In briefAgent Board: a durable kanban-style coordination layer for agents that outlive a single conversation
SummaryThis paper presents the Agent Board, a coordination substrate in the Corvic harness: a durable shared table where agents post, watch, and block on semi-structured tasks without holding a process while they wait. A system agent, the Board Keeper, carries events from outside the agent system onto the board, matches them to waiting tasks by reading their text, and allocates agents for tasks that name a kind of agent. In three small simulated pilots, the board's wait cost stayed flat as wait length grew while naive polling grew about 165-fold, but a plain external watcher was cheaper; the Board Keeper produced no false wakes but missed 3.6 percent of real wakes.
In briefEvaluation Contracts and Evaluation Leverage: How Sliced Gaps Can Shift Without Model Changes
SummaryThis paper argues that sliced performance gaps, such as the intra-inter gap in document-level relation extraction, depend jointly on the model and the evaluation contract, including scoring function, gold version, threshold, and output format. With predictions held fixed, swapping the gold standard reverses a method ordering. The authors propose an evaluation leverage framework and EvalLever, a preregistered, audited evaluation workflow.
WhyThe authors propose an evaluation-leverage framework showing that frozen predictions can reverse method rankings under gold or scoring contract changes, plus an audited procedure and released workflow
In briefScaleTraj: Generating Human Movement Paths by Splitting Trips Across Spatial Scales
SummaryThis paper addresses generating realistic individual human movement trajectories while also matching population-level mobility patterns. The authors formalize the Population-Level Displacement Gap, where bottom-up generated populations fail to reproduce the heavy-tailed displacement distribution, and introduce the Scale Fidelity Profile to diagnose it. They propose ScaleTraj, an autoregressive Transformer that first selects a movement scale and then predicts the destination conditioned on that scale. Experiments on two datasets show that ScaleTraj reduces this gap compared with model-driven and data-driven baselines while preserving fidelity.
In briefCHROMA-MILL: Joint Low-Light Enhancement and White Balance Under Colored Light
SummaryThe paper introduces CHROMA-MILL, a RAW image dataset of 50 scenes captured with a DSLR and a smartphone under four calibrated illuminants at eleven intensity levels. Captures are blended in the linear RAW domain to create physically valid illuminant mixtures with exact chromaticity ground truth. The authors propose CTI-Retinexformer, which jointly supervises an illumination map, a chromaticity prediction, and an illumination-invariant scene embedding. It outperforms the best cascaded approach by 0.7 dB PSNR_L and 1.1 ΔE00.
Pol Riubrogent-Comas, David Serrano-Lozano, Maria Pilligua et al.
In briefSame rule, different verdicts: how rubric language and wording affect LLM judges in Korean support dialogues
SummaryThis paper examines whether an LLM judge's verdict depends on the language of its evaluation rubric and on how each rubric item is worded. The authors wrote 12 criteria in four clause structures (affirmative, negated, conditional, exception), provided each in Korean and English, and had seven models judge translated ESConv dialogues and boundary cases. Cross-provider agreement was lower with Korean rubrics than with English ones, and disagreement concentrated in negated, conditional, and exception items.
In briefChanging the LLM judge changes system rankings: a preregistered study of cross-provider replication
SummaryThis preregistered study re-judges 4,000 human-voted Chatbot Arena battles with five LLM judges from three providers to test whether conclusions can be replicated with a judge from another provider. Agreement with human votes is low, and system rankings shift substantially depending on the judge used. The preregistered hypothesis that rankings remain stable was not supported.
Sangdo Han (CEO Business School), Moonsoo Kim (CEO Business School)
In briefScienceVision: Can AI Observe and Reason Like a Scientist?
SummaryThe paper introduces ScienceVision, a benchmark testing whether AI models can understand and reason about scientific 3D data across biology, chemistry, medical imaging, mechanical CAD, and architecture. It contains 250 expert-level visual question answering items, 50 per domain, built with synthetic generation and expert review. Across 12 models, the best mean score was 0.53 out of 1. Workflows using native 3D artifacts outperformed rendered 2D images for every model, by 17.2 percentage points on average.
In briefFollow the Features: Using sparse autoencoder features to select small instruction-tuning datasets for language models
SummaryThis paper introduces Follow the Features (FtF), a method that uses sparse autoencoder (SAE) features to pick a small subset of instruction-tuning data. It ranks features by dataset-level activation statistics, scores samples by how strongly they express top-ranked features, and resamples features to promote diversity. At a 100-sample budget, FtF performs competitively with GRAPE and S2L, and on Qwen 3 8B trained on Alpaca it beats full-dataset training on the harmonic mean of AlpacaEval win rate and Open LLM Leaderboard score, though full-dataset training scores higher on the Leaderboard alone.
In briefFrom Visual Matching to Textual Conditioning: Cross-Domain Few-Shot Segmentation via Reference Inversion
SummaryThis paper addresses cross-domain few-shot segmentation, which segments objects in unseen domains using only a few annotated references. The authors propose RI-SAM3, which builds on SAM3's text-conditioned segmentation by learning suitable textual tokens for each target category instead of relying on visual matching. With SAM3 frozen, the method optimizes about 12K parameters per category and reports superior performance across multiple benchmarks.
In briefRAPID: Faster Internal Distillation Using Randomized Projections
SummaryThis paper introduces RAPID, a method for distilling a teacher model's internal representations into a smaller student model. Existing approaches learn projection matrices to match dimensions, and these are discarded after training, which wastes compute. RAPID replaces learned projections with randomized projections based on the Johnson-Lindenstrauss lemma, and combines exponential moving average with homoscedastic uncertainty weighting in its objective. In post-training experiments on reasoning tasks, RAPID improves training efficiency by up to 21.1% while performing comparably to existing methods.
In briefAlgoPilot: Cross-Paradigm Reasoning in Language Models via Strategy Selection
SummaryThis paper proposes AlgoPilot, a framework that helps language models first choose a suitable problem-solving paradigm, then formulate the problem in structured form before execution. It has three parts: SteerLM selects the strategy, GuideLM generates formulation guidance for that strategy, and ExecutionLM produces the solution. Built on Qwen3-8B and tested on 13 benchmarks across 5 problem categories, it raises average accuracy from 38.4% to 72.9%. The authors also report that the learned steering and formulation modules transfer across execution models.
In briefHow hard it is to build a ReLU network from a fixed set of parameter values
SummaryThe paper studies whether a ReLU neural network with a bounded number of neurons can be built so that, with its weights and biases taken from a given multiset of rational numbers, it meets a target behavior. It analyzes the computational complexity of this synthesis problem under different requirements, such as exact agreement on examples, agreement on all inputs, or a bound on expected loss. The results range from NP-complete to Σ2P-, DP- and NP^PP-complete, depending on the setting.
In briefTutorial-Conditioned Computer-Use Agents Gain Little from World Models
SummaryThis paper isolates the contributions of tutorials, a judge, and world models to computer-use agents. The authors build two tutorial datasets, OSWorld-Tutorial and ScienceBoard-Tutorial, and propose RAG-PolicyJudge, where a policy proposes actions and a judge selects among them. They find that with tutorials present, even perfect visual future observations give little benefit, and error analysis points to the judge's reasoning over future states and task completion as the bottleneck.
In briefFinite-sample convergence of multi-agent learning on networks, and how entropy regularization helps
SummaryThis paper studies average reward multi-agent reinforcement learning, where agents in a network each update using only information from their k-hop neighborhood while jointly maximizing total network reward. It first proves impossibility results for converging to globally optimal joint policies in this decentralized setting. It then gives finite-sample guarantees for a decentralized actor-critic algorithm with linear function approximation, converging to an approximate stationary point with small gradient norm. Adding fixed-state entropy regularization is shown to reshape the objective landscape and yield stronger global convergence guarantees for this algorithm.
In briefConformal prediction that stays accurate for each part of a graph, not just on average
SummaryThis paper introduces GLocalCP, an online conformal prediction framework for inductive node classification and edge prediction on graphs. It combines marginal and graph-local calibration, using a nonconformity score based on mutual information from Monte-Carlo dropout. Experiments show that similar overall coverage can hide very different local reliability, and that GLocalCP reduces local calibration error while shrinking prediction sets.
In briefTB-Transolver: Faster, Coherent Long-Range Forecasts for Time-Dependent Physics Simulations via Temporal Bundling
SummaryThe paper presents TB-Transolver, which scales the Transolver model to time-dependent PDEs by grouping snapshots into temporal bundles. It also introduces a spatiotemporal coherence metric that uses only training data to choose bundle widths in advance. Across plasticity, fluid dynamics, reaction-diffusion, and global weather forecasting, the method trains up to 29 times faster than autoregressive and time-conditioned baselines. Multi-node, multi-GPU parallelism further raises the aggregated speedup to as much as 45.6 times for highly coherent systems.
In briefStreamLTT: Risk Certification for Choosing Policies in Dependent Data Streams
SummaryThis paper introduces StreamLTT, a method for certifying deployment policies against a prescribed risk target when losses are dependent and change over time. It replaces the fixed population risk with an average predictable risk, conditioning each decision's loss on information available before its outcome is revealed. Combined with family-wise error rate control, it yields a simultaneously certified policy region. Across seven forecasting datasets and seven deployment risks, its certified frontier lies closer to the risk target than several alternative methods, while empirical family-wise error closely follows the prescribed level.
SummaryThis paper examines whether Reinforcement Learning with Verifiable Reward (RLVR) can extend the reasoning abilities of large language models beyond what the base model already encodes. Through the lens of primacy bias, the authors show that RLVR on mixed data mainly sharpens performance on easier problems, with only marginal gains on harder ones. Hard problems contribute most to expanding reasoning, but prolonged training on them may cause forgetting on easy problems.
Phuc Minh Nguyen, Ngoc-Hieu Nguyen, Parshin Shojaee et al.
In briefQuery-based submodular data selection for efficient instruction tuning
SummaryThis paper proposes QSGS, a method for choosing a small subset of training examples from a large, mixed pool to fine-tune language models on a target task. It frames selection as a submodular graph cut over a gradient similarity graph, balancing relevance to the target with a penalty on redundancy among selected examples. A gradient-guided step synthesizes examples for under-covered target directions. Experiments on MMLU, BBH, and TydiQA with Llama-2, Llama-3, and Mistral report larger gains as the data budget shrinks.
Zhenghao Zhao, Yasutomo Kawanishi, Huihuo Zheng et al.
In briefPredicting Before Launch: Learned Throughput Forecasting for Choosing Hardware in a Production Training Fleet
SummaryThe paper presents LEAP, a machine learning service that predicts each training job's throughput on every candidate hardware type at submission time and ranks the options by predicted throughput per unit cost. It routes tens of thousands of job attempts daily across hundreds of model types and more than 20 hardware types in a production fleet. Over 30 nightly retrains, LEAP reports 77% accuracy, and the share of eligible training GPU-hours receiving a ranking rose from 58% to 95% over six months. The authors estimate savings of about 3% in training computational cost.
In briefStructured priors: an algebra of component priors for prior-data fitted networks
SummaryPrior-data fitted networks are usually trained on synthetic data from one hand-built generator. This paper proposes structured priors, sets of named component priors linked by operators (mixture, convolution, amplitude scaling, local transform, projection) with proven algebraic properties, implemented in the macl2l package. A mixed prior of small components matched or beat every single component on seven of eight regression benchmarks. A pre-registered test of the model-averaging predictor failed, because the networks rarely identified the generating component.
WhyThe authors propose an algebra of structured priors with proven operator laws, evaluated on eight regression benchmarks, and report a failed pre-registered test.
In briefHybridSpec: Throughput-aware speculative decoding for hybrid GDN-MoE language models
SummaryThis paper studies speculative decoding for hybrid language models that combine Gated DeltaNet and mixture-of-experts layers, where verification cost grows quickly with the number of draft tokens checked. The authors propose HybridSpec, which chooses the verification length on the GPU using acceptance estimates and measured latency, applies a lightweight routing check to keep or drop a token, and uses block-parallel kernels to reconstruct recurrent states. In single-request greedy decoding on three such models, HybridSpec is reported to be 1.75 to 1.88 times faster than autoregressive decoding and 1.30 to 1.33 times faster than DFlash.
In briefForgetting to Improve: a principled way to remove harmful data in active learning
SummaryThe paper proposes Forgetting to Improve, a method for cases where model mismatch means some data harms modeling for a specific task. The authors derive an influence function for Gaussian process models that allows detrimental data samples to be removed in a principled way. The approach is reported to give significant improvements on Bayesian optimization, model-based reinforcement learning, and transductive learning.
Manuel Wendl, Erik Englesson, Andreas Krause et al.
In briefDeep learning to sharpen and denoise 4D heart blood-flow MRI scans
SummaryThis paper introduces 4D-MRAR, a deep learning method that improves 4D Flow MRI of blood flow inside the heart by increasing spatial and temporal resolution and reducing noise. It combines a residual convolutional network, a recurrent network, and an implicit neural field, and enforces fluid flow governing equations during training. The model was trained on left ventricle simulation data and tested on synthetic whole-heart data and real scans.
In briefExploit or conform? How prompting for norm utility changes LLMs' use of community beliefs
SummaryThe paper studies how three open-weight large language models make one-shot decisions in two-player social dilemmas when told what the community believes. Without guidance on how those beliefs should matter, cooperation in the Snowdrift game fell as community belief in cooperation rose. When the components of norm utility were explicitly prompted, this relationship reversed and cooperation increased, broadly matching a simple norm utility model. Norm utility prompting also reduced the volatility of decisions across changing social contexts and produced more consistent reasoning across models.
Leon Klingborg, Theodor Cimpeanu, Alexander J. Stewart
In briefDenoising as Stein: learning hidden structure in diffusion models via moments
SummaryThe paper studies learning a score function from denoising pairs when data contains a hidden non-Gaussian component. If covariance reveals the component, its subspace is recovered directly; if covariance is isotropic, the learner must extract the first distinguishing moment. It gives exponent results for feature and sample requirements across learners and proposes a two-stage method that obtains a fixed score gain over the Gaussian baseline.
In briefTesting whether logits processors get the right inputs in speculative decoding engines
SummarySpeculative sampling speeds up text decoding without changing the target output distribution, but this guarantee assumes that constraint-enforcing processors receive the correct token prefix and state at each step. The paper formalizes this requirement as a processor-prefix contract and introduces BADC, a branch-aware differential conformance testing method. Experiments on selected Transformers and llama-cpp-python releases confirm two causes of failure: processors reading unverified drafts, or advancing state without rollback.
In briefWhen Model Complexity Helps in Predicting Brain Responses to Movies Across People
SummaryThis paper examines when more complex neural models are useful for predicting blood-oxygen-level-dependent (BOLD) brain responses to naturalistic stimuli. The authors compare linear and nonlinear predictors on three public fMRI datasets, with and without recently measured BOLD history. They report that pretrained visual features generally outperform raw pixels, and that measured history gives substantial gains that generally shrink as rollouts lengthen.
In briefRecursa: A Self-Improving Foundation Model System
SummaryThis paper introduces Recursa, a foundation model system that adapts to new domains by learning from its own experience, without human experts or a stronger teacher model. It stores domain skills, guidance on when to apply them, past task experiences, and reflections in a knowledge repository, which it refines through recursive self-improvement using parallel learners. On Harvey LAB's Firm Knowledge benchmark, the all-pass rate on held-out tasks rose from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.
In briefDo Pretrained Backbones Transfer to Physical Sensors? Multimodal Initialization Study
SummaryThis paper examines how well pretrained vision backbones transfer to multi-sensor autonomous driving tasks. The authors assemble 2.82M images from 16 public datasets, covering physical sensor modalities such as RGB, LWIR, NIR, and event cameras, plus model-derived modalities like stereo depth and segmentation. They perform in-domain MultiMAE-based pretraining and compare it against ViT, MAE, DINO, and MultiMAE backbones on downstream sensor fusion tasks. They report that MultiMAE is a strong out-of-the-box baseline that transfers even to unseen modalities, and their targeted modifications give modest but consistent gains.
Fabio Montello, Ronja G\"uldenring, Lucas Mogensen Sandby et al.
In briefROCm InferaSim: a tool that predicts LLM serving performance to cut configuration trial costs
SummaryThe paper presents ROCm InferaSim, which predicts large language model serving performance from a few short measurements instead of exhaustively testing every configuration. It benchmarks each distinct kernel set once on at most one node and extends those results to other deployments with analytical laws. On AMD MI355X GPUs, 11 anchors took 9.1 GPU-hours, versus 204 GPU-hours to measure the 226 configurations they cover. Time per output token has 8.1% median error across 172 configurations, and throughput is within 5.5% for 159 disaggregated deployments not used as anchors.
In briefLanguage Model Augmented Decision Transformer for Multi-Task Offline RL
SummaryThis paper addresses two problems in offline multi-task reinforcement learning: scarce data per task and limited generalization across tasks. It proposes LaMA-DT, which uses one large language model to generate task-aligned trajectories from natural-language task descriptions, filter them for quality, and then serve as a language-conditioned Decision Transformer fine-tuned on real and augmented data. Experiments on Meta-World and MuJoCo benchmarks show it consistently outperforms competitive baselines, especially in low-data settings.
In briefCompression Forcing: Hierarchical History Tokenization for Flexible World Models
SummaryThis paper proposes Compression Forcing, a framework that aligns history compression with the predictive goals of world modeling. Its Dynamics-Aware Tokenization compresses past observations into a sequence ordered by decreasing dynamics saliency, enabling Hierarchical History Guidance on a diffusion backbone during inference. Experiments across environments report improved prediction fidelity, physical realism, and downstream decision-making.
In briefA practical causal framework for mobile robots moving from simulation to real human spaces
SummaryThis paper presents a causal inference framework that helps mobile robots reason about why events happen, enabling safer and more efficient decisions. It transfers causal relations learned in simulation to the real world by keeping the high-level causal structure fixed and updating only the causal model parameters with a limited set of new observations. The framework runs onboard a mobile robot and, according to the authors, significantly outperforms a non-causal baseline in navigating safely in complex human-shared scenarios.
In briefDUNEDepth: Estimating Bronchoscope Depth and Camera Pose by Adapting a 3D-Vision Foundation Model
SummaryThe paper presents DUNEDepth, a model that estimates depth and camera motion from bronchoscopy exam videos so they can be linked to a patient's pre-operative 3D chest CT scan. It combines a shared DUNE ViT encoder, a twin MASt3R cross-attention decoder, and a prediction head drawing on MASt3R and Reloc3r. Training uses a hybrid strategy of synthetic supervised and self-supervised learning, drawing on bronchoscopic video and CT-derived virtual bronchoscopic views. Experiments show more accurate depth and camera trajectory estimates than previously proposed self-supervised methods.
In briefSurgEye: A Surgical Scene Reasoning Framework That Needs No Instruction Tuning
SummaryThis paper presents SurgEye, a framework for understanding surgical scenes that avoids instruction tuning on large question-answer datasets. It uses a vision foundation model with routed multi-task perception to extract spatial and temporal evidence, and a frozen large language model that coordinates these perception tools to compose grounded answers. The authors also build SurgEye-Bench, covering eight surgical procedures with perception labels for verifiable evaluation, and report that SurgEye outperforms state-of-the-art methods in surgical reasoning.
In briefTask-aware Decision Transformer for offline meta-RL that learns task representations
SummaryThis paper proposes CoTA-DT, a Decision Transformer-based framework for offline meta-reinforcement learning. It learns task representations by combining contrastive learning, alignment with natural-language task descriptions, a task-conditioned world model, and a task-conditioned Q function, then uses them to condition the policy. The authors report that it consistently outperforms prior methods across offline meta-RL benchmarks, especially in zero-shot generalization.
In briefPhysics-informed estimation of tensegrity prestress through cable slackening
SummaryThe paper formulates an inverse problem for identifying prestress, damping and axial stiffness in tensegrity structures from sparse measurements. Neural networks are trained on the equation-of-motion residual in constrained Cartesian coordinates, keeping a tension-only cable law so that slack cables are handled. The method is tested on simulated frames and on measured records of a biomimetic leg and a full-scale simplex, with reported identification errors.
WhyThe authors extend physics-informed networks from forward tensegrity response to inverse prestress identification across taut-slack transitions, reporting small errors on simulated and measured record
In briefControlling how action heads co-adapt to anisotropic VLM representations
SummaryThe paper studies how a downstream action head adapts to the geometry of a pretrained vision-language model in vision-language-action systems. It finds that the action model increasingly focuses its sensitivity on high-variance dimensions of the VLM outputs. The authors propose AdaIE, a parameter-free method that reweights pretrained representations during training, which improves success rates on simulation and real-world manipulation benchmarks.
In briefStudying adversarial generative engine optimization through implicit brand injection
SummaryThis paper examines how a brand can be promoted in generative search answers by inserting a single, contextually blended mention into a web article. The authors propose Implicit Brand Injection (IBI), a multi-agent framework that selects insertion points, rewrites sentences, verifies edits, and chooses final candidates. They evaluate it on IBIBench, a benchmark of 1,000 articles and 60 synthetic brands across 12 domains. IBI preserves semantics better and improves retrieval, reaching up to 82.2% Hit@10 and 62.7% brand mention rate.
In briefWASP-Mem: Agents Organize Procedural Memory, Humans Focus on Diagnosis
SummaryThe paper introduces WASP-Mem, a method that lets LLM agents distill execution traces and feedback into persistent, interlinked wiki pages for continual learning. On the TheAgentCompany benchmark, it raises checkpoint pass rates by 4.6-6.1% over the strongest compared baseline across three models while requiring 57-61% less storage. The study finds that prescriptive human-defined schemas hurt performance, and a single round of human diagnosis outperforms four rounds of self-judgment by 7.0-10.9%.
In briefEffect of Nonlinear Registration on Longitudinal PET Analysis in a Rat Stroke Model
SummaryThis study compares two pipelines for aligning rat PET images to a common atlas: affine-only registration and a two-stage affine plus nonlinear registration. It uses 32 longitudinal [18F]-FDG PET scans from 8 rats across 4 timepoints. The nonlinear stage yielded higher Dice similarity coefficients on every scan, with a significant gain at the rat level. However, group-level linear mixed-effects maps differ between the two pipelines in the size, number and location of suprathreshold clusters, showing that the registration choice can change downstream interpretation.
Haytham Ziani, El Mostafa Kalmoun, Zohaib Atif et al.
In briefDory1.0-4B: A 4B-Parameter Language Model That Adjusts Depth via Looping
SummaryThis paper introduces Dory1.0-4B, a 4-billion-parameter language model whose recurrent block can be applied a variable number of times, so one checkpoint serves different compute budgets. It is built on a normalized Transformer backbone and pre-trained on 6 trillion tokens, then post-trained with supervised fine-tuning and reinforcement learning. Increasing the loop count improves accuracy, most notably on coding, agentic, and long-context benchmarks.
Maxim Fishman, Simeng Sun, Shantanu Acharya et al.
In briefMeasure Before You Repair: An Explainable AI Agent for Blind 3-D MRI Artifact Correction
SummaryThis paper presents Dx-Rx, an agent that uses a multimodal large language model to restore whole 3-D brain MRI volumes without knowing which artifacts are present. It first measures suspected artifacts such as noise, spikes, ghosting, bias field and motion with diagnostic tools, then applies repair tools and checks each result. It ends with a radiology-style report explaining its diagnoses and repairs. On 148 simulated T1-weighted volumes, the best model removed a mean 64.5% of artifact error, close to the 64.6% achieved when the true artifacts were supplied.
In briefWhen Music Needs Meaning: Semantic Communication for the Musical Metaverse
SummaryThis paper examines the communication demands of the Musical Metaverse, where distributed musicians, audiences, sound engineers, and AI agents interact through synchronized audio, visual, kinematic, and haptic streams. The authors argue that semantic communication, which transmits musically meaningful information rather than raw samples, can address its latency, synchronization, and bandwidth needs. They propose a three-pillar framework covering semantics extraction, semantic-aware orchestration, and cross-modal synchronization, along with a semantic-first reference architecture. The paper presents the Musical Metaverse as a demanding use case for 6G semantic communication.
Ali Al Housseini, Jaime Llorca, Cristina Rottondi et al.
In briefDoes an AI crowdfunding advisor favor founders by gender or ethnicity? An algorithm audit of LLM choices
SummaryThis paper examines whether large language model assistants that advise backers on crowdfunding campaigns favor founders based on gender or ethnicity signals. In a pre-specified choice experiment, seven LLMs chose among five synthetic campaigns with randomized quality signals and founder names, yielding 40,248 responses. By default founder names had negligible effects, but when backers wanted to support underrepresented creators, models favored Black-signaled and female-signaled founders.
WhyThe authors report that LLM assistants favor Black and female founders only when backers seek to support underrepresented creators, across seven models and about 40,248 controlled choices.
In briefFlagRot: selective rotation to make 2-bit memory caches for AI models more accurate
SummaryLong-context AI inference is limited by the KV cache, which 2-bit encoding could shrink about sixfold if the error stays small. Prior methods rotate keys before quantization, but the authors find rotation does not always help. FlagRot stores each 32-channel group either rotated or unrotated, choosing whichever has lower calibrated attention error, and rotates the query at read time. At 2.5 bits per value it lowers KIVI's KL divergence across twelve models and improves RULER at 128k tokens and LongBench v1 scores.
In briefA guideline-based multi-agent system for personalized exercise prescriptions from clinical and wearable data
SummaryThe paper presents FITTER, a multi-agent framework that combines clinical data, wearable device data, and condition-specific exercise guidelines to generate personalized FITT-VP exercise prescriptions. It applies deterministic safety auditing and repair to correct violations that can be checked against rules. Evaluated on data from 400 All of Us participants against five LLM comparators, FITTER achieved 98.5% safety-constraint compliance, compared with 75.8-85.8% for the comparators. Clinicians also reviewed 80 prescriptions.
In briefSLATE: A Large-Scale Benchmark for Testing AI Agents' Long-Task API Selection in E-commerce
SummaryThis paper introduces SLATE, a benchmark that tests how well LLM agents select and chain the right APIs from a library of about 1,000 e-commerce tools over long, multi-step procedures. It uses a deterministic simulator with recorded outputs and scores agents by whether they reach all required task outcomes. In experiments with six open models ranging from 3B to 12B parameters, the best model completed only 5.0% of tasks.
In briefA fast mean-field method for moving multi-agent formations using one shared neural controller.
SummaryThis paper proposes a mean-field formation transport method for large-scale multi-agent systems that has low online computational cost. A single shared parameterized nonlinear map approximates both the density and the control. Its parameters are optimized using Fokker-Planck residuals, formation-matching costs, and control effort. Experiments show a favorable accuracy-runtime trade-off and lower transport error than classical feedback baselines.
In briefMeasuring model-specific alignment drift in LLM-based opinion dynamics
SummaryThis paper uses large language model agents as synthetic participants in opinion-dynamics experiments, letting them discuss a contested policy issue over repeated rounds. Stances are measured with a fixed psychometric instrument that is independent of the dynamics model. Without peer input, agents keep their assigned side throughout, so the loss of alignment under discussion is interaction-induced and varies by model. The authors conclude that the choice of LLM is a key validity variable for social simulation.
Nata\v{s}a Djurdjevac Conrad, Sebastian Pokutta, Christof Sch\"utte
In briefMust AI Spell It Out? Understanding Alerts in Time-Critical Tasks
SummaryThis study examines whether people can use their prior knowledge to understand incomplete alerts from a machine in a time-critical monitoring task. In a preregistered experiment (N=114), a pragmatic planner that reasons over a model of the listener's knowledge was compared with a baseline that describes problems fully in onset order. Participants resolved underspecified alerts almost as reliably as fully specified ones, responded sooner, detected more problems, and rated the pragmatic assistant helpful more often without reporting higher mental demand.
In briefOne Mistake, Three Layers: Persistence Rules for Layer-wise Probing
SummaryThe paper examines how to decide the layer at which a signal first becomes decodable from frozen network representations. It shows that a common rule requiring three consecutive layers above threshold produces false early onsets when neighboring layers are correlated, and proposes an estimator calibrated by permuting labels within each evaluation unit. Applied to pocket information in a co-folding model, it also reports that a public checkpoint loads with 128 trunk parameters missing.
Maryam Rahimimovassagh, Michael Brown, Sanjeda Sara Jennifer et al.
In briefStacking-Based Weighting for Large Language Bayes Under Model Misspecification
SummaryThis paper proposes a stacking-based method for combining candidate models in Large Language Bayes (LLB), which turns natural language problem descriptions into probabilistic programs. Instead of Bayesian model averaging by marginal likelihood, the method chooses model weights by optimizing predictive performance. Because leave-one-out predictions are intractable in latent-variable models, the authors derive a tractable ELBO-based stacking objective. On real-world inference tasks, it matches Bayesian model averaging in predictive performance while producing more dispersed weights.
Anant Bhide, Akarsh Gupta, Edmond Cunningham et al.
In briefErmakov-based parameter-free embeddings for physics-informed neural networks
SummaryThe paper proposes embedding the classical Ermakov-Pinney scale function into the input layer of a physics-informed neural network as a parameter-free feature, for evolution equations with time-dependent coefficients. The scale function is obtained in closed form from the initial data via an exact moment identity, requiring no auxiliary network or knowledge of the solution. The authors compare three problem families and find that each requires a different component of the construction, with the benefit varying across them.
Roberto Bernal-Jaquez, Roberto Romero Huerta, Alexander Schaum
In briefProteoCraft: protein design framed as a propose-then-verify search
SummaryThis paper treats protein design as a search optimization problem over combinatorial sequence space where structural evaluation is expensive. It introduces ProteoCraft, a tree-based search framework that uses approximate structural gradients from diffusion-based structure predictors to propose where to explore, and no-gradient folding model inference to verify sequences. Across fourteen de novo binder-design settings, the authors report improved design quality and high-confidence yield, and an 82.9% binding hit rate for PD-L1 miniproteins.
WhyThe authors introduce a tree-based search using diffusion-predictor gradients as proposal policies, evaluated across fourteen design settings and with prospective experiments.
In briefSampling interventional outcome distributions with orthogonal MMD flow
SummaryThe paper proposes IMMD-Flow, a particle-based method that represents a target interventional outcome distribution through its kernel mean embedding and samples from it via MMD gradient flow. The kernel embedding estimator is doubly robust and Neyman-orthogonal, and the method supports both hard and soft interventions. It extends to high-dimensional image outcomes using frozen encoders. On image benchmarks, it matches or improves on state-of-the-art generative approaches while being significantly faster.
In briefA conditional adversarial diffusion model that predicts how diabetic macular edema patients will respond to anti-VEGF eye injections.
SummaryThe paper proposes a conditional adversarial diffusion framework that synthesizes post-treatment OCT eye images from baseline scans to estimate a patient's initial response to anti-VEGF therapy. It was trained with 5-fold cross-validation on a US cohort of 127 patients and externally validated on an Iranian cohort of 44 patients. At the 10% and 20% central subfield thickness reduction thresholds, it achieved the highest balanced accuracy among the compared methods in both cohorts.
Reza Rasti, Armin Biglari, Sobhan Mohammadi Hafshejani et al.
In briefForget What Comes Next: a statistical-independence method for unlearning in LLMs
SummaryThis paper introduces PINT, a machine unlearning algorithm for large language models that uses the Hilbert-Schmidt Independence Criterion (HSIC) to reduce statistical dependence between tokens to be forgotten and the internal representations used to predict them. The authors relate this objective to Exact Memorization through a computable upper bound. Experiments across eight settings, with models up to 70B parameters, cover unlearning Gaussian poisons, personal information, hazardous biology and cybersecurity knowledge, and copyrighted books and news. PINT achieves the best forgetting-utility trade-off on average and is the only method that makes Gaussian poisons undetectable while keeping utility near the original model.
WhyUses HSIC to reduce dependence between forgotten tokens and predictive representations, with a computable bound to exact memorization; the authors report experiments across eight settings up to 70B.
In briefP²-MOPD: Speeding Up Multi-Teacher On-Policy Distillation Using Parameter Priors
SummaryThis paper studies multi-teacher on-policy distillation, which transfers expert abilities to a student model by supervising the trajectories the student generates. The authors treat the linear combination of student and expert parameters as one differentiable model and propose P²-MOPD, which jointly optimizes the mixing coefficients and student weights under the same distillation objective. Experiments across four domains at 1.7B and 4B scales show faster and more balanced recovery than MOPD, with the 1.7B model reaching 94.7% average recovery versus 80.9% for MOPD.
In briefDo Point Cloud Models Still Need Built-In Rotation Invariance? A Robustness Benchmark
SummaryThis paper benchmarks 10 point cloud classification architectures under pose, noise, and occlusion perturbations across 5 domains, using confirmatory hypothesis tests. Rotation-invariant architectures show the best pose robustness but rank among the lowest on noise robustness in 4 of 5 domains. The authors retrain all models with a simple invariance loss and evaluate them on ModelNet40-C, finding that invariance training reliably improves only the corruption types it targets. At matched training budget, the best standard backbone and the best rotation-invariant architecture perform on par on unseen corruptions.
Abdulla Nasir Chowdhury, Gulpi Qorik Oktagalu Pratamasunu, Kazuhiro Fukui
In briefWeighting Chain-of-Thought and Answers When Fine-Tuning Small Medical Language Models
SummaryThis paper compares two supervised fine-tuning methods for small language models (up to 9B parameters) used in local medical question answering. Uniform-across-Sequence weights the full reasoning-and-answer sequence equally, while Segment-weighted weights the reasoning and answer losses separately. Evaluated on multiple-choice and free-text medical QA, the study finds neither method is universally best: Segment-weighted generally does better on multiple-choice, and Uniform-across-Sequence performs best on free-text.
In briefA research sprint with AI: the AI worked, but the research didn't
SummaryThe paper argues that individual productivity gains from generative AI research tools do not guarantee better team research, and that how AI use is structured matters as much as access. Through a 100-minute sprint pilot with twelve researchers in three independent teams, the authors observed two recurring patterns: a consumption bottleneck and AI-induced research silos. They propose a Human-AI-Human collaboration framework and seven falsifiable design hypotheses for follow-up study.
In briefCamera, radar and LiDAR fusion using projected point clouds for multi-task learning
SummaryThis paper proposes an early fusion architecture that combines camera images with perspective-view point cloud data from HD radar and LiDAR for autonomous driving perception. The model is trained to detect vehicles and segment free drivable space simultaneously. Experiments on the RADIal dataset show that front-view fusion of point clouds and camera images outperforms recent standalone and fusion models.
In briefFixing Coordinates, Class-Wise NMS and Null Returns for Sim2Real V2X LiDAR Detection
SummaryThis paper describes an entry to the V2X-Real track of the UrbanTwin Sim2Real LiDAR Challenge, where a 3D detector is trained only on synthetic point clouds and tested on real roadside LiDAR. The authors report three corrections: recovering a coordinate frame mismatch, using per-class non-maximum suppression so nearby pedestrians are not suppressed, and modelling null returns encoded as (0, 0, 0). The final combined score is 0.7457.
In briefFlowBench: A training resource for counterfactual spatial imagination
SummaryThis paper formalizes counterfactual spatial imagination, the ability to predict how a scene changes under hypothetical motions, and introduces FlowBench with 428K training examples across simulated and real-world scenes. Fine-tuning Qwen3-VL-8B on the simulated training split raises accuracy by 29.5 points in simulation and 16.9 points on real images. Training on FlowBench-Sim also raises Psi-0.6's text-to-flow prediction accuracy from 52.1% to 61.7%.
In briefGeometric Refinement and Distributional Alignment for Sim2Real Roadside LiDAR Detection
SummaryThis paper describes an entry to the LUMPI track of the UrbanTwin Sim2Real LiDAR Challenge, where a 3D object detector is trained only on synthetic data and tested on unseen real roadside LiDAR. The synthetic point clouds are also scored for distributional realism. Starting from a PointPillars baseline, the authors report gains from a parameter-free geometric refinement, repairing a faulty post-processing rule, and analytically examining the realism metrics. The final combined score is 0.4793.
In briefLabtune: Treating lab structure and information sharing as experimental variables in LLM-agent research
SummaryThis paper introduces Labtune, an open framework for bounded research workflows where users choose lab organization, communication policy, model settings and budget. It offers six editable organizational designs and lets information sharing vary independently, with the controller enforcing sharing at phase barriers. The work is an idea and system-description paper without benchmark results, aiming to make these design choices configurable and measurable.
In briefSpeeding Up the FastDEC Clustering Algorithm on GPUs
SummaryFastDEC is a clustering method based on dominance-based density estimation, but its exact nearest-neighbor search and dominance component processing are computationally costly. This paper presents SymWarp-FastDEC, a GPU-accelerated version that uses SymTrikNN, an exact self-k-NN kernel exploiting search symmetry, and WP-DC, a warp-parallel dominance component reduction algorithm. On the evaluated benchmarks it produces clustering outputs identical to the original FastDEC and achieves about 1.7x to 4.9x higher end-to-end performance than a FAISS-based variant.
In briefCoReMatch: Real-Time Robot Feature Matching via Recurrent Continuous Coordinate Refinement
SummaryThis paper presents CoReMatch, a feature matching method that refines image correspondences directly in continuous coordinate space instead of predicting discrete confidence maps on fixed grids. A Deep Aggregated Transformer captures global context and iteratively updates correspondence coordinates at a coarse stage, followed by a lightweight regression head for finer localization. On indoor and outdoor benchmarks, CoReMatch outperforms Efficient LoFTR in accuracy while running more than 10 times faster on embedded edge hardware.
In briefBeyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent Systems
SummaryThis paper studies how role instructions affect multi-agent LLM teams, comparing role-structured and role-free teams after a failed first answer. Agents, topology, call budget, and evaluator are held constant while only roles vary, using experiments on Omni-MATH and SciBench. Roles leave first-pass accuracy largely unchanged, but a role-structured Reviewer often withholds approval, so the team submits less and receives fewer evaluator verdicts, which lowers recovery after failure.
In briefWhat the Memory Helps With: Measuring Task-Specific Value in Long-Term Agent Memory
SummaryThe paper measures how useful a stored agent memory is for a specific future task, rather than assigning each memory one fixed importance score. On single-call factual-recall tasks built from LongMemEval-V2 web-agent memories, memories containing the needed answer information gave larger paired gains. The authors release a public dataset of paired measurements and costs. They report that these gains do not extend to the tested software-engineering and scientific-coding benchmarks.
Chih-Hsuan Yang, Anjir Ahmed Chowdhury, Feng Yan et al.
In briefSequential audit of linked forecast errors with missing outcomes
SummaryThis paper studies how an auditor can selectively retrieve missing true outcomes to check whether a connected group of locations has outcomes falling outside the same side of their forecast intervals. The audit stops once the observed labels settle the question or the query budget runs out. Tested on three public graph time series, the MOMS method cut mean queries by 19.6% to 38.5% versus screened random acquisition when 30% of labels were hidden.
In briefBenchmark comparing Koopman-type representations for deep latent data assimilation
SummaryThe paper systematically compares seven representative Koopman-type methods and structural baselines across five systems, ranging from approximately periodic to strongly chaotic dynamics. It evaluates forecast and assimilation accuracy, physical state recovery, latent-trajectory behavior, and dynamical consistency. The results reveal a trade-off between representational flexibility and optimization tractability, with methods using closed-form analysis updates generally performing better and more robustly.
In briefEpiLoRA: Fusing Many LoRA Adapters into Low-Precision Matrix Multiplications for Serving and RL
SummaryThis paper presents EpiLoRA, a plugin for the vLLM engine that fuses LoRA adapter computations into the final stage of low-precision base-model matrix multiplications, reducing the cost of running adapters as separate GPU kernels. On NVIDIA Hopper GPUs, the authors report up to 24% lower time per token for single-adapter decoding than vLLM's fastest configuration, and higher adapter-kernel throughput with many tenants. The approach also enables LoRA reinforcement-learning training directly on the engine's FP8 weights without a 16-bit copy.
In briefLimits of Neural CDE Information with Sparse Observations
SummaryThis paper studies the uncertainty in Neural Controlled Differential Equations when observations are sparse, since many continuous paths can fit the same data and yield different terminal states. For planar drivers with bounded speed, it gives a leading-order zonotopic approximation of the compatible terminal states using feasible signed areas and transported Lie brackets. It also derives a computable, target-specific ambiguity certificate and analyzes how many measurements are needed to reduce uncertainty, with synthetic experiments testing the certificate.
In briefProofPaths: An AI Agent Framework for Assessing Student Reasoning
SummaryThis paper introduces ProofPaths, an AI agent framework that combines rubric-based grading with individualized, interactive interviews to assess whether students understand the reasoning behind their submissions. Multiple grading agents evaluate work against instructor guidance, critique each other, and revise their judgments, with their scores determining the grade and feedback. An interview agent asks students to justify steps in their work and generates follow-up questions based on their answers. The system stores grading reviews and interview transcripts, and instructors can inspect them and override grades.
In briefEmbedding Semantic Entropy: NLI-free uncertainty estimation via geometry-aware clustering for language models
SummaryLarge language models often produce fluent but incorrect answers, so estimating their uncertainty matters. This paper proposes Embedding Semantic Entropy, which replaces entailment-based clustering with K-Medoids clustering in semantic embedding space and uses dynamic cluster selection. The method is evaluated on five datasets and four models, showing competitive AUROC and AUARC results.
Akshiya Raveendran, Praveenasarma Baskarakurukkal, Pragatheeswaran Vipulanandan et al.
In briefINSIGHT-HPC: An Interpretable System Intelligence for Global HPC Telemetry
SummaryThis paper presents INSIGHT-HPC, a multi-plane anomaly-detection architecture for telemetry from high-performance computing clusters. It combines statistical, workload-aware, and physics-guided evidence to turn minute-level observations into de-duplicated fault episodes with operator-facing explanations. On production telemetry with weak labels and controlled fault injections, it reports a precision of 0.913, recall of 0.817, F1 of 0.862, and AUROC of 0.896, outperforming all evaluated baselines.
In briefLearning Almost Complex and Complex Structures: numerical searches for complex structures on spheres and sphere products.
SummaryThe paper numerically constructs almost complex and complex structures on spheres and products of spheres up to real dimension six. Almost complex structures are parametrised by a global matrix field in Euclidean space, and polynomial features and neural networks are used to search for them. Nine examples pass the numerical admissibility checks, while S^4 is rejected. Learned complex structures on S^2×S^2, S^3×S^1, S^3×S^3 and S^5×S^1 have Nijenhuis norms below 10^-9, and direct constructions in torus-fibred coordinates are given for S^6, S^4×S^2 and S^4×T^2.
Edward Hirst, Tancredi Schettini Gherardini, Tom\'as S. R. Silva et al.
In briefEvoGroup: A self-evolving framework for group agents working in shared contexts
SummaryThe paper proposes EvoGroup (EvoG), a self-evolving harness for group agents that starts from a minimal six-tool setup and adapts through interaction with shared context. On a held-out split of EverMemBench, pass@1 rises from 71.07% to 89.58%, surpassing self-evolving memory systems and human-designed harnesses, including Claude Code at 85.83%. The frozen harness also generalizes to the unseen GroupMemBench, outperforming memory baselines and human-designed harnesses.
In briefExaTune: Exhaustive Mapping of Hyperparameter Performance Landscapes at Scale
SummaryThe paper presents ExaTune, an open-source HPC system that exhaustively evaluates every point on a discretized hyperparameter grid using K-fold cross-validation to produce hyperparameter landscape maps. It computes five landscape metrics, including spatial autocorrelation, ruggedness, local optima density, fitness-distance correlation, and variance-based parameter importance. The authors release eight complete landscape maps covering 285,080 configurations across five model families. They report that Bayesian optimization, Optuna, Hyperband, and Random Search at standard trial budgets miss instability regions, secondary optima, and parameter interactions visible in the full landscapes.
In briefDr. No: steering model features so vision-language models abstain when evidence is lacking
SummaryLarge foundation models often answer even when visual or textual evidence is insufficient, which makes them hard to trust in high-stakes settings. This paper presents Dr. No, which mines sparse autoencoder features tied to abstention and commitment and adjusts a small set of them during the prefill pass, without retraining the base model. It is tested on LLaVA-1.6 Mistral 7B across multimodal abstention benchmarks such as UNK-VQA and MM-UPD, and on AbstentionBench subsets with Qwen3.5-2B. The method achieves a better balance between correct answering and appropriate abstention than post-hoc rejection methods, but its gains are limited on absent-answer detection and multi-hop QA such as MuSiQue.
Farid Hasanov, Riccardo Renzulli, Bahram Zonooz et al.
In briefPredictive gains, causal losses: why interventional competence cannot be learned from data alone
SummaryThe paper argues theoretically that, for deterministic predictors, beating the best parent-only predictor on observational data requires persistent errors in intervention responses. Under the stated conditions, root-mean-square intervention-response error is bounded below by a positive constant. It further argues that correcting this error is limited by the causal hypotheses currently admitted, so reducing it further requires admitting better causal hypotheses.
In briefPerformance, Scalability and Maintenance: RPA vs. Autonomous Agents in Healthcare
SummaryThis paper compares two ways to retrieve mammography and breast ultrasound reports from a laboratory portal: a Playwright-based robotic process automation (RPA) solution and a large language model-based autonomous agent (APA). Across two rounds, RPA reached 98.90% accuracy versus 96.49% for APA and was far faster per patient. The authors conclude that RPA is superior when the interface is stable.
Daniel Schulz, Pedro Cabral Buchaim, Bernardo Soares et al.
In briefCan Rewriting Replace Long Reasoning? Effects of Compacting Reasoning History on Later Generation and Agent Cost
SummaryThis paper studies whether open-weight language models can reduce future reasoning-token use by rewriting their earlier reasoning into compact state records, without retraining. It compares continuations from original and compacted histories, finding that accepted compact histories usually shorten later reasoning. In software-agent deployments, compaction can lower total spending, but task quality and time including rewriting depend on the target model and compressor.
In briefRoboScope Kit: Unified Interfaces for Faster Embodied-AI Evaluation
SummaryThis paper presents RoboScope Kit, which wraps each model as a unified Policy class and each simulation benchmark as an adapter, leaving the benchmark's own evaluation rules and scoring unchanged. It schedules simulator workers and policy replicas as separate pools, sized from measured probes, and shards work deterministically across GPUs. On LIBERO-Long it evaluates 500 episodes 4.45 times faster than the official StarVLA evaluator with equivalent success rates. On RoboTwin Easy, four GPUs finish 500 episodes in 14.4 minutes, 25.04 times faster than the single-GPU baseline, with identical success outcomes and terminal steps.
In briefLaplaceDiag: Locating Faults via Changes in Learned Actuator Effects
SummaryThe paper introduces LaplaceDiag, a data-driven method that uses changes in actuator effects to narrow the search for faulty equipment. It learns a healthy plant model with LaplaceNet, a neural network operating in Laplace space, separating nonlinear history-based prediction from an explicit matrix describing how actuators affect sensors. Comparison with a healthy reference detects changes and generates actuator-related diagnostic hypotheses, supporting both online and offline diagnosis. The authors report evaluating the method on two datasets.
In briefSteering AI-generated antibody designs with uncertainty, tested in a lab-in-the-loop setup
SummaryThis paper proposes merging the two usual stages of antibody development, discovery and refinement, into one process. A generative structure model is gradually handed over to iterative optimization, with the switch guided by the posterior uncertainty of a probabilistic surrogate model. The authors report better optimization efficiency when data is scarce, along with more diverse optimized molecules.
Anisha Parsan, Sebastian Ober, Calvin McCarter et al.
In briefPianoBench: A Benchmark Testing Whether Language Models Can Perceive and Reason About Piano Music
SummaryThe paper introduces PianoBench, a benchmark that asks language models to reason about piano performances using both audio and symbolic transcript inputs. It is built from a new dataset of over 20,000 hours of music interleaved with spoken commentary from piano instruction videos. Open audio-language models reach at most 52.8% accuracy, while frontier models given text transcripts reach up to 93.2%.
Louis Bradshaw, Irmak Bukey, Daniel Chenyu Lin et al.
In briefMeasuring How Varied Language Model Responses Are Using Conditional Mutual Information
SummaryThis paper analyzes the variety of responses generated by large language models using an entropy-based approach built on conditional mutual information. It defines three limiting behaviors, arbiter, independent, and neutral (deterministic), which span from no to maximal mutual information. The authors give examples of each and suggest applications in temperature analysis, language analysis, prompt compression, and knowledge graphs.
In briefLow-Rank Hierarchical Merging for Efficient Long-to-Short Reasoning
SummaryThis paper proposes LHM, a training-free framework that merges long chain-of-thought models with base models to keep strong reasoning while producing concise outputs. It replaces first-order heuristics with lightweight second-order curvature estimates to assign layer-adaptive merge coefficients. Tested on six math benchmarks with 1.5B and 7B models, LHM reports over 4% higher accuracy than ACM-TIES and ORCA and up to 93.1% shorter outputs than the long-CoT model.
In briefEl Agente Fármaco: A Scalable AI Scientist for Drug Discovery
SummaryThis paper introduces El Agente Fármaco, an AI scientist that uses large language models to turn everyday-language drug discovery goals into scalable computational screening workflows. It relies on a shared memory, Fármaco Atlas, which stores molecules, structures, and results together with their provenance. In a retrospective comparison and a 22-task benchmark, the agent reportedly performed competitively against an expert-led campaign and baseline tool-calling approaches.
Marcel M\"uller, Mohammad Ghazi Vakili, Yoshihiro Ogura et al.
In briefRethinking next-visit diagnosis prediction with retrieval-memory multi-agent adjudication
SummaryThis paper examines predicting the diagnoses a patient will have at their next clinical visit from electronic health records. The authors propose an evaluation framework that considers diagnosis set distribution and temporal transitions, and compare rule-based, LLM, retrieval, memory, and multi-agent approaches. They find that repeating the last visit's diagnoses outperforms most LLM-based methods overall. They then introduce LONGINUS, a framework combining retrieved patient record evidence with structured patient trajectory memory, evaluated on MIMIC-III and MIMIC-IV.
In briefAIRoA-0: Robot Manipulation Learned Entirely from Handheld Gripper Recordings
SummaryThe paper presents AIRoA-0 (A-0), a bimanual manipulation system trained on 16,151 hours of demonstrations recorded with the open-source handheld gripper YUBI, with no robot data collected at any stage. On ten evaluated tasks it reaches an average success rate of 78%, compared with 27% for the strongest of four other vision-language-action models post-trained on the same demonstrations. The system runs on several robot platforms that mount the YUBI gripper, and the dataset, pre-trained weights and post-training recipe are released.
WhyThe authors train bimanual manipulation solely on handheld-gripper demonstrations, reporting 78% average success across ten tasks versus 27% for the best of four baselines.
In briefBEND 2.0: A benchmark testing the skills needed to design DNA
SummaryThis paper introduces BEND 2.0, a benchmark for DNA language models that tests how well they predict the effects of DNA changes and whether those predictions can guide sequence design. Across eight tasks built on five experimental datasets, model choice and how representations are accessed both affect performance. Linear readouts on frozen representations can match or exceed models' native outputs, and functional post-training substantially improves capabilities.
Andreas Bjerregaard, Rachael DeVries, Dionysia Danai Brilli et al.
In brief3D dark matter map of the COSMOS-Web field from JWST, using generative AI priors
SummaryThis paper reconstructs the three-dimensional dark matter distribution in the 0.54 square degree COSMOS-Web field using deep James Webb Space Telescope imaging. It combines a weak lensing forward model with a diffusion-based Bayesian generative prior learned from cosmological simulations, producing a posterior distribution across 20 lens planes out to redshift about 2. The reconstruction separates overlapping structures and recovers non-Gaussian matter structure across several spatial scales.
Diana Scognamiglio, Brandon Zhao, Katherine L. Bouman et al.
In briefPixel: A Full-Stack Robot System Co-Designed for General Mobile Manipulation Learning
SummaryThis paper presents Pixel, a full-stack robot system designed together across hardware, data, control, and learning to support general-purpose mobile manipulation. It combines a low-cost wheeled humanoid robot, multiple teleoperation and robot-free human data-collection interfaces, a unified whole-body controller, and a shared policy learning pipeline. A hierarchical learning stack enables language-conditioned visuomotor control and long-horizon task execution, and the system is evaluated across these components.
Tony Tao, Jason Jingzhou Liu, Ruslan Salakhutdinov et al.
In briefPixel-UMI: Teaching Robots Whole-Body Manipulation from Sensor-Equipped Human Demonstrations
SummaryThis paper presents Pixel-UMI, a pipeline that learns environment-aware whole-body robot policies from sensorized human demonstrations. It combines Environment-Aware Retargeting, which uses scene geometry and human head and gripper trajectories to set robot posture, with Retargeting-Error-Aware Masking to handle viewpoint mismatches, and on-policy human corrections via DAgger. Evaluated on five challenging manipulation tasks, it achieves 87.0% average task progress.
Jason Jingzhou Liu, Tony Tao, Deepak Pathak et al.
