In briefRecovering Missing Relational Facts in Knowledge Graphs with Latent Diffusion Generation
SummaryThe paper addresses triple set prediction, which recovers missing relational triples in a knowledge graph by predicting the whole triple set rather than isolated links. It proposes LatentDiff, which encodes the entire graph into compact relation-aware latent embeddings with a residual-quantized autoencoder and then recovers missing triples through an observation-anchored latent graph diffusion model. On three benchmark datasets, LatentDiff outperforms the state of the art in F_tsp by 5.4 percentage points on average.
In briefCritic-guided synthetic experience generation in latent world models
SummaryThis paper proposes CGSE, a method that uses a latent world model as a local transition generator to create synthetic offline experience. Candidate actions from an offline policy are refined within a trust region to maximize a first-order Q-function approximation, then a frozen world model predicts the next state and the critic selects the best action for the replay buffer. On three OGBench manipulation tasks, CGSE improves the offline policy's average success rate by a relative 7.1%.
Jorge de Freitas, Christos Ziakas, Alessandra Russo
In briefPALATE: Pretrained Autoregressive Latent Embeddings for Local-Commerce Ad Personalization
SummaryThis paper addresses personalized ad ranking on local-commerce and food-delivery marketplaces. A GPT-style generative model is trained on each consumer's engagement sequence to predict their next engagements, and its last-token hidden state serves as a frozen consumer embedding refreshed daily and used as a dense ranker feature. The authors add an exploration-weighted loss, a two-stage click-then-conversion curriculum, and a designed negative set. Online tests on two ad surfaces reported gains in revenue and conversion rate.
In briefHALO: Human-Grounded Granular Evaluation of Proactive Agents
SummaryThis paper introduces HALO, a framework for evaluating proactive agents that decide whether, when, and how much to assist without an explicit request. The HALO-Arena benchmark covers four need types (Dependency, Blockage, Risk, Opportunity) across 10 everyday domains and 50 personal environments. The authors release HALO-8K, 8,000 expert-annotated responses from 10 models, and find that over-proactivity, with premature timing or excessive intensity, accounts for 61.20% of rejected responses.
In briefTactus: A Glove for Recording Natural Hand Skills With Touch and Pose
SummaryThis paper presents Tactus, a textile-integrated glove with thin 3D-knit construction, capacitive fingertip tactile sensors, and IMU-fused pose tracking for recording human manipulation data. It introduces standardized benchmarks for tactile force regression and hand pose estimation and compares Tactus with two commercial tactile gloves. Reported results include a minimum detectable force of 0.59 N at 95% confidence, 1.6 to 2.7 times lower than the baselines, a cross-participant consistency of 0.87, and an 11.0 mm fingertip keypoint error during heavy occlusion.
In briefHERA: An Autonomous AI Agent for Testing Quantum Devices
SummaryThis paper presents HERA, an autonomous agent that performs cryogenic transport measurements on quantum devices using a knowledge base of protocols, software tools, and literature. It was evaluated through a blinded trial against an expert on two multilayer graphene devices, an ablation of the knowledge base, and offline replays across models. HERA's conclusions matched the expert's sealed conclusions, and the knowledge base improved measurement design and reasoning across models. HERA was also applied to find and characterize superconductivity in a different material system after a few hours of expert setup.
Zachary W. Gomez, Nathan Bittner, Jordan T. McCourt et al.
In briefCertified Access: Test-time compute on a physics-structured AI memory
SummaryThe paper proposes a theoretical and empirical framework for test-time compute (such as retries, confidence-based escalation and non-local routing) applied to a Latent Dynamics Model used as memory. Under a damped symplectic update of a learned Hamiltonian, it derives exact phase-volume Jacobians, bounded energy changes and latch-transport laws for mechanisms that expand access. The certificate stack is verified on an analytic testbed, with the main contribution being formalization and verification rather than benchmark performance.
Pratik Jawahar, Riccardo Maggioni, Jonas Petersen et al.
In briefPhysicsNeMo-Mesh: GPU-native mesh processing and simulation cases for physics AI
SummaryPhysicsNeMo-Mesh is the mesh module of the open-source PhysicsNeMo library, offering GPU-based, batched mesh processing inside PyTorch autograd. It attaches names and shapes to fields and keeps full PDE case specifications, including boundary conditions. The paper reports large speedups over VTK for mesh operations and case loading, and shows one stored dataset serving several models.
Peter D. Sharpe, Corey Adams, Negin Sobhani et al.
In briefWhy the same LLM prompt can give different answers, and why a same answer may still be wrong: a taxonomy of failure mechanisms, tests and mitigations for LLM read-and-evaluate tasks in evidence synthesis
SummaryThis paper proposes TRAILS, a taxonomy of ways that large language model judgements can fail in evidence synthesis tasks such as study screening, data extraction, risk-of-bias assessment and critical appraisal. It presents 38 mechanisms that can cause inconsistent judgements across repeated runs or systematically affect outputs in ways not visible from the judgement itself, along with deployment settings, testing approaches and mitigation strategies. The authors argue that repeatability or accuracy does not rule out hidden judgement failures, so human evaluation should check whether judgements are grounded in the supplied evidence, free from inappropriate external influence, and produced through an examinable and verifiable process.
In briefClosed Gates Leak: How Gated Language Models Cancel Adjacent-Layer Outputs
SummaryThis paper studies how adjacent layers in transformer language models with gated MLPs write partly against each other into the residual stream. The authors split this cancellation into the next layer's normal output and a damping effect caused by the previous layer's write, traced to neurons with closed gates. They conclude that adjacent-layer cancellation is weak evidence about which features a model rejects, since it arises from a contraction gated MLPs apply to the directions they read on average.
In briefTabular foundation models as calibrated clinical risk predictors: a simulation test
SummaryThis paper compares three tabular foundation models (TabPFN-3.5, TabICLv2, Mitra-v2) with Firth logistic regression with intercept correction (FLIC) using simulations where the true risks are known. The foundation models were not better calibrated than FLIC, and even with 3,000 patients many scenarios failed to yield accurate risk estimates. On four clinical data sets, no model stayed accurate when event rates changed between populations, and recalibration on 500 local patients largely fixed this.
In briefHEARSAYBENCH: Testing whether AI can turn human rights into practical advice for real lives
SummaryThis paper introduces HEARSAYBENCH, a dataset of 400 human-authored test cases drawn from United Nations and other human rights reports, covering over 80 regions. The scenarios involve structural barriers such as war, statelessness, and displacement that limit people's ability to exercise their rights. Using Amartya Sen's Capabilities Approach, it tests whether models distinguish formal entitlements from actual abilities. Across 14 models, every model inferred users' situations more accurately than it advised on them, and although stronger models narrowed this gap, their recommendations remained unsafe to follow.
Ava Iranmanesh, Sobhan Lotfi, Ali Iranmanesh et al.
In briefHow Retrieved Papers Shape AI-Generated Research Ideas
SummaryThis paper examines how the literature retrieval strategy used by large language models affects the research ideas they generate from scientific papers. The authors compare five retrieval strategies against a no-literature baseline, generating 45,000 ideas across 15 machine learning topics with ten open-weight models. They also introduce an attribution method that traces how individual input papers influence each idea's components.
Zhipin Wang, Koel Dutta Chowdhury, Christoph Leiter et al.
In briefWhen Closed-Loop Curriculum Adaptation Helps GRPO: A Countdown Study
SummaryThe paper proposes CLOSEDLOOPGAUSSIAN, a lightweight adaptive curriculum method for reinforcement fine-tuning on ordered reasoning tasks. It adjusts the pace and spread of an Easy-to-Hard Gaussian sampler using periodic probe estimates of per-difficulty accuracy, instead of a fixed schedule. On a symbolic arithmetic reasoning task, it improves exact-match accuracy over vanilla GRPO and fixed open-loop curricula, with the largest gains at intermediate difficulty.
Yuv Raj Pant, Moshiur Rahman Tonmoy, and Kaiqun Fu
In briefBASILISC: Bayesian Matching of Single-Cell Data Across Modalities with Uncertainty
SummaryThis paper presents BASILISC, a probabilistic siamese network framework that matches single cells measured in two different molecular modalities and attaches an uncertainty score to each match. It works at the cell-type level, asking whether two cells share a type, and at the cell level, asking which cell is the partner in the other modality. The authors report that the uncertainty score identifies less reliable matches and that filtering by it raises accuracy.
In briefReinforcement learning for Balatro with a fast simulator and two-policy training
SummaryThis paper applies reinforcement learning to Balatro, a 2024 roguelike deckbuilder with a large state-dependent action space, long sparse-reward horizons, and stochastic card draws and shop generation. The authors present a fast Rust simulator synchronized with the live game, a condensed action interface with dynamic legal-move masking, and a multi-agent PPO setup that splits play into a tactical blind policy and a strategic shop policy. In joint training, the first-blind clear rate exceeded 99% and the mean maximum ante plateaued at 3 of 8, while single-agent baselines failed to stably clear the first blind.
Ajay Vikram Periasami, Anjaneya Gupta, Ethan Cheng et al.
In briefWhen AI Should Decide, Defer, or Deviate: Responsible Automation in Deceased-Donor Kidney Allocation
SummaryThis paper presents the Arbitrator framework, which determines when an AI trained on past decisions should recommend an action, defer to humans, or flag patterns that deviate from past practice. Using US deceased-donor kidney allocation data from 2016–2019 and 2023–2025, the framework resolves 93.5% of placement decisions and defers the remaining 6.5%. It correctly identifies 70.1% of historically discarded low-KDPI kidneys as candidates for out-of-sequence allocation.
Madeleine Pollack, Thomas Daillak, Mehmet Toner et al.
In briefModular Bayesian foundation models: Bayesian time-series encoders for detecting, routing, and forecasting
SummaryThis paper proposes a modular time-series foundation model that trains a collection of compact Bayesian encoders using Singular Bayesian Neural Networks (SBNNs). Their epistemic uncertainty is used to select the appropriate encoder for unseen data without labels. Across seven datasets, uncertainty-based routing matches distance-based routing on forecast MSE while improving calibration and interval coverage. The SBNN pipeline reaches forecast error comparable to deep ensembles and MC Dropout with substantially fewer parameters.
In briefExponential-Nemytskii Neural Operators: Learning How Fields Evolve Across Step Sizes
SummaryThis paper proposes a neural operator whose dependence on step size follows a first-order Duhamel approximation of semilinear dynamics, combining exact linear propagation with a local nonlinear function of the field and its derivatives. Under suitable conditions and fixed spatial resolution, the authors prove that agreement between models at two step sizes controls their prediction differences over a prescribed range. The paper also gives conditions for identifying both components from two step sizes and evaluates the model on synthetic and real datasets against several baselines.
In briefLightGenBench: A Benchmark for 3D Emission Generation
SummaryThis paper introduces emission generation, the task of predicting spatially varying RGB emission across a 3D asset's surface from its geometry, non-emissive materials, and a rendered view. The authors release LightGenBench, a dataset of 37k artist-authored emissive assets, and establish baselines across UV, sparse-voxel, multi-view, and segmentation-based methods. The sparse-voxel baseline TRELLIS.2-Emission performs best, yet reaches only 0.271 IoU, so the task remains largely unsolved.
In briefFairSplit: A Fair Classification Method That Splits the Embedding Space
SummaryThis paper proposes FairSplit, which decomposes an embedding space into a Fair Space used for prediction and a Sensitive Space that isolates sensitive information. A learnable mask and the Hilbert-Schmidt Independence Criterion (HSIC) guide this decomposition. The authors prove that the demographic parity gap is upper-bounded by the Fair Space's dimensionality and its HSIC with sensitive attributes. Experiments on established fairness benchmarks show competitive accuracy-fairness trade-offs.
In briefMCCL: Training PyTorch models across a fleet of Macs at Skywalker Sound
SummaryThis paper introduces MCCL (Metal Communication Collective Library), the first open-source PyTorch distributed backend for Apple's Metal Performance Shaders, enabling multi-node training over switched Ethernet. The authors evaluated it on 24 nodes, finding that large all-reduce operations reach 98–99.7% of TCP goodput. A 113.5M-parameter Vision Transformer achieved 95.9% parallel efficiency, and the paper compares Mac and NVIDIA A100 throughput.
In briefRepresentation complexity during learning: trading width for precision and time
SummaryThis paper studies the role of overparameterisation in neural network learning by comparing the representation complexity needed for prediction with that needed to make an unknown target visible through gradient feedback. The authors introduce an oracle model of learning and quantify effective representation complexity with Kolmogorov n-widths. For k-sparse parity, they show a polynomial separation between representing a predictor and discovering it, and they derive a width-time frontier. Experiments support the predicted dimension dependence.
Ryan Singh, Miguel De Llanza, Christopher L. Buckley
In briefDistilling the Shape of Uncertainty: What On-Policy Distillation Learns and How to Predict Its Gains
SummaryThe paper studies on-policy distillation (OPD) for large language models and finds that students learn not only which tokens to consider but also where to keep reasoning paths open. It introduces the entropy shape and an Entropy Shape Correlation (ESC) metric to measure student-teacher agreement in this respect. Interventions on teacher entropy show that where to branch causally drives OPD gains, and pre-OPD ESC selects the best teacher more accurately than existing predictors.
WhyThe authors introduce entropy shape and ESC, separate where-to-branch from how-to-branch, and report intervention experiments and teacher selection results.
Shiping Yang, Chuanqi Tang, Hossein Hajimirsadeghi et al.
In briefAuthor-Informed Stance Detection: A Dataset and Findings
SummaryThis paper reframes political stance detection on social media as a task that uses author context, inferring both the author's political leaning and their stance toward a target. It finds that humans and pretrained large language models struggle with single tweets, but that author context improves stance detection for both, especially for sarcastic tweets. The authors also release a dataset of stance-labeled tweets enriched with author-context information.
In briefAdjustable spatial task allocation for maritime swarm robots
SummaryThis paper presents a task allocation method for teams of autonomous agents, built on a generalized Nonlinear Opinion Dynamics model, that lets members redistribute across spatially embedded tasks. The authors couple opinion dynamics to spatial movement and show in simulation that social coupling controls group cohesion, producing states from fully dispersed to fully consensus-based allocations. For distance-based preferences with equal numbers of agents and tasks, the final allocation is close to the Hungarian optimum, and performance is also tested under communication constraints in the MOOS-IvP-Multi environment.
In briefImplicit reasoning in transformers: closing the compositionality gap in GPT-2-like models
SummaryThis paper studies a weakness of transformer language models: they can combine known facts correctly on familiar data yet fail when the same combination is applied to out-of-distribution knowledge. The authors train a GPT-2-like model from scratch on a controlled two-hop reasoning task. They present in-distribution and out-of-distribution atomic facts in a format aligned with the second retrieval step of the learned circuit, without changing the architecture or adding compositional supervision. The model then transfers its learned compositional behaviour to held-out out-of-distribution facts, closing the gap in this controlled setting.
In briefCan generated images be made statistically indistinguishable from real ones? A study testing this with neural classifiers
SummaryThe paper trains deep neural networks to tell real images from fake ones produced by generative models. It then applies progressive abstractions such as blurring, downsizing, cropping, and conversion to edge or semantic maps to see whether fake images become indistinguishable. Almost all abstractions fail to make the fakes indistinguishable, except when images are reduced so far that the original content is mostly lost, such as a 256×256 face shrunk to 4×4. The pattern holds across MNIST, CelebA, FFHQ, and GenImage and across GANs, VAEs, and diffusion models.
Mohaimin Rafi, Otavio Moraes Rosa, Anshuman Chhabra et al.
In briefFed-IGTD: a federated learning framework that turns tabular data into images while keeping client data private
SummaryThe paper introduces Fed-IGTD, a federated learning approach that converts tabular data into image-like representations and trains a convolutional neural network without centralizing raw data. Clients share only local statistics, from which the server rebuilds a global correlation matrix and derives a feature-to-pixel mapping. The authors evaluate two differential privacy mechanisms on five datasets under varying non-IID severity and privacy budgets. They find that privacy protection on the mapping step has little effect on prediction, while DP-FedAvg noticeably reduces utility at moderate budgets with five clients.
David Chushig-Muzo, Eva Milara, Francisco J. Lara Abelenda et al.
In briefCombining formal specifications and data-driven analysis for signal monitoring
SummaryThe paper proposes a hybrid method for analyzing and monitoring real-valued signals that combines specification-based and data-driven techniques. It uses regular expressions with interval operations to find regions of interest, classical signal processing to extract features, and semi-supervised machine learning to learn relationships separating correct from faulty behavior. Evaluated on a voltage regulator, the approach is shown to effectively detect undesired system behaviors.
Roderick Bloem, Benedikt Maderbacher, Cristinel Mateis et al.
In briefBlind Spots: Where Small Models Lag Frontier Models on Terminal Tasks
SummaryThis paper examines why small language models lag behind frontier models on terminal tasks, splitting the gap into solvability, verification, and confidence. Across two standard terminal benchmarks, repeated sampling narrows but does not eliminate the solvability gap, while a persistent gap remains in verification. Small models are also more overconfident and poorly calibrated in predicting their success. The authors argue the gap spans generating, recognizing, and anticipating success, and hope this decomposition guides future development of smaller terminal agents.
In briefSelf-Supervised Doppler-Guided RF Odometry: Radio-only robot motion estimation without pose labels or auxiliary sensors
SummaryThis paper presents RFWalk, a self-supervised odometry method that estimates a robot's ego-motion using only radio-frequency (RF) sensing, so it works in darkness, smoke, and fog. It models correspondences between consecutive RF frames as random walks on a space-time graph and learns them from unlabeled sequences via temporal cycle consistency. A Doppler consistency loss, based on radial velocity measured by the RF sensor, rules out physically implausible correspondences. The authors report competitive performance with supervised baselines in static scenes and a 58% reduction in absolute trajectory error in dynamic scenes with moving pedestrians.
In briefHow model design, language structure, and adaptation methods affect speech recognition for low-resource African languages
SummaryThis paper systematically compares speech recognition models on five low-resource African languages, with Swahili as a mid-resource reference, fine-tuning Phi-4-multimodal-instruct, Whisper-large-v3-turbo, and MMS-1B-all. It examines data sufficiency, monolingual versus multilingual adaptation, domain fine-tuning, and language-adapter transfer. Phi-4 achieves the best overall accuracy, while Whisper shows high character error rates linked to hallucinations and repetition loops. Proxy language adapters outperform training MMS adapters from scratch.
Mercy Muchai (1), Kevin Chege (1), Nick Mumero Mwangi (1) et al.
In briefTask-aware projection selection for sparse-view CT that predicts blood flow in brain aneurysms
SummaryThis study examines how to choose which X-ray projection views to acquire in sparse-view CT angiography so that blood flow wall shear stress in intracranial aneurysms can be predicted directly. The authors used fluid dynamics simulations to create synthetic projections and trained a Fourier neural operator to predict the TAWSS field and its spatial average. They compared fixed sampling with three task-aware selection methods, including an agentic AI selector, and found that optimized selection helps most at extreme sparsity. The selected view subsets differ from those suited to image reconstruction.
Mahmoud Elhadidy, Roshan M. D'Souza, Amirhossein Arzani
In briefAlbumVSR: Identity-Preserving Face Video Super-Resolution Using Personal Photo Albums
SummaryThis paper presents AlbumVSR, a method that restores degraded face videos by using photographs of the same person as an external identity source. It conditions a frozen, one-step streaming diffusion transformer on album tokens and texture tokens derived from a fused UV atlas. On an identity-disjoint VoxCeleb2 benchmark with heavy synthetic degradation, ArcFace similarity to the ground truth rises from 0.586 to 0.670 and face PSNR from 24.98 to 26.71 dB. The authors report higher mean identity similarity than all 14 baselines on 91 test identities.
In briefRethinking GFlowNet Training: Negative Reinforce Dynamics
SummaryGenerative Flow Networks (GFlowNets) learn to sample discrete objects in proportion to a target reward, training by minimizing balance errors over sampled trajectories. The paper links the full gradient to min-min dynamics between the GFlowNet and a decoupled sampler, and proposes Negative Reinforce Dynamics (NRD), which extends exploration while maintaining correctness. The authors report that NRD outperforms existing approaches in convergence speed and exploration efficiency on standard benchmarks, without training a separate teacher or auxiliary exploration policy.
Pedro Dall'Antonia, Daniel Csillag, Diego Mesquita
In briefSplitting and Recombining Game Levels Using Perceptual Distance
SummaryThis paper explores ways to automatically extract reusable level chunks and a higher-level graph structure from existing videogame level designs. It introduces "perceptual distance," drawing on Space Syntax analysis from architecture, and uses it to segment levels through hierarchical clustering. The authors compare their approach with prior StarCraft analysis, show results on Super Metroid rooms, and demonstrate a proof-of-concept level generation algorithm built on this segmentation.
In briefBenchmarking graph neural networks for AC power flow prediction, focusing on design choices and physical consistency
SummaryThis paper benchmarks representative graph neural network architectures for predicting AC power flow under a common experimental framework. All models use identical preprocessing, training procedures, and comparable parameter budgets to isolate the effects of message-passing mechanisms and inductive biases. Tests on the IEEE 14-bus and IEEE 118-bus systems assess predictive accuracy and physical consistency. The results indicate that anisotropic message-passing architectures offer the best trade-off, and that explicit edge features encoding line impedances and global context aggregation are important, providing practical guidelines for graph-based power system models.
Federico Clerici, Ferran Bohigas-Daranas, Hamid Latif-Mart\'inez et al.
In briefWhy speculative decoding beats predictions on sparse MoE models
SummaryThis paper studies speculative decoding, which speeds up language model inference by verifying several draft tokens in one forward pass, on sparse mixture-of-experts (MoE) models. Evaluating 10 models ranging from 30B to 1T parameters, the authors find that recent fine-grained models achieve higher speedups at low batch sizes, contrary to prior predictions. They attribute this to routing overlap across draft tokens and GPU execution behavior, and propose a maximum-entropy Iterative Proportional Fitting model that predicts unique expert counts with less than 4% error.
Ekagra Ranjan, Komal K. Teru, Bharat Venkitesh et al.
In briefContext Compaction for World Models: Helping Models Remember Rooms They Have Generated
SummaryThe paper addresses a problem where world models often fail to regenerate previously visited regions consistently, framing it as a context retrieval issue. The authors build a benchmark that generates new worlds for every episode to measure context retrieval, and they design an agent harness that performs context compaction for world models. Their experiments show that context compaction lets models retrieve rooms up to 4 times beyond their context window without retraining, whereas models using sliding-window attention lose most of their retrieval ability when dreaming.
In briefBeyond Passing Tests: Reviewing and Re-evaluating Methods for Better LLM-Generated Code
SummaryThis paper systematically reviews 628 studies on techniques that improve the quality of code generated by large language models, such as readability, efficiency and security. It then re-evaluates ten techniques under matched compute budgets across four models, and compares 66 techniques head-to-head. The authors find that most gains are measured only on the targeted dimension, and that best-of-n sampling with an analyzer-weighted ranker is a strong baseline.
In briefDoes Physics-Informed Machine Learning Help Solar Power Forecasting?
SummaryThis paper tests whether physics-informed machine learning and hybrid models deliver their usual benefits for large-scale solar power forecasting. The authors combine the ADR physical PV efficiency model with the TSMixer neural architecture in two hybrid designs, evaluated on over 3 million samples from 3,713 PV installations across 33 European countries. The physics-informed models converge slightly faster but do not consistently beat the purely data-driven baseline in test accuracy, generalization, or limited-data settings. The authors conclude that explicitly adding the ADR model offers limited practical benefit for day-ahead PV forecasting with modern deep architectures trained on large datasets.
Peter Pavl\'ik, Franz M. Rohrhofer, Bernhard C. Geiger
In briefAnalyzing Dungeons & Dragons encounters through semi-random simulation in Foundry VTT
SummaryThis paper presents Mage-Hand, a tool that semi-randomly simulates Dungeons & Dragons combat encounters on the Foundry Virtual Tabletop platform. Existing difficulty metrics sum predefined per-enemy difficulty values and ignore positioning, dungeon layout, monster ability synergies, and party composition. The authors report that the simulated winrates correlate strongly with D&D's published difficulty metrics and better match some players' reported experiences.
In briefCommunication-efficient clustered federated learning for non-IID wireless networks
SummaryThis paper proposes a Clustered Federated Learning (CFL) framework that assigns clients to specialized models during training to handle non-IID data in wireless networks. Data-domain clustering initializes model-based clustering, exploiting client similarities without sharing raw data while reducing communication cost. Evaluated on DeepMIMO, FEMNIST, and Rotated CIFAR-10, the best configurations outperform IFCA in accuracy on DeepMIMO and FEMNIST, and cut FEMNIST downlink data volume by 83.1%.
Martin Isaksson, Rickard C\"oster, \v{S}ar\=unas Girdzijauskas et al.
In briefTableTune: Self-correcting draft-length tables that measure speculative decoding speed.
SummarySpeculative decoding speeds up LLM serving, but the best draft length shrinks as batch size grows, and vLLM relies on hand-written tables that nothing measures or corrects. TableTune times decoding for each batch size and draft length on vLLM's serving path and builds a throughput-maximizing table in minutes. A vLLM plugin tracks accepted guesses in live traffic and rebuilds the table while serving. On an H100, the tuned table reaches at least 94% of the best fixed draft length's throughput in most drafter settings and outperforms vLLM's documented table on a mixture-of-experts model and in llm-d deployment.
Abhishek Malvankar, Paul Castro, Nick Mitchell et al.
In briefEfficient agentic audio-visual question answering via toolset pruning and modality consolidation
SummaryThis paper studies agent-based audio-visual question answering, where models actively call perception tools to gather evidence. Starting from OmniAgent, which uses many specialized audio and visual tools, the authors progressively prune the toolset and merge modality-specific tools into omni-modal ones. The result is a two-tool agent with omni-modal captioning and omni-modal QA. Across three widely used AVQA benchmarks, this compact agent matches or exceeds the accuracy of the specialized-tool agent while cutting inference cost by about 70%.
In briefCan AI models find a specific item among many look-alikes? MIMO-Bench tests visual reasoning.
SummaryThis paper introduces MIMO-Bench, a benchmark for testing multimodal large language models on images where several objects of the same category appear. It contains 3,802 image-question pairs covering attribute binding and numerical counting. Both open-source models such as InternVL and QwenVL and proprietary models such as Gemini and GPT-5 perform substantially worse on this benchmark than on existing ones.
Saumya Mundra, Ajoy Mondal, Vineeth N Balasubramanian et al.
In briefPlanned Restlessness: Budgeted, Delayed Meta-Bandit Learning Framework
SummaryThis paper studies episodic finite-horizon Bayesian bandits, or meta-bandits, in which each episode draws arm qualities from unknown recurring priors, observations arrive after known delays, and plays are subject to per-episode caps and total budgets. The authors propose a planning optimization whose linear programming relaxation guides decisions and certifies admissible policies, including hierarchical Thompson sampling. For stationary priors with bounded report lag and nonbinding budgets, posterior sampling is shown to bound the cost of not knowing the priors by O(√N) over N episodes. For a single undelayed episode, they obtain an improved approximation factor with a matching lower bound on the leading term.
In briefGradSplat: a feed-forward 3D Gaussian Splatting model that refines its shapes with gradients inside the forward pass
SummaryGradSplat is a feed-forward model that predicts 3D scenes from a few images as Gaussian splats. The authors find that its accuracy gap relative to per-scene optimization mostly comes from the Gaussians' shape, meaning scale and rotation. The method applies log-anisotropy correction and learned gradient-based refinement inside the forward pass, without a target view. It reports the best PSNR and SSIM on RealEstate10K among nine methods, and on DL3DV and ScanNet++ without fine-tuning.
In briefMAS-Loop: Moving Beyond One-Shot Orchestration to Self-Evolving Multi-Agent Systems
SummaryThis paper presents MAS-Loop, a framework in which an Orchestrator builds and runs a multi-agent system, and a trajectory-conditioned Refiner diagnoses failures and proposes bounded patches that are re-executed to compare outcomes. The two components co-evolve across training iterations. Experiments report the strongest overall performance among evaluated baselines, and the authors also release MAS-JudgeCorpus, a dataset of execution-grounded source-updated pairs.
In briefPlanning safe lane changes in mixed traffic using neural network surrogates embedded in optimization
SummaryThis paper proposes a data-driven optimization framework for lane-changing trajectory planning of connected and automated vehicles in mixed traffic. Neural network surrogates for safety and lateral comfort costs are embedded in a Bézier-curve-based planner using linear programming, avoiding the binary variables of conventional mixed-integer programming. Experiments on a hybrid real-world and simulated dataset show competitive surrogate accuracy and improved optimization efficiency. Closed-loop simulations indicate the planners reduce weighted planning cost and outperform baseline methods.
In briefStudy of how many parameters invariant convolutional networks need to memorize two-class data.
SummaryThe paper studies convolutional networks invariant to a given finite group of orthogonal transformations, such as translations, reflections, or rotations. It proves upper bounds on the size of such networks with threshold activations needed to interpolate a two-class dataset. The bounds rely on an instance-specific complexity measure reflecting how the classes are arranged geometrically, and the proof is constructive via a randomized algorithm whose practical effectiveness is tested numerically.
In briefOneScene-Bench: Do Vision-Language Models Combine Evidence Across Multiple Views?
SummaryThis paper introduces OneScene-Bench, a controlled benchmark that tests whether vision-language models integrate multiple views of a scene into one coherent representation. It assesses unique object counting, cross-view consistency, and spatial reasoning on real-world multi-view scenes, controlling object overlap between views to separate tasks needing cross-view correspondence. Evaluating twelve open-weight and proprietary models, the authors find counting accuracy drops sharply on complementary views and cross-view consistency stays around 50%, despite stronger spatial reasoning performance.
In briefTreasuryQA Bench: Testing whether AI agents can find and read figures in scanned financial archives
SummaryThis paper introduces TreasuryQA Bench, a benchmark of 70 questions over 1,435 scanned U.S. Treasury publications. Agents must search for the right file through a Dropbox MCP server, read the scanned pages with OCR, and reply with one exact figure. Ten agents were tested, with success rates ranging from 64.8% for the best to 16.7% for the worst. Performance drops as scan quality worsens, and most failures come from misread digits.
In briefTrussTutor-Bench: A Benchmark for LLM Solution Accuracy, Sycophancy, and Error Detection in Statics
SummaryThis paper introduces TrussTutor-Bench, a benchmark of 200 verified two-dimensional truss problems paired with automatically generated student solutions. It evaluates thirteen large language models on solving accuracy, sycophancy, and error detection in statics. Baseline accuracy ranged from 64.5% to 97.0%, and error detection rates were uniformly high, while sycophancy varied widely, with models rejecting their own correct solutions in 1.6% to 50.0% of cases.
Evangelos Apostolakis, Erfan Hamdi, Sara Jahedazad et al.
In briefLev-3B: A Data Recipe for Compact Probabilistic Log Decision Models
SummaryThis paper presents Lev-3B, an open 3B-parameter model built on Granite-4.1-3B-Base for choosing among known options in log analysis, such as anomaly detection and attack-technique identification, with calibrated probabilities. Its two-stage training recipe first uses non-log natural-language decisions, then mixes target log tasks with general log data from 35 auxiliary tasks. Lev-3B outperforms prompted frontier models on anomaly detection, including on an excluded logging system, but trails them on attack-technique classification. Its probabilities are well calibrated on familiar sources, yet it confidently misses anomalies on the excluded logging system. The checkpoint, recipe, data manifests, and prediction artifacts are released.
In briefTraining Language Models to Report Internal Signals Without Causal Shortcuts
SummaryThis paper studies whether language models can accurately report their own internal signals. The authors trained hundreds of introspectors across three low-level signal types and eight language models. They found that straightforward supervised fine-tuning leads to causal bypass, where self-reports rely on signals correlated with the target but causally unrelated to it. They propose interventional training, which trains models to report counterfactually perturbed values, and show it improves introspection of both endogenous and injected signals, including on out-of-distribution data.
In briefStructural-test geometry as an inductive bias for partially observable world models
SummaryThis paper studies how to make latent geometry in partially observable world models reflect distinctions needed for later control, using privileged simulator access during training. The authors define a pseudometric over histories from the response distributions of posterior-marginalized structural tests, and train quotient-preserving world models by aligning latent Euclidean distances to this metric. On a six-task diagnostic suite, the method achieves higher planning success than metric decoding and structured belief distillation, especially under held-out structural variation.
In briefHEDGE: Gated Hypergraph Memory for Scalable Multi-Agent Debate
SummaryThis paper proposes HEDGE, a selective memory framework that represents multi-agent debate trajectories as semantic hypergraphs linking related claims across agents. A learned memory selection module decides whether each memory candidate is stored, backed up, or omitted. Across seven benchmarks, HEDGE improves accuracy by 5.0 points on average over the strongest multi-agent system with GPT-4o-mini, reaching 8.2 points on conflict-heavy tasks. It also reduces memory usage by 48% and debate rounds by 34%.
In briefGeometric Time Score Estimators: Linking Self-Predictive and Generative Occupancy Models in RL
SummaryThe paper observes that self-predictive representations give poor estimates of occupancy density in reinforcement learning, while generative occupancy models produce accurate samples but do not easily yield densities. It proposes Geometric Time Score Estimators (GTSE), which learn occupancy density estimators along probability paths to bridge these two paradigms. In experiments on locomotion and manipulation tasks, GTSE improves goal-conditioned RL performance by 2x to 6x over state-of-the-art methods, and its learned representations transfer effectively to downstream tasks.
Aravind Venugopal, Grace Liu, Jeff Schneider et al.
In briefAligning Autonomous Research Agents with Humans Through Co-Improvement
SummaryThis paper studies how autonomous research agents collaborate with human researchers. In benchmark creation, the authors find humans still play a pivotal role, especially during ideation. They introduce HackBench, which shows current agents are not well aligned and struggle with hack localization. The paper also proposes training agents with reinforcement learning using human simulators and expert-aligned rubrics to improve human interaction and communication.
In briefAuditable Belief Updates: A Missing Safety Layer for LLM Multi-Agent Systems
SummaryThis paper proposes the Explicit Belief Multi-Agent System (EB-MAS), which records agents' belief states as a structured substrate, uses belief operators to propose candidate updates, and applies a Cognitive Firewall to reject updates that are weakly evidenced but driven by vulnerability cues. It also discusses group-level risks such as manufactured consensus, collusion, and cascading misinformation. The authors argue that scalable agent safety requires explicit runtime interfaces for tracing and constraining belief revision.
In briefContextual embeddings that retrieve both answers and the evidence needed to verify them
SummaryThis paper introduces pplx-embed-v2-context, a family of contextual embedding models trained by distilling graded chunk-level relevance from a query-aware context compression model, removing the need for gold-chunk labels. The models are released in 9B and 0.6B sizes, and two benchmarks are introduced: Context-bench, a controlled diagnostic, and Q2D-Chunk, built from real-world search queries and published as a public leaderboard. The 9B model sets new state-of-the-art results on both benchmarks and ConTEB, and extending chunking to image inputs yields top performance among dense retrievers on visual ViDoRe-V3.
Sedigheh Eslami, Markus Krimmel, William Barber et al.
In briefLearning reasoning from reference answers alone via adversarial RL for LLM post-training
SummaryThe paper proposes Reinforced Adversarial Learning (RAL), where a co-trained LLM judge uses only reference answers as the reward for on-policy post-training of reasoning LLMs. It is first validated on three verifiable tasks (integer multiplication, arithmetic-expression synthesis, and SMILES-to-formula translation), where it matches direct RLVR given sufficiently capable judges and enough training. RAL is then applied to scientific abstract generation, including producing NIH grant proposal abstracts from titles, and human scientists and an LLM judge ensemble find its outputs harder to distinguish from real abstracts than frontier-LLM baselines.
Albert Bou, Javier Valdes-Aleman, Jon M. Laurent et al.
In briefAutonomous quantum measurement using agentic AI for materials characterization
SummaryThis paper introduces QUANTA, an agentic AI framework for autonomous characterization of quantum materials. It uses a Physical Property Measurement System as the experimental platform, combining retrieval from scientific literature, protocols, and prior measurements with specialized agents for planning, troubleshooting, and data interpretation. The framework is demonstrated on transport measurements of the Weyl semimetal TaP and nonlinear response studies of Co2MnGa.
Eunbi Rha, Abhijatmedhi Chotrattanapituk, Mouyang Cheng et al.
In briefGenerating diverse, realistic financial tail-risk scenarios with generative flow networks
SummaryThis paper presents GRID, a method that uses Generative Flow Networks (GFlowNets), which sample in proportion to a reward, to systematically generate diverse and realistic financial tail-risk scenarios. It introduces a mixture-policy architecture for high-dimensional sequential macro-financial data along with plausibility constraints. On three real-world crises, GRID improves scenario diversity, realism, and tail-risk coverage over reinforcement learning and MCMC baselines.
In briefWhich randomization matters: a controlled study of data volume, axis and representation in small-data molecule generation
SummaryThis paper examines how different kinds of randomization affect whether chemical language models trained on small datasets generate genuinely new molecules. Using 1,000 seed molecules, the authors compare canonical and randomized SAFE, isolate atom-order and fragment-order randomization, and compare randomized SAFE with randomized SMILES on matched training corpora. Randomized SAFE raises the yield of novel molecules more than 450-fold, and the two randomization axes act jointly rather than additively. At matched corpus size, randomized SAFE yields 2.8 times the genuine output of randomized SMILES.
Neerav Kaushal, Alisa Teige, and Ajay MNV Penmatsa
In briefDiscovering Jailbreaks via Curriculum of Models: training an investigator model against progressively harder target versions to find prompts that elicit prohibited behaviors
SummaryThis paper introduces Curriculum of Models (CoM), a method that trains a reinforcement learning investigator model to find prompts that elicit prohibited behaviors from safety-trained language models. Training starts against a target version that rarely refuses harmful requests and moves to harder versions, ending with the original target, once the investigator reliably elicits the behavior. Evaluated on CBRN, explosives, and illicit drug elicitation tasks with gemma-4-31b-it and gpt-oss-120b, CoM finds successful jailbreaks more reliably than the PRBO and direct-training baselines when trained on a single behavior. When trained on multiple behaviors, PRBO-trained investigators achieve higher success rates on held-out behaviors.
Micha{\l} Burzy\'nski, Noah T. Rossi, Shomit Basu et al.
In briefBeyond Region Memory: Local Point Tracking for Locating Vocal Cords in Video Laryngoscopy
SummaryThis paper proposes Local Point Tracking (LPT), a method for locating vocal cords in video laryngoscopy. It first uses a single-frame detector to generate sparse anchors, then tracks local query points through frames the detector misses and derives temporal bounding boxes from them. A stateless box-snap step then refines the boxes to the target boundary. On two datasets, LPT achieves the best mAP_50, success AUC, and missing rate, and ablations show point tracking mainly improves temporal robustness while box-snap improves boundary tightness.
In briefJudging AI-Generated Video Realism by Tracking Where People Look
SummaryThis paper studies eye tracking as a supplementary signal for evaluating AI-generated videos, going beyond explicit ratings. Two Visual Turing Test experiments with 90 participants yielded 4,286 valid eye-tracking trials across 438 videos. The authors find gaze patterns align with explicit human feedback and propose GazeEval, a method that predicts perceived visual realism from eye-tracking data with 71.3% accuracy.
In briefCopyLoop: faster decoding for looped Transformers using context copying and shallow screening.
SummaryThis paper introduces CopyLoop, a training-free decoding algorithm for looped Transformers that combines low-cost context-copy proposals, shallow-depth screening, and continuous recurrent scheduling. It proposes multi-token suffixes from repeated context and evaluates them early at a shallow recurrent depth to filter out incorrect branches. Evaluated on six checkpoints from the Ouro, Raven, and Huginn families for code generation and multi-step math reasoning, CopyLoop achieves 2.98x to 10.70x speedups on Base models and 2.74x to 3.52x on Thinking models over full-depth autoregressive decoding.
In briefUnEZ: A label-free neural network that estimates galaxy redshifts and spectral line strengths without templates.
SummaryThis paper presents UnEZ, a fully unsupervised, physics-informed neural network that predicts spectroscopic redshifts and line strengths without any redshift or line strength labels. An encoder estimates emission and absorption line amplitudes and widths, and a fixed decoder builds a template, redshift, confidence curve, and spectral model. Tested on SDSS Data Release 18, UnEZ reaches outlier fractions of 0.459% at SNR > 10 and 3.89% at SNR > 3, with NMAD values of 1.237×10⁻⁴ and 1.471×10⁻⁴, competing with modern redshift estimation methods.
WhyThe authors present a fully unsupervised, template-free model estimating redshift and line strengths, reporting results on SDSS DR18 at two SNR thresholds against existing algorithms.
In briefCausal Attention on CPU: A Cache-Aware, Load-Balanced Kernel for Faster Prefill
SummaryThis paper presents CACPU, a high-performance CPU attention kernel for causal full-sequence prefilling. It uses tiered cache blocking and an in-kernel task dispatching mechanism to schedule threads effectively. The authors report up to a 2.01x speedup over PyTorch SDPA, FlexAttention, and SGLang attention kernels. Integrated into SGLang, it cuts Time to First Token for Llama-3.1-8B by up to 1.74x while preserving accuracy.
In briefFinding the Key Decision Points: The DECIDE Framework for Ranking Data Analysis Choices
SummaryJudgment calls made during data analysis can change conclusions, and different analysts given the same data often reach different results. This paper introduces DECIDE, a model-free framework that measures each analytical decision's importance directly from actual outcomes across the plausible decision space. The authors apply DECIDE to real-world multiverse analyses and also give a formal definition of the multiverse.
Zachary T. Rewolinski, Jan Simson, Frauke Kreuter et al.
In briefRouted Continual Learning: routing agent experience to different memory clusters
SummaryThe paper proposes Routed Continual Learning (Routed CL), which groups past agent trajectories into clusters, each keeping state with one memory system from a fixed library. An LLM-based router decides before each run whether to reuse an existing cluster or create a new one. The authors also introduce a Context Memory Adapter that connects the memory system to the agentic loop. On a mixed stream of CL-Bench and τ³-Bench tasks, it reports higher normalized reward than a single global in-context learning history and lower estimated API cost.
Michael Philip, Mahesh Balakrishnan, David Geraghty et al.
In briefTranslated vs. Native Benchmarks: Does Translated MMLU Measure Language, Knowledge, or Locality?
SummaryThe paper examines whether DeepL-translated MMLU serves as a valid proxy for evaluating large language models in non-English languages, comparing it with six native benchmarks across five Nordic and Baltic languages and 29 models. It tests whether translated scores add predictive value for native performance after controlling for model size and language proficiency. The translated benchmark proves more useful when its source content overlaps with the native target, such as Skolprov, but provides no significant added value for Sweden-specific factual knowledge in SwedishFacts.
Klaudia-Doris Thellmann, Bernhard Stadler, Dan Saattrup Nielsen et al.
In briefLoss Does Not Equal Behavior: A Unified Output-Space Analysis of Gradient-Based Machine Unlearning
SummaryThis paper shows that gradient-based machine unlearning methods can have similar losses yet produce different output behavior. For linear regression, the authors derive closed-form expressions for the output-space gap between the unlearned model and a model trained only on retain data, extended to nonlinear models through approximations. They introduce two leverage matrices capturing the entanglement between forget and retain data, which can be precomputed to audit forget sets.
Xingjian Zhao, Mohammad Mohammadi Amiri, Malik Magdon-Ismail
In briefA 211K-Parameter Medical Segmentation Model Using Genetic Prompting in Cellular Automata
SummaryThe paper introduces Foundational-NCA, a lightweight foundation model for medical image segmentation built on Neural Cellular Automata. It uses a compact task vector and a genetic prompt mechanism to adapt to different segmentation tasks without retraining. With only 211 thousand parameters, it achieves an average Dice score of 84.5% on trained tasks and 60.4% on unseen tasks from the TotalSegmentator dataset.
In briefFitzRoy: a reinforced diffusion model for realistic, accurate precipitation forecasts
SummaryThis paper presents FitzRoy, a precipitation forecasting system for the continental United States. It refines a conditional discrete diffusion model with reinforcement learning to balance pointwise accuracy with structural realism. The authors report that it matches or exceeds state-of-the-art models for lead times up to 24 hours and captures extreme rainfall events.
In briefAuditable hierarchical inference combining multimodal evidence with Bayesian questioning
SummaryThis paper presents a synthetic evaluation framework that separates objective multimodal classification from active questioning. An objective model assigns eight syndrome families from synthetic tongue, pulse, and acoustic observations, while a Bayesian questioning component narrows 33 fine-grained terms. In 600 synthetic cases, an expected-information-gain strategy raised family accuracy from 92.7% to 95.5% and fine-term accuracy from 33.0% to 42.5%, but asked more questions on average than a fixed order. The authors state the results establish a reproducible synthetic workflow, not clinical diagnostic or treatment validity.
In briefRennala NSGD: Asynchronous stochastic optimization for non-Euclidean geometries
SummaryThis paper analyzes minibatch stochastic gradient descent for smooth nonconvex optimization under an arbitrary norm on a finite-dimensional space. Using the framework of κ-regular normed spaces, it derives high-probability oracle complexity bounds under a light-tail condition in the dual norm, making the dependence on geometry explicit. In the Euclidean case the bound matches standard minibatch SGD up to logarithmic factors. The paper also introduces Rennala NSGD, a semi-asynchronous distributed algorithm, with wall-clock time guarantees under fixed, heterogeneous worker computation times.
In briefLeaving, Staying, or Living There? Duration Commonsense Reasoning with Human-LLM Co-annotations
SummaryThis paper examines whether vision-language models can infer how long a person stays in a place from geotagged social media posts, since duration is rarely stated explicitly. The authors built a corpus of tweets co-annotated by humans and LLMs as a testbed. They found that state-of-the-art models struggle with duration inference even with multi-turn Chain-of-Thought prompting, often missing contextual information or overrelying on salient visuals while ignoring hashtags and subtle cues.
In briefAlpha-TCM: A Nine-Layer Framework for Auditable TCM-Guided Multimodal Inference
SummaryThis paper proposes a nine-layer system architecture for multimodal inference guided by Traditional Chinese Medicine, separating domain knowledge, case evidence, probabilistic inference, and downstream authority to make the system auditable. The authors build a controlled synthetic prototype to test how contextual knowledge interacts with evidence in low-data settings. Contextual structure helped relative to marginal fusion but not relative to equal-information neural controls, and the work makes no claims about TCM theory or clinical validity.
In briefAtomTree: An LLM-based hierarchical framework for state-aware embodied instruction following
SummaryThis paper presents AtomTree, a hierarchical framework that uses large language models to plan tasks for embodied agents following natural language instructions. The LLM incrementally builds a tree of subgoals and actions, deciding at each step whether to decompose further or output a grounded low-level action based on the current symbolic state. When a low-level action stalls, the system backtracks to the parent node and replans the subtree from the latest state. In zero-shot evaluation on the ALFRED benchmark, the authors report that AtomTree outperforms prior LLM-based planners.
In briefRunning a speech recognizer on a microcontroller using 2- and 4-bit weights on the ESP32-P4
SummaryThe paper ports Kroko, a 65.7M-parameter streaming Zipformer2 transducer for English, to the ESP32-P4 microcontroller. The encoder weights are quantized to 2 or 4 bits using Hadamard rotation and GPTQ, shrinking 71 MB of int8 weights to a 26.65 MB file. On LibriSpeech, word error rates are 4.34% on test-clean and 10.26% on test-other, compared with 3.94% and 9.17% for the released model.
WhyThe authors port a 65.7M streaming transducer to an accelerator-free ESP32-P4 using training-free 2/4-bit quantization, reporting WER on LibriSpeech, FLEURS and Common Voice, with GPTQ credited for mo
In briefRuleForge: Learning hidden visual rules and turning them into executable programs
SummaryThe paper introduces Bongard-Forge100k, a controlled benchmark of 108,797 problems, and RuleForge, a framework that jointly learns to understand visual rules and generate executable programs conditioned on those rules. Training combines rule-understanding, joint understanding and scene-program supervision, and execution-grounded reinforcement learning. With UniLIP as the backbone, RuleForge raises multiple-choice accuracy by 20.93 points on familiar rules and 28.42 points on held-out rule compositions, and increases scene-level rule satisfaction on familiar rules from 2.91% to 70.99%.
In briefAdversarial Self-Distillation: Keeping Language Models' Refusals Under Attack
SummaryThis paper addresses the problem that safety-aligned language models can still be bypassed by adversarial perturbations, and that standard adversarial training on template refusals does not reliably teach refusal. It proposes Adversarial Self-Distillation (ASD), in which the model given a clean prompt acts as teacher and the same model under attack acts as student, distilling safe behavior at the prefixes the student visits. According to the paper, ASD achieves lower attack success than state-of-the-art defenses with little loss in multiple-choice accuracy, and can recover a refusal even after a compliant opening.
Mohammad Sadegh Akhondzadeh, Soroush H. Zargarbashi, Aleksandar Bojchevski
In briefFrom Grind to Grasp: How Work, Knowledge and Self-Improvement Relate in Intelligent Systems
SummaryThe paper introduces a formalism that compares achievements under limits on both work and information, showing how learning and persistence can make deep achievements shallow by reusing earlier work. It defines "grasp" as performance across different work budgets at a fixed information allowance, and uses this to give a criterion for recursive self-improvement. The paper also discusses collectives and organisation, and cautions against letting an emergent agent society take authority over social coordination.
In briefQuine Test: How Much of a Model's Self Survives Its Own Description?
SummaryThe paper asks whether a language model's self-description can serve as a portable specification of its preference profile, testing whether the text can regenerate its behavior in other models. Ten models wrote self-descriptions, which were placed in the prompts of other models and fresh instances of the author, and their choices on a 110-item preference battery were compared to the author's. Native self-descriptions produced no reliable reconstruction gain, while explicitly listing the author's answers reproduced its profile with high fidelity.
WhyThe authors propose a quine test of self-descriptions as portable preference specifications, reporting experiments on 110 items across ten models.
In briefinkGround: A benchmark testing how well vision-language models trace handwritten words in photos
SummaryThis paper introduces inkGround, a benchmark for recovering pen stroke trajectories from full-page, camera-captured handwriting. It contains 537 page images with about 18.7k words and 163k strokes, with reference trajectories generated by a pretrained offline-to-online model, filtered by OCR agreement, and checked by human raters. The task asks models to locate a target word in the image and output its ordered stroke points, which is evaluated on inkGround and on the IAM-OnDB dataset across four open-source and three closed-source vision-language models. Closed-source models, especially Gemini-3.1-Pro, produce legible trajectories close to the reference, and this ranking also holds on IAM-OnDB.
In briefLearning scientific trajectories to predict future biomedical discoveries
SummaryThis paper introduces ForeSCITE, a framework that groups published biomedical conclusions into themes and projects them along directions separating earlier from more recent findings. A language model then decodes these projections into natural-language hypotheses. In retrospective time-sliced tests across 407 clusters from about three million conclusions, it outperformed five baselines on semantic coverage, entity-pair novelty, and post-cutoff re-observation. Experts identified 37 CAR-T hypotheses as both novel and biologically plausible.
In briefTraining clinical reasoning models with guideline-based decision flow rewards
SummaryThis paper introduces GUIDER, a method for training clinical reasoning models using decision flows derived from clinical practice guidelines. It jointly optimizes a decision flow controller and a reasoning policy, computing process rewards over several generated flows and gating them on answer correctness to address path multiplicity and the path-answer gap. On the MedGUIDE benchmark, GUIDER with Qwen3-14B reports 82.5 weighted accuracy and 73.5 guideline adherence, and it also outperforms CRPO, MRPO, and FaithMed on out-of-domain benchmarks.
In briefSatOps-Decision: Fast Decision-Making Method for Onboard Satellite Communication Operations
SummaryThis paper proposes SatOps-Decision, a method for making low-latency decisions in satellite communication operations. It uses a locally adapted Laya encoder with a discriminative decision head that selects among supplied operational answers in a single forward pass. On the same 1,024 held-out questions, its local scorer has a median latency of 90.3 ms per decision and a mean raw accuracy of 85.3%, outperforming the compared generative models and frozen Laya.
SummaryThe paper argues that packing a 3D volume into a token sequence removes the geometric meaning of sequence order, so position mechanisms indexed by sequence position can learn memory layout rather than the object's structure. It proposes PACK3D, a local-global attention operator that keeps local neighborhoods exact, compresses distant context into contiguous 3D block summaries, and applies axial RoPE on true grid coordinates. Experiments on a controlled synthetic task and on Allen WTC-11 hiPSC microscopy data show gains over baselines, and a fused Triton implementation is reported for scalable training.
WhyThe authors identify that packed 3D data makes positional encodings learn memory layout, and propose axial RoPE on grid coordinates; the authors report validation on a synthetic task and WTC-11 micros
In briefBeyond Fabrication: Numbers Misattributed in AI-Agent-Written Papers
SummaryThe paper re-analyzes 540 judgments from a public benchmark of AI research agents and manually labels 256 numerical claims from 18 submissions. It finds that most unsupported claims stem from misalignment between the paper and the underlying work, with about four in five matching values from the wrong experiment, metric, or run in the agent's files. The authors propose a deterministic claim-to-evidence checker that certifies claims with over 90% precision, while noting that failing to verify a claim does not mean it is false.
WhyThe authors attribute failures to paper-work misalignment and propose a deterministic claim-evidence checker calibrated with numerical decoys. They report validation on 256 labeled claims and 99 submi
Mosab Alfaqeeh, Omar Hourani, Muath Hourani et al.
In briefDeogen: automatic benchmark generation for testing rule-based reasoning
SummaryThis paper presents DeoGen, a scalable and parametrizable tool that automatically generates benchmarks for deontic reasoning, which is applying rules to facts to reach a verdict. It autoformalizes policy rules into first-order logic and builds decision trees whose paths become benchmark entries. Frontier models gave correct verdicts 89-92% of the time on entries with enough information, but only 22-40% correct on entries where they should abstain, even when instructed to do so.
In briefGeneralization gap of federated learning on Markovian streaming data
SummaryThis paper studies the expected generalization gap of a shared model trained by streaming federated learning when client data follow Markov processes. It uses information-theoretic and rate-distortion techniques to derive bounds under cross-entropy loss, specialized to linear and self-attention next-token prediction. The bounds reveal a communication–generalization tradeoff, where frequent aggregation may weaken generalization. Experiments use GPT-style language-model pretraining on TinyStories and real-world time-series forecasting.
Masoud Kavian, Milad Sefidgaran, Youssef Attia El Hili et al.
In briefRanking merger that learns from selective LLM pairwise judgments for document reranking
SummaryThis paper introduces LLM-IRA, an extension of a quantum-inspired interactive ranking aggregation framework to document reranking. It uses a large language model only for pairwise judgments on high-disagreement document pairs to learn fusion weights, and the final ranking comes from fast pointwise linear score fusion. Experiments on BEIR and TREC Deep Learning show it matches or numerically exceeds strong static fusion baselines on 5 of 6 benchmarks, with statistical significance on 4 of them.
In briefWeaken to Strengthen: A Self-Distillation Method for Reasoning Models
SummaryThis paper introduces WISD, a self-distillation method that improves large reasoning models without an external teacher or privileged information. It builds a weaker version of the original model through low-bit quantization or magnitude pruning, then uses the contrast between the two as a training signal. The authors report average accuracy gains of 4.0 points for quantization and 5.5 points for pruning across math and coding reasoning benchmarks.
