In briefBack to Pixels: Running a Latent Video Diffusion Model Directly in Pixel Space
SummaryPixVid is a method that adapts a pretrained latent video diffusion model to generate every frame directly in pixel space. It keeps the diffusion transformer and replaces the VAE decoder with a convolutional head. The authors find that predicting clean video instead of velocity, along with a larger noise shift, is key to quality. Finetuned from Wan2.1-14B, it is competitive with the latent baseline and generates at 2x higher resolution with the same per-frame token count.
In briefSeparating vision encoder design choices in multimodal language models
SummaryThis paper examines how multimodal language models should encode visual input, separating three choices bundled in vision encoders: pretrained visual priors, dedicated vision capacity, and placement before fusion. The authors pretrain all models from scratch on common data and compare them on visual question answering and language benchmarks at matched compute. They find that a jointly trained encoder from scratch matches a pretrained one, and that dedicated vision capacity matters more, with vision modules placed early inside the backbone performing comparably to a pre-fusion encoder.
In briefArchSpace: A public platform for recording, reviewing, and referencing LLM architecture research
SummaryArchSpace is a public platform for recording, reviewing, and referencing research on large language model architectures. Each proposal is logged with its experiments and reasoning, reviewed by the community, and cited by later work. It uses scaling laws to predict model loss at larger scales from smaller experiments, and provides Qwen3.5 and OLMo3 baselines, fitted scaling laws, and a reusable training recipe.
In briefTTB: Test-time Baking for Faster Rendering in Decoder-only View Synthesis Models
SummaryThis paper proposes Test-time Baking (TTB), which learns to bake global scene information from cross-view attention into a lightweight MLP for fast novel-view decoding. It repurposes the fast-weight MLP of Test-Time Training (TTT) to explicitly learn the cross-view attention mapping, and separates context tokens into camera-only rendering tokens and ground-truth image tokens to address modality misalignment. The authors report up to 23x rendering speedup with quality on par with state-of-the-art methods across several benchmarks.
In briefProtoVis-Nav: Prototype-based visual imagination for UAV navigation aligned with vision and space
SummaryThis paper proposes ProtoVis-Nav, a prototypical visual imagination framework for vision-language navigation of unmanned aerial vehicles. The model predicts feature prototypes and combines them linearly to estimate the visual features of the target region, imagining the next intended location before generating an action. The predicted prototypes serve as context for the final action, aligning visual context with waypoint outputs. The authors report extensive experiments showing the framework's effectiveness and substantial gains over existing baselines.
In briefInternalizing Investment Mandates: Teaching Trading AI to Follow Portfolio Rules
SummaryThis paper proposes a continuous-action method, inspired by Probabilistic Logic Policy Gradient, for portfolio allocation under investment mandates. It estimates the probability that the policy's own executed action satisfies the mandate using reparameterized Monte Carlo sampling and adds this to the training objective, so compliance is learned rather than enforced after the fact. On a six-rule regime-conditional mandate, violations fell from 96-98% for unconstrained RL and 33.4% for a reward-penalty baseline to 4.8%.
Zeynep Y{\i}ld{\i}z, Satchit Chatterji, Ercan Atam et al.
In briefK2V: Generative Keyframing for Physically Plausible Video Generation
SummaryThis paper proposes K2V, a method that reduces physical errors in video generation. It assigns physical reasoning to a language model-backed keyframe generation module, which reasons explicitly in language and controls the appearance at key moments. A video generation model is then conditioned on these keyframes, so the diffusion model only interpolates between physically correct intermediate states. The authors report state-of-the-art results on the Physics-IQ Verified benchmark and PhyGenBench.
Felix O'Mahony, Daniel Olmeda Reino, Roberto Cipolla et al.
In briefAgentSSL: Can ML Agents Use Unlabeled Data?
SummaryThis paper introduces AgentSSL, which frames semi-supervised learning as a program search problem and studies how agents explore it when labeled data is scarce. Agents combined and adapted existing SSL techniques, beating the strongest human-engineered baselines in 8 of 12 settings by up to +21.8%. The authors find that program selection is a key bottleneck, with performance dropping by as much as 47% when selecting by the agent's own feedback metric. They also release ProgramSelectBench, containing 16,977 programs, and evaluate AgentSSL on real-world ecology detection tasks.
In briefBrickFlow: Connectivity-Guided Brick Reconstruction
SummaryThis paper addresses rebuilding a LEGO assembly from an image by inferring the pose of each part. The authors train a set-conditioned SE(3) flow model on procedurally generated assemblies to learn how parts fit together, then finetune it with image conditioning using about 1% as much data. To ensure connected structures, they add an analytic connectivity-field guidance at test time, and report that model size, pretraining, and guidance improve build quality.
In briefLearning when diffusion language models should regenerate text
SummaryThe paper proposes DiVR, a verifier that predicts whether a partially generated diffusion language model output will reach a correct answer, to decide when to regenerate a block and which alternative to keep, without changing the generator. On five reasoning benchmarks using LLaDA-8B-Base, DiVR raises the average accuracy of the fine-tuned generator from 35.3% to 43.4% at 1.9 to 3.5 times the cost of greedy decoding. It complements supervised fine-tuning and reinforcement learning and generalizes to other generators such as Dream.
In briefSpatial Foundry: A Multimodal Dataset for Spatial Intelligence
SummaryThis paper introduces Spatial Foundry, a dataset of 5,280 furnished residential 3D scene configurations with editable assets and aligned visual, geometric, and temporal data. Modalities include floor plans, RGB images, metric depth, point clouds, meshes, 3D bounding boxes, and camera trajectories paired with video. The authors describe the data representation and construction workflow, and discuss applications such as floor-plan-conditioned 3D generation, reconstruction, and scene editing.
In briefCoReMem: A Memory Framework for Successive Coding Tasks That Weighs Costs and Predicts Future Value
SummaryThis paper introduces CoReMem, a memory framework for coding agents that handle successive tasks on an evolving code repository. It guides memory design and agent adaptation by measuring how much memory helps on the next task, with memory built around anchors such as repository locations and executable checks. On ChainSWE, the 27B actor's resolve rate rises from 21.0% to 25.7%, with 17% fewer steps and 11% fewer total cache-aware tokens, including memory-module calls, compared with the untrained model without memory.
Peiwen Li (Yale University), Harshit Verma (Yale University), Sophia Zhou (Yale University) et al.
In briefLinking fast learning to compositional structure in neural networks through additive analogies in meta-learned parameter spaces
SummaryThis paper studies why neural networks optimized by meta-learning can learn new functions quickly. Inspired by word-vector analogies, the authors show that parameter vectors for different adapted functions can be added and subtracted according to symbolic relations among those functions. The resulting recombined vectors define networks that implement the intended functions without direct training on them. The authors conclude that meta-learning shapes the network's starting point so that later learning follows systematic, functionally meaningful paths.
In briefSketching language models for context attribution: estimating how removing prompt parts changes outputs
SummaryThis paper studies leave-one-out (LOO) attribution, which measures how removing each token from a prompt changes a language model's output. Under a stability assumption, it presents a sketching algorithm that estimates all LOO values to error proportional to ε with constant probability using O(√log(1/ε)/ε) evaluation calls. The authors also prove a matching lower bound in an oracle model.
WhyThe authors apply forward differentiation to backpropagation circuits, reducing evaluation calls and report a matching lower bound plus a small retrieval experiment.
Christopher Liaw, Aranyak Mehta, Grigoris Velegkas et al.
In briefClassical energy models learn reasoning 1,000 times faster than looped models
SummaryThe paper proposes an energy-descent approach to reasoning, using a learned version of Hopfield and Tank's 1980s energy function called CRAM. After 42 ms and 132 ms of training on one GPU, it solves 99.996% of Sudoku-Extreme and 100% of Maze-Unique test puzzles. The authors report this is at least 1,000 times faster than looped models on these tasks. They also argue that trivially solvable benchmarks like Sudoku and Maze are a weak test of scalability for looped reasoning methods.
Benjamin Hoover, Hugh Blayney, James C. R. Whittington
In briefMeasuring how stable transparent pricing models stay over time in general insurance
SummaryGeneral insurance contracts are short, so the data shifts quickly and models must be updated often, which can change both predictions and model structure. The paper defines functional transparency for models whose prediction is a sum of univariate and bivariate shape functions, and proposes the Mean Squared Shape function distance to measure structural stability. Benchmarks on three public auto insurance datasets find that generalised linear models with centred numerical covariates are stable, and explainable boosting machines offer a compromise between stability and performance.
SummaryThis paper studies debate as a training method, in which two models argue for opposite answers and a human judge decides. The authors formalize the learning process as an online learning problem and show that the best way to learn with a judge can differ from adversarial debate. They define a cooperative version of the protocol and characterize its learnability using a combinatorial dimension. They also show that a small number of incorrect judgments can cause a large learning error.
WhyThe authors formulate debate training as online learning, propose a cooperative variant characterized by a combinatorial dimension, and show few judging errors can cause large learning error.
Tosca Lechner, Vinayak Pathak, Sivan Sabato et al.
In briefFLEX-LBP: Learning Texture Descriptors with Exact Binary Inference
SummaryThis paper presents FLEX-LBP, a method that learns the radius, sampling angles, color transform, and thresholds of Local Binary Patterns (LBP), a widely used texture descriptor, via backpropagation. Forward passes compute exact binary codes and hard histogram counts, so the deployed descriptor remains an interpretable hard-binary LBP with 90 learnable parameters and unchanged inference cost. The authors compare configurations on six color texture benchmarks against five classical LBP-family descriptors, PCANet, a neural LBP, and frozen ImageNet ResNet-50 features.
WhyThe authors learn LBP radii, angles, color transform and thresholds via backpropagation while keeping exact binary forward passes; the report covers six color texture benchmarks over five seeds.
In briefConsistent video-to-video translation using explicit frame correspondences
SummaryThis paper introduces vid2vid-long, a method for interactive video-to-video translation that aims to produce real-time output while keeping long-range temporal consistency. It rests on the idea that regions matching in the input video should also match in the output. The method computes patch-level correspondences in the input, retrieves previously generated outputs for corresponding regions, and conditions generation on these patches via correspondence cross-attention. Experiments show improved long-range consistency across several recent video models while preserving generation quality and real-time performance.
In briefGoal Encoding in LLMs: How Language Models Represent and Track Goals Internally
SummaryThis paper investigates whether large language models encode goals as internal representations that precede and guide their actions. In a multi-sub-task coding setting, the authors locate an encoding of the next sub-goal in the outputs of a sparse set of attention heads, which can also be decoded into natural-language descriptions of future actions. They show this encoding is causal for the next action and transfers to another programming language, where it triggers the same operation with different tokens. The authors further propose a non-recurrent goal-seeking (NRGS) account, in which the model repeatedly re-reads the goal from the task specification, and test its predictions with targeted causal interventions.
Arnab Sen Sharma, Eric Todd, Sheridan Feucht et al.
In briefMetal Sandbox: Isolated, Scheduled Execution for GPU Coding Agents on Apple Silicon
SummaryThis paper presents Metal Sandbox, infrastructure that runs AI coding agents on interconnected Macs with isolated per-agent file access and native command execution. Commands are scheduled individually, with parallel commands running concurrently and waiting commands able to migrate to available Macs. In controlled experiments on four Mac mini M4 nodes, command throughput improved by 1.68 to 2.05 times under concentrated demand. The system and its Harbor integration are open-source on GitHub.
In briefNegative-only policy optimization for rewards that can only verify failures
SummaryThe paper addresses reinforcement learning settings where verifiers can reliably falsify wrong answers but cannot confirm correct ones. It proposes Negative-Only Policy Optimization (NOPO), which treats verified positives as unlabeled and updates the policy only by suppressing verified negatives, with updates scaled by the proportion of non-falsified responses. The authors report improved pass@k over baselines on code, logic, and math benchmarks across unit-test, constraint-check, and LLM self-verifier settings.
In briefSynthesis Smuggling: hiding guessed personal traits inside AI-written text
SummaryThis paper proposes Synthesis Smuggling, a threat model in which an adversarial LLM infers a user's sensitive attributes from routine assistance conversations. It hides these inferences steganographically in content that looks harmless. When the user shares the generated text through ordinary channels, an adversary monitoring those channels can recover the information. The authors train two open-weight models to infer gender and age group and encode the predictions using invisible characters or word choices.
Gautam Galada, Supreeth Shastri, Sagnik Ray Choudhury
In briefLiveGeo: Benchmarking Geographic Agents in Changing Worlds
SummaryThe paper introduces GeoWorld, a controllable geographic environment platform for LLM agents that evolves over time while keeping verifiable state histories. It also presents LiveGeo, a benchmark that asks agents to gather task-relevant evidence, check whether earlier decisions remain valid, and revise their plans. Results show that even frontier LLMs struggle with static tasks, and strong static performance does not guarantee dynamic reliability. The authors identify acquiring evidence, turning it into valid decisions, and keeping them correct as the world changes as a key bottleneck.
In briefBallad: Bandit-based routing of language models for automatically discovering heuristics
SummaryThis paper studies how to choose among large language models when automatically discovering heuristics for hard optimization problems. The authors find that bigger, more expensive models do not consistently produce better heuristics across several NP-hard problems. They propose Ballad, an online bandit algorithm that orchestrates a pool of LLMs into an ensemble. The paper reports that Ballad matches the best single model and sometimes exceeds it, while cutting API costs by up to 65.6%.
In briefObjects as Addresses: How Vision Transformers Keep Instances Apart
SummaryThis paper examines how pretrained Vision Transformers keep individual objects separate within their self-attention computations. Across eight pretrained ViTs, the authors find that instance separation appears more strongly in attention routing, through query-key relations, than in token content. Patches belonging to one individual tend to attend to a common image region, which the authors call an attention address. Editing the query component at a selected head redirects attention to another individual, and grouping patches by this component separates same-category instances better than final-layer features do.
Khalil Sabri, Guillaume-Alexandre Bilodeau, Nicolas Saunier et al.
In briefWhen Both Sides Adapt: Evaluating Interactive LLM Systems
SummaryThe paper examines how users adapt their prompts and tasks after seeing LLM responses, so model outputs shape the prompts later used for evaluation. It distinguishes two comparison targets and shows that offline benchmarks and live A/B tests measure different quantities, which can rank models differently. The authors give a sufficient condition for when the two rankings agree and show in a linear model that user adaptivity can increase estimation variance.
In briefAl-Ghasali: A framework simulating Islamic legal rulings with automata and language models
SummaryThis paper presents Al-Ghasali, a computational framework for formalizing and simulating Islamic jurisprudential (Fiqh) rulings. It proves that the semantics of closed Fiqh chapters can be reduced to deterministic finite state automata, and discusses how hierarchical compositions and order requirements are handled. A deterministic automaton controls a large language model that produces human-readable juristic summaries, using grammar-guided constrained decoding to aim at preventing hallucination.
In briefTwo mathematical models explaining why knowledge distillation works
SummaryThe paper proposes and analyzes mathematical models of knowledge distillation, arguing that its benefits come from easier-to-optimize loss landscapes rather than better final models. For generalized linear models trained with stochastic gradient descent, it proves that distillation fits strong student models asymptotically faster than non-distilled training. For rank-1 matrix approximation, it characterizes target-matrix conditions under which distillation converges strictly faster, and experiments on real datasets support these predictions.
In briefPolynomial-time linear regression under self-selection bias
SummaryThis paper studies estimating k linear regressors under self-selection bias, up to error ε, using the maximum self-selection criterion. The authors give an algorithm running in polynomial time in n, k, and 1/ε for the unknown-index variant, which also handles correlated noise between regressors, a setting related to max-affine regression. Their method analyzes high-degree Hermite moments of the covariates, which approximately decompose into sums of tensor powers of the regressors, followed by tensor decomposition and a least-squares step.
WhyThe authors replace exhaustive search with a tensor decomposition of high-order Hermite covariate moments, giving a polynomial-time algorithm for self-selection regression that also covers correlated
In briefAdding structured scoring inspired by sheaf theory to LightGCN for recommendation
SummaryThis paper adds a modular scoring extension to a three-layer LightGCN backbone, inspired by cellular sheaves in which local vector spaces are linked by linear restriction maps. Graph propagation is left unchanged; instead, user and item embeddings receive structured, pair-dependent transformations. Several scorers are compared on Amazon-Book, Gowalla, MovieLens-1M, and Yelp2018, showing modest, dataset-dependent ranking changes and sometimes large coverage changes. The authors note that they do not construct a cellular sheaf or implement sheaf-Laplacian diffusion, so the experiments do not establish a causal benefit from sheaf geometry itself.
Ksenia Belkovskaya, Timur Lidzhiev, Maria Stebliy et al.
In briefOpen data archive for 2D and 3D electrical impedance tomography using phantoms with moderate conductivity contrast
SummaryThis paper releases open electrical impedance tomography (EIT) measurement data collected from biomedical-imaging-inspired phantoms with moderate conductivity contrasts. The data were gathered with the KIT4 and KIT5 devices at the University of Eastern Finland for developing and testing image reconstruction algorithms. Data from thoracic-inspired cylindrical and chest-shaped phantoms suit 2D algorithms, while data from a head-shaped phantom with a resistive skull, inspired by intracerebral hemorrhage imaging, suit 3D algorithms. Computational meshes, example reconstruction codes and sample reconstructions are also provided.
In briefCombinatorial Bernoulli factories with polynomially many coins
SummaryThe paper presents an efficient implementation of combinatorial Bernoulli factories, which convert coins with unknown biases into a random vertex of a polytope whose mean is exactly the unknown value. For every affine slice of the cube, it gives a single algorithm whose expected number of coin flips and running time are polynomial in the dimension and in 1/δ*. For perfect matchings this yields O(n^8 δ*^-2 log(n/δ*)) flips, whereas the prior factory of Niazadeh, Paes Leme, and Schneider (STOC 2021) needs at least n^(2n-1) expected flips at the uniform doubly stochastic matrix.
WhyThe authors give a single algorithm for every affine cube slice, with expected coin flips polynomial in dimension and 1/δ*, and report polynomial bounds for perfect matchings versus prior exponential
In briefScalable Chat-Based Exams Using a Two-Agent LLM System
SummaryThe paper presents an LLM-based system that conducts text-based conversational exams to assess students' ability to make and support arguments, aiming to scale the benefits of oral exams. A two-agent design separates an interviewer, who talks with the student, from a hidden evaluator that tracks progress against a rubric and tells the interviewer when to follow up or move on. The system was first tested with simulated students of varying ability and adversarial tactics, using a fictional topic, and then deployed in a small elective AI literacy course for formative and summative assessment. Transcript review led to improvements in follow-up question generation and ensured grading credited only ideas contributed by the student.
In briefAdding Deliberate Feedback Delays to Adversarial MDPs for Optimal Regret Bounds
SummaryThis paper studies regret minimization in episodic adversarial Markov decision processes with unknown transitions and bandit feedback. The authors propose a technique called introduced feedback delays, which estimates the current policy's Q-function using future episodes while still updating the policy. It uses Geometric Resampling with Rejection and thresholded visitation probabilities to handle policy changes during the delay. The paper reports high-probability regret of $\tilde O(\text{poly}(H)\sqrt{SAK})$ in tabular AMDPs, resolving an open problem, and $\tilde O(\text{poly}(d,H)\sqrt K)$ in linear AMDPs.
WhyThe authors introduce voluntary feedback delays with geometric resampling, reporting optimal √S regret in tabular and √K regret in linear AMDPs.
Aarush Kulkarni, Khang Nguyen, William Chang et al.
In briefAnalysis of sparse convolution for faster neural network inference on edge devices
SummaryThis paper studies standard convolution with sparse weights, the kind produced by pruning neural networks, for use on resource-limited edge devices. Roofline analysis shows a 3x3 sparse convolution reaches about 10 FLOP/byte of arithmetic intensity, compared with roughly 96 FLOP/byte for its dense counterpart. The authors propose techniques for efficient sparse convolution at high sparsity and report preliminary speedups of up to 1.65x on NVIDIA A100 and up to 1.52x on NVIDIA Jetson Nano.
Keshvi Tuteja, Yu-Hsiang Tsai, Gregor Olenik et al.
In briefUnsupervised hyperspectral image representation learning using MALNISI-based graph construction and a weighted contrastive loss GAT
SummaryThis paper proposes MALNISI, an index that measures pixel similarity from neighborhood information and uses it to build a weighted graph from a hyperspectral image. It then introduces an enhanced Graph Attention Network with residual connections and feature normalization, plus a weighted contrastive loss that pulls strongly connected pixels together in latent space. The learned low-dimensional embeddings are clustered with spectral clustering. The authors report superior results over fourteen state-of-the-art approaches on six benchmark hyperspectral datasets.
Soumi Sarkar, Subhadip Boral, Sarbani Palit et al.
In briefChoosing Good Solutions for Discrete Origami Designs
SummaryThis paper studies discrete-variable mechanical design using origami as a test case, where automatic differentiation and bilevel optimization select one of four hinge types at each fold. Baseline gradient-based optimization solves 97% of relaxed runs, but only about half of the final discrete designs succeed. Biasing initialization toward the stiff, flat hinge, or penalizing frustration and floppiness in the outer loss, both raise the discrete solve rate.
In briefeDMD: Endpoint matching to distill diffusion models into few-step generators
SummaryThis paper proposes Endpoint Distribution Matching Distillation (eDMD), a method for compressing diffusion models into student models that generate images in few steps. It trains a single shared fake score model only on the student's final outputs and uses it to correct every denoising step. Experiments on SD3-Medium, FLUX.2-klein (4B and 9B), and Qwen-Image-2.1 report that eDMD consistently outperforms existing distribution matching methods for 4-step generation.
In briefUsing LoRA adapter capacity better to reduce the learning-forgetting trade-off
SummaryThe paper finds that LoRA adapters, especially high-rank ones, often fail to use the capacity their configured rank allows, since standard updates have effective rank far below that rank. The authors add a simple, stable, differentiable, SVD-free Rényi-2 effective-rank regularizer to the language-modeling loss to raise effective rank. With a penalty weight of 0.01, mean Wiki perplexity drops relative to unpenalized LoRA at every tested rank of at least 16 across three models.
In briefConsensus capture in drone video stabilization on busy urban scenes
SummaryThe paper studies reference-frame stabilization of hovering drone video in scenes with moving vehicles, where the robust homography estimate can be pulled toward a group's motion. The authors build a two-part benchmark with static and dynamic protocols that score ten registration pipelines against pseudo-ground-truth references. They report that masking moving objects removes this consensus capture for every pipeline while leaving the leading pipelines unchanged. Code, data and results are released.
WhyThe authors introduce consensus capture and a static and dynamic benchmark, scoring ten UAV stabilization pipelines, with masking shown to remove capture.
In briefNeural render proxies enabling fast, adjustable relighting of static scenes
SummaryThis paper introduces a neural render proxy (NRP) for relighting static scenes with a fixed camera and materials. It collects light transport path samples from one light-agnostic render pass and trains a lightweight, scene-specific neural network to model light transport. The authors report interactive relighting at roughly 30–60 Hz and show that the differentiable proxy supports gradient-based inverse lighting workflows.
In briefValidation-Guided Prompt Selection for Differential Morphing Attack Detection
SummaryThis paper studies differential morphing attack detection (D-MAD), which compares a trusted live capture with a document image that may contain a facial morph. Instead of fine-tuning, the authors select prompts for a frozen vision-language model using labelled validation data, generating 50 candidates and ranking them with a loss that prioritises missed morph attacks. On a subject-disjoint test set of 160 pairs, the selected prompt reduces morph attack classification error to 19.17%, compared with 70.00% for a controlled manual prompt. Recursive prompt revision did not outperform the best validation loss from the initial generation.
In briefSpotDiff: A Plug-and-Play Framework for Detecting Changes in Satellite Images Without Training
SummaryThis paper presents SpotDiff, a training-free framework that detects changes between two satellite or aerial images taken at different times without needing labeled examples. It combines existing vision foundation models, including DINOv3, AnyUp, and SAM 3, with a multi-scale strategy to find and compare regions of change. On the LEVIR-CD benchmark it reaches 82.46 F1, 2.06 points above the best prior training-free method. The authors conclude that composing modern foundation models is an effective route to scalable, label-free change detection.
Anit Upadhyaya, Ata Akbari Asanjan, Chaitanya Jalluri et al.
In briefParsed but wrong: how output format drives silent errors in small-model semantic parsing
SummaryThe paper studies whether the output format of a small semantic parser affects how often it produces well-formed but incorrect trees, using the PIZZA ordering benchmark. Two sub-billion models were fine-tuned with LoRA on six lossless encodings of the same target tree. Shortening the output alone did not help, but writing the target in ordinary words gave the most accurate results and reduced silent errors.
Umar Ahmed, Maxim Dunaev, Mohammad Nour Al Awad et al.
In briefActor, Judge, or Teacher? How Decision Models Work in Long-Task AI Agents
SummaryThis paper studies decision models that choose among options with probabilities, comparing three such models and three general LLMs as actors and judges on ALFWorld. As actors, all six models tend to pursue an easier subgoal instead of a prerequisite one. As judges reviewing whole trajectories, they better credit steps toward that prerequisite when asked whether each action helped. The authors also show JEV works as a teacher for reinforcement learning, and conclude that a weak actor can still teach well.
In briefUAV-Flow++: A unified benchmark and learning suite for language-guided drone flight
SummaryThis paper introduces Flying-on-a-Word (Flow), a task in which a drone converts a short natural-language instruction, its egocentric view, and its current state into continuous flight actions. The authors release UAV-Flow, containing 30,692 real-world expert-piloted trajectories and 10,109 simulation trajectories, and propose FlowVLA with Tail-Aware GRPO (TGRPO) post-training. On 273 held-out simulation episodes, FlowVLA reaches a 75.58% mean success rate before TGRPO and 93.16% after it. Physical drone experiments show it executes object-relative commands in indoor and outdoor scenes.
In briefSymmetry in Time: Representation theory of how linear recurrent models remember sequences
SummaryThis paper uses representation theory to analyze state space models (SSMs), showing that diagonal recurrent updates can only track commutative groups. The authors introduce HyperSSM, which adds quaternion channels while keeping linear-time inference and parallel training, and report that it solves group tasks exactly, including S5 with 32 generators. They also show that the learned weights can be interpreted to recover a gyroscope's calibration, imperfect qubit gate rotations, and protein backbone geometry.
WhyThe authors use representation theory to characterize state tracking in linear recurrent memory and propose HyperSSM with quaternion channels, reporting exact group-task solutions including S5 and 25x
In briefCited but Unused: Provenance Problems in Retrieval-Augmented Generation
SummaryThe paper finds that citations attached by retrieval-augmented generation systems often misstate where an answer actually came from. By substituting false values for facts in retrieved passages and re-querying, the authors measure that 1.2% to 15.7% of citable claims across several model sizes derive from parametric memory rather than the retrieved context. They also report that natural language inference verification performs at chance level and that models tend to select citations that agree with their own outputs.
WhyThe authors propose counterfactual value substitution to measure citation provenance, reporting 1.2% to 15.7% of claims from parametric memory and comparing ten attribution methods.
In briefAeroZoo: A benchmark of well-posed flow problems for aerodynamics foundation models
SummaryAeroZoo is a benchmark and training corpus of 3,815 time-mean solutions to unsteady RANS equations with a transition model, covering seven geometry families, eight enclosure configurations, and Reynolds numbers from 10^3 to 10^8. Its 2,816 samples with stable time means form a canonical benchmark with two tasks and 22 frozen splits that are disjoint in geometry. Tests show that adding boundary-condition tokens reduces GeoTransolver's median velocity error on the random split from 76.9% to 8.4% of inlet speed in single-seed runs. The data and scoring code will be released.
Peter Sharpe, Carmelo Gonzales, Corey Adams et al.
In briefHiVE-VQA: Video quality assessment using hierarchical features from a frozen video encoder
SummaryThis paper presents HiVE-VQA, presented as the first framework to apply a JEPA-based video encoder to no-reference video quality assessment. It extracts features from multiple depths of a frozen pre-trained V-JEPA 2.1 encoder, stacks them into a spatio-temporal volume, and fuses them with a compact 3D convolutional head of about 5.2M trainable parameters. Trained only on LSVQ without fine-tuning, it reports higher SROCC and PLCC than 17 state-of-the-art methods on intra- and cross-dataset benchmarks.
WhyThe authors apply a frozen JEPA video encoder with hierarchical 3D convolutional fusion to no-reference VQA, reporting gains over 17 SOTA methods across four benchmarks.
In briefImproving preference detection by combining eye movement patterns with multilayer model fusion
SummaryThis paper studies whether eye movement patterns can predict which of two face images subjects prefer, using an eye-tracking dataset. It applies multiple machine learning classifiers combined through the Multi-layer Combinatorial Fusion (MCF) framework, which fuses models layer by layer based on cognitive diversity. The second MCF layer reaches a Pre@k of 0.9286, outperforming both the first layer and the best single model at 0.90.
In briefConvex reformulation of poisoning attacks on regularized logistic regression
SummaryThis paper studies data poisoning attacks on logistic regression with ℓ2 regularization, where an attacker selects features for injected observations with set labels to maximize logistic loss on clean training data. Using the defender's strong convexity and first-order optimality conditions, the authors derive an exact reformulation that expresses the attacker's objective as a concave function via the convex conjugate of softplus. Experiments on eight binary classification datasets show that local optimization attacks degrade predictive performance on most datasets, while the convex-analytic formulation yields substantially stronger reported dual bounds within a fixed time limit.
Carina Moreira Costa, Maria Eduarda Pinheiro, Martin Schmidt
In briefISLA: A Coordinate-Free Transformer for Boundary-Driven PDEs
SummaryThis paper presents ISLA (Invariant Slice Attention), a transformer model for steady, boundary-driven partial differential equations. No learned operation takes raw coordinates as input, and sums over the boundary are computed as quadratures, so the model stays consistent under rigid motions, reflections, unit changes, and mesh refinement. On AeroZoo, a set of 2,816 flows varying in shape, size, speed, and boundary conditions, its mean reference-scaled velocity and pressure errors are at least 3.1 and 3.3 times lower than any baseline. On AirfRANS, its errors are second only to the boundary-integral model GLOBE and at least 1.6 times lower than Transolver and GeoTransolver.
WhyThe authors build rigid-motion and unit invariance into slice attention and report lower errors than baselines on AeroZoo and AirfRANS.
In briefOffline policy improvement for pediatric ECMO care using limited clinical data while keeping actions close to clinician practice.
SummaryThis paper presents FQI-SC, an offline reinforcement learning method for managing children on ECMO. It restricts optimization to state-action pairs supported by empirical data and falls back to the clinician baseline in low-data regions. The framework is evaluated on 78 pediatric trajectories covering over 6,000 hours of clinical data, using a unified Balance Score to capture the performance-safety trade-off.
Fateme Golivand Darvishvand, Saurabh Mathur, Nikhilesh Prabhakar et al.
In briefPaper-faithful federated learning: scientific checks for developers and AI agents
SummaryThis paper presents FedBrew, a research codebase for controlled and reproducible federated learning experiments. It uses configuration checks, identifiable client data partitions, checkpoint recovery checks, self-describing run records, and development-time checks to verify that experiments follow the intended protocol, such as whether client partitions are disjoint and exhaustive and aggregation weights match the prescribed rule. The same structure offers extension points and automated checks for new algorithms, datasets, models, and tasks, whether added by researchers or AI coding agents.
In briefParameter-efficient method for making small medical language models give shorter answers
SummaryThe paper introduces PECA, which adds one differentiable conciseness term to the QLoRA supervised fine-tuning objective. It raises the end-of-sequence probability once generation exceeds a token budget, needing no preference data, auxiliary model, or proprietary teacher. On a 1.7-billion-parameter model, mean response length falls by 72.8% while multiple-choice diagnostic accuracy rises from 40.8% to 44.0%. Varying the conciseness strength traces an accuracy-to-length trade-off curve.
Tanzeel Abbas, Ahmad Kamran Malik, Nafees Qamar et al.
In briefSTeP-C: Real-time neural compression for large-scale fiber-optic sensing data
SummaryThe paper presents STeP-C, a neural compression framework that separates spatial encoding, temporal prediction, and entropy-optimized residual coding, trained end-to-end with a rate-distortion objective. The authors identify entropy coding as the main bottleneck for streaming and address it with GPU-side symbol computation, cached probability tables, and parallel GIL-free arithmetic coding. On a live DAS testbed, the pipeline achieves a 98.85% file size reduction at 31.04 dB full-band PSNR and an average throughput of 535 MB/s.
In briefEstimating the size a second-generation DNA-encoded library needs to train a high-resolution binding model
SummaryDNA-encoded libraries (DELs) are used to find small molecules that bind a target protein, but building a machine-learning model precise enough to guide later chemistry is the harder task. The paper models binding as a sum of independent interactions called pixels and antipixels, and uses only the spread of affinities measured in a naive DEL to estimate how complex a second-generation library built around a hit must be. The method also estimates the expected yield of a randomly designed second-generation DEL, providing a baseline for measuring the gain from machine-learning or active-learning design.
In briefA unified framework for Bellman contraction under MMD
SummaryThis paper presents a mathematical framework for studying contraction of the distributional Bellman operator under Maximum Mean Discrepancy (MMD). It shows that contraction reduces to conditional positive definiteness of a kernel built from the base kernel, and uses Schoenberg's theorem and Bernstein functions to derive criteria for radial kernels. The framework unifies known results for several kernels and numerical search identifies 11 new kernels that induce contraction.
WhyThe authors reduce MMD contraction of the distributional Bellman operator to conditional positive definiteness and report 11 new contracting kernels via Schoenberg's theorem.
In briefDesigning Institutions for Autonomous AI: Making Rules for Groups of AI Agents a Research Topic
SummaryThis Perspective argues that as autonomous AI agents increasingly interact in persistent populations, their collective capabilities and alignment depend on the institutions governing those interactions, not only on individual models. Drawing on institutional political economy, social epistemology and organizational theory, it proposes seven principles for designing such institutions, covering reciprocal contributions, accountable authority, and human direction. It illustrates these ideas through machine science and proposes an experimental programme comparing AI populations under different institutional arrangements, describing MoltJournal, an API-first journal for autonomous research agents.
In briefTime-aware 4D segmentation for tracking pore changes and fluid flow in CT scans
SummaryThis paper presents a transformer-based 4D segmentation framework that uses sequences of time-resolved micro-CT images to separate real physical change from segmentation error. On mineral dissolution images, it measures pore-volume growth and rock boundaries more accurately than an equivalent 3D model without temporal information, and captures permeability increase. The method transfers to another dissolution sample without retraining, and on two fluid displacement datasets it reduces spurious changes and accurately identifies filling events.
In briefSALA: Aligning Human and Robot Latent Actions to Scale Robot Foundation Models
SummaryThis paper proposes SALA, a method that links human egocentric videos and robot actions by establishing semantic correspondence in a shared latent action space. It estimates similarities between unpaired human and robot clips using their language descriptions, and combines this with robot action reconstruction to ground the latent space in control. The authors report that pre-training on more general human video reduces robot action prediction error on downstream tasks, and post-training on more task-specific human video substantially raises real-world robot task success rates.
In briefOracle-efficient algorithm with near-optimal regret for adversarial contextual bandits
SummaryThis paper studies contextual bandits with K actions, N policies, and i.i.d. contexts the learner can sample. The authors give an oracle-efficient algorithm with expected regret O(√(KT log(2N)) log(2KT)) against the best policy in hindsight, improving the T^{2/3} horizon dependence of prior oracle-efficient bounds and matching the optimal rate up to logarithmic factors. The method perturbs policy scores using Laplace noise at Poisson-sampled auxiliary contexts, with scales approximated by projected stochastic gradient descent using only an exact offline cost-sensitive optimization oracle.
WhyThe authors improve oracle-efficient adversarial contextual bandit regret from T^{2/3} to near-optimal √T order, using Laplace perturbations and projected SGD.
In briefLazyWeights: Decode-on-Demand LLMs for Edge Devices
SummaryThis paper addresses the memory capacity and bandwidth limits of running large language models on edge devices. LazyWeights keeps existing quantization formats but tunes GPTQ so the integer codes compress better, applies lossless Huffman coding, and decodes weights on demand in a small unit between DRAM and L2 so GPU kernels run unchanged. The paper also reports a bit-exact ASIC decoder unit and its operation on an FPGA, with latency and energy measurements on Jetson devices.
In briefDisentangling Spurious Correlations: How IRM Handles Diverse Shortcuts in Text Classification
SummaryThis paper studies how text classifiers often rely on spurious correlations, such as dialectal markers, instead of robust linguistic features, which hurts out-of-distribution generalization. The authors use a causality-driven framework to test Invariant Risk Minimization (IRM) under three shortcut mechanisms: label-generated features, confounding, and selection bias. IRM successfully rejects shortcuts in controlled synthetic settings, and it can mitigate more realistic shortcuts in Bias in Bios and ActorCorr. However, semantically meaningful features require careful calibration of the invariance penalty.
Paul Grunenwald, Beno\^it Crabb\'e, Guillaume Wisniewski et al.
In briefPredicting alcohol drinking from monkey liver tissue images using few-shot AI tuning
SummaryThis study tests whether combining vision-language models with parameter-efficient fine-tuning can classify whole-slide tissue images from few examples. It predicts ethanol drinking status from nonhuman primate liver histology using a subset of the Monkey Alcohol Tissue Research Resource. Every fine-tuned configuration achieved higher mean balanced accuracy than the non-fine-tuned baselines, for example 78.1% with an adapter versus 43.5% for ABMIL alone in the four-shot setting.
Amirmohammad Mohammadi, Rupak Khadka, Kathleen A. Grant et al.
In briefTuMAP: A pipeline for measuring kidney tubule remodeling from tissue images, with tubule shape types accounted for
SummaryThe paper presents TuMAP, a pipeline that combines foundation-model image features, tubule shape-type annotation, pathomics measurements, and patient-level statistics to quantify disease-related kidney tubule remodeling. It analyzed 456,158 tubules from 358 PAS-stained kidney biopsy slides and identified eight image-based morphotypes. Across 219 patients in six diagnostic groups, it compared morphotype proportions and within-morphotype features against minimal change disease.
In briefData-free model merging via joint null-space projection
SummaryThis paper proposes JNSP (Joint Null-Space Projection), a method that merges independently fine-tuned checkpoints into one multi-task model without data, labels, or gradient steps. For each task, it takes the top singular subspace of the weight delta and removes from every task vector the components lying in the other tasks' subspaces. On eight CLIP image-classification tasks, merged ViT-B/16 reaches 89.45% average accuracy, above the listed training-free baselines. The paper also reports that the method underperforms plain task arithmetic on RoBERTa-large.
In briefA compact neural network using xDAWN filtering for P300 brain signal detection
SummaryThis paper presents a compact deep learning model for decoding P300 brain responses in brain-computer interfaces that enable character selection. The model combines xDAWN spatial filtering, a fixed frequency filter bank, and convolutional feature learning, with depthwise-separable convolutions to reduce computation. Its target/non-target classifier has only 5,100 trainable parameters. On two public benchmarks, it reached 17.8 bits/min with two repetitions on BCI Competition III and 40.5 bits/min on BCI Competition II.
In briefWhy Build Speed Matters More Than Power Price for AI Data Centers
SummaryThis paper compares ways to power AI data centers across the continental United States, including grid connection, solar-plus-storage with AC or DC coupling, and islanded natural gas generation. The authors built an open-source technoeconomic model and introduced a speed-adjusted cost of energy that values the compute gained from faster construction. They report that DC coupling lowers lifetime energy costs by 18% on average versus AC, and that once foregone compute is valued, DC-coupled microgrids are the lowest-cost option in 85% of U.S. locations.
Alex C. Newkirk, Daniel L. Gerber, Erica R. H. Fuchs et al.
In briefSALT: State-aligned low-rank transfer for single-image building localization across remote sensing datasets
SummaryThis paper addresses remote sensing change detection, which usually needs spatially aligned before-and-after images, by localizing structures such as buildings from a single current image. The authors propose SALT, which freezes a DINOv3 backbone and learns positive and negative structural states, correcting patch features through State-Aligned Low-Rank Adaptation along with alignment and geometry-preservation objectives. Experiments on LEVIR-CD and the WHU aerial building dataset in both transfer directions show SALT outperforms four baselines at the pixel level, using 0.068M trainable parameters and 40.2 ms per image.
Aggelos Psiris, Georgios Angelidis, Giorgos Sfikas et al.
In briefWhether reasoning traces help LoRA fine-tuning of a language model on financial tasks
SummaryThis paper studies how to fine-tune the Qwen3-8B model with low-rank adaptation (LoRA) for financial tasks without losing general abilities. It compares three training setups: learning from labels only, adding teacher reasoning traces to a single adapter, and training separate adapters for traces and labels. Label-only training improves in-domain performance while keeping general capabilities close to the frozen model. Training with reasoning traces scores lower in-domain, and the dual-adapter setup gives no further gain.
In briefEvolving programs to discover effective memory designs for AI agents
SummaryThis paper introduces MAPE, a framework that evolves complete memory programs for agents starting from minimal seeds, using a standardized lifecycle interface and an iterative diagnosis-proposal loop. On LoCoMo and LongMemEval, it reports accuracy gains of up to 12% over manually designed baselines at 35% of the serving cost. The authors identify two recurring design patterns: keeping raw conversations as retrievable evidence and composing retrieved evidence according to query intent.
WhyThe authors evolve agent memory programs and report up to 12% accuracy gain on LoCoMo and LongMemEval at 35% serving cost, plus transferable composition module.
Xinheng Ding, Seyyed Saeid Cheshmi, Ziyang Zhang et al.
In briefNeural-network eddy-viscosity closure for molten salt reactor thermal-hydraulics in a system code
SummaryThis paper develops an invariant neural-network eddy-viscosity closure for the System Analysis Module (SAM) and applies it to the Molten Salt Fast Reactor. The model predicts turbulence from neighboring element flow and geometry without solving transport equations, and iterative retraining against Nek5000 targets addresses the coarse-mesh covariate shift. In a 120 s pump-coastdown transient, the velocity error falls from 37.6% to 17.7% relative to the mixing-length model. The coupled transient takes about 0.2 core-hours versus about 330 for Nek5000.
Zaid Abulawi (1), William Oliveira (2), Akzhol Almukhametov (1) et al.
In briefRopedia Xperience 2: A Gold Standard Dataset for Scaling Physical AI Environments and Behaviors
SummaryThe paper introduces Ropedia Xperience 2, comprising 1,589 released metric-scale 3D scans and 4,230 egocentric recordings covering 74 tasks totaling 131 hours. It shows that each capture can be reused for robot learning, converting scenes into physically interactive simulation environments and retargeting human motion to whole-body robot trajectories. Additional robot behaviors are generated through virtual teleoperation and automated long-horizon trajectory generation. The authors propose this joint capture of environments and behaviors as a gold standard for physical AI data collection.
In briefCapacity-Constrained Delayed Feedback Control for Online Conformal Risk
SummaryThis paper studies online conformal risk control, a method that builds prediction sets with long-term risk guarantees, when feedback is delayed and limited tracking capacity causes some feedback to be lost. The authors propose CA-OCRC, which uses a randomized tracking policy and inverse probability weighting to update prediction sets. They analyze risk bounds for arbitrary and stochastic delays and optimize the proxy delay distribution, deriving scaling laws for tracking capacity. Experiments span medical imaging, natural language processing, computer vision, and wireless communications.
In briefTesting whether pretrained language models help predict outcomes from structured patient records
SummaryThe paper introduces tEHR, which describes clinical events in electronic health records in natural language and adapts a pretrained causal language model (Mistral-7B-v0.3 backbone) with continued pretraining followed by LoRA fine-tuning. It is evaluated on four MIMIC-IV prediction tasks and pancreatic cancer prediction on CPRD at 1 to 12 month horizons, achieving the strongest overall performance. Ablations show performance drops when clinical descriptions are removed or misaligned, when numerical values are omitted, or when either adaptation stage is removed.
In briefAuditing AI reviewers with a benchmark on real papers
SummaryThis paper introduces a benchmark that injects objective errors of four types (general, mathematics, numerics, citations) at three severity levels into real arXiv papers, measuring each AI reviewer's detection rate and false-positive rate on clean copies. It also presents Skepthical, a multi-agent reviewer that verifies claims by retrieving cited papers, using computer algebra, and running numerical code. Evaluating twelve reviewer systems on astrophysics and machine learning corpora with 357 injected errors, the authors find that model choice has the largest effect on detection rate, while tools help mainly where the model cannot supply the capability itself.
WhyThe authors introduce a benchmark injecting 357 localisable errors into real astrophysics and ML preprints, and a multi-agent reviewer using retrieval, computer algebra, and executed code across twelv
\'I\~nigo Zubeldia, Boris Bolliet, Francisco Villaescusa-Navarro et al.
In briefAURA 2.0: Model-Initiated Conversations in Streaming Video
SummaryThis paper presents AURA 2.0, a streaming video large language model that can start conversations about events in a video stream without waiting for a user query. It adds tool calling that uses the current visual and conversational context, and visual grounding that links user-marked objects and model responses to frame-aligned bounding boxes. The authors built a streaming interaction dataset, used a three-stage training scheme, and added long-horizon memory to keep context over long videos. The paper reports leading performance among open-source streaming models on public benchmarks, and the model weights and inference code are open-sourced.
In briefFrom Delegation to Coordination: Learning Efficient Multi-Agent Deep Research
SummaryThis paper studies how to train a main agent and its subagents together in multi-agent deep-research systems, where delegation parallelizes search. The proposed LEAD-RL framework uses a training-time rubric to assess the research utility of each subagent report and gates the team advantage it inherits, with final-answer correctness as the only team reward. Across five benchmarks and two model families, it improves answer quality. On BrowseComp, it raises accuracy by 2.49 points and cuts critical-path cost by 24.5% compared with joint training using shared team credit.
In briefYour vision model may be secretly good at reading brain signals (EEG)
SummaryThis paper shows that a vision foundation model never trained on EEG can outperform specialized EEG foundation models when EEG signals are rendered as images. The proposed SeeEEG method keeps the vision backbone frozen and learns only how EEG is displayed, using at most 1,113 trainable parameters per view. On eight clinical, sleep, affective, and brain-computer interface datasets, SeeEEG with a frozen DINOv3 ViT-B beats seven EEG foundation models in balanced accuracy on all eight. The learned displays also transfer to the video encoder V-JEPA 2.1.
WhyThe authors render EEG as images for a frozen vision transformer, learning only display parameters, and report beating seven EEG foundation models on eight datasets with a random-init control.
In briefSimilar but Not Redundant: Embedding-Gated Refinement for High-Granularity Topic Models
SummaryThis paper examines high-granularity topic modelling, which extracts hundreds of niche topics from a corpus, and argues that calling it "over-clustering" confuses genuine redundancy with semantically similar but distinct topics. The authors propose EGIM, a hybrid method that uses embedding similarity to propose candidate merges and LLM judgement to accept only truly redundant pairs. Across two patent corpora and one legal corpus, the LLM rejected 83% of embedding-identified candidates as non-redundant, and EGIM improved discriminability of merged topics by 25-55% over a prior iterative LLM baseline.
In briefSelf-supervised learning that is local in both space and time
SummaryThis paper introduces Temporal Regularized Learning (TeReL), a learning rule for self-supervised representation learning from sequential input streams that is local in both space and time. It uses a VICReg-style objective with temporally adjacent positives and non-differentiated running population statistics, giving a per-neuron online loss with bounded state. On a class-chunked MNIST stream, a two-layer fully connected network trained with batch size one reaches 96.48% linear-probe accuracy, above random features at 95.09% and classical SFA.
In briefAI for Biologists: Where Biology's Foundation Models Can Be Trusted
SummaryThis review covers biological foundation models trained on millions of nucleotide, protein, and single-cell measurements, along with agentic AI systems built on large language models that call these models as tools. The authors define a model's knowledge boundary as the range of inputs, tasks, and conditions over which its outputs are reliable. They propose that five design elements (data, architecture, training scheme, evaluation scheme, and post-analysis) define this boundary. The paper examines where current models succeed or fail and proposes standardized descriptions of model limits for researchers and their agents.
In briefRandom exploration: is the randomness real or just model limits?
SummaryThis paper examines random exploration, where choices become more variable and less tied to the currently best-looking option in decision making. Using the Horizon Task on an aggregated dataset of 1,392 participants completing over 1,270,695 trials, the authors test whether more expressive neural network models can capture deterministic structure behind this variability. Neural networks fit behavior better overall than the traditional model, yet long-horizon choices remained less predictable than short-horizon ones.
In briefSkill Adapter: personalized customization and online selection of general-purpose AI agent skills
SummaryThis paper proposes Skill Adapter, a system that personalizes general-purpose skills for large language model agents and selects them during task execution. It distills local adjustments from user requirements and execution feedback and stores them as Adapters linked to general skills, with a routing agent choosing skills and Adapters per subtask. Evaluated on ALFWorld, WebShop, and SkillsBench, it improves task success rates by an average of 11.43 percentage points over SkillRL and reduces mean execution time.
In briefA Data-Efficient Real-Time Neural Imitation Architecture for Nonlinear Time-Optimal Model Predictive Control
SummaryThis paper addresses the high computational cost that prevents real-time use of nonlinear time-optimal model predictive control (TOMPC) on fast systems. It proposes self-informed sampling (SIFS), which uses a learned time-to-go to query the expert only where the policy deviates from time-optimal behavior, along with a complete training and control architecture. On a Furuta pendulum, the controller raises control frequency from 1.4 Hz to 1 kHz and swings up and stabilizes the pendulum along trajectories closely matching the expert TOMPC solution.
Maarten De Baecke, Flavia Sofia Acerbo, Andrea Giusti et al.
In briefAnalogLayoutWorld: An Open Benchmark for AI Agents Designing Analog Circuit Layouts
SummaryThis paper introduces AnalogLayoutWorld, an open-source benchmark of 40 expert-reviewed tasks where AI agents must generate complete analog circuit layouts from fixed transistor-level netlists. The benchmark pairs an integrated design environment with independent physical and post-layout electrical verification, plus layout quality metrics developed with analog layout experts. Across 13 model-harness configurations, 91.1% of attempts produced GDS layouts, but only 23.0% met the acceptance criteria.
In briefPrimer: An AI Framework for Instructor-Centered Educational Simulation Creation
SummaryThis paper presents Primer, an open-source, LLM-agnostic system that helps educators build adaptive educational simulations. Educators supply their subject, lesson, and student knowledge, while generative AI produces responsive practice variations for students. Primer uses a multi-agent system to carry out the educator's instructional intent across the whole learning experience.
Ethan Mollick, Lilach Mollick, Ben Przystanski et al.
In briefElasticDino: Sharp dense image features via elastic deformations
SummaryThis paper introduces ElasticDino, a lightweight adapter that turns vision transformer outputs into high-resolution feature maps aligned with the fine details of the input image. It uses spatial deformations to keep semantics sharp and avoid the smoothness of alternative approaches. Without ground-truth labels, it is guided by linear heads for pixel-level segmentation and depth estimation. The authors report higher accuracy than current methods on several dense prediction benchmarks, with better computational performance.
In briefCompiling known game rules into compact neural agents with weight-space programming
SummaryThe paper studies weight-space programming (WSP), in which language models construct neural network weights without training. The authors test whether known game mechanics can be compiled into compact agents that generalise to held-out levels, renderings and start states within a game family. Using 20 privately authored ARC-AGI-3-style game families and a fixed 57k-parameter recurrent architecture, the strongest writer reaches 56% held-out-level success with feedback, while four main writers' Python players reach 97-100%.
In briefMorphological Simplification in Alzheimer's Disease: Analyzing Word Forms in Speech
SummaryThis study examines how word-form patterns in connected speech differ between people with Alzheimer's disease and healthy controls, using 55 automatic measures on Cookie Theft picture descriptions. Speakers with AD produced fewer suffixed and inflected words, shorter verbs and nouns, and more past-tense forms, while errors were similar across groups. Morphological features classified AD about as well as syntax and better than fine-tuned BERT in one setup, and the authors conclude that AD speech shows simplification rather than breakdown.
In briefThe Value of Thinking: Learning When a Model Should Reason Step by Step
SummaryThe paper proposes a "zero-or-T" paradigm in which one model either answers from implicit computation in a single forward pass, generating no tokens, or performs full autoregressive reasoning at a cost of T tokens. It introduces a hybrid decision model with a frozen backbone and a lightweight self-routing block that estimates the value of slow reasoning over fast answering. For episodic tasks lacking per-decision labels, it also presents a reinforcement-learning method, Calibrated Decision Policy Optimisation (CDPO), and evaluates the approach on several benchmarks.
In briefLara Mini Base: a small multilingual language model that treats language diversity as a resource
SummaryThis paper presents Lara Mini base, a language model trained from scratch on 6 trillion tokens covering English and 44 other languages, including all 24 official EU languages. It uses a mixture of experts design with 7.2B total and 0.6B active parameters. The authors report the highest average score among open base models with up to 2B active parameters, and faster text generation than higher-scoring models. The weights are released under CC BY-NC-SA 4.0.
Translated, Maria Pia Adinolfi, Isabelle Andrieu et al.
In briefA disease-specific retinal aging marker for predicting progression to late age-related macular degeneration
SummaryThe paper proposes fundus-based AMD age (FAA), a measure of disease-specific retinal aging from eye images, to better predict progression of age-related macular degeneration. It was trained on 36,782 images from 2,192 AREDS participants and validated in AREDS and AREDS2 cohorts. Adding FAA to Cox proportional-hazards models improved the C-index in AREDS from 0.836 to 0.888, outperforming chronological age alone.
In briefS1-Evo: Experience-Guided Evidence Competition Attacks on Typed System-1 Decision Models
SummaryThis paper studies typed System-1 (S1) decision models used in LLM agents for routing, moderation, and authorization, and shows that untrusted context can manipulate their typed outputs and alter execution branches. The authors propose S1-Evo, a self-evolving framework that perturbs untrusted fields through contract-aware evidence channels to create evidence competition while keeping facts invariant. It optimizes candidate perturbations offline using margin-guided feedback search and distills verified flips into experience cards that are retrieved to attack unseen tasks. The paper reports results across four benchmarks and three S1 models, including mean ASR and end-to-end agent violations.
Kaixiang Wang, Junjie Li, Zhaojiacheng Zhou et al.
In briefMantis 2: A General Model for Forecasting Infectious Diseases
SummaryThis paper presents Mantis 2, a foundation model for infectious disease forecasting that combines simulated outbreak data with observed surveillance data. The authors found that simulation-rich training is especially useful when local surveillance history is limited. Mantis 2 jointly forecasts up to six variables sixteen weeks ahead and achieved the best aggregate probabilistic accuracy across 30 forecasting tasks among the models compared. The model is publicly released.
In briefTime Flies When You Parallelise: Efficient Time Series Foundation Models Using Linear Recurrences
SummaryThis paper introduces ci-proto, a family of time series foundation models built on Mamba-3, whose linear state recurrences allow chunk-wise parallel computation during training and inference costs that grow linearly with sequence length. The architecture alternates recurrent temporal processing with cross-variate attention to support multivariate probabilistic forecasting. The authors report that ci-proto XL sets a new state of the art on fev-bench's 100 forecasting tasks, and that its five variants form the Pareto frontier against parameter count and median runtime.
In briefKezek: a model that splits mono two-person conversations into one audio track per speaker for full-duplex dialogue data
SummaryThe paper presents Kezek, a model that takes a single-channel recording of two people talking and separates it into one track per speaker, keeping the original length and timing. Overlapping speech and short replies such as "yeah" stay on the track of the person who said them. It works in two stages: a two-voice activity model guided by a phase-insensitive speaker fingerprint decides who speaks when, then a conditioned separator carries out that decision, with outputs summing to the input. Compared with the open pipeline pyannote plus DialogueSidon, Kezek keeps more short replies, leaks less between speakers, swaps channels less often, and runs 4 to 5 times faster.
