In briefSemantic communication framework that sends only key traffic information from edge cameras to keep a city traffic digital twin updated in real time.
SummaryThis paper proposes a semantic communication-based edge-cloud framework for real-time urban traffic digital twins. Instead of transmitting full video feeds, edge devices extract and send only task-relevant traffic semantics to reduce communication overhead. Experiments report over 99% transmission savings and an average trajectory reconstruction error of 0.12 m.
In briefVisualizing the World Model Inside an RL Agent's Value Function
SummaryThis paper asks whether model-free reinforcement learning agents secretly learn a model of their environment. The authors apply counterfactual interventions to a Proximal Policy Optimization agent in the Crafter environment and visualize its value function. Although trained only to maximize reward, the value function appears to encode concepts such as when diamonds can be collected, how water and food affect health, and inventory heuristics.
In briefULPack: Lossless compression of model checkpoint sequences using a reference weight
SummaryThis paper addresses the high storage cost of saving many model checkpoints during training, which limits how finely training can be studied over time. It proposes ULPack, a lossless codec that conditions entropy models for integer residuals on the reference weight's exponent, which predicts the displacement bit length without extra side information. Across bfloat16, float16, and float32 trajectories, ULPack reduces storage by up to 27% versus ZipNN, 21% versus zstd-19, and 11.5% versus FM-Delta.
In briefOpen-weight models tested on CTI-REALM without the message limit
SummaryThe paper compares 26 language models, 19 open-weight and 7 closed, on CTI-REALM, a benchmark where an agent turns threat-intelligence reports into working detections. It finds that with enough allowed turns, an open-weight model can match the strongest closed models at less than a tenth of their cost. The authors also introduce stage-level metrics that execute and judge submitted detection queries, showing different models lead different stages.
Sahil Garg, Ruhi Gottumukkala, Vishal Agarwal et al.
In briefCan models "see" what I see? A multi-panel chart question-answering dataset for LLM analysis
SummaryThis paper introduces a visual question answering dataset of synthetic multi-panel scientific charts to test whether large language and multimodal models can reliably analyze figures. Questions follow a four-level description model, from basic concepts to interpretive and analytical understanding. The authors find model accuracy declines as panel count, plot elements, and question difficulty increase, and they consider cost relative to performance. The question-answer pairs, generation code, and chart-element bounding-box metadata are publicly released.
Daniel J. Evans, J. P. Naiman, JooYoung Seo et al.
In briefCan Language Models Play the Wikipedia Game?
SummaryThis paper tests whether language models can navigate from a start page to a target page using only article links, as in the Wikipedia Game. The authors compare sentence embedding models and LLMs across five Wikipedia languages and five academic-literature datasets, using locally built graphs from OpenAlex. No model significantly outperformed humans, though the top models performed comparably well.
In briefPhysically Grounded 3D Articulated Assets: A Multi-Agent Pipeline and Benchmark
SummaryThis paper studies generating articulated 3D assets that capture how objects move and function, not just how they look. It presents Artisan, which extends Articraft with a multi-agent pipeline using physical simulation feedback to yield collision-free motion, gravity stability, and accurate mass and material properties. The authors also build a unified benchmark covering spatial fidelity, physical plausibility, downstream task performance, and overall quality, and evaluate more than 10 representative models.
In briefSplat2Act: Personalizing Robot Pickup Policies from Scanned Scenes
SummaryThis paper presents Splat2Act, a pipeline that uses a phone scan of a building and images of a user's objects to build a photorealistic simulation of the environment. It automatically generates pickup tasks and simulated demonstrations, which are used to fine-tune a pretrained vision-language-action policy, π0.5. The personalized policy reaches 62.5% pickup success in the target scene and 63.9% on a 12-task subset on a real Stretch 3 robot.
Xiaomeng Ye, Krishna Jaganathan, Gyubeum Lim et al.
In briefInstantFusion: A shared latent interface linking different image generators
SummaryThis paper proposes InstantFusion, a lightweight shared latent interface that connects the denoising processes of image generators with different architectures. Model-specific latent autoencoders map intermediate states into a shared latent space and decode them into a target generator's latent space, so denoising can move between frozen generators without a dedicated mapping for each model pair. The authors report competitive results across acceleration, multi-preference composition, cross-architecture on-policy distillation, and cross-model composition.
In briefImages Talk Back: Suppressing Image-to-Text Feedback to Guide MMDiT Generation
SummaryIn multimodal diffusion transformers, the evolving image updates internal text representations through joint attention, and these representations guide later image updates. The authors find this image-to-text feedback can preserve errors as well as correct them, and propose Semantic Feedback Guidance (SFG) and its in-place variant iSFG to suppress it. Experiments on three image backbones report better alignment and preference-model scores, and the control also improves VBench scores for video generation.
WhyIdentifies image-to-text feedback preserving errors in MMDiTs, proposes suppression method SFG, validated across image, video backbones with consistent gains.
In briefHide and Seek: a black-box method that explains vision models using concepts
SummaryThis paper proposes HnS, a concept-level attribution method for interpreting black-box vision models without needing model weights or labeled concept datasets. It retrieves concept-aligned images with a vision-language model, localizes and masks the concept, and measures attribution from the change in model output. On CelebA and COCO-Stuff, HnS agreed better with ground-truth attributions than existing methods, and it was applied to image captioning and satellite image segmentation.
In briefUltraData-Code: Tiered code training data, from raw repositories to task-oriented exercises
SummaryThis paper presents UltraData-Code, a tiered code dataset with four levels: L0 repository archives, L1 standardized code, L2 algorithmically relevant files, and L3 structured programming exercises. In controlled 10B-token continual pre-training of a 1B model, L2 improves pass@1 over L1 on EvalPlus and MultiPL-E, and replacing half of L2 tokens with L3 yields further gains. The authors have released the datasets and classifiers.
In briefConstructive upper bounds on neural networks' topological expressivity
SummaryThe paper proposes measuring a neural network's topological expressivity via the Betti numbers of its output rather than of a decision boundary, making it applicable to regression. It proves that a ReLU network with a specified architecture can represent spaces homotopy equivalent to wedge sums of spheres with arbitrary Betti numbers. For products of spheres, it gives tighter architectural bounds. The authors also conjecture a stronger tight bound for a single sphere.
In briefChronos Policy: Teaching Robots to Learn with Real Time
SummaryThis paper introduces Chronos Policy, a robot control method that explicitly uses observation capture times and action execution times instead of relying on fixed-rate sequence positions. The authors compare timestamp embeddings with time-aware rotary positional encodings in a diffusion transformer and find the latter more robust. Experiments in simulation and on real robots report better results on dynamic and multisensory tasks, similar performance on quasi-static tasks, and improved robustness to changes in sensing and inference timing.
In briefSame Space or Not: A Benchmark for Spatial Intelligence Using Cross-View Contradictions
SummaryThe paper introduces SSoN, a benchmark of 1,000 questions testing whether models keep a consistent scene representation across views. Each question shows three unchanged views of an indoor scene and one edited view, and the model must find the inconsistent view and explain the contradiction. Humans reach 83.3% accuracy while the strongest vision-language models reach 39.7%. Targeted fine-tuning with rationales raised Qwen2.5-VL-7B to 41.79% on the held-out test set.
SummaryModern machines handle equations and language-based logic well but still struggle with visual non-verbal reasoning. The paper introduces GenVerOpt, a framework that generates candidate solutions, verifies them, and iteratively refines them for each sample. The authors report that it surpasses humans on abstract IQ puzzles and matches or exceeds clinicians and experts in medical imaging tasks, including prostate cancer segmentation and liver vessel reconstruction.
Shaheer U. Saeed, Zion Tse, Gregory Slabaugh et al.
In briefKL Divergence Before and After Mixing: Round Complexity Gap in Parallel Sampling
SummaryThe paper studies parallel samplers that randomly choose which coordinates to generate together, comparing the schedule-averaged KL divergence with the KL divergence of the mixed output law. For the uniform distribution on balanced binary strings, the expected number of rounds needed grows as (log n)^2 under the first criterion but only as log log n under the second. The separation is also extended to weighted fixed-count targets and a full-support Beta-Bernoulli target.
In briefReinforcement learning for goal-oriented bundle and flow control in deep-space relays
SummaryThis paper addresses relay-side bundle scheduling, dropping, and flow control in deep-space networks, where contacts are intermittent, delays are long, and relay storage is limited. The authors measure information freshness with an excess Age of Information metric and model relay control as a constrained semi-Markov decision process solved with masked Lagrangian PPO. Simulations in DTNSim are reported to improve age-based utility and inter-flow fairness over baseline schedulers.
Sajjad Baghaee, Aimin Li, Shanmugarajan Balasubramanian et al.
In briefDyssomniaNN: Spiking Neural Model of Sleep Disorders and Feeling Rested
SummaryThe paper presents DyssomniaNN, a multi-scale biophysical model combining sleep homeostasis and circadian regulation, a hypothalamic flip-flop circuit, and a corticothalamic spiking network to simulate five clinical groups over a 72-hour photoperiod. Its synthetic local field potential closely matches human EEG spectra from the PhysioNet Sleep-EDF database (spectral mean absolute error 0.33 to 0.35). Simulations show that total sleep duration does not predict restorative recovery: paradoxical insomnia had normal total sleep time (21.15 h) but a Well-Rested Index of 37.6%.
In briefProtein structure design using natural language prompts
SummaryThe paper introduces Prompteina, an all-atom protein structure generation model that accepts design objectives written in natural language. It was trained on a dataset of 1.8M paired structures and over 50M captions, adapting a pretrained model via text-conditioned training. Prompteina outperforms baselines on 15 of 18 structural properties and on at least 13 of 20 binder design targets. Experimental validation against GM2A yielded nine binders, the strongest with an apparent KD of 107 nM.
WhyUses natural language to condition protein structure generation, with a 1.8M-structure dataset and experimental validation yielding nine binders.
In briefAutoEnvScaling: Automating Data Generation for Terminal Agents
SummaryThis paper introduces AutoEnvScaling, which treats the design of reinforcement learning environments as a terminal task that agents perform themselves. A proposer agent builds new environments from the solver's rollouts, and the solver's new rollouts guide the next round. The authors report up to 25 times higher valid-environment rate and up to 75% lower cost per valid environment compared with prompt-only generation. They also report benchmark gains from training and note that the pipeline generalizes to other agentic domains.
WhyTreating environment design as a terminal task and automating RL environment generation is novel, but experimental scale and baseline details appear limited.
In briefWhere to look: rethinking how 3D medical AI samples image patches during self-supervised pretraining
SummaryThis paper studies how sub-volumes are chosen when pretraining models on 3D medical images without labels. The authors show that better-chosen sub-volumes lead to better transfer to downstream tasks, and propose VolumeProbe, which replaces random cropping with a coarse-to-fine probe bank guided by anatomy, informativeness, and diversity. The lightweight VolumeProbe-Lite consistently outperforms random sampling across multiple backbones and downstream tasks.
In briefRSI-Anything: A Domain-General Framework for Recursive Self-Improvement Using World Models
SummaryThis paper introduces RSI-Anything, a framework built on JEPA-Anything that uses predictive world models to guide action and learning-data selection. Observed feedback is used to improve policies or workflows and to adapt the world model under conditions, with checks governing which updated policy-model pairs are kept. The authors report experiments across code repair, GUI tasks, laboratory planning, and reaction optimization, including a 96.15% mean best observed yield versus 82.16% for Bayesian optimization under the same 40-experiment budget.
In briefDo vision-language models lose what they know? Spatial relations are encoded less faithfully than words
SummaryThe paper introduces EICS, a metric that tests whether a model's representations consistently encode concepts under controlled one-concept interventions, and CASP-VL, a benchmark of spatial perturbations built from Visual Genome scene graphs. Across several VLM architectures, spatial relations are represented substantially less faithfully than lexical information, even when the models score well above chance on compositional benchmarks. The relation is identifiable at the predicate's own tokens but is lost when the sentence is pooled into a global representation via the EOS token or mean pooling.
In briefNext-Concept Modeling: Faster Text Generation Using Coarse-to-Fine Steps
SummaryThe paper introduces Next-Concept Modeling (NCM), a two-stage framework in which a multiscale residual VQ-VAE encodes text into hierarchical discrete codes and a transformer generates one level at a time, predicting all codes in a level in parallel. The authors report that NCM runs substantially faster than diffusion language model baselines on LM1B, OpenWebText, and XSum while matching or exceeding their quality. On Sudoku reasoning, NCM outperforms autoregressive and diffusion baselines across difficulty levels with fewer function evaluations.
Hermann Kumbong, Ishan S. Khare, Iddah Mlauzi et al.
In briefRevisiting Deep Regression Objectives for Tabular Data and Global Early Stopping Limits
SummaryThis paper systematically re-evaluates regression objectives for deep learning on tabular data across many established datasets and neural backbones. It finds that most proposed alternatives to Mean Squared Error offer limited or inconsistent gains over a tuned MSE baseline, while quantile loss consistently improves aggregate results. The authors show that regions with different target noise peak at different training epochs, so a single global early-stopping checkpoint forces a compromise, and they propose a heteroskedasticity-aware quantile loss using input-dependent label smoothing that improves on several backbones.
Danil Rudenko, Georgii Bulgakov, Nikolay Kartashev et al.
In briefHitting-time bounds for single-trajectory asynchronous Q-learning
SummaryThe paper studies the last iterate of standard asynchronous Q-learning run along one trajectory of a finite Markov decision process. It proves that zero-initialized Q-learning with a constant step size reaches sup-norm error at most epsilon with high probability using roughly the stated number of transitions, which includes a hitting-time term. An empirical count rule is also proposed to select a constant step size without prior model knowledge, and matching lower bounds are discussed.
In briefAutomata that keep neural meaning parsers' outputs valid
SummaryNeural semantic parsers produce meaning representations one token at a time, but validity depends on references, scope, and graph relations that emerge during generation. The paper studies structural automata at three stages of DRS parsing (decoding, post hoc processing, and training) and also evaluates decoding on AMR. It introduces accepting-path decoding (APD), which keeps a verified completion within the length limit and guarantees an accepted output without retracting tokens. Under tight length limits, APD improves Smatch++ over local masking on both DRS and AMR.
In briefHow task context changes the roles of components in language models
SummaryThis paper studies how the same attention heads in large language models play different roles across tasks. Using identical graph facts as inputs, the authors measure how interventions affect attention heads and later-layer activation states. Across five open-weight models, task changes redistribute this influence. In Gemma-2-9B and Qwen3-14B, editing an early activation vector switches which fact the model reads, and the same heads respond to different facts.
In briefBeyond the Snowflakes: Separating Weather Effects from Confounding Factors in LiDAR Detection
SummaryThis paper evaluates LiDAR object detection on the WADS-3D and CADC datasets and finds that accuracy improves as snow intensity rises. Ablation and meta-information analysis indicate that snow acts as noise, but a concurrent increase in object point count outweighs its negative effect. The authors argue that grouping evaluations by weather severity alone can mislead, and that atmospheric noise should be separated from surface snow accumulation effects.
In briefRATED: Reliability-aware retrieval of tools for AI agents, using execution feedback
SummaryThis paper proposes RATED, a framework that helps AI agents retrieve external tools by combining semantic matching with real-time tool reliability. It uses a frozen dual-encoder for semantic matching and a lightweight contextual bandit that re-ranks candidates based on execution feedback, avoiding model retraining. The authors also introduce EcoAgent, a benchmark of 5,172 real-world tools and over 210,000 query-tool pairs.
In briefX-AgentRed: Stress-testing personal AI agents across evolving multi-service tool ecosystems
SummaryThis paper presents X-AgentRed, a self-evolving red-teaming framework that finds harmful tool-call paths personal agents might take across multiple web and app services. It builds a directed graph of tool dependencies, where one tool's effects satisfy another's preconditions, and samples realistic scenarios from it. Using execution feedback, it diversifies harmful paths through tool substitution and refines blocked steps. Across four frontier agentic models, the authors report 1.77 times the harmful path rate of STAC and 2.99 times the average tool coverage.
In briefDisentangling task-relevant dynamics to learn reliable video-based rewards for robot manipulation
SummaryThis paper proposes DiTaR, a framework that learns dense rewards from heterogeneous videos to help reinforcement learning in robotic manipulation. It uses an asymmetric two-stage procedure to suppress task-irrelevant visual variation and learn a task-relevant latent representation, then models task dynamics with a history- and language-conditioned rectified-flow model. The reward is derived from temporal changes in dynamics compatibility. The method improves policy learning on MetaWorld-v3 and ManiSkill over video-based reward baselines and transfers to real-world offline RL through reward relabeling.
In briefVideo-Arena: A scalable game-based benchmark testing whether AI models track changing world states in videos
SummaryVideo-Arena is a benchmark that uses competitive game footage to test whether models can understand and track evolving world states in videos. It evaluates three task types: perceiving critical events, interpreting state updates, and tracking the resulting states, with reference labels generated at scale from game records. A curated golden set of 180 tasks across 9 games was evaluated under native visual understanding and tool-assisted agentic analysis. Native models struggled to perceive events reliably, while tool-assisted systems improved event recovery but still had difficulty interpreting state updates and tracking states consistently over time.
In briefPrediction Alone Is Not Enough for a Sufficient Memory State
SummaryThe paper studies a memory that reproduces every current prediction yet may have lost what is needed to learn from the next observation. It characterizes this gap for exact cumulative learning and attention. Results include the state dimension required for polynomial losses, the hidden evidence cost in ridge regression, and the error scale under bounded compression. The work clarifies what future updates reveal rather than proposing a new regression algorithm.
In briefOpenMobile-2: A Versatile Mobile Agent Built with Scalable Environments and App-Native Tools
SummaryOpenMobile-2 is a near-frontier mobile agent whose training environments and methods are fully open. The work introduces MobileGym++, a simulated environment with 35 commercial-style apps that together with Android emulator apps covers over 110 apps, plus about 12K training trajectories and over 2K verifiable reinforcement learning tasks. It also explores combining GUI actions with app-native tool calls, and reports competitive results on AndroidWorld (79.9) and MobileWorld (50.4), with SPA-Bench rising from 31.9 to 59.6.
In briefTracking the Invisible: Self-Supervised Wheel Track Detection Using Thermal-RGB Fusion
SummaryThis paper proposes a low-cost, self-supervised method that detects wheel tracks on snow-covered roads using synchronized RGB and long-wave infrared images. Training labels are generated automatically by projecting future wheel trajectories from recorded GNSS data. The authors report that the approach works in visually featureless winter conditions and release the ThermalTrack dataset.
In briefZeroth-order residual tracking for optimization with partial gradient access
SummaryThe paper proposes PISO, a stochastic optimization method for problems where only part of the gradient is directly available and the rest must be estimated from function evaluations. It centres zeroth-order observations on a running residual estimate so that probing variance depends on tracking error. For smooth nonconvex objectives, the authors give convergence guarantees and compare PISO with baselines across five settings.
In briefAsk Your Advisor: Teaching AI Agents When to Consult a Stronger Model
SummaryThis paper introduces ARIA, a meta-controller that learns when a cheaper executor model in an AI agent workflow should consult a stronger advisor model. It treats consultation as a sequential resource-allocation problem, balancing task success against consultation cost. According to the paper, across SWE-bench Verified, Terminal-Bench 2.1, and τ²-bench, ARIA recovers 44-81% of the executor-advisor success gap at less than half the advisor's cost.
Pascal Sowietzki, Steffen Holter, Yannick Metz et al.
In briefAnimating water in static 3D Gaussian Splatting scenes for realistic flow
SummaryThis paper proposes a method to add plausible water motion to static 3D Gaussian Splatting (3DGS) scenes without dynamic capture or physical simulation. It estimates 2D optical flow from reference images or renderings, lifts it into a consistent 3D Eulerian motion field, and uses it to deform Gaussian primitives into seamlessly looping animations. The authors report real-time rendering and free-viewpoint navigation are preserved, with improvements over prior work on a newly captured waterfall dataset.
Kjell Keune, Florian Hahlbohm, Jannis M\"oller et al.
In briefLapMerge: A differentiable, exact-budget token merging method running in linear time
SummaryThis paper introduces LapMerge, a differentiable operator that merges tokens in ordered sequences. A learned mass distribution under a Laplace-kernel density splits the sequence into exactly N equal-mass segments, which are aggregated with prefix and suffix scans, so both forward and backward passes run in O(M+N) time. The authors report that it matches or improves on strong token reducers on vision and vision-language tasks, and is up to two orders of magnitude faster at large sequence lengths.
Magdalena Tr\k{e}dowicz, Aleksandra Nowak, Tomasz Steifer et al.
In briefRethinking satellite-based assessment of building damage after disasters
SummaryThis chapter revisits automated building damage assessment (BDA) from satellite imagery, organized around data, evaluation, and methods. It proposes a sensor-agnostic, instance-based task definition that tolerates missing pre-event imagery, and argues that the prevailing single aggregate score is theoretically fragile. The authors lay groundwork for a ranking-based, multi-score evaluation protocol. They also review methods from task-specific deep networks to Earth observation foundation models, noting that generalization to unseen events and regions remains an open difficulty.
Anthony Cioppa, Virgile Blua, S\'ebastien Pi\'erard et al.
In briefAction recognition from event camera data using multi-frame pre-training
SummaryThis paper studies action recognition from event camera data, which captures fast motion and offers better privacy than conventional cameras. Single-frame event representations lose discriminative power on long and complex sequences, so the authors propose a multi-frame representation. They pre-train a Vision Transformer with a masked auto-encoding framework, using sparsity-aware masking, token selection, and multiple position-aware mask tokens. The method reports state-of-the-art results on THU-EACT-50 CHL and DailyDVS-200.
In briefFinite-Sample Error Guarantees for Fully Discrete Conditional Diffusion Models
SummaryThis paper provides finite-sample guarantees for conditional diffusion models that are learned from data and sampled with a finite number of denoising steps. The authors align the score-training objective with the sampling grid so that statistical estimation error and numerical discretization error can be analyzed together. Under a smooth conditional density model, they derive high-probability bounds on the error of the generated distribution and quantify the number of score evaluations required. The total-variation error bound matches existing conditional distribution-estimation rates up to logarithmic factors.
In briefDynamic compute allocation for more efficient nanobody design across scaffolds and epitopes
SummaryThis paper studies how to allocate computing resources during generation for de novo nanobody design. The authors treat choosing conditions such as scaffold and target epitope as an inference-time resource-allocation problem. They propose a joint Global Sequential Monte Carlo method and a decoupled method adapted from Batched Successive Elimination. According to the paper, these allocators raise the yield of designable candidates per GPU-hour nearly 5 times without reducing diversity across targets and configurations.
In briefReturning to Development States: Whether Full Sandbox Snapshots Help Coding Agents
SummaryThe paper studies whether coding agents can recover from failures by restoring intermediate environment states and branching, rather than restarting from the initial state. The authors implement Go-Explore-style search over full Daytona snapshots. They find that higher and more efficient solve rates depend mainly on progress outside the repository, such as databases, services, and other runtime state.
In briefLarge-Scale AI Study of Comics: The PAGE Framework Tested on American Comics and Japanese Manga
SummaryThis paper proposes PAGE, a framework that uses AI to parse comic scans into panels, text and bibliographic records, and annotate pages with a vision-language model following a codebook. The authors check samples against human annotation and apply it to 3,929 American comic books from 1938 to 1956 and 87 Japanese manga volumes. Three studies cover how political power was portrayed, the visual form of pages and artists' styles, and who speaks across both traditions.
In briefAxiom-0: Hierarchical Self-Refinement for Vision-Language-Action Robot Policies
SummaryThis paper presents Axiom-0, a vision-language-action policy for robot manipulation that refines intermediate representations during training and recursively updates conditioning before generating actions. On the RoboCasa365 benchmark it reports 62.32% overall success, 4.22 points above the highest published leaderboard score (Paimon-0, 58.1%). Under the evaluated profiling setup, mean action-chunk generation latency rises by 1.7% over Xiaomi-Robotics-1.
In briefSpecialists Learn Faster: how organizing knowledge during pre-training affects later model updates
SummaryThe paper compares language models trained on knowledge in chronological order, which become specialists for recent information, with models trained on shuffled data, which become generalists performing evenly across periods. Specialists need far fewer update steps to acquire new knowledge. The authors derive a law relating update cost to optimization dynamics and design a pre-training data organization strategy that keeps stable historical knowledge while enabling efficient updates.
In briefA study of design choices for continuous diffusion language models
SummaryThis paper examines two key levers in training continuous diffusion language models: self-conditioning and embedding geometry. It introduces DoubleCat, which improves TinyGSM accuracy more than threefold without an extra forward pass, and SemiCat, which reaches discrete diffusion-level performance. The authors also find that cross-entropy shapes embeddings so their decoding error rate closely matches one-hot embeddings.
In briefSpatialASM: Training Vision-Language Models to Imagine Viewpoints for Spatial Reasoning
SummaryThe paper introduces SpatialASM, a framework that teaches Large Vision-Language Models to actively revisit scenes from task-relevant viewpoints and mentally reconstruct spatial layouts. It uses an allocentric novel-view generator during training to produce explicit viewpoint-planning trajectories and automatically built perspective-taking tasks. Across seven spatial benchmarks and four model backbones, the authors report higher average accuracy while general vision capabilities are retained.
In briefMC-EvoBench: Testing whether Minecraft agents truly learn reusable skills from experience
SummaryThe paper introduces MC-EvoBench, a controlled Minecraft benchmark for checking whether agents turn past interactions into capabilities that transfer to new worlds. It uses 19 evolution chains and three paired transfer probe families, comparing no-prior, related, unrelated, and shuffled experience under matched step budgets and five seeds. Results show a consistent direct-capability gap between controller and planner-backed variants, but related experience yields no stable transfer, with all intervals including zero.
In briefTrajCFG: Per-token guidance in image and video generation, adjusted using past denoising steps
SummaryThe paper introduces TrajCFG, a training-free method for flow-matching generative models that sets guidance strength separately for each latent token. It uses the history of changes in guided velocity direction, giving stronger guidance to tokens whose direction changes more and weaker guidance to those with consistent directions. Experiments on text-to-video, image-to-video, text-to-audio-video, text-to-image, and image-to-3D generation report gains over base samplers, including a 0.30 to 0.75 point VBench Total improvement for text-to-video across three backbones.
In briefOpenPA: Multi-level pronunciation scoring without forced alignment, using mixed transcripts
SummaryThis paper introduces OpenPA, a framework for automatic pronunciation assessment that uses Whisper's cross-attention to predict word- and utterance-level scores without forced alignment. Its mixed-transcript adaptation and adaptive substitution weighting let one model handle both closed and open responses. Trained only on speechocean762, it performs best overall among the evaluated systems in in-domain closed and simulated open-response tests and stays zero-shot on MultiPA. A cross-task test on oral-cancer speech suggests it may apply beyond non-native speech.
In briefSame Query, Different Answer: Run-to-Run Variance in Semantic Query Processing
SummaryThis paper measures how much LLM-based semantic query engines change between repeated executions at temperature 0. The authors re-ran text and tabular workloads from three benchmarks, logged about 1.2 million model calls, and reanalyzed SemBench's published results. They report that some rows change output across runs, and that query plans amplify these changes at the answer level. The paper closes with reporting rules for benchmarks and design rules for engines.
WhySystematically measures run-to-run variance in LLM semantic queries via 1.2 million logged calls, reanalyzing published benchmark results with strong evidence and practical implications.
In briefCan Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection
SummaryThis paper examines whether language models can learn when to answer factual questions and when to abstain. Using Qwen3-14B and a scoring scheme that rewards correct answers, penalizes mistakes, and allows abstention, the authors compare training the model directly with training a separate answer selector. On TriviaQA, a learned selector beats the original model's calibrated confidence filter, raising coverage and lowering the error rate. On NQ-Open it still underperforms always abstaining, so reliable abstention across datasets remains unresolved.
Juan Terven, Diana-Margarita C\'ordova-Esparza, Ruben Gomez et al.
In briefAuditing Predictive Signal Beyond Expert Features via Shared Component Removal
SummaryThis paper studies whether black-box models trained on raw inputs capture predictive signal beyond the expert features that clinicians typically use. It proposes Shared Canonical Component Removal (SCCR), an audit method that removes only the directions strongly shared between raw and expert embeddings. The method is evaluated on synthetic data, coronary artery disease detection, antidepressant treatment dropout prediction, and skin lesion classification with SkinCon annotations. The authors report that SCCR identifies where expert views are insufficient and recovers residual predictive signal with less redundancy.
Jose Roberto Tello Ayala, Akl C. Fahed, Roy Perlis et al.
In briefTabMF: Learning Dataset Representations with Tabular Foundation Models
SummaryThis paper introduces TabMF, a framework that adapts pretrained tabular foundation models into encoders for supervised datasets, using contrastive post-training on synthetic dataset views. It is instantiated on TabICLv2 and TabPFN-2.5 and evaluated on six tasks using unseen real datasets. The two TabMF encoders achieve the strongest aggregate performance among the compared dataset representations, notably in dataset retrieval, compact benchmark construction, and distribution-shift severity assessment.
Rafiq Kamel, Nicolas Atienza, Aziz Fouch\'e et al.
In briefDream to Remember: Continually Improving Generalist Robot Policies with World Models
SummaryThe paper proposes Dream to Remember (D2R), which steers a frozen generalist robot policy toward new tasks with a small steering module while avoiding catastrophic forgetting. A frozen pretrained world model replays earlier tasks through imagined rollouts, requiring no replay buffer or new data. On CALVIN, D2R improves eight tasks in sequence entirely in imagination, raising average success from 63.7% to 74.8%, while LoRA fine-tuning drops it to 11.7%.
Bahey Tharwat, Iman Nematollahi, Marvin Alles et al.
In briefUnified Diffusion and Flow Posterior Sampling via Exact Gaussian Conditioning (β-Flow)
SummaryThis paper proposes β-Flow, a plug-and-play method for linear inverse problems that unifies flow, diffusion, and decorrelated trajectory priors through a correlation parameter β, with a confidence parameter κ controlling how strongly the path distribution is enforced. Each step uses one denoising network evaluation followed by an exact Kalman-style Gaussian conditioning update. Across 7 inverse problems on 4 datasets, it reports state-of-the-art-level PSNR and SSIM, and finds strong priors work best when observations are less informative while low correlation is preferred when observations are more informative.
In briefFiltering hallucinated query expansion terms from LLMs while keeping retrieval quality
SummaryLLM-generated query expansion terms often include non-existent entities or off-topic words that add ranking noise. The paper proposes SAFE-QE, a training-free, retriever-agnostic method that down-weights terms reproduced across randomized contexts but unsupported by pseudo-relevant documents, and selects terms via constrained coverage maximization. On TREC and BEIR benchmarks, it matches baselines on nDCG@10 while injecting far fewer unsupported terms.
In briefError rates alone miss the costs: augmentation trade-offs in low-resource speech recognition
SummaryThis paper studies data augmentation for low-resource automatic speech recognition across four low-resource languages and English. It compares methods that modify existing audio (speed perturbation, SpecAugment) with methods that modify transcriptions and synthesize audio from them. Phoneme-level error analysis shows the methods help for complementary reasons, and combining them lets each reduce the errors the other introduces.
In briefFORTE: A production framework for evaluating and removing features in industrial recommender systems
SummaryThe paper presents FORTE, a model-agnostic production framework for discovering, validating, and removing features in recommender systems. It uses grouped ablation to evaluate candidate feature sets that account for interactions, and lineage tracing to identify all downstream consumer models. Validation on two large-scale recommender systems found deprecable feature sets that caused no online regressions. On a large ranking model, the launch reduced trainable parameters by 29.9% and improved serving throughput by 19.8–25.5%.
In briefConstrained Probabilistic Numerics: Gaussian process solvers for PDEs with inequality constraints
SummaryThis paper extends Gaussian process solvers for linear PDEs so that solutions can also satisfy linear inequality constraints, such as nonnegativity, monotonicity, or no-arbitrage conditions. The authors prove that joint conditioning on equality and inequality information is equivalent to sequential conditioning, and the combined posterior is computed using truncated-Gaussian moments. Experiments on a Poisson problem, S&P 500 option prices, and air-quality fields report reduced errors and fewer constraint violations, though the gains vary by data regime.
WhyProves equivalence of joint and sequential conditioning for inequality constraints in GP PDE solvers, with experiments on Poisson, option, and air-quality data.
In briefAdaptive variance reduction that works without knowing smoothness constants
SummaryThis paper studies variance reduction for stochastic non-convex optimization and proposes a two-phase algorithm that adapts to average smoothness, standard smoothness, and noiseless settings without knowing smoothness constants or the variance bound. With gradient budget T known, the rates are O(T^{-1/3}) under average smoothness, O(T^{-1/4}) under standard smoothness, and O(T^{-1/2}) without noise. When T is unknown, a two-path method achieves these rates up to logarithmic factors in stochastic settings.
In briefREFLECT: Role-aware detection of sentiment, emotion, politeness, and formality
SummaryThis paper proposes a role-aware transformer that links four communicative dimensions (sentiment, emotion, politeness, and formality) to semantic role labeling configurations and incorporates them into the self-attention mechanism. A learned gate fuses these role-aware representations with contextual embeddings. Across four benchmark datasets, the model shows statistically reliable gains over BERT-based baselines for emotion, politeness, and formality, while sentiment gains are marginal and non-significant. Evaluation on real-world mental-health data indicates improved predictive performance and traceability.
Sangpil Youm, Annabel Stocks Natalias, Hayden Heninger et al.
In briefBeyond Evaluation: How Language Models Respond to Discourse Particles like "Just" and "Only"
SummaryThis study examines how large language models respond to the English discourse particles "just" and "only" in natural conversation, using behavioral and mechanistic analyses. Both models and humans prefer these particles in context, with models more sensitive to "just" than "only". Instruction tuning increases this sensitivity, and the internal overlap between the two particles is asymmetric. The authors release their code, data, and human annotations.
In briefOccludeBench: Can Vision-Language Models Infer Hidden Objects?
SummaryThis paper introduces OccludeBench, a manually curated benchmark testing whether vision-language models can recover fully occluded objects from remaining visual cues. Across five models, accuracy averages 42.7% on direct cues such as shadows or reflections, versus 62.7% on inferential cases requiring scene context, a 20.0-point gap, while humans show no significant gap. The authors propose OccludeTrace, a four-stage diagnostic, which locates the main bottleneck in recovering the relevant object from the full image. An inpaint-then-confront intervention, generating a candidate object and checking it against the visible cue, reduces the gap to a statistically non-significant difference for models with image-generation capabilities.
Jasin Cekinmez, Addison J. Wu, Ryo Mitsuhashi et al.
In briefRandomized momentum transport: skipping some costly gradient corrections to cut optimization expense
SummaryThe paper proposes randomized momentum transport, which computes the momentum transport correction with probability q and otherwise uses a cheap prediction, with importance weighting that preserves the conditional mean. Under stated assumptions it gives a stationarity bound for linear minimization oracle updates. Experiments on MNIST1D and CIFAR-10 at roughly equal oracle cost report lower training loss and higher test accuracy for some settings.
In briefOptimal algorithms for constrained nonsmooth nonconvex stochastic optimization
SummaryThis paper presents new algorithms for stochastic optimization problems whose objective is neither smooth nor convex, under a convex constraint set. The algorithms return (δ,ε)-Goldstein stationary points and are the first to achieve optimal convergence rates for this task in several settings. Stochastic-gradient methods reach an Õ(δ^{-1}ε^{-3}) rate, and zeroth-order methods reach Õ(dδ^{-1}ε^{-3}), both shown to be optimal.
WhyFirst optimal rates for constrained nonsmooth nonconvex stochastic optimization, with matching lower bounds via rigorous complexity analysis.
In briefPaper showing discrete graph diffusion models secretly memorize training graphs
SummaryDiscrete graph diffusion models corrupt graphs by adding independent noise to nodes and edges. Analyzing models trained on two-block stochastic block models, the authors find that when the reverse process is conditioned on a corrupted training graph, the model almost perfectly recovers it, while held-out graphs are not recovered similarly. They call this conditional memorization, which is hidden during unconditional inference, grows with a generalization gap, and is not prevented by early stopping. It appears before community structure emerges in generated graphs.
In briefUkrMeme: Explainable Multimodal Detection of Persuasion in Ukrainian Memes
SummaryThis paper studies detecting persuasion techniques in Ukrainian memes using vision-language models. It introduces UkrMeme, a dataset of 200 wartime Ukrainian memes annotated for persuasion techniques, and PTM-UA, a Ukrainian translation of a subset of the Propaganda 950 benchmark that keeps the original images. Models are evaluated zero-shot on image, text, and multimodal inputs using hierarchical F1, and two explanation methods, Layer-wise Relevance Propagation and model-generated token attributions, are compared by the semantic categories of evidence they emphasize.
In briefSimVTA: A simulation-based method that aligns driving video and trajectories for self-driving world action models
SummaryThis paper proposes SimVTA, a post-training framework for driving world action models that renders future videos for candidate trajectories using a simulator. It optimizes trajectory planning with reinforcement learning and uses video supervision to update shared representations. On NAVSIM v1, PDMS improves from 86.2 to 91.3, and the method reaches 88.2 EPDMS on NAVSIM v2.
In briefComparing methods for detecting very rare dead cells in fluorescence microscopy images
SummaryThis paper compares computational methods for automatically detecting apoptotic (dead) cells in fluorescence microscopy images. It evaluates heuristic intensity-based methods, classifiers using numerical and image features, and anomaly detection approaches on 258 images containing 65,527 cells or nuclei, with 243 manually annotated dead cells. ResNet-18 and XGBoost performed best, significantly outperforming the heuristic and anomaly detection methods.
In briefReplaySSM: Caching SSM Inputs Instead of States to Speed Up Hybrid Model Decoding
SummaryThis paper introduces ReplaySSM, which represents a state-space model's recurrent state as a periodically updated checkpoint plus a buffer of recent inputs, reducing state reads and writes during decoding. Implemented in vLLM for Mamba-2 and Gated DeltaNet, it reports 1.20 to 1.47 times faster standard decoding across four hybrid models at batch size 256. On two 120B-class models, the authors report about 1.8 times the throughput of vLLM's speculative decoding.
In briefREAD: A Single-Pass Method for Explaining Video Predictions in State-Space Models
SummaryThis paper proposes READ, a parameter-free attribution method for selective state-space models used in long-sequence video modeling. It unrolls the recurrence at the classifier's readout and traces the contributions accumulated in the recurrent state back to the video tokens behind the final prediction. READ needs only one forward pass to produce a spatiotemporal relevance map, and it achieves faithfulness and localisation competitive with far more expensive methods while avoiding systematic relevance biases of existing approaches.
Magdalena Tr\k{e}dowicz, Piotr Kubaty, Arkadiusz Lewicki et al.
In briefBeyond Gate Weights: Decision-Level Radiomics-Image Fusion Interpretation for Kidney Cancer Classification
SummaryThis paper examines a fixed 3D MedVAE-radiomics model that classifies clear cell versus non-clear cell renal cell carcinoma from contrast-enhanced CT. The binary decision logit is decomposed exactly into a gated image term, a direct radiomics term, and a bias. The authors show that structural gate weighting and effective branch contribution differ substantially, so they should be interpreted separately.
Yuan Liang, Sourav Bhattacharjee, and Abraham Campbell
In briefNorm-aware multi-objective optimizer with entropic task weighting and convergence guarantees
SummaryThis paper extends multiple-gradient descent, usually formulated in Euclidean geometry, to general finite-dimensional norms by combining a norm-constrained linear minimization oracle with entropic updates of task weights. The authors prove an O(T^{-1/2}) ergodic stationarity rate in the deterministic setting and an O(T^{-1/4}) rate with stochastic gradients using a persistent momentum buffer. They also analyze regret and inner-gap control for the weighting objective, and compare the method with existing multi-task optimizers in experiments.
In briefMARGoS: A configuration-driven platform for reproducible multi-agent reinforcement learning experiments in the ARGoS robot simulator.
SummaryThis paper introduces MARGoS, a lightweight platform that manages the settings of multi-agent reinforcement learning experiments run in the ARGoS swarm robotics simulator. It provides configuration hashing, end-to-end seed propagation, frozen configurations, structured logging, environment fingerprinting, and exportable experiment bundles. The authors evaluate the platform across five dimensions and argue that managing experiment variation this way makes experiments easier to reuse, reproduce, and hand off. The code is publicly available on GitHub.
Tom Philipp Backert, Mohammad D. Soorati, Javad Ghofrani
In briefFrom Pixels to Privacy-Sensitive Features: An Information-Theoretic Bound on Reconstruction Error
SummaryThis paper studies how much an attacker can recover from training data when a model is trained with DP-SGD. The authors derive a distortion bound that can be computed before training and holds even if the attacker observes the parameter trajectory during training. The framework combines information leaked by training with uncertainty already present in the data distribution to lower-bound the expected error of any reconstruction attack. It is instantiated in pixel space and LPIPS feature space.
In briefDualQuant: A calibration-free method for quantizing LLM weights and activations to low bit widths
SummaryThis paper introduces DualQuant, a post-training quantization framework that quantizes both weights and activations of large language models without calibration data, targeting accelerators with native 4-bit support. It minimizes weight-reconstruction mean squared error, and the resulting scaling factors serve as per-channel activation rescalings. The authors report that DualQuant matches or exceeds calibration-based SmoothQuant on most Llama and Qwen models and formats, and achieves up to 2.61x memory reduction and 2.46x speedup over BF16 on an RTX 4090.
In briefShortcut Learning in Disaster Damage Assessment Models
SummaryThis paper studies geospatial foundation model embeddings for assessing disaster damage on the xView2 and BRIGHT datasets. Holding out entire disasters, rather than random image tiles, substantially lowers balanced accuracy, suggesting models rely on shortcuts such as event context and building count. The authors also show that feature transformations learned only from training disasters can improve generalization to unseen disasters.
In briefWhen Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors
SummaryDiscrete graph diffusion models corrupt edges independently, so their terminal prior is an Erdős–Rényi random graph, which makes it hard to recover node heterogeneity without node features. The authors propose a Latent Sociability Prior and a joint diffusion process in which latent node sociability variables and graph structure co-evolve, removing edge independence while preserving exchangeability. Experiments on several graph generation datasets reportedly show that the method captures graph distributions more accurately than baselines without node features.
In briefBeyond Discrimination: Probability Calibration in Radiomics-Image Fusion for Kidney Cancer Classification
SummaryThis study uses 396 patients from the KiTS23 dataset to test whether adding radiomics features to four image-model families improves probability predictions, not just discrimination. It compares 17 individual models and four probability-averaging ensembles. In 12 matched comparisons, fusion improved AUC and AP in all cases, and Brier score and negative log likelihood in 11 of 12, while calibration error improved in eight. Probability averaging gave better point estimates on all five primary metrics, with stronger bootstrap support for Brier score and NLL than for calibration error.
In briefHow well prior-data fitted networks approximate Bayesian prediction
SummaryThis paper examines how accurately prior-data fitted networks (PFNs) approximate the Bayesian posterior predictive distribution of a synthetic prior. The authors split deployment error into prior, context, and amortization error, measuring amortization error against a reference distribution under a realistic prior. They find PFNs recover the target distribution well overall, but residual error persists in rare settings and shows up as disagreement between independently pretrained PFNs. Several Bayesian model averaging methods consistently reduce this uncertainty and improve predictions, including for full-size TabICLv2 models.
Manuel H\"ulskamp, Julius Kobialka, Emanuel Sommer et al.
In briefNeurons++: A unified framework reconstructing images and videos from fMRI, with neural-to-visual interpretability analysis
SummaryThis paper introduces Neurons++, a framework that reconstructs static images and dynamic videos from fMRI brain recordings. It first pretrains on image-fMRI data for a stable spatial-semantic starting point, then fine-tunes separately for each subject's image and video reconstruction, using a serial, cortex-inspired decoding design. The authors report state-of-the-art video reconstruction and competitive image reconstruction, along with analyses of controllability under neural perturbation.
In briefGaussian splatting that uses visible scene cues to model near-field reflections
SummaryThis paper proposes a novel view synthesis framework built on 2D Gaussian Splatting to better render view-dependent specular reflections, especially near-field ones. It uses environment information that can be directly observed in the input images, and combines a hybrid reflection appearance representation with view-dependent specular modeling. Experiments on public benchmarks and a newly curated near-field reflection dataset report better rendering quality and geometric accuracy.
In briefVETNET: A framework of benchmark datasets for AI-driven veterinary innovation
SummaryVETNET is an ecosystem of live benchmark datasets and evaluation frameworks for advancing AI in animal science and veterinary medicine. It uses standardized, high-quality datasets and collaboration among academic and non-academic institutions. The framework aims to close the gap between AI progress in human medicine and the fragmented state of veterinary AI development.
Katherine J. Koebel, Renata Ivanek, Parminder Basran et al.
In briefMatched-Context Group Policy Optimization (MCGPO): training LLM agents to credit intermediate decisions
SummaryThe paper proposes MCGPO, a method that builds weighted action preferences from already collected rollouts to assign credit to intermediate decisions from sparse terminal feedback. It is critic-free and is applied alongside GraphGPO task learning. Across six environment-backbone settings with three seeds each, MCGPO reports higher mean success than matched GraphGPO on ALFWorld and WebShop.
In briefStudying how task variety shapes neural networks using small math worlds
SummaryThis paper introduces two procedurally generated mathematical task ecosystems, IntWorld and PermWorld, built from related but distinct tasks on integers and permutations. The datasets simulate task diversity within a cohesive mathematical world and allow the number of tasks to be precisely varied. This enables the authors to begin separating the role of task variety from other training factors. Initial results support the idea that greater task diversity enriches a model's internal representations.
In briefA paper on improving reasoning by having a teacher critique a model's own solution before scoring it
SummaryThe paper examines whether scoring signals in on-policy self-distillation can locate errors in a model's reasoning, finding that they cannot reliably do so. The authors propose critique-conditioned self-distillation, which first generates a critique of the current rollout and then conditions the scoring teacher on that critique. The method localizes errors better, shifts probability toward backtracking tokens, and is reported to help at test time and in cold-start GRPO.
Mayug Maniparambil, Ayush Rai, Ekansh Verma et al.
In briefSmolWAM: A Latent World-Action Model for Edge-Device Robot Manipulation
SummaryThis paper presents SmolWAM, a latent world-action model for robot manipulation designed to run on edge devices. It is derived from the 2B-parameter Cosmos Policy, keeping its pretrained trunk while using 12 of 28 transformer blocks and predicting actions and future visual latents in one rectified-flow step. On 400 LIBERO episodes it succeeds 95.5% of the time at 57 ms per 16-action chunk, about 11x faster than its teacher.
Timi Owolabi, Hikmah Olawore, Frances Adelakun et al.
In briefSafeDiSTL: a method that splits shared safety constraints among agents in multi-agent reinforcement learning.
SummaryThis paper presents SafeDiSTL, a constrained multi-agent reinforcement learning framework for cooperative tasks with shared safety constraints. It measures episode-level violations using Signal Temporal Logic robustness and uses action sensitivities to assign each cost increment to individual agents. The framework was tested on payload transport, connected navigation, and networked traffic control. Reported results show higher safe-success rates and traffic throughput than constrained baselines.
In briefOrthros: Phase-Aware Heterogeneous Attention for Faster Transformers
SummaryThe paper introduces Orthros attention, a unified module that uses linear-complexity computation during prefilling and full attention during decoding to reduce Transformer cost on long contexts. The authors report language modeling competitiveness, a data-efficient conversion from Transformers that reportedly surpasses the original model, and near-perfect 128K needle-in-a-haystack results. They also report a 21.9x prefilling speedup over Transformers.
WhyProposes phase-switching unified attention with linear prefilling and full decoding; broad experiments including 128K retrieval and 21.9x prefill speedup support strong evidence and wide impact.
In briefKVeil: First-Place Solution to the ECCV 2026 LifeGenIP Unlearnable Videos Challenge
SummaryThis report presents first-place solutions to both phases of the ECCV 2026 LifeGenIP Challenge, which aims to stop unauthorized customization of facial videos using small, hard-to-see perturbations. Phase 1 uses Reference Capacity Collapse losses and a Motion-Guided EOT mechanism, while Phase 2 adds a two-layer framework that protects facial and background regions separately. The authors report nearly perfect protection against white-box reference-based customization, with effectiveness also shown against other customization methods, models, and temporal attacks.
In briefLAPT-Scale: Scaling Robot Policies via Latent Action Pre-Training on Human Videos
SummaryThis paper presents LAPT-Scale, a framework that scales generalist robot policies by learning latent actions from large-scale egocentric human videos that lack robot action labels. It uses a continuous latent action model combined with perceptual, semantic, and optical-flow alignment signals to supervise transitions beyond pixel reconstruction. The authors report that real-robot success rose from 63.3% to 85.0% as annotated human video grew from 1K to 5K hours. The learned policy also generalizes to unseen visual conditions.
In briefSeoulMMOD: A Large-Scale Public Benchmark for Multimodal Origin-Destination Flows in Seoul
SummaryThis paper introduces SeoulMMOD, a public benchmark of three years of hourly origin-destination flows covering citywide mobility in Seoul across 6 travel modes, 25 districts, and 426 sub-districts. It also includes travel-time, distance, calendar, rainfall, and point-of-interest signals. The authors evaluate 14 forecasting baselines under a common protocol and study spatial scaling, joint versus mode-separate training, and cross-year generalization. They report that some district-level spatio-temporal models do not scale to sub-district forecasting, and that naive joint multi-mode training does not consistently improve accuracy.
In briefLong-Horizon-Terminal-Bench 2.0: A Benchmark for Measuring Sustained AI Agent Progress
SummaryThis paper presents Long-Horizon-Terminal-Bench 2.0, an updated benchmark of 77 terminal tasks across 13 domains, keeping 46 original tasks and extending 31 others. Tasks are generated from working artifacts using parameterized reference implementations and scored with dense progress ladders that reward partial progress. Evaluation runs now last three hours with checkpoints every 30 minutes. Across 1,145 scored runs from 15 frontier models, binary verifier scores averaged 0.221 while dense ladder scores averaged 0.653, indicating that outcome-only evaluation misses substantial intermediate progress.
In briefAmortized Counterfactual Prediction with Graph Neural Networks
SummaryThis paper introduces the Counterfactual Graph Neural Process (CFGNP), which answers counterfactual questions directly from observational data without first estimating a causal graph, avoiding error propagation from graph estimation. It uses graph neural networks to pass messages across worlds only between corresponding factual and counterfactual variables. The authors report stronger predictive performance than existing methods on synthetic and image benchmarks, reducing MSE on the German Loan benchmark from 4.51×10^-2 to 0.19×10^-2.
In briefKnowledge graph-guided multi-task pre-training to improve legal reasoning in language models
SummaryThe paper proposes a multi-task pre-training method for German traffic law (StVO) that combines Next Sentence Prediction, Knowledge Graph Guided Clustering, and Knowledge Graph Aware Paragraph Classification to encode structured legal knowledge into model weights. Across nine Transformer encoders and 324 experimental runs, the models gain 10.3% to 155.0% relative improvement on synthetic StVO question answering. They also improve over prior LexGLUE results on four of seven benchmarks and achieve 3.4% to 15.8% zero-shot transfer gains on Austrian driving regulations.
