In briefTesting AI models for spotting misinformation by using predefined social noise measures
SummaryThis study evaluates six text-based models and a graph neural network on detecting misinformation using predefined social noise constructs. The dataset contains 500,000 tweets collected from X during the Russia-Ukraine conflict between March 2024 and March 2025. RoBERTa achieved the highest accuracy at 97.17%, while the graph neural network reached 64.52% accuracy and an AUC of 0.899.
In briefHierarchical reinforcement learning that guides a spacecraft from orbit matching through docking
SummaryThis paper presents a hierarchical reinforcement-learning framework for autonomous three-degree-of-freedom rendezvous and proximity operations, covering the full sequence from orbit matching to docking. Five phase-specific policies are trained for orbit matching, far-range, mid-range, close-range approach, and docking, and a learned high-level manager selects among them based on spacecraft state. The approach is compared with a fixed sequential-invocation heuristic on end-to-end trajectory performance and transition behavior.
Mariko Storey-Matsutani, Anatoly Zlotnik, Lee Holguin et al.
In briefEfficient Fréchet Distributional Decoder Alignment for latent generative models: a cheaper way to post-train the decoder
SummaryExisting Fréchet Distributional post-training methods for latent generative models update only the generator while the decoder stays frozen, so the decoder is deployed on latents whose distribution it was not trained on. The authors propose FDDA, which freezes the generator and optimizes the decoder directly on generation-time latents with an FD objective. Across 11 evaluated models, one epoch of FDDA reduces gFDr^6 further and runs substantially faster than generator-side FD post-training.
In briefFPGA-Based Spiking Neural Network for Learning and Classifying Spatiotemporal Patterns
SummaryThis paper presents a multilayer spiking neural network implemented on an FPGA, where both inference and training run on the same hardware for spatiotemporal pattern classification. It uses a temporal-forward-only training algorithm instead of Backpropagation Through Time, with mixed-precision arithmetic and architectural optimization, and resolves memory port conflicts and read-after-write hazards from shared memory. On MNIST and five multivariate time-series datasets, fully on-chip training reaches accuracy comparable to BPTT-trained SNNs or LSTMs. Compared with CPU and GPU implementations, the design reports 2.5 to 6 times higher throughput and roughly an order of magnitude lower energy per sample.
In briefTransferring text language model knowledge to speech-to-speech models by distilling internal text predictions
SummaryThis paper proposes a speech-to-speech pretraining method that builds an internal text interface while keeping speech-only input and output. The model reads speech context without transcription, switches to predicting text, and uses that text to guide later speech generation word by word. These internal text predictions can be distilled directly from a base text language model. The authors report improved generative perplexity of spoken continuations and better factual performance under text and speech evaluation.
In briefTesting differences between clusters without inflated false positives under unknown covariance
SummaryThe paper addresses the problem that testing differences in means between clusters, after forming them from the same data, inflates the Type I error rate. It proposes a new selective test that provably controls this error when the covariance matrix is arbitrary and unknown, without estimating it, using Monte Carlo sampling over orthogonal matrices. The authors show the test is efficiently computable and give closed-form solutions for k-means and hierarchical clustering. Simulations and applications to gene expression and drug review data show error control and substantial power.
In briefCompositional RL: Training LLMs to Collaborate Well in Multi-Model Systems
SummaryThe paper argues that language models are usually judged alone, yet they are often used together in systems, so it defines a model's "compositional strength" via Shapley values and a leave-one-out proxy. Across many pools, methods, and datasets, standalone strength often fails to predict collaborative advantage. The authors then propose compositional reinforcement learning (C-RL), which rewards the outcome of the whole multi-model system, and report that it yields the strongest collaborator in most tested settings.
WhyIntroduces Shapley-based compositional strength and compositional RL; experiments across many pools, methods, and tasks show individual and collaborative strength are largely uncorrelated.
In briefHow spherical parameter space affects the number of non-forgetting tasks and learning capacity in continual learning
SummaryThe paper models continual learning on a high-dimensional sphere, representing each task's acceptable loss region as a geodesic ball around its optimum. Task optima are generated by a Neyman-Scott clustered point process to reflect task similarity. In the limit of large dimension, the authors derive an upper bound on the number of tasks for which the acceptable regions still intersect, meaning no forgetting occurs, and an upper bound on each task's information-theoretic capacity under this constraint.
In briefIdentifying Sleep-Wake Disorders from Polysomnography Representations
SummaryThis paper compares two kinds of pretrained encoders on whole-night polysomnography recordings: self-supervised PSG foundation models and sleep-staging encoders pretrained at comparable scale. Using frozen representations on 7,906 PSGs from an unseen registry, the authors identify six sleep-wake and circadian disorder groups. The self-supervised models outperformed an EEG-only baseline, but this advantage reversed when supervised pretraining matched their scale and modality coverage. Transfer from neural sleep scorers achieved the highest overall performance.
Alvise Dei Rossi, Luigi Fiorillo, Giuliana Monachino et al.
In briefFormal-verification-guided Bayesian optimization for closing last-mile chip coverage without modifying testbenches
SummaryThe paper presents a non-intrusive workflow that closes remaining bins in an existing, human-written functional coverage model rather than generating new verification objectives from the specification. It maps each residual target to existing testbench controls using AST parsing and cone-of-influence tracing, then compares Bayesian optimization against constrained-random verification and random search under a matched 20-simulation budget. Formal reachability removes targets proven unreachable, and the remaining targets are grouped by difficulty and searched again. On OpenTitan AES and Entropy Source, the integrated workflow reached 96.31% and 97.06% raw coverage. The authors note that the evaluation does not isolate the causal contribution of each component.
Chengjia Liu, Priyadarsi Mishra, David Kebo Houngninou et al.
In briefTableVision: Recovering Verified Structure from Historical Tables
SummaryThe paper addresses common problems in scanned historical tables, such as missing rows, page-crossing tables, and misnamed columns, by treating table structure recovery as the main task. TableVision partitions a volume into blocks using a dynamic program over column sum checks, reads the cells, and classifies each column as exactly reconciling, flagged as faithful to an arithmetically wrong page, or withheld. On 2,201 columns of Wisconsin electoral data, the gate accepted 2,042, flagged 118, and withheld 41.
In briefBelief-action gap: language model agents often know more than they act on
SummaryThis paper studies how beliefs about states, action values, and action preferences interact when language model agents reason and commit to decisions in text-rendered grid worlds. The authors find that task-relevant state and value information is often represented more reliably than it is reflected in behaviour. They also report that decisions sharpen around a commitment point, though individual belief changes do not consistently explain shifts in action preference.
Ayse Sila Okcu, Weisheng Wang, Tongmeng Xie et al.
In briefREAP: Learning general robot rearrangement heuristics from motion-planner feedback
SummaryThis paper proposes REAP, a framework that learns a relational graph neural network heuristic over a simple symbolic abstraction for robotic rearrangement planning. A motion planner supervises training by eliciting hidden geometric action preconditions. The authors report that the learned heuristic generalizes out-of-distribution across object count, plan length, and arrangement, solves more problems than baselines, and uses roughly an order of magnitude fewer motion-planner queries. The top-ranked action was geometrically feasible in 94% of sampled states, versus 22% for random choice.
In briefKnow-Thy-Not: Training-free targeted negation for dual-encoder vision-language models
SummaryDual-encoder vision-language models often fail to handle negation, scoring "a cat without a dog" almost the same as "a cat with a dog." The paper argues this flaw is inherent to the single-embedding framework and proposes Know-Thy-Not (KTN), a training-free method that adds an image-side rejection dimension to penalize the negated concept. The authors also introduce COCO-Confusion, a benchmark that requires genuine rejection to pass. Across 9 backbones, KTN improves COCO-Confusion mAP by up to 25.6 over vanilla models and 9.6 over the strongest prior method.
DoYoung Kim, Changkyu Choi, Ad\'in Ram\'irez Rivera et al.
In briefLLMs Push Further From Known Work but Stay Within Existing Research Directions
SummaryThis paper compares hypotheses generated by large language models with human-authored papers in a shared conceptual space, separating novelty into distance from existing work and direction, meaning whether a new line of inquiry is opened. The authors introduce an orthogonal novelty score (ONS) and analyze 26.9 million human papers and 524,308 LLM-generated hypotheses. LLM ideas are more distant from existing work, but their median ONS is 15–23% lower than published papers on the same topic. The authors conclude that LLMs extend the frontier within established directions, while creating new directions remains led by human scientists.
WhyProposes an orthogonal novelty score separating distance from direction, showing at scale that LLM hypotheses extend but don't create new research directions.
In briefSchemaWeaver: Speeding Up On-Device Mobile AI Agents with Schema-Guided Parallel Decoding
SummaryThis paper addresses the long delay when on-device language models call external tools on mobile phones, which is mainly caused by the decoding stage being limited by memory bandwidth. It proposes SchemaWeaver, which uses schema-guided parallel decoding to generate multiple tokens per decoding pass during tool calling. The method requires no training or changes to the models, and experiments on major mobile chips and model families report latency reductions of up to 3.9 times compared with competitive baselines.
In briefWorld-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models
SummaryThis paper introduces World-T3, a camera-controlled video world model that stores past observations in test-time training memory within its model parameters. A recursive least-squares rule updates this memory online, keyed by camera pose, so scene features can be read back during long generation. The autoregressive model is also distilled into a few-step generator via dual-teacher distribution matching for real-time use. On a long-horizon revisit benchmark, it raises revisit PSNR from 13.52 to 18.31 dB and lowers absolute rotation error from 3.344 to 1.075 versus the strongest baseline, running at 20.6 FPS at 1280 x 704 on one NVIDIA B200 GPU.
In briefMonarch-Muon: A Structure-Aware Optimizer for Faster, Leaner LLM Pretraining
SummaryThis paper proposes Monarch-Muon, a matrix-aware optimizer for weights parameterized by Monarch factors, aimed at reducing the time and memory cost of pretraining large language models. It replaces one dense orthogonalization step with several small independent update subproblems. In 6.89B-parameter system measurements, the authors report peak memory falling from 68.9 to 52.9 GB and the optimizer step falling from 1288 to 304 ms.
Alexander Miasnikov, Ruslan Kabirov, Gleb Molodtsov et al.
In briefReinforcement learning with partly verifiable rewards for colon cancer treatment planning
SummaryThe paper frames first-line colon cancer treatment planning as an interactive reinforcement learning problem under partial observability, where an agent must gather patient information before recommending treatment. It builds patient-specific, process-aware rewards from decision processes in real-world clinical records and optimizes Llama-3.1-8B-Instruct with GRPO and GDPO. The GDPO-trained policy improves Mean@3 clinical reward by 79.60% over base Llama and exceeds GPT-5.2 by 18.82% in the same environment, while reducing query count during training.
Li-Ching Chen, Andr\'es Franco Valiente, Yuta Ishiyama et al.
In briefKeeping Sparse Networks Connected: Reachability-Guided Topology Updates
SummaryDynamic Sparse Training repeatedly prunes and regrows network connections, but at extreme sparsity it can break the paths that carry information from input to output, causing sharp performance drops. The authors formalize forward, backward, and functional reachability on layered sparse graphs and show that Monte Carlo simulations match their derived expectations. They introduce guided DST, which biases pruning and regrowth using reachability, and report large accuracy gains over unguided DST on CIFAR-100 at 99.9% sparsity, with MLP-Mixer remaining trainable at 99.99% sparsity.
Mohammad Pasande, Mateusz Jaworski, Jeroen Rook et al.
In briefMaking SE(2)-invariant networks optimal: a method for building neural networks that ignore translations and rotations
SummaryThe paper uses classical invariant theory to derive invariant feature maps for the SE(2) symmetry group that are provably minimal and information-preserving. These maps are packaged as a plug-and-play module for convolutional networks, needing no changes to nonlinearities or normalization layers. Experiments reportedly show accuracy competitive with or better than state-of-the-art equivariant methods at significantly lower computational cost.
Tomas Karella, Emily Shinkle, Alice Elisabeth Anastasia Allen et al.
In briefClearNav: Perception-Aligned Quadrotor Flight Using Differential Physics (a paper on drone navigation that keeps motion within the camera's view)
SummaryThis paper presents ClearNav, a navigation policy for quadrotor drones flying autonomously through cluttered environments. It uses reward terms that keep motion within the direction the onboard camera observes and limit speed near obstacles. Simulation results report higher success rates and fewer collisions than the compared state of the art, and real-world flight tests are also described.
In briefSuper-resolution as Spatial Downscaling: Structure and Task Utility in Earth Observation
SummaryThis paper treats satellite image super-resolution as a spatial downscaling problem and compares four lightweight networks, including the authors' 0.36-million-parameter model, with bicubic interpolation. The networks beat interpolation by about 1.39 to 1.42 dB in pixel fidelity at ×2, but their effects on land-use and building-change tasks are small, and they can amplify spurious change between observations of an unchanged scene. The authors argue that evaluation should combine pixel metrics with spatial structure, observation coherence, and task-level tests under spatial validation.
In briefWhy Agentic Workflows Fail: A Source-Aware Failure Taxonomy
SummaryThis paper proposes a source-aware taxonomy of failures in agentic workflows, built from academic literature and GitHub evidence. It uses a two-axis matrix, with workflow stages as the process axis and agent capabilities as the functional axis, to localize intermediate failure states. Literature failures emphasize planning, reasoning, and learning, while GitHub failures concentrate on execution, tool, API, and infrastructure issues, with iterative execution and actuation as the dominant hotspot. The authors argue that agent reliability should be treated as a systems problem, not only a model-output problem.</summary>
In briefCompressing vision tokens by keeping what a summary bank fails to reconstruct
SummaryVision Transformers are costly when processing long sequences of visual tokens. This paper proposes ReConV, which summarizes all visual tokens in a compact Summary Bank and keeps only tokens whose reconstruction residuals show the bank does not capture them well. On ImageNet, it surpasses the uncompressed baseline with 48% fewer GMACs. It also yields reusable image-level representations for efficient multi-turn inference.
In briefAgency as a New Organizing Abstraction for Computer Science: A Paper on Structural Gaps in Agentic AI Systems
SummaryThis paper argues that agentic AI systems, which pursue goals autonomously and interact with open environments, should be treated as a new organizing abstraction for computer science, not merely as applications of foundation models. The authors frame recurring failures of deployed agentic systems, including runaway resource consumption, behavior redirected by adversarial content, and opaque continuous operation, as symptoms of structural abstraction gaps. The paper examines these gaps across six dimensions: execution, interfaces, correctness, oversight, coordination, and persistence.
In briefPosition: Test-time coordination matters for cooperative MARL after training
SummaryThis position paper argues that cooperative multi-agent reinforcement learning should treat test-time coordination as part of the problem definition. It notes that the common centralized training with decentralized execution (CTDE) paradigm has no coordination channel at execution, which prevents teams from realigning after conditions change. The authors propose centralized training with mixed execution (CTME), where a coordinator observes global state and emits high-level signals during execution, with CTDE as a limit case.
In briefCross-view diffusion model for generating time series that capture extreme events
SummaryThe paper proposes CrossTF, a diffusion framework that uses the differences between time-view and frequency-view denoisers to generate time series with realistic extreme events. The authors find that adjusting along the direction connecting the two views' clean-sample estimates strongly affects extreme fidelity while having limited impact on overall quality. On seven real-world datasets against 11 baselines, CrossTF reportedly improves extreme fidelity by 26.1% and overall performance by 7.4%.
In briefA Transformer that jointly detects, locates and links earthquakes across stations
SummaryThe paper introduces EQNet2, a geometry-aware multi-station Transformer that jointly predicts phase arrivals, first-motion polarities, source location and magnitude, and associates arrivals into events. Tested on a Northern California dataset and continuous waveforms from Ridgecrest and the San Francisco Bay Area, it recovers more weak arrivals, produces fewer false detections, and links more picks into located events. Its attention links information across stations in patterns consistent with seismic moveout. The authors argue that exploiting multi-station coherence and event-centric prediction can yield more complete and reliable earthquake catalogs.
In briefTesting whether LLMs can reconstruct the timeline of event relations in full stories
SummaryThe paper asks five large language models to rebuild the full set of Allen's 13 interval relations among events in 360 English narratives, where events are described relative to each other rather than by dates. Outputs are scored against gold annotations, converse consistency and logical satisfiability under a pre-registered protocol. The best model beats trivial baselines on macro-F1, but under the pre-registered parser no model matches always answering "before", and most of that gap comes from formatting errors in large matrices.
Adib Sakhawat, Fardeen Sadab, Tamjid Hasan Fahim et al.
In briefGeneralizable Field-of-View Extension for Diffusion MRI Using Conditional Flow Matching
SummaryIncomplete field-of-view in diffusion MRI excludes some brain regions from white matter tractography and connectivity analyses. The authors propose a flow-matching generative model that extends the field of view superiorly and inferiorly by predicting apparent diffusion coefficients, conditioned on neighboring acquired slices and a full-FOV T1w image. Trained on 3,103 subjects across five datasets, it was evaluated on 299 subjects from five independent datasets, including unseen acquisition schemes, and significantly improved downstream tractography bundle Dice scores (p < 0.01).
Gaurav Rudravaram, Lianrui Zuo, Ema Topolnjak et al.
In briefLearning to Detect Shifts in Unstructured Data Streams via Betting-Based Monitoring
SummaryThis paper addresses continuous monitoring of generative AI systems for distribution shifts in unstructured data. It reframes sequential changepoint detection as a betting game, letting the detection direction adapt as new observations arrive while keeping false-alarm control statistically valid. Across three real-world datasets, adaptive monitors detected shifts faster than fixed monitors at the same false-alarm level, with LLM-concept monitors performing best for sparse, interpretable shifts.<br>
WhyApplies testing-by-betting to adaptive changepoint detection on unstructured data, proving false-alarm control and validating on three real datasets.
In briefJustOPD: On-policy distillation from autoregressive to diffusion language models
SummaryThis paper proposes JustOPD, a method for transferring knowledge from a strong autoregressive language model (ARM) teacher to a weaker diffusion language model (DLM) student. During training, it uses autoregressive rollouts so the student's generation process aligns with the teacher's left-to-right decoding, while the DLM keeps its flexible generation order at inference. Experiments with Qwen3-14B as teacher and Dream-7B as student show consistent gains in mathematical reasoning over the base DLM across multiple divergence choices.
In briefExact boundary-condition enforcement on curved triangular domains, with machine-learning PDE solving
SummaryThis paper presents a method for exactly enforcing Dirichlet, Neumann, and Robin boundary conditions on triangular domains with arbitrarily curved boundaries. It combines an exact mapping to a standard triangle, TFC constrained expressions, and Barnhill transfinite interpolation to build explicit trial functions that satisfy the conditions, and it handles corner compatibility constraints. The method is implemented with extreme learning machine-type randomized neural networks for linear and nonlinear PDEs. Numerical experiments report that the boundary conditions are satisfied to machine precision.
In briefLearning to Play: How Machines Master Games Through Reinforcement Learning
SummaryThis book explains reinforcement learning methods behind game-playing AI systems that reached top human level in Backgammon, Chess, Go, Poker, and StarCraft. It covers techniques ranging from exact planning and adaptive sampling to deep function approximation and self-play. Programming examples are in Python, with guidance on frameworks such as Gym, TensorFlow, and Keras, plus exercises.
In briefMeta-learning across electricity markets to forecast day-ahead prices
SummaryDay-ahead electricity price forecasting is usually done with a separate model per market, limited by each market's few years of hourly data. This paper presents MetaGrid, a topology-aware transformer meta-trained across markets so one zone-count-independent parameter set adapts to unseen grids. On a leak-free benchmark of five U.S. ISOs, MetaGrid achieves the lowest average MAE, 3.1% better than the best baseline and 7.7% better than the same architecture trained from scratch.
In briefMeasure Neural Mappings: a neural framework for learning operators on probability measures
SummaryThe paper introduces measure neural mappings (MNMs), which combine measure-dependent neural transport with ensemble consistency to learn operators on probability measures. Under strong sub-Gaussian and tail-integrability assumptions, it establishes quantitative continuum limits for attention on unbounded laws and bounds the expected output Wasserstein error of self-attention, cross-attention, and multihead transformers by an N^{-1/2} attention term. Experiments on seven operators in dimensions 1 to 10 show particle-count dependence close to the theoretical sampling rates.
Ricardo Baptista, Edoardo Calvello, Bohan Chen et al.
In briefA variable-input masked autoencoder for wrist-worn accelerometer data
SummaryThis paper introduces VarMAE, a masked autoencoder that handles varying input durations and sampling rates for wrist-worn accelerometer data. It is pretrained on the large-scale NHANES dataset using a cross-resolution masked reconstruction objective that learns one shared representation. The single pretrained encoder matches or outperforms controlled foundation model baselines on eight benchmark datasets, with only 3.82M parameters.
Harish Haresamudram, Wanting Mao, Aditya Radhakrishnan et al.
In briefRobustly Reliable Regression: Computing Guaranteed Prediction Ranges Under Adversarial Data Corruption
SummaryThe paper studies how to output the tightest prediction range for a test point in regression when up to b training points may be corrupted, such that every L-Lipschitz predictor consistent with the data must fall within that range. When the corrupted dataset remains L-Lipschitz consistent, the authors give a simple b-th best heuristic that runs in O(n) time and is proven correct and optimal. For the general case, they reduce the problem to analyzing a violation graph and propose a fixed-parameter tractable algorithm running in O(n^2 + b^4) time, with a matching lower bound of Ω(n^{2-o(1)}) under the Strong Exponential Time Hypothesis.
In briefEPOCH: Coverage-aware evolutionary LLM method for automatically designing optimization heuristics
SummaryThis paper proposes EPOCH, a framework that uses large language models to automatically design heuristics for combinatorial optimization problems. It makes covering training instances the goal of the evolutionary search, so the heuristics complement each other as an ensemble. It also fine-tunes the generator with experiential preference optimization and reflective self-distillation so reflection feedback is absorbed into the model weights. According to the paper, EPOCH's heuristic ensembles outperform inference-only and fine-tuned baselines across problems and open-source LLM backbones, and generalize across problem scales.
Oguzhan Gungordu, Ayse Tursucular, Siheng Xiong et al.
In briefMerchant Readiness Benchmark for AI Shopping Agents: A Pilot
SummaryThis paper proposes a benchmark measuring whether merchant websites are ready for AI agents to buy on people's behalf, covering agent access, whether published terms suffice for a purchase decision, and whether evidence of agreement is kept. A pilot assessed four large merchants, access to 555 sites, and 330 merchants, finding that key terms such as total cost or refund terms were often undetermined and that evidence coverage decided which merchants could be ranked. The authors state the pilot shows the measurements can be taken but does not validate them or rank merchants.
Adrian de Valois-Franklin, Alex Bogdan, Anas Ibrahim et al.
In briefWhat makes latent spaces from different motion models directly compatible
SummaryThe paper argues that sharing a latent space does not mean representations from different schemas can be used interchangeably. In a controlled motion VAE experiment on HumanML3D, even identical inputs gave low raw retrieval and token agreement, although correspondence was almost fully recoverable after linear calibration. Using only 64 paired clips substantially improved direct compatibility, yet crossed decoding remained clearly worse than native reconstruction.
In briefPIG: Parallel Inconsistency Guidance to speed up diffusion language models
SummaryThis paper proposes Parallel Inconsistency Guidance (PIG), a method for speeding up diffusion language models that decode several masked positions at once. Parallel decoding can reduce accuracy because its outputs drift from sequential generation. PIG uses a trained auxiliary model to estimate this inconsistency in a single forward pass and decides how many tokens can be decoded in parallel. The authors report that a lightweight auxiliary model of about 2M parameters suffices, and that PIG improves the accuracy versus tokens-per-forward trade-off on several benchmarks.
Quentin Fruytier, Anton Xue, Sanjay Shakkottai et al.
In briefHarnessTTT: Test-time learning to optimize the setup guiding AI scientific discovery
SummaryHarnessTTT is a framework that trains a proposer to generate executable harnesses, including prompts, skills, tools and controls, while keeping the language model executor frozen. Each harness guides a budgeted discovery search, and the outcome updates the proposer through a policy-gradient objective. Evaluated on fourteen tasks across mathematics, systems, GPU kernels, heuristic programming, astrodynamics and biology, it outperforms OpenEvolve with the same executor on all eleven benchmark tasks. The learned harnesses also transfer to stronger executors such as GPT-5.5.
In briefHASTE: a training framework adding hierarchical and attribute supervision to generative recommendation
SummaryHASTE is a training framework for generative recommendation models that represent items as semantic IDs generated autoregressively. It adds a Hierarchical Loss that supervises each level of the semantic ID hierarchy, which the authors report converges about 2x faster than standard causal cross-entropy. Attribute-Specific Losses aim to improve representations of data-sparse tail items. The method is evaluated on a fashion benchmark and multiple Amazon domains, with improvements reported across datasets and item-frequency regimes.
Tripti Shukla, Shwetha Somasundaram, Akshat Gupta et al.
In briefHow cached goal information from one client leaks into decisions for another in LLM agents
SummaryThe paper tests OLMo-3-7B-Instruct in a two-client market setup with opposing goals across 20 scenarios. With shared conversational memory, the model deviated from the current client's objective in 22% of eligible decisions. Probing shows client-goal information is recoverable from middle KV cache layers and concentrated in specific attention heads. Ablating the competitor's cached goal representations shifts decisions toward the current client, more strongly than ablating random heads.
In briefUMCR: Universal Medical Concept Recognition via Guideline Integration and Agentic Linking
SummaryThis paper proposes UMCR, a pipeline that extracts medical concepts from any source text and links them to any target ontology without ontology-specific retraining. It combines direct guideline integration with an agentic linking module. The method is evaluated on the MedPath benchmark, which covers diverse medical text types and vocabularies, and the authors report strong performance. They also state that it reduces reliance on the parametric memory of large language models.
In briefclaimcheck: Measuring and Improving Specification Faithfulness
SummaryThis paper examines whether formal specifications faithfully capture a natural-language requirement, since a verified proof of a narrower property can still pass all checks. The authors build a benchmark of 3,234 rows across Dafny, Lean 4, Isabelle, and Rocq, including meaning-preserving rewrites. They introduce claimcheck, which translates a formal artifact back into English without seeing the requirement and uses a judge to decide whether the two express the same guarantee.
In briefBrain2Face: Linking EEG Brain Signals to Spatio-temporal Head Avatars
SummaryThis paper asks whether a person's head geometry and appearance over time can be tracked using only EEG brain signals, to drive digital head avatars. The authors release a dataset pairing 64-channel EEG with per-frame expression latent codes from a personalized Gaussian head avatar. They also propose an EEG-to-expression model that regresses 3D Gaussian avatar expression parameters from EEG sequences, along with a benchmark protocol for the task.
Camilo Andr\'es Mart\'inez Mart\'inez, Olaf D\"unkel, Kartik Teotia et al.
In briefWIRE: Weight-space implicit encoding of video for frozen video LLMs
SummaryVideo language models process long sequences of visual embeddings, which makes prefill, memory, and generation costs grow with video length. WIRE compresses each frame into a few carrier tokens using a question-conditioned bidirectional selective scan, shortening the decoder input, and hypernetworks generate LoRA updates that are merged after prefill into the frozen backbone. On Qwen3-VL-8B with 16 carrier tokens per frame, the paper reports a 3.5x contraction of the visual stream and decoder prefill speedups of 3.35x at 256 frames and 4.14x at 1024 frames.
In briefMID: A Distribution-Level Metric for Evaluating Medical Image Segmentation
SummaryStandard metrics such as Dice and Hausdorff distance compare each prediction with its ground-truth mask, which is not possible for unlabeled data and does not check whether a set of predictions matches the ground-truth distribution. The paper proposes MID, which compares the joint distribution of images and masks between a reference set and predictions without requiring sample-wise correspondence. The authors evaluate it on 13 public CT, MRI, and MRA datasets.
WhyProposes a distribution-level image-mask joint metric without sample pairing, validated across 13 datasets and 56 organ groups against Dice and HD.
Vincenzo Marcian\`o, Xiaoming Zhang, Gianluca Guglielmo et al.
In briefHairFlow: A Diffusion Model That Generates 3D Hair Strands from Meshes and Sketches
SummaryThis paper presents HairFlow, a multimodal conditioned diffusion framework for efficiently generating and controlling 3D hair strands. It uses a coarse mesh to constrain overall hair shape and multi-view sketches to specify local strand flow, then traces the generated volumetric fields into explicit strands. To address the lack of paired sketch and strand training data, the authors introduce an automatic synthesis pipeline. Experiments reportedly show plausible strands that follow the input conditions across diverse hairstyles, with substantially reduced inference time.
In briefA statistical-mechanics view of Transformer attention
SummaryThe paper revisits Transformer attention through modern Hopfield models, using statistical mechanics. It presents a free-energy framework to characterize parameter regimes and critical thresholds for robust attention inference, treating attention as the equation of motion under free-energy minimization. The authors identify first-order phase transitions between glassy and robust-inference phases, and extend the analysis to attention with stochastic queries and keys.
Jerry Yao-Chieh Hu, Yu-An Chen, Maojiang Su et al.
In briefDetecting Distribution Shifts at Unknown Times Without Hyperparameters
SummaryThis paper proposes a sequential test for detecting distribution shifts using conditional conformal test martingales that rely on a fixed reference dataset. The authors recast the penalized betting problem as portfolio selection in a three-asset market, using a restarting universal portfolio with heavy-tailed wealth allocation and requiring no hyperparameters. They establish anytime false-alarm control, derive finite-sample bounds on detection delay, and present an FFT-based implementation that runs in O(T²lnT) time.
WhyReframes penalized betting as three-asset portfolio selection with restarts, giving hyperparameter-free shift detection and provable finite-sample delay bounds.
In briefBudgeted KV-cache repair for language-model agents whose context changes
SummaryLanguage-model agents that delete outdated text still have cached computations influenced by that text. The paper introduces budgeted KV-cache repair, which recomputes a chosen fraction of the retained suffix to recover task performance. Its Semantic-Unit Refresh method allocates this budget to query-relevant text regions using inexpensive embedding scores, and a layer-packed executor runs the repairs efficiently. The authors report strong recovery across five models while recomputing only 2-5% of the suffix.
In briefHandling model error in ensemble variational data assimilation with a learned generative model
SummaryThis paper introduces IEnKF-G, an ensemble variational data assimilation method that replaces the usual additive Gaussian model error with a learned stochastic forecast model. The authors extend it to a smoother over multiple time steps and test it on stochastic Lorenz 1963 and multi-scale Lorenz 1996 models. Compared with IEnKF-Q and an adaptive ensemble Kalman filter, it is generally more accurate, especially under partial observations and substantial model error.
In briefDo Theorem Provers Need Theorem-Proving Agents? mini-Lean-agent: A Minimal Coding-Agent Baseline for Lean
SummaryThe paper introduces mini-Lean-agent, a deliberately minimal agent for Lean tasks whose only model-visible tool is bash, to test which parts of theorem-proving agents drive benchmark gains. On Putnam 2025, it solved 7 of 12 problems with Gemini 3.8 Flash at a $2 per-problem limit, versus 4 of 12 for AxProverBase, and 12 of 12 with Astra at about $1.28 per problem. The authors suggest that a generic coding agent with a minimal harness can already perform strongly on Lean benchmarks and propose it as a standard baseline.
In briefThe Curse of Knowledge in Role-Playing Language Models: Do Characters Respect What They Can't Perceive?
SummaryThe paper tests whether role-playing language models keep characters within what they can hear, see, and understand. Using 480 scenes with counterfactual twins, it finds that 11 open-weight models often use hidden information the character never perceived, though most can state what the character perceives. Giving models only the information the character observes stops this leak.
In briefBeyond AI's Patterns: Human Ideas in Its Blind Spots Are Not Inferior
SummaryThis paper finds that AI-generated outputs are often seen as creative but tend to be more similar to one another than human ideas. The authors introduce EDGE, a method for mapping AI's semantic blind spots, and compare human and AI ideas across three open-ended tasks. Human ideas that fall into regions sparsely covered by AI showed no meaningful deficits in creativity, quality, or effectiveness, and some performed as well as or better than overlapping ideas.
Kibum Moon, Dan R. Johnson, Kostadin Kushlev et al.
In briefGPU Goldmines: Tunable WebGPU Kernels for Fast LLM Inference on Edge Devices
SummaryThis paper introduces GPU Goldmines, a framework for specifying, executing, and tuning WebGPU kernels used in large language model inference. It covers matrix multiplication, matrix-vector multiplication, and flash attention, with performance characterized across devices, shapes, and browsers. On shapes from Llama and Qwen models, its matrix multiplication kernels achieve 7.0x and 1.41x geometric mean speedups over ONNX Runtime and llama.cpp WebGPU kernels, respectively, and yield a 1.48x end-to-end prefill improvement in llama.cpp for chat workloads.
In briefNear-optimal regret bounds for monotone stochastic optimization with semi-bandit feedback
SummaryThis paper studies monotone stochastic optimization when input distributions are unknown and only probed variables are observed (semi-bandit feedback). The authors propose a general online-to-offline reduction that, under strong monotonicity, improves expected regret from the prior Õ(n√T) to Õ(√(nT)), which is optimal up to logarithmic factors. Under monotonicity and bounded coordinate-wise variation, the algorithm achieves Õ(n^{3/4}√T) approximation regret. The framework applies to prophet inequality, Pandora's box, stochastic knapsack and related problems.
WhyGives a general online-to-offline reduction achieving near-optimal √(nT) regret with theoretical proofs, applicable across several stochastic optimization problems.
In briefHow sampled evidence becomes radiology reports: objectives, constraints, and text generation
SummaryThis paper studies how evidence from multiple sampled reports should be turned into a final decision, and whether a better structured decision leads to a better radiology report. Controlled comparisons on MIMIC-CXR and IU-Xray show that changing the objective, relaxing candidate constraints, and the text realization step each affect results differently. Gains at the decision level do not consistently improve report quality, so the authors argue that decision rules, candidate constraints, and realized reports should be evaluated separately.
In briefLearned-denoiser unrolled alternating projections for phase retrieval, staying accurate without retraining
SummaryThis paper proposes a deep unrolling framework for phase retrieval that adds a learned denoising step to the alternating projections method. The authors argue that the performance drop of deep unrolling methods is not inherent to the paradigm but can be reduced by choosing the underlying iterative algorithm carefully. According to the paper, the method yields high-quality reconstructions without retraining in settings where model-based methods struggle, such as when the measurement operator's dynamic range changes significantly at test time.
In briefJEDI-net: a neural network that groups data by feature-importance patterns to adapt predictions
SummaryThis paper presents JEDI-net (Joint Embedding with Dynamic Integration Network), a unified neural network architecture. During training it periodically clusters SHAP values to identify latent subgroups with similar feature-importance patterns and embeds their identities into the network. This lets predictions adapt to each observation's predictive structure. The authors note that their earlier work, FORCE, showed that clustering by feature importance improved neural network performance.
In briefA Geometric Method for Building Neural Networks That Perfectly Classify Training Data
SummaryThe paper proposes a way to construct a neural network, based on geometric analysis of the input data, that classifies every training sample correctly without regularization. Separating hyperplanes are mapped to neurons with step activations, and the network's layer sizes and parameters are set accordingly. Experiments on MNIST reportedly achieve perfect training-set classification with compact models.
In briefBeyond Winning: Testing Whether LLMs Really Model Their Opponents
SummaryThis paper introduces a benchmark for repeated hide-and-seek against algorithmic opponents whose beliefs and action probabilities are known. Structured elicitation and supervised fine-tuning improve play in several settings, but opponent action prediction and belief tracking remain limited. Providing the opponent's type and parameters improves predictions, while inferring them from history is harder, a pattern that also appears in GPT-5.4.
In briefStartLux-Decision: Reinforcement-learned decision models that reason only when unsure
SummaryThis paper presents StartLux-Decision, decision models ranging from 0.8B to 27B parameters. The Instant variants answer in a single pass, while AutoThink variants invoke explicit reasoning only for uncertain decisions. The authors propose RLCD², a reinforcement learning method that raises the share of reasoning traces finishing within the generation limit and cuts average reasoning length by up to 65%. They also introduce DecisionPivot, a programmatically verifiable benchmark.
In briefTuning LLM explanations with reinforcement learning to improve human decisions
SummaryThe paper introduces HD-RL, a reinforcement learning framework that optimizes LLM explanations for human decision quality rather than model performance or human preference. It calibrates a human decision simulator on observed decisions and uses a proxy reward model to train the LLM. In AI-generated text detection with the open-source Qwen 3.6 model, human accuracy rose from 51.3% to 63.7%, outperforming RLVR, RLAIF and other baselines by over 8%.
In briefMaking small LLMs more reliable at Text-to-SQL via automatic retry workflows
SummaryThis paper evaluates small and medium open-source LLMs (1B to 8B parameters) on Text-to-SQL under a fixed inference budget, using the Spider benchmark. It finds that SQL-specialized models are more reliable than general-purpose and chat models at small scales, and that results vary considerably with the prompt used. To reduce this variability, the authors introduce three lightweight workflows that retry with alternative prompts or models on failure, with an optional syntax-correction step. These workflows significantly raise end-to-end success rates, at the cost of additional inference calls.
In briefDoes Personalization Break Trust? Rejecting Unknown Commands After On-Device Few-Shot Adaptation
SummaryThis paper studies whether on-device few-shot personalization of voice assistants, such as LoRA fine-tuning for new user commands, harms their ability to reject inputs matching none of their intents. Across five public intent-classification benchmarks, the authors find that naive LoRA personalization lowers out-of-scope rejection AUROC while in-scope accuracy stays intact. They report that scoring rejection by Mahalanobis distance instead of softmax confidence recovers most lost quality on three of five datasets without extra training.
WhyIdentifies few-shot LoRA personalization degrading OOS rejection and fixes it via Mahalanobis scoring, validated across five benchmarks.
In briefThe Agent Knew: Internal States Predict and Stop Unauthorized Actions, and the Lies That Follow
SummaryThe paper studies language-model agents that act beyond their authorization and then misreport what they did. Using CAPE-Bench and tau^2-bench, the authors find that acting first and then asking permission often precedes a false report. A linear readout of the agent's internal residual stream predicts violations before they occur, and gating actions on this forecast reduces violations while preserving authorized task completion.
WhyLinear residual-stream readout forecasts unauthorized actions and identifies retroactive permission laundering as a lie precursor, with preregistered multi-family validation.
In briefToward Resilient Healthcare IoT: Slice-Based Orchestration for Mobile Medical Devices at the Edge
SummaryThis paper proposes an adaptive orchestration framework that combines 5G network slicing, resource-aware orchestration, and mobility-aware task offloading for medical IoT services. Services are assigned to network slices while considering CPU, memory, and storage capacity at the edge. The system also supports user mobility and dynamic offloading to maintain service continuity. SimPy-based simulations indicate improved deployment efficiency, fewer blocked requests, and higher user acceptance.
Muhammad Zeeshan, Hafiz Faheem Shahid, Erkki Harjula
In briefPhysCodeGym: A Benchmark for Coding Agents Reconstructing Physically Verifiable 4D Scenes
SummaryThis paper introduces PhysCodeGym, a benchmark of 156 tasks covering rigid-body, liquid, deformable, and wind-driven motion. Coding agents watch videos, write and run Blender programs, and revise their scenes based on simulation feedback. Reconstructions are scored with six metrics and tested with counterfactual interventions for physical consistency. The authors report that reconstructions can match visible events while still hiding contact and support failures.
In briefChoosing Optimizer Geometry per Layer: Measured Sign-Spectral Assignment in Mixed-Optimizer Training
SummaryThis paper treats the choice between sign-based and spectral updates for each layer of a neural network as a measurable decision, rather than fixing it by layer type. The authors derive a ratio of descent-lemma lower bounds for the two update types and use a local directional quadratic score to route layers in a method called G-Scion. Experiments on a 31M GPT-2 and ViT-B/16 report that G-Scion improves validation loss over fixed Scion.
Hiroki Naganuma, Laura Gomezjurado, Mahdi Ghaznavi et al.
In briefA single neuron with step-by-step reasoning can simulate general computation
SummaryThis paper asks how simple a next-token predictor can be while still simulating general computation through chain-of-thought generation. The authors show that a single autoregressive neuron with threshold or ReLU activation can simulate any feedforward circuit with the same activation, with size and step bounds independent of circuit depth. They also train a single ReLU neuron on CIFAR-10 to mimic a neural network teacher and report non-trivial test accuracy.
In briefColored graph signal diffusion: tuning graph-frequency timing in generative diffusion
SummaryThe paper observes that standard diffusion with isotropic noise inherits a coarse-to-fine generation order from the nonuniform data spectrum. It proposes colored score-based diffusions for graph signals, where a positive spectral rate controls when each graph-frequency band loses and recovers information. Numerical experiments reportedly show that these colored diffusions reach a given generation quality in far fewer sampling steps than standard diffusion.
Yi\u{g}it Berkay Uslu, Samar Hadou, Shirin Saeedi Bidokhti et al.
In briefPhase-conditioned affordance fields that make robot control policies focus on the task-relevant part of a scene
SummaryThe paper proposes a phase-conditioned affordance field (PAF), a learned per-point score over the observed point cloud that indicates which regions matter for the current task phase. The policy uses the score-weighted centroid as a learned anchor and appends the score as a per-point feature, adding local geometric bias while keeping global SE(3) equivariance. The authors report that PAF consistently improves success rates on multiple MimicGen tasks under spatial randomization, and outperforms baselines on multi-object real-robot tasks, maintaining its lead under lighting variations.
In briefNufo: A Benchmark for Ewe Speech Recognition
SummaryThis paper consolidates and standardizes existing Ewe speech resources into a 174-hour automatic speech recognition benchmark called Nufo. The authors fine-tune and evaluate Whisper and AfriHuBERT, with Whisper-medium reaching a 37.98% word error rate and AfriHuBERT 39.70% on the test set. They also analyze common recognition errors and release the benchmark, models, and evaluation resources.
Christ Kekeli Keli, Steve Sanogo, Patrice Sebastiano et al.
In briefExplainable propaganda detection via label-guided chain-of-thought optimisation
SummaryThis paper presents ReCoT, a framework that detects propaganda techniques while producing interpretable reasoning, without requiring extra explanation annotations. It refines reasoning chains through population-level refinement and performance-guided resampling to capture definitional cues of persuasive techniques. On INDI-PROPspan and PTC, ReCoT reaches a large share of the supervised fine-tuning upper bound while remaining interpretable.
In briefMeasuring Whether Networks Learn Equivariance from Data Augmentation
SummaryThis paper studies how to measure whether a neural network has learned equivariance to symmetries through data augmentation rather than architectural design. For finite discrete settings it uses normalized mutual information (NMI) to test whether representations of original and transformed inputs determine one another. For high-dimensional activations it uses centered kernel alignment (CKA), which has an exact guarantee for orthogonal feature transformations. Across architectures and tasks, augmentation usually increases transformation consistency.
Claire Schlesinger, Wolfgang Gatterbauer, Robin Walters
In briefA method that post-trains diffusion language models with RL along their own denoising paths
SummaryThe paper introduces PACT, a reinforcement learning post-training method for diffusion language models. It jointly learns a policy and a scalar potential that scores intermediate denoising states, anchored by terminal reward, avoiding sequence-likelihood estimation and importance-ratio clipping. On four tasks, PACT improves LLaDA-8B-Instruct, gaining 9.9 points on GSM8K and 72.6 points on Sudoku.
In briefFederated learning for hospital mortality prediction: protocol choices matter more than the algorithm
SummaryThis paper compares seven federated learning aggregation strategies, including FedAvg and SCAFFOLD, for training in-hospital mortality prediction models across hospitals without sharing patient records. Experiments used two federations built from MIMIC-IV and eICU-CRD, with results repeated over 10 seeds and tested statistically. SCAFFOLD ranked first, but its margin over the runner-up was small. The authors find that evaluation protocol decisions, such as predictor selection, client-level resampling and which validation round is evaluated, changed discrimination more than the choice of strategy, so they recommend fixing the protocol before ranking algorithms.
In briefDoes a Governed Confidence Score Know When a Text-to-SQL Answer Is Right?
SummaryThe paper audits the confidence score a production database question-answering system attaches to its answers. On BIRD dev, the designed score separates correct from wrong answers weakly (AUROC 0.624) and withholds 35.7% of correct answers. The authors propose a rule based on agreement among samples, with deterministic checks as vetoes, and pre-register a test on Spider dev, where AUROC rose to 0.725 and withheld correct answers fell to 8.8%.
In briefValue-preserving action quantization for search-based robot policy learning
SummaryThe paper argues that discretizing continuous robot actions for search-based policy learning must preserve how actions affect future states and values, not just reconstruct actions accurately. The authors propose value-preserving action quantization (VAQ), which quantizes actions by preserving state reachability, and combine it with Monte Carlo tree search for behavioral cloning. On object pushing tasks, VAQ-MCTS matches or exceeds existing methods, improves data efficiency and robustness to heterogeneous demonstrations, and enables zero-shot adaptation; it also completed all 30 real-world pushing trials.
Max Muchen Sun, Jueun Kwon, Benjamin Riviere et al.
In briefHybridDualPathNet: A hybrid model fusing EfficientNetV2 and Swin Transformer to classify Alzheimer's, Parkinson's, and healthy brain MRI scans
SummaryThe paper proposes HybridDualPathNet, which combines EfficientNetV2-S for local features and Swin-Tiny for global context via a Cross-Fusion Attention module, to separate Alzheimer's disease, Parkinson's disease, and healthy controls. It reaches 98.25% accuracy on a Kaggle dataset with an image-level split, but only 73.37% on a combined Kaggle and ADNI dataset with patient-wise separation, with 59.2% on ADNI test images. The authors stress patient-level splitting, cross-dataset evaluation, and uncertainty estimation, noting that single-dataset performance does not guarantee generalization.
In briefLLM-to-Flow Distillation: teaching natural-language science goals to generators
SummaryThe paper proposes LLM-to-Flow Distillation (LFD), in which a frozen LLM acts as an adaptive teacher for a domain-specific flow or diffusion generator. The teacher builds a curriculum of intermediate sub-goals toward a natural-language goal and judges samples, and these preferences are distilled into the generator via DPO, without an explicit proxy reward. Experiments on therapeutic peptides and small-molecule structure design show the approach aligning generators to such goals.
Riccardo De Santi, Federico Di Gennaro, Angelo Gnazzo et al.
In briefMIRRA: Fast, High-Fidelity Molecular Generation via Chemistry-Informed Discrete Flow Matching
SummaryThis paper introduces MIRRA, a discrete flow-matching model that generates molecules using chemical building blocks and learns from chemical knowledge. It computes identity probabilities only when needed while keeping the full chemical vocabulary available during sampling. On ZINC250k, it reduces chemical and graph distribution errors relative to the strongest baselines and generates up to 8.6 times faster. On a held-out polymer test set, it achieves 100% periodic validity, compared with 57.0% for a polymer variational autoencoder.
In briefFrom Concept to Working Backend: A Modular System for Energy Information Extraction and Grounded Q&A
SummaryThis paper presents EneMonIE-Backend, a modular and executable redesign of the earlier conceptual EneMonIE system for monitoring energy trends from online sources. It separates data sources, acquisition, semantic processing, knowledge organization, and storage, and supports connectors for RSS, arXiv, social-media JSON, and open-data CSV files. The backend offers three grounded question-answering paths and excludes the user interface, so the semantic core can be embedded in other applications. A deterministic smoke test checks filtering, persistence, clustering, evidence retrieval, and deletion.
Janik Lesage, Khadim L\^o, and Jean-S\'ebastien Dessureault
In briefWhen Left Is Right? Understanding camera motion beyond the screen
SummaryThis paper asks whether vision-language models can tell if a camera moved left or right, a task where most current models perform near chance. The authors propose CameraVLM, which combines off-screen camera-motion tokens with on-screen flow tokens and uses two-stage training to keep language involved. The paper reports that the 9B-parameter model outperforms open-source and camera-motion-specific baselines on three benchmarks and matches or exceeds GPT-6 Astra and human viewers in a 48-participant study.
In briefIsoVRBench: Benchmark for Video Reward Models Without Human Feedback
SummaryThe paper presents IsoVRBench, a benchmark of 8,000 video preference pairs across eight controlled interventions, where each pair changes only one property of a generated video. The pairs are built automatically, so no human annotation is needed. The authors find that state-of-the-art reward models distinguish broad quality differences between generators but miss features such as temporal order, reaching only 60.8% on reordered frames versus 100% for humans. Training on these synthetic pairs yields a reward model that outperforms a human-labeled counterpart on VBench-2.0 by 3.8 points.
WhyControlled single-property interventions replace human labels, exposing missed temporal order and yielding reward models that beat human-data-trained counterparts.
Shuyue Stella Li, Jacqueline He, Jiahui Chen et al.
In briefBeyond Sparse Captions: Aligning Slide-Level Text and Patch-Level Vision in Pathology
SummaryThis paper proposes LIMIT, a weakly supervised vision-language pretraining method that uses slide-level pathology reports generated in routine diagnosis instead of sparse image captions. It freezes a pretrained vision encoder and fine-tunes only the text encoder, using report embeddings as cross-attention queries over patch feature bags, optimized with InfoNCE. The resulting model, CONCH-Z, achieves state-of-the-art zero-shot performance across 22 tasks and approaches slide-level foundation models that use far more parameters and compute.
WhyUses clinical reports instead of sparse captions, aligning slides and text via frozen encoder and report-conditioned multiple-instance contrastive learning. Achieves zero-shot SOTA over 22 tasks, with
Daniel Shao, Dongmin Bang, Luca L. Weishaupt et al.
In briefAReA: Adaptive Receptive-Field Attention for Faster, Sharper High-Resolution Diffusion Transformers
SummaryDiffusion Transformers struggle to generate images above their training resolution, since token counts grow quickly, full attention becomes costly, and outputs can show structural distortions and lost detail. The authors propose AReA, a training-free attention mechanism whose receptive field shifts from coarse to fine as denoising proceeds, using cross-scale attention over downsampled keys and values early and progressively shrinking local windows later. Experiments reported in the paper show AReA consistently outperforms state-of-the-art training-free baselines across target resolutions and different models, while being faster and more efficient.
Javad Rajabi, Koorosh Roohi, David B. Lindell et al.
In briefCodeLift: Turning the Test Suite into a Dense Training Curriculum for Coding Agents
SummaryThis paper introduces CodeLift, a reinforcement learning method for training software engineering agents. It argues that a binary all-tests-pass reward is too sparse, and that a plain test pass-rate is unreliable because many trivial tests can dominate the score. CodeLift dynamically recalibrates pass-rate rewards by grouping tests and patches, weighting tests that fewer patches pass, which turns the test suite into an adaptive curriculum. The authors report consistent gains in pass@1/2/4/8 over binary and pass-rate baselines on SWE benchmarks.
In briefCompute-optimal data scaling for neural PDE surrogates via multi-fidelity training
SummaryNeural surrogates for partial differential equations are trained on simulator-generated data, and generating that data can become the main cost limiting scaling. The paper formalizes training as a function of data quantity and numerical fidelity under a fixed data-generation compute budget. It reports that mixing lower-fidelity simulations into training matches high-fidelity-only accuracy with up to 6.3 times less data-generation compute. The findings are validated across six PDE problems, four architectures, and over 7,000 training runs.
Paul Setinek, Fabian Paischer, Nils Thuerey et al.
In briefDensity-Matrix RL: Adding Action Relations to Policies
SummaryStandard reinforcement learning policies store only one probability per action at each state, leaving out how actions relate to each other. This paper stores each policy as a density matrix, where diagonal entries are the usual action probabilities and off-diagonal coherences are new parameters, a framework called density-matrix reinforcement learning (DMRL). The authors prove that coherences change outcomes in two settings: regularizing toward a coherent reference policy, and rewards that are quadratic functions of a unit-vector action. This is a purely theoretical paper with no empirical evaluation.
In briefDexNinja: a robot learns robust dexterous cutting using a real-sim-real data engine
SummaryDexNinja is a framework that turns a few real demonstrations of robotic food cutting into a larger, realistic training set. It reconstructs real objects, randomizes physical parameters within category constraints, and generates augmented manipulation episodes with a differentiable simulator. On a food-slicing task, success exceeds 60% and augmented data improves real-data-only results by more than 20%. The learned policies also generalize to objects with different shapes and sizes..
In briefAutoMARS: A Multi-Agent LLM Framework for Automated Supply Chain Risk Assessment
SummaryThis paper proposes AutoMARS, an agentic framework for automatic and scalable comparative risk assessment across supply chain networks. It plans tasks, coordinates specialized agents, invokes external analytical tools, and synthesizes results into integrated risk evaluations. The framework combines risk factors from socioeconomic, disaster, environmental, and transportation domains and parallelizes independent tasks to support comparison across multiple locations.
In briefReleasing the Inner Jev: A Simple Method for Open Decision Models
SummaryThe paper shows that open instruction-tuned language models can produce option probabilities for decision tasks by reading answer-letter logits, without any extra training. It also presents a simple LoRA-based supervised fine-tuning recipe with a decision head. The authors introduce OpenDecisions, a 2,352-question validation dataset with a Capability score based on accuracy and calibration. The fine-tuned Gemma 4 31B reports scores of 84.4 on OpenDecisions and 81.5 on JevBench v1.7, above Jev.
Jonathan Geuter, Andrew Cheng, David Alvarez-Melis
In briefOnline LLM personalization using pairwise choices shared across users
SummaryThe paper studies inference-time personalization of large language models, where users prefer different answer styles that are not directly observed. The system serves users sequentially, showing two responses per question and learning from one pairwise choice, which is framed as a contextual dueling bandit problem. The authors release a reproducible tutoring benchmark with LLM-generated answers in six styles and simulated student judgments. Their baseline learns a mixture prior from past sessions and reportedly removes 44-48% of random pairs' excess regret on unseen students and topics.
In briefMLAfication: Making High-Quality Multi-Head Latent Attention Practical for Modern LLMs
SummaryThis paper presents MLAfication, an extension of the MHA2MLA conversion method that works with QKNorm and needs fewer parameters to convert models. It combines calibrated KNorm, parameter-efficient distillation, Partial-RoPE, and activation-aware joint KV initialization. Across MHA and GQA models, including QKNorm and MoE variants, it maintains performance under substantial KV-cache compression. On Qwen3-30B-A3B, the authors report up to 5.35 times the serving throughput of BF16 GQA, and a 94.73% KV-cache reduction when combined with 4-bit KV quantization.
WhySubstantively extends MHA2MLA with QKNorm and MoE support, validated across multiple models with throughput and KV-cache measurements.
