicit affordance representation: a scene-conditioned model first predicts where and how an interaction can feasibly happen, then an affordance-conditioned model generates the human-object motion. This split lets scene understanding and interaction dynamics be learned from separate data sources, without paired human-object-scene data. Experiments reportedly show fewer object-scene penetrations while preserving interaction quality.
In briefWorldCast: A Distributed Multiplayer World Model That Generates Consistent Views Per Player
SummaryThe paper presents WorldCast, a distributed world model for multiplayer games in which each player runs a local client with a video generator and a state model. Clients exchange only player and scene states rather than relying on joint multi-view generation. Experiments on Counter-Strike 2 report faster player rendering than joint-generation methods, better visual consistency across a round, and real-time operation.
In briefPredicting How Value Training Generalizes in LLMs
SummaryThis paper studies alignment generalization prediction, forecasting how fine-tuning a model on one value changes its behavior on other held-out values. The authors analyze 66 values from modern alignment targets and benchmark representation methods. Activation-based representations, computed when values are applied in context, substantially outperform text-description baselines, reaching correlations of 0.45 versus 0.05.
WhyProposes alignment generalization prediction and validates on 66 values; activation-based representations greatly outperform description baselines (0.45 vs 0.05), with solid evidence.
In briefTesting whether an AI safety benchmark measures one thing: a psychometric audit of harmful refusal
SummaryThe paper asks whether "harmful refusal," a model's tendency to decline dangerous or policy-violating prompts, is a single measurable attribute in HELM Safety. Three of the four candidate datasets are saturated, so the authors analyze HarmBench with item response theory and differential item functioning. They find HarmBench does not appear to measure a single attribute, and argue that aggregated safety scores can hide saturation and mix distinct behaviors.
Christopher M. Stewart, Preston Botter, Natalie Sarabosing et al.
In briefLeWAM: A JEPA-based world action model with diffusion-steered MPC planning
SummaryThe paper introduces LeWAM, a bidirectional transformer that predicts forward, backward, inverse dynamics, and policy outputs in a decoder-free JEPA latent space, trained end-to-end. The authors report that its latent better captures robot and object state than a forward-only JEPA world model and ignores visual distractors better than a reconstruction-based model. Closed-loop results match a flow-matching policy of the same size while also serving as a world model. Planning in the policy head's noise space is reported to improve closed-loop performance over sampling raw actions in MPC.
Shashank Hegde, Alexander Popov, Elie Aljalbout et al.
In briefViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills
SummaryThe paper proposes ViSkill, a framework that stores successful interactions as visual skill cards that vision-language model (VLM) agents can access directly, rather than text-only descriptions. Retrieved skills guide both inference and reward shaping, and successful trajectories are fed back into the library, creating a loop where skill accumulation and policy improvement reinforce each other. Evaluated on Sokoban, FrozenLake, and PrimitiveSkill, ViSkill reports an overall success rate of 0.89, rising to 0.91 with cold-start initialization, and the authors state it outperforms the evaluated baselines and converges faster than standard PPO.
In briefSpaceCast-Bench: A Benchmark for Predictive Spatial Reasoning in Vision-Language Models
SummaryThis paper introduces SpaceCast-Bench, the first benchmark that directly tests predictive spatial reasoning, which means building a scene from observations and reasoning about unseen outcomes of an intervention. It contains 3,862 questions across 182 real-world scenes and 16 task types, organized into static perception, local prediction, and global prediction levels. Evaluating 21 models, the strongest reaches 58.0% versus 87.2% for humans, and the authors also report that fine-tuning on their generated data raises Qwen3-VL-4B from 34.0% to 65.7%.
WhyProposes a new benchmark for predictive spatial reasoning, with systematic evaluation of 21 models, controlled analyses, and fine-tuning validation.
In briefLearning kilometer-scale weather forecasts by aligning global and regional models (ScaleCast)
SummaryThis paper proposes ScaleCast, a framework that uses Global-Regional Alignment to guide high-resolution regional weather forecasts with large-scale predictions from pretrained global weather models. Experiments on ERA5 and CERRA data report improved regional forecasts for surface and upper-air variables, with one model working with Pangu-Weather, GraphCast, and HRES without specific retraining. The authors also fine-tune on 3 km HRRR data and evaluate windstorm cases and station observations.
In briefSpaceFlow: A Training-Free Method for Locally Controllable 3D Generation
SummaryThis paper presents SpaceFlow, a training-free pipeline that generates 3D objects from text and geometric primitives, with each primitive assigned a local control level. Users can specify whether regions should strictly follow the input shape or allow generative completion. The authors report that high-control regions preserve the specified geometry while low-control areas allow plausible shape variation. On fixed geometry, text-conditioned routing achieves strong prompt faithfulness and color/material accuracy.
Neil De La Fuente, Joan Lafuente, Mukhammadali Sayfiddinov et al.
