In briefShadowMiner v1: A Report on Building and Measuring a System That Finds Research Problems and Hypotheses
SummaryShadowMiner v1 is a nine-stage system that automatically discovers research problems and generates hypotheses from AI papers. It structures documents into a knowledge graph, identifies structural gaps in that graph, and feeds them into an LLM generation prompt. Each generated hypothesis is checked for prior coverage, scored for quality, and verified against its source facts. The report does not propose new techniques; it describes implementing prior ideas and measuring whether each one contributed.
In briefS³N: A Spherical Spiral Scanning Network for Weather Forecasting (a machine-learning global weather prediction model)
SummaryThis paper proposes S³N, a Spherical Spiral Scanning Network for global weather forecasting. It uses L2Proj to map atmospheric fields bidirectionally between latitude-longitude and HEALPix grids, and an AQSS block that propagates information along four pole-to-pole spiral paths, enabling continuous flow across HEALPix base-face boundaries. Experiments report better results at 4-, 7-, and 10-day lead times and slower error growth in long-range forecasts.
In briefFLAT: a method that smooths rugged error landscapes to calibrate traffic models with fewer simulations
SummaryThis paper presents FLAT, a method for calibrating microscopic traffic models used in digital twins, which requires costly full simulation runs. It uses an eight-dimensional behavioral fingerprint to smooth the parameter-error landscape and an annealed lower-confidence-bound acquisition to decide where to sample next. Across six real-world scenes under a matched budget, FLAT with a Gaussian process surrogate achieves the lowest average behavioral error. The authors attribute the gains to objective geometry and sequential allocation rather than surrogate capacity.
In briefA differentiable model linking 3D Gaussian Splatting with SPH for solid mechanics simulation and parameter inference
SummaryThis paper proposes a differentiable model that couples 3D Gaussian Splatting, a method for reconstructing object appearance from images, with smoothed particle hydrodynamics (SPH) for simulating solid deformation. The object is first reconstructed from calibrated multi-view images, then separate SPH particles are generated and their deformation is mapped back to the Gaussian positions and covariances. The framework supports several elastic and viscoelastic material responses and boundary conditions. Numerical studies compare it with finite-element results and show forward simulations and estimation of material and boundary parameters from rendered deformation observations, including noisy cases.
In briefSynthetic-to-Real ViT Pose Estimation for Non-Cooperative Drones
SummaryThis paper studies whether a Vision Transformer model trained only on synthetic images can estimate the pose of a non-cooperative UAV from real photos. The approach uses a self-supervised DINOv2 backbone, pose-ambiguity-aware training and inference strategies, and an α-β filter at inference. On a real-world set of 77,077 labeled UAV images, the mean angular error is 19.18° before filtering and 8.74° after, with inference times of 13.25 ms and 13.42 ms respectively.
Krishnanujam Srinivas, Hanish Acharla, Brij Agrawal et al.
In briefFactorized Delayed Streams Modeling (F-DSM) for LLM-based streaming speech recognition
SummaryDelayed Streams Modeling (DSM) lets large language models perform streaming speech recognition by aligning audio and text on a shared timeline. This paper shows the word-start token <w> can be removed without hurting recognition, and proposes Factorized DSM (F-DSM), which separates the waiting probability for <p> from the original vocabulary distribution. In experiments on the Corpus of Spontaneous Japanese and LibriSpeech, F-DSM outperforms DSM in recognition, greatly reduces GPU memory use with similar training throughput, slightly speeds up inference by skipping softmax, and lessens the text-only perplexity degradation seen with DSM.
Tatsunari Takagi, Kai Washizaki, Atsushi Kojima et al.
In briefWhen Robots Crash: Optimal Asynchronous Gathering at Weber Meeting Nodes
SummaryThis paper studies how asynchronous, anonymous, oblivious mobile robots on an infinite grid can gather at one of several designated meeting nodes when some robots may crash. The goal is to bring all non-faulty robots to a Weber Meeting Node while minimizing the total Manhattan distance traveled. The proposed algorithm, CrashTolerantWeberGathering(), lets each robot independently pick the same target and uses target-dependent restricted shortest paths so the target remains a Weber Meeting Node. The authors prove that optimal gathering is impossible from certain fully symmetric configurations under strong multiplicity detection, and guarantee finite-time gathering for all remaining configurations.
Animesh Maiti, Prakhar Shukla, Abhinav Chakraborty et al.
In briefSuppressing Pressure, Amplifying Evidence: Self-Guided Attention Steering Against Sycophancy and Stubbornness
SummaryThis paper examines two failures of language models: sycophancy, where models yield to unsupported user pressure, and stubbornness, where they fail to revise answers when context warrants it. The authors introduce CoPE-Bench, a six-condition benchmark for measuring both failures together, and propose SPAE, a training-free method that uses the model's own judgments to steer token-level attention away from user pressure and toward contextual evidence. Across five backbones, the authors report an 18.8 percentage point reduction in pressure following and a 5.5 point gain in joint-condition updating relative to the strongest baseline.
In briefBidirectional Preference Synthesis: Learning Prompt-Dependent Preferences from Boundary Failures
SummaryThe paper argues that correction-based offline preference training treats model failures only as rejected responses under the original prompt, which is incomplete. The authors propose Bidirectional Preference Synthesis (BPS), which for each validated boundary failure keeps the forward pair under the original prompt and adds a reverse pair under a synthesized achieved prompt, without changing the DPO objective or training a reward model. On Qwen3-4B-Instruct-2507, BPS keeps original-side pairwise ranking while raising achieved-side ranking accuracy from 6.8% to 62.3% on held-out crossed anchors.
In briefWhen Satellite and Street Views Disagree: Fusing Views for Disaster Damage
SummaryThis paper studies how satellite and street-view images complement each other when assessing building damage after disasters, focusing on cases where single-view models disagree. On paired data from the 2025 Eaton wildfire and Hurricanes Ian and Milton, fusion improves accuracy on these conflict cases over either view alone, but still falls well short of an oracle that picks the correct view. A linear fusion gate gives gains where ground photos are aimed at the building.
In briefLyapuFlow: Steering Generative Flows with Lyapunov Feedback for Inverse Problems
SummaryThis paper introduces LyapuFlow, a method that guides pretrained flow models at inference time to satisfy constraints without retraining. At each sampling step, it predicts the terminal sample the flow is heading toward and measures the constraint violation on that prediction. It then computes the minimum-norm control that meets a prescribed Lyapunov decrease condition, applying a corrective update only when needed within a trust region. The authors report results in data and latent spaces, outperforming alternative test-time constraint methods on scientific machine learning and image inverse problems.
Minseon Gwak, Hans Hao-Hsun Hsu, Danielle C. Maddix et al.
In briefUniPro: One model for 2D-to-3D medical image segmentation through propagation
SummaryThis paper presents UniPro, a unified model that handles semantic, in-context, interactive, and propagation-based medical image segmentation within one slice-based framework. It extends 2D segmentation to 3D volumes by propagating predictions from neighboring slices, and adds bidirectional and 3D supervision to improve propagation reliability. The authors report strong performance across diverse modalities and anatomies, with annotation-efficient 3D segmentation from sparse 2D initialization.
In briefQiYao-I: A Manifold-Based Foundation Model for Irregular Time Series Forecasting
SummaryThis paper presents QiYao-I, a foundation model for forecasting irregular multivariate time series, where observations are unevenly sampled and asynchronous across variables. It maps real timestamps into a learnable temporal manifold space and uses sampling-conditioned attention to handle irregular intervals. A frequency-aware mechanism selectively passes information between variables under asynchronous observations. Experiments on real-world benchmarks reportedly outperform existing models, with strong zero-shot and few-shot generalization.
In briefRethinking Long-Video Efficiency: Joint Allocation of Frames, Pixels, and Decoding Latency
SummaryThis paper examines how to make vision-language models more efficient on long videos. The authors find that at matched token budgets, dense low-resolution sampling beats sparse native-resolution sampling, and that front-end decoding dominates processing time for hour-long videos. They propose LoHi, a training-free single-pass framework that combines a dense low-resolution video stream with sparse high-resolution image frames. Across three long-video benchmarks, LoHi improves average accuracy by 10.6 points over the native-resolution baseline and reduces front-end decoding latency by up to 7x on hour-long videos.
In briefMeasuring Accessibility of Safety Content in Model Internals Versus Outputs
SummaryThis paper studies how safety-relevant content inside a language model can be read out through Jacobian-space maps, and how that accessible content relates to the latent-space safety geometry used in representation engineering. The authors propose a layer-wise transport-amplification profile and a paired base-versus-tuned protocol, finding that training can widen the gap between internal recognition and behavioral safety. A deployment audit tests monitoring, evasion, and training-time defenses, and the authors conclude that safety monitoring should target accessibility itself rather than outputs alone.
In briefMOIRA: Adaptive Sparse Attention for Faster Long-Context Decoding
SummaryMOIRA is a training-free sparse decoding method that adapts the number of cached memory pages read per request, layer and KV head, keeping the smallest set whose estimated attention mass reaches a fraction γ. A new self-planning attention kernel lets the whole decode step run inside a CUDA graph. On an H200 at RULER 128k context, γ=0.99 matches dense accuracy while reading about 30% of pages and cutting time per output token by 2.2–2.5×. Under high serving load, throughput rises by up to 51%.
In briefML-OPF-Bench: A Unified Benchmark for Machine Learning in Optimal Power Flow
SummaryThis paper introduces ML-OPF-Bench, a unified benchmark for AC and DC optimal power flow that evaluates representative machine learning methods under a consistent pipeline. The methods are stress-tested across system sizes, distribution shifts, and resource budgets, then ranked with a multi-objective framework. The authors find that prediction accuracy alone does not reliably indicate operational feasibility, and that under congested conditions feasibility largely comes from post-processing rather than the pure ML predictor. The benchmark is released as an extensible open-source Python package.
In briefComparing Methods for Classifying Job Experience Levels from Job Postings
SummaryThis paper studies how to automatically predict the experience level a job posting requires. The authors propose a Section-Aware BERT that encodes titles, responsibilities, and requirements separately before combining them, and compare it with rule-based systems, TF-IDF baselines, and large language models in few-shot and fine-tuned settings. Results show that explicitly using text structure improves prediction, especially when job titles are ambiguous. Error analysis highlights confusion between Entry and Senior levels and unclear boundaries for Mid-level positions.
In briefWhat to Preserve in Recursive Computation: A Local Predictive Sufficiency Principle
SummaryThe paper argues that intermediate states in recursive computation can lose information needed for later predictions as they are repeatedly compressed. It proposes local predictive sufficiency: controlling predictive deficiencies at each interface controls the discrepancy at the root. The authors turn this into a training procedure that projects parameter updates to preserve predictive information. Across temporal graphs, language memory, vision-language-action control, and recursive self-improvement, it matches or improves host models under matched compression budgets.
In briefEnvDreamer: Generating Large-Scale Simulated Environments for Embodied AI Training
SummaryThis paper presents EnvDreamer, a framework that uses large language and vision language models to generate Unreal Engine 5 environments for training embodied AI and robots. It can sample large, diverse, interactive, and validated virtual environments for navigation, interaction, and manipulation tasks. The authors report that policies trained on these generated environments achieve competitive results on several embodied benchmarks, and they release EnvDreamer-20k, a dataset of 20,000 validated environments.
In briefGeometric-transformation method for stable open-vocabulary segmentation of remote sensing images
SummaryObjects in remote sensing images keep their meaning under dihedral-group geometric transformations, yet SAM3-based open-vocabulary segmentation methods respond inconsistently to such transformations. The paper proposes a feature-adaptive manifold repair method that selects complementary transformed views, uses the original view as an anchor to repair unreliable features, and adapts the SAM3 pixel decoder online at test time. It reports an average mIoU of 55.6% across eight remote sensing segmentation benchmarks and consistent gains under different SAM3-based inference frameworks.
In briefQuestioning the Questions: Keeping Reasoning Models Improving Themselves
SummaryThe paper finds that reasoning models that learn from their own generated questions can degrade over repeated self-training. It attributes this to growing numbers of invalid questions and to mathematically equivalent duplicates that lexical-similarity diversity controls miss. The authors propose R-Quest, which uses validity and novelty feedback to guide self-evolution, and report the highest average performance across 12 benchmarks and stable gains over ten rounds.
In briefLatent Safety Filters: When a Lossy Encoder Admits a Transferable Certificate
SummaryThe paper studies latent safety filters that certify safety on a learned low-dimensional state representation. Because the encoder is lossy, a filter can report safe while the physical state is unsafe, without any detectable model error. The authors show that whether a safety certificate transfers to the physical system depends on the detectability of the discarded safety-relevant dynamics.
Johannes Mootz, Zahra Nili Ahmadabadi, Reza Akhavian
In briefRRM-GPT: A Framework for Radio Resource Management Foundation Models
SummaryThis paper proposes RRM-GPT, an autoregressive framework that generates radio resource management decisions field by field, much as a language model generates text. The model is pretrained on unannotated network logs and then adapted to specific operator goals through imitation or reinforcement learning. In a 5G New Radio case study, a single model generates complete scheduling grants covering user selection, timing, link adaptation, resource allocation, and control signaling. It also transfers learned behavior to an unseen scenario without adaptation.
Ahmed Aboulfotouh, Akram Bin Sediq, Koosha Pourtahmasi Roshandeh et al.
In briefLanguage-dependent token and reasoning efficiency in LLMs: a paired cross-lingual study protocol
SummaryThis paper presents a prospective paired study protocol that separates the effects of input language, observable reasoning-trace language, and answer generation on large language model costs. The design covers 240 items rendered in English and seven other languages, three open-weight checkpoints, several trace-token budgets, and pre-specified statistical tests. The protocol measures tokens, correctness, latency, memory, and energy. No confirmatory model results are reported.
In briefLMBuild: Testing whether AI agents can design structures that can actually be built and work
SummaryThis paper introduces LMBuild, a benchmark for checking whether LLM agents can generate 3D structures that can be physically built and perform their intended functions, not just look good geometrically. It provides an interactive environment where agents use tools to retrieve, create, and place parts, along with a benchmark built from CAD datasets augmented with Wikipedia knowledge. Tests on 30 systems found that functional affordance and physical operability remain much harder than structural soundness and alignment.
In briefSparse causal discovery using local Fisher information, without high-dimensional density estimation
SummaryThis paper introduces FiCS, an algorithm for sparse causal discovery in additive noise models that uses one local Fisher score for both variable ordering and parent selection. The authors prove under regularity conditions that the approach identifies parents and recovers the true DAG, and they establish high-dimensional consistency for a regularized kernel Stein estimator. Experiments show the largest gains when sample size is small relative to the number of variables, and competitive graph recovery on three real-data benchmarks.
WhyProposes local Fisher-based causal discovery, proves parent sets are unique minimal Fisher completions, and establishes high-dimensional DAG consistency with benchmark experiments.
Byeongguk Kang, Donghyeon Lee, Euijong Song et al.
In briefScaling wireless foundation models through diverse self-supervised objectives and multi-branch designs
SummaryThis paper studies a new way to scale wireless foundation models: combining representations learned from reconstructive, predictive, and contrastive self-supervised objectives. At an equal output embedding dimension, the fused representations outperform the evaluated single-objective encoders on six downstream tasks spanning communication, sensing, and positioning, while using roughly one-third of their parameters. The authors also propose a jointly trained multi-branch architecture with a shared trunk that retains much of this benefit within a single encoder's parameter budget.
Ahmed Mohamed, Ahmed Aboulfotouh, and Hatem Abou-Zeid
In briefVIGIL: A Gated Learning Loop That Improves Spreadsheet Question Answering from Delayed Feedback
SummaryThis paper introduces VIGIL, a framework that lets an enterprise spreadsheet question-answering system improve from delayed labels while limiting how much each correction can change its behavior. The base model, prompts, retriever and candidate pool stay fixed, and only a contract calibrator and a query/result-column selector are updated through promotion gates. On 79 documents, forward accuracy rises from a 75.8% static baseline to about 82%. However, some gate-approved updates reduce held-out accuracy, showing that replay-buffer checks do not guarantee generalization.
In briefGrid-free method for stochastic control policy learning using coupled actor-critic flows
SummaryThis paper presents an amortized, grid-free policy-value iteration method for entropy-regularized, infinite-horizon relaxed stochastic control problems, based on continuous Langevin dynamics. It projects the dynamics onto a shared conditional sampler for the actor and a parametric critic, yielding coupled actor-critic flows with gradient and Hessian residuals. In the linear-quadratic class the projections are exact, and numerical experiments on general models demonstrate coupled learning in high dimensions.
In briefFirst-Order Steering: Turning Weight Changes into Steering Vectors for Controlling Model Behavior
SummaryThis paper introduces First-Order Steering, which treats activation steering as a first-order approximation of weight update matrices controlled by a vector of steering strengths, and derives theoretical bounds on the approximation error. It also proposes HeRD-Merging, a model merging procedure that minimizes these first-order approximation errors to improve steering accuracy. The authors report that their steering vectors control both single and composed behaviors more accurately than existing activation steering methods, while HeRD-Merging matches conventional model-merging baselines.
Sri Pranav Kunda, Alexander Kurz, Tomas Dominik et al.
In briefOctMesh: A Unified Octree Framework for Lossless Triangle Mesh Compression
SummaryOctMesh is a learned framework that encodes both geometry and connectivity of triangle meshes on a shared octree hierarchy for lossless compression. It groups candidate child edges by parent type and shared-parent status, then uses three neural predictors to estimate probabilities that guide arithmetic coding. On 256 frames from eight MPEG V-DMC test sequences, it averages 7.033 bits per face, 12.8% below V-Mesh, and supports nine levels of progressive refinement.
In briefDense Neuro-Symbolic Reasoning in a Unified Geometry State
SummaryThis paper proposes a geometry problem-solving framework in which a neural component and a symbolic engine share one typed state and exchange executable actions at every search step. Neural proposals supply theorem instances, constructions, and algebraic bridges, while the symbolic runtime applies rules, propagates exact constraints, and records provenance. The framework is implemented as OmniGeo, a single solver covering plane, analytic, and solid geometry, and the authors report accuracy on several benchmarks and 21 of 30 IMO-AG-30 problems solved.
In briefDo RUL explanations hold up? Checking whether attribution maps for engine life prediction are faithful and stable.
SummaryThe paper trains a 1D CNN, an LSTM, and a small Transformer to predict remaining useful life on NASA C-MAPSS. It compares Integrated Gradients, occlusion, and attention maps using deletion/insertion faithfulness, stability under input noise, and agreement across training seeds. The authors recommend occlusion or Integrated Gradients for maintenance reports and treat raw attention weights as visualization only.
Manh Hien Nguyen, Ngoc Thanh Nguyen, Isabella Mendoza Cortes et al.
In briefModel-based reinforcement learning steers droplets autonomously on open surfaces
SummaryThe paper presents a robotic platform that uses model-based reinforcement learning to move a liquid droplet in closed loop across an open surface. A two-axis tilting board drives the droplet, and an overhead camera provides real-time feedback. The learned policy was trained on 50 to 150 physical episodes without simulation or analytical models, and the authors report zero-shot transfer to unseen geometries and an autonomously discovered oscillatory strategy to free a stuck droplet. The full training pipeline finished in under 90 minutes.
In briefJoint Imputation and Classification for Supervised Learning with Missing Data
SummaryThis paper studies supervised classification when some feature values are missing. The authors propose the Integrated Imputation and Classification Network (IICN), which adversarially trains an imputer together with an (n+1)-class discriminator that separates the n true classes from an extra "imputed" class, using a single supervised objective. Evaluations on FashionMNIST, CIFAR-10, and tabular datasets with natural missingness report better accuracy and robustness than classical impute-then-classify pipelines and recent generative baselines.
In briefRethinking Self-Distillation for Combining Multiple Expert Models
SummaryThis paper compares multi-teacher on-policy distillation (MOPD) with off-policy methods such as supervised fine-tuning and soft-label distillation for merging expert model capabilities. In a controlled study across two multi-teacher settings, four models, and eleven benchmarks, all four methods achieve nearly identical accuracy, while MOPD uses 14.8 to 23.1 times the training GPU-hours of SFT. The authors argue that much of MOPD's reported gain stems from training design and hyperparameter choices, and suggest that well-tuned off-policy baselines are a viable, more efficient alternative.
In briefA review of AI methods for checking multiple-choice test question quality
SummaryThis focused narrative review examines research on automated detection of item-writing flaws, question revision, psychometric screening, and NLP benchmark auditing. The authors reviewed fourteen research reports in full text and mapped a 19-criterion rubric to detection methods and reported evidence. They find that high label-level accuracy often coexists with weak detection of positive cases, and that rubric definitions and reference standards vary. Evidence on whether revisions improve quality is mixed. The authors propose treating quality assurance as a sequence of independently validated decisions.
In briefAdaptive projection method for feasible Gromov-Wasserstein distribution matching
SummaryThis paper studies computation for the Gromov-Wasserstein problem, which compares structured distributions without a shared feature space or known correspondences. The authors propose A-KL-BAPG, which follows a fixed-penalty burn-in with a guarded increasing-penalty phase and uses backtracking with a projective-diameter safeguard. Experiments report that, compared with fixed-penalty KL-BAPG, it keeps similar accuracy and objective values while reducing marginal feasibility gaps and stationarity residuals.
In briefMultimodal dual-encoder retrieval for automated ICD coding
SummaryThis paper addresses three challenges in automated ICD coding: handling both structured and unstructured patient data, scaling to over 9,000 ICD-9 codes, and providing clinical transparency. The proposed two-stage framework first retrieves ICD codes with a multimodal dual-encoder that fuses structured and unstructured data through gated fusion. An LLM-based re-ranker then refines the top-k candidates and provides explanations. The experiments reportedly show gains in Micro-F1 and Precision over a multimodal dual-fusion classifier baseline.
In briefCADForge: Agentic Single-View CAD Reconstruction with Explicit Geometry Reasoning
SummaryCADForge is an agentic framework that turns a single image into editable parametric CadQuery programs. It splits an object into CAD-meaningful parts, reasons explicitly about each part's geometry, and uses a review agent to give feedback for iterative refinement. Experiments on single- and multi-part objects report higher reconstruction fidelity and perceptual quality than existing baselines.
In briefTraining language models with reinforcement fine-tuning to play social deduction games
SummaryThis paper studies how reinforcement fine-tuning (RFT) changes the social behaviour of large language models, using hidden-role social deduction games. Optimizing only final win-loss outcomes did not reliably teach social deduction, but RFT notably improved social reading, such as inferring roles from discussion. The authors also propose multi-agent social-cognitive fine-tuning, which combines social-reading and social-influence signals and receives more favourable human evaluations.
In briefCross-Trait Transfer in Subliminal Learning
SummarySubliminal learning is when a student language model picks up a teacher model's behavioral traits from training on semantically unrelated outputs. The paper introduces a directed trait-transfer matrix that measures how data from one teacher trait shifts the student's preferences for other traits, using log-probability gains. The matrix reveals clusters of related traits, and multi-trait experiments suggest opposed traits can sharpen this differentiation.
In briefFloVMos: An optical flow-based method for stitching medical videos
SummaryThis paper presents FloVMos, a deep learning framework based on optical flow for real-time video mosaicking across various biomedical imaging modalities, producing wide field-of-view composites. It fine-tunes an optical flow model on synthetic training data that simulates tissue motion and imaging distortions. The authors tested it on seven imaging modalities and report better accuracy, robustness, and speed than conventional baselines.
In briefA hand-checkable proof that two hidden ReLU layers can compute the maximum of six numbers
SummaryThe paper studies how deep ReLU networks must be to compute the maximum function exactly. Computer-assisted constructions were previously known to need only two hidden layers for up to twelve inputs. For six inputs, the authors give an explicit hexagon identity and prove the depth bound analytically through cancellations that can be checked by hand. The identity yields an explicit network with hidden widths 17 and 41, zero biases, and rational weights.
In briefGiT: A Benchmark for Robot Manipulation That Depends on Past Events
SummaryThis paper introduces GiT (Grounded in Time), a dataset and benchmark for robot manipulation decisions that depend on past events, spanning biolab, household, and industrial scenarios. It includes real-robot and UMI-style demonstrations for 18 bimanual tasks, simulation data, and a ManiSkill-based evaluation suite with nine tasks, plus fine-grained subtask annotations and counterfactual task pairs. Evaluations of representative end-to-end VLA models show substantial room for improvement in history-dependent manipulation.
In briefSpec2Game: Can AI Models Build Complete, Playable Games From Detailed Rules?
SummaryThis paper introduces Spec2Game, a benchmark asking large language models to generate complete Pygame projects from detailed natural-language game specifications. It covers 15 game families with 150 task instances across three complexity levels. Generated projects are evaluated on executability, specification realization, code quality, and user-facing quality, across 3,330 projects from 14 models. The findings show that executable code does not guarantee faithful implementation, with rule and termination logic being the hardest to get right.
In briefHow optimizer update geometry controls when singular modes emerge and fade in matrix factorization
SummaryThis paper studies how optimizer update geometry and network depth jointly shape the emergence, persistence, and disappearance of singular modes in deep matrix factorization. Under explicit balance and alignment assumptions, it derives birth, saturation, and decay laws for gradient descent, SignGD, an Adam approximation, Muon, cumulative Shampoo, and K-FAC. The low-rank bias across optimizers does not follow a simple ordering, and normalization, accumulated state, and damping are interpreted as controls on mode dynamics.
In briefQuanLing: Cross-Branch Test of Measuring Language Distance in Western Romance
SummaryThis paper extends the QuanLing framework, previously validated on North Germanic, to Western Romance languages (French, Portuguese, Spanish, Italian). Using 150 four-language parallel sentences, it measures distances with LaBSE sentence embeddings, tokenization fragmentation rates from four monolingual BERT tokenizers, and mBERT masked-language-model intelligibility. Portuguese and Spanish are closest and French and Italian most distant, with LaBSE and mBERT rankings agreeing on 4 of 6 pairs.
In briefBenchmark for Psychological Dynamics in Generative AI Personas
SummaryThis paper introduces a psychometric benchmark to test whether large language models can realistically simulate how human psychological states change over time. It scores models on internal validity, covering distributions, variance between and within persons, temporal dependence, and range, and on external validity, covering established trait, state, and indicator relations. Across 36 models, most recover the direction of known relations and the variance partition, but the within-person correlation and distributional structure remain hard to reproduce. Internal validity does not track model size or capability.
In briefHow Many Directions Control Refusal in Language Models
SummaryThis paper studies how many activation-space directions are needed to reliably control two types of refusal in language models: refusal triggered by safety alignment and refusal in general contexts. The authors introduce steering dimensionality, the minimum subspace dimension required to reliably control a behavior. They find that safety-triggered refusal can be steered with one dimension, while refusal in general contexts is not fully captured even by 5-dimensional steering subspaces.
In briefStochastic Adaptive Fourier Decomposition for Operator Learning: a neural operator using adaptive instead of fixed spectral bases
SummaryFourier Neural Operators rely on fixed Fourier bases and hard frequency truncation, which limits how well they capture non-periodic, localized, and fine-scale PDE solutions. The paper proposes SAFDNO, which uses a neural pole predictor to build an input-adaptive Takenaka-Malmquist orthonormal system and learns filtering on analytic-branch coefficients in the Hardy space. Across nine PDE benchmarks it performs best on all six regular-grid problems among strong baselines, with notable gains on Darcy, Burgers, and Navier-Stokes, and stronger zero-shot super-resolution.
In briefAttention-based surface representation learning for robot state prediction and open-ended surface classification
SummaryThis paper addresses how outdoor ground robots can recognize the type of terrain they travel on. Most prior methods classify surfaces into a fixed set of classes defined during training, whereas this work allows new surface classes to be added as trainable vectors for use in higher-level tasks. The approach learns by predicting the robot's next state and uses attention blocks, reaching 98.56% accuracy on the Belyaev-Kushnarev dataset and 94.8% on BorealTC.
In briefConformal Prediction with Paraphrase-Aware Scoring for LLM Uncertainty Quantification
SummaryThis paper proposes an uncertainty quantification framework for large language models that stays stable under meaning-preserving rewordings. It trains a lightweight proxy on LLM hidden states, aggregates its predictions across paraphrases to build label-wise nonconformity scores, and applies conformal calibration. Evaluated on seven multiple-choice QA benchmarks across several model families, the method yields compact prediction sets with empirical coverage generally near the target, mainly due to the learned proxy.
WhyAdds paraphrase-robust nonconformity scores to conformal prediction with test-side rewording guarantees, validated across seven benchmarks and several models with ablations.
In briefHumanoid Rickshaw Pulling: Whole-body control for robots pulling wheeled loads
SummaryThis paper presents a whole-body control framework that lets humanoid robots haul heavy payloads by pulling passive wheeled vehicles. A privileged teacher policy is distilled into a history-conditioned student, then fine-tuned with reinforcement learning. Compared with baselines, the policy tracks vehicle motion more accurately while reducing oscillation, torso tilt, and actuation cost. On hardware, a single policy handles loads up to 115 kg without load-specific retuning.
WhyNovel whole-body control for humanoids pulling rickshaws via teacher-student distillation and RL fine-tuning, validated against baselines and on hardware up to 115 kg; impact is mainly within humanoid
In briefPaLoRA: Paced Low-Rank Adaptation for Continual Learning
SummaryThis paper examines LoRA-based continual learning, where small learning rates are commonly used as an unexplained heuristic to limit forgetting. The authors derive a pacing law showing that the optimal gradient step scale grows with the effective rank of past updates. They propose PaLoRA, which uses adaptive SVD truncation, nullspace projection, and rank-aware adaptive pacing. The paper reports accuracy gains of about 4% on 50-task ImageNet-A and ImageNet-R benchmarks.
In briefFlashGaze: A Training-Free Method for Faster Video Understanding by Pruning Redundant Image Patches
SummaryThis paper introduces FlashGaze, a training-free method that removes redundant spatiotemporal patches from videos before vision transformer encoding, without using auxiliary networks. It uses pixel-space differences as a proxy for information loss and applies Quadtree Dynamic Programming to jointly decide which patches to drop, merge, or keep under a fixed budget. On Qwen3-VL-8B, it keeps 98% of full-input accuracy on LongVideoBench while achieving up to 5.4x ViT encoding speedup, 17x MLLM prefill speedup, and 1.8x lower peak GPU memory.
In briefRADC: Risk-aware dual caching to improve vision-language models at test time
SummaryThe paper addresses cache-based test-time adaptation for vision-language models, which can suffer from background bias in global representations and unreliable cache admission. It proposes RADC, which adds a Semantic Foreground Cache built from category-consistent spatial evidence and a Gaussian Risk Admission module that models multi-view features as Gaussians to weigh class separation and uncertainty. The authors report consistent state-of-the-art results on cross-domain and out-of-distribution benchmarks.
In briefAdaptive insertion-operator selection for the uncertain electric autonomous dial-a-ride problem (paper)
SummaryThis paper studies the electric autonomous dial-a-ride problem, which adds battery and charging constraints for electric vehicles to the classical dial-a-ride problem. It compares six insertion-operator selection methods within a bilevel large neighborhood search framework, for both deterministic and chance-constrained versions where travel times are uncertain. The methods include fixed greedy and regret-based insertion, random selection, a state-based rule, performance-adaptive ALNS, and LLM-based selection. Results show similar performance on small instances, with larger differences on bigger, more constrained ones, and no single strategy performs best across all instances.
In briefTCMClinicalReason-Bench: Testing whether language models can reason from pathogenesis to prescription on real clinical cases
SummaryThis paper introduces a benchmark built from 2,000 multicenter electronic health records to test whether language models reason from etiology and pathogenesis to prescriptions in traditional Chinese medicine. Seven models were evaluated zero-shot, with an evidence-constrained rubric and clinician ratings. Structural completeness was nearly perfect, but content scores ranged from 40.7% to 54.1%, with the weakest performance in prescription generation and analysis.
In briefFTD-GNO: Memory-Efficient Graph Neural Operators via Tensor-Decomposed Kernels
SummaryThe paper proposes FTD-GNO, a graph neural operator framework that splits the high-dimensional continuous integral kernel into low-dimensional mode-wise functions using CP, Tensor-Train, or Tucker decompositions. This allows the integral operator to be reconstructed algebraically without explicitly building full edge-wise kernel tensors. Theoretical analysis indicates lower parameter and activation-memory costs. Experiments report lower peak memory and shorter training times than unfactorized graph-integral baselines.
In briefCurveCodec 2: A skeleton-agnostic animation compressor with a learned entropy model
SummaryThis paper presents CurveCodec 2, a codec for skeletal animation that works on any skeleton and enforces a stated error bound. It codes each sub-track as a curve in the log map, quantizes it in closed loop, keeps only rate-distortion-selected keys, and entropy-codes the residuals with a small learned model. On 4,472 held-out clips, it needs about 0.37x ACL's bytes under the worst-case contract at 0.01 cm and about 0.22x at 0.1 cm under the mean-error contract.
In briefCan AI Tell Pasted Text from a User's Own Comment? Absorption at Unmarked Boundaries
SummaryThis paper studies a failure where language models treat a user's comment typed after pasted text as part of that pasted text and return it in the edited output. The authors introduce SEAM, a benchmark of 300 editing examples tested under six conditions across 20 models. At a bare newline, absorption ranges from 7.7% to 66.7%. Explicit boundary markers reduce absorption but do not eliminate it.
In briefNear-optimal sample complexity for recursive entropic risk RL with a generative model
SummaryThis paper studies how many samples are needed to learn values and policies in finite discounted MDPs under recursive entropic risk preferences, assuming access to a generative model. It gives a refined analysis of an existing model-based risk-sensitive Q-value iteration method and proves (ε,δ)-PAC guarantees for the optimal Q-function and an ε-optimal policy. The bounds improve the exponential dependence on the effective horizon and match known lower bounds up to logarithmic factors in the exponential terms and other parameters.
Amirparsa Bahrami, Oliver Mortensen, Mohammad Sadegh Talebi
In briefFine-tuning vision-language models to improve AI spatial reasoning about 2D and 3D rotations
SummaryThis paper examines how to improve vision-language models' spatial reasoning so they can handle practical spatial tasks better. The authors fine-tuned models on several object-rotation datasets built for training and evaluation. Fine-tuned Gemma-4 mixture-of-experts models outperformed fine-tuned generalist Gemma-4 models at predicting rotations defined by axes and angles, and fine-tuning also substantially improved 2D angle estimation.
In briefClean: Memory-efficient second-order LLM training via Nyström sketching
SummaryThe paper introduces Clean, an optimizer that uses the randomized Nyström method to approximate SOAP's preconditioners, reducing memory from quadratic to linear in model dimensions. It also reintegrates off-subspace curvature components and proposes Q-Clean, a low-precision variant that compresses optimizer states. The authors report over 50% lower optimizer memory than Muon when pre-training LLaMA-1.3B, and that Clean reaches AdamW's final performance 26% faster in wall-clock time. They also state that a 13B-parameter model can be pre-trained on a single 80GB GPU.
Beheshteh T. Rakhshan, Sahar Rajabi, Maziar Sargordi et al.
In briefSparse-GS2Mesh: Surface reconstruction from few views using stereo-guided 3D Gaussian Splatting
SummaryThis paper presents Sparse-GS2Mesh, a framework for reconstructing 3D surfaces from sparse input views. It initializes 3D Gaussian Splatting with epipolar depth priors and adds adaptive baseline selection, stereo-network fine-tuning, and 2D/3D co-regularization. The authors report a 15% improvement over state-of-the-art methods in little-overlap settings and comparable results in large-overlap settings.
In briefCellSplat4D: PSF-aware 4D Gaussian splatting for sparse robotic live-cell imaging
SummaryRobotic microscopes can only image living cells sparsely because each volume costs time and light. The paper fits a 4D Gaussian model to the acquired volumes to reconstruct missing frames, measuring the point-spread function from data and supervising unmeasured timesteps with a 3D U-Net. On two Cell Tracking Challenge sequences, scored per cell nucleus, the reconstruction reports the highest fidelity at every distance from an acquired frame.
In briefALoDLM: adaptive looped diffusion language model that allocates computation per token to close the quality gap with autoregressive models
SummaryDiffusion language models can predict multiple tokens in parallel for fast generation, but their output quality has lagged behind similarly sized autoregressive models. The authors attribute this to a mismatch between computation and token difficulty, and propose ALoDLM, which iteratively refines latent states at each denoising step and allocates computation according to how difficult each token is. At 1.7B and 8B parameter scales, ALoDLM achieves higher average scores than the evaluated diffusion models and corresponding autoregressive baselines across eleven benchmarks, while retaining fast parallel decoding.
In briefSparse tensor methods for detecting anomalies in multivariate functional sensor data
SummaryThe paper proposes two unsupervised sparse tensor decomposition methods for detecting anomalies in densely sampled multi-sensor process trajectories. ES-CP uses an entrywise penalty to find localized anomalies, while FG-Lasso adds a fiberwise penalty to also detect anomalies concentrated within temporal fibers. In a multichannel forging case study, FG-Lasso and ES-CP achieve macro F1 scores of 0.85 and 0.82, above TRPCA and PCA-based detectors (0.69 or lower).
In briefAsynchronous evolution strategies add almost no cost for long-horizon LLM agent tasks
SummaryThe paper notes that synchronous policy updates waste compute in post-training long-horizon LLM agents, because trajectories finish at different times and faster workers sit idle. The authors propose bounded-staleness asynchronous evolution strategies (ES) and test them on WebShop, varying lag and the fraction of stale fitness measurements, plus two stress tests for extreme staleness. Asynchronous ES closely tracks synchronous ES, and evaluation accuracy drops only slightly even when all fitness measurements are stale. The results suggest ES tolerates substantial policy staleness without explicit off-policy correction, making it suitable for asynchronous post-training.
In briefMemLeak: Cross-User Memory Leakage in Shared AI Agent Memory
SummaryThis paper studies how personal AI agents in multi-tenant enterprise systems that share one vector store for long-term memory can leak memories between users. Ordinary similarity search can return another user's semantically related memories without any attack. Incidental leakage reached 70% to 100% in team-pooled retrieval, and among three mitigations, only post-retrieval ownership gating consistently restored the clean baseline, at roughly 1.4 ms per query.
WhyFormalizes cross-user memory leakage in shared vector stores, validated by multiple experiments, ablations, and comparison of three mitigations.
In briefStructured Chain-of-Thought Prompting to Improve AI Spatial Reasoning About 3D Rotation
SummaryThis paper tests how well a generative agentic AI model (GPT-5.6) understands rotations of objects in 3D space. The authors adapted the Revised Purdue Spatial Visualization Test: Visualization of Rotations by adding graphical and contextual features, then compared several chain-of-thought reasoning strategies. Structured chain-of-thought reasoning improved performance, especially when relevant contextual information was included, but without that context the gains were limited and the model still struggled with spatial transformations.
In briefMetonymic circuits that help vision models understand abstract concepts
SummaryThis paper studies how Vision Transformers recognize abstract concepts such as "angry" when training data offers little direct evidence. The authors propose that concrete anchor concepts, such as "fire", link visual signals to abstract meaning. Using Transcoders on CLIP and DINO encoders, they find structured circuits on a curated icon dataset and show through causal interventions that these intermediate concepts are functionally involved.
In briefTracking referred objects in moving-camera videos by removing camera motion
SummaryThis paper addresses referring multi-object tracking (RMOT), which tracks all objects matching a language expression in a video. In videos from a moving camera, camera motion can make stationary objects appear to move, so the authors estimate camera motion in driver-view videos to extract residual object motion. This motion is compared with the query expression and fused with the predictions of existing RMOT methods through late fusion. The motion compensation module is shown to improve performance when used as a plug-in across different RMOT hosts.
Hsin-Chen Pai, Jyun-Kai Wang, Yi-Cheng Peng et al.
In briefEvalResearchBench: Can AI Agents Design Their Own Evaluation Feedback for Self-Improvement?
SummaryThis paper introduces EvalResearchBench (ERB), a benchmark testing whether AI agents can autonomously design evaluators. Under fixed time and API budgets, a researcher agent selects or synthesizes tasks, implements graders, and freezes an executable evaluator after pilot tests. Across 9 researchers and 13 candidate models, the best evaluators ranked about 75% of model pairs as target benchmarks did, below the roughly 91% ceiling set by agreement among the targets. A human-designed sample of public tasks remained a strong baseline.
In briefLanguage model activations sit in special error-correcting basins
SummaryThe paper studies the geometry of language model activation space and finds that activations cluster in distinct attracting basins, which separate natural activations from distributionally similar synthetic ones. It also shows that semantic steering moves activations between feature-specific basins. Adaptively adjusting steering strength to move activations across basins improves inter-language steering, raising the probability of sampling tokens from the target language compared with fixed-strength steering.
Matthew Finlayson, Francisco Pernice, Eric Todd et al.
In briefSHarP: Pruning Redundant Modules from AI Agent Frameworks
SummaryThis paper studies agent harnesses, systems that coordinate language model calls, tools, and task execution, which often grow more complex as they are repeatedly patched. Inspired by neural network pruning, the authors propose SHarP, which measures each module's saliency by disabling it and assessing its effect on task performance and token cost, then removes the least useful or most costly modules. Evaluations on held-out validation sets found that most harnesses studied were highly redundant and kept comparable performance and efficiency even after pruning a substantial portion of their modules.
In briefAttSVD: Prompt-adaptive low-rank KV cache compression using attention-guided SVD
SummaryThe key-value cache of autoregressive transformers grows with context length and dominates memory at long contexts. Instead of evicting tokens, this paper keeps every token and stores it more cheaply along the feature dimension. The proposed AttSVD computes a per-prompt truncated SVD online, retaining only the directions attention actually reads. Across several models, on an agentic benchmark and the full LongBench suite, it reportedly matches the dense cache while using up to 50% of the KV-cache memory.
In briefDimSteer: Guiding LLM Writing with Automatically Found Style Sliders
SummaryThis paper presents DimSteer, a writing interface for large language models. It samples several completions near the current prompt, finds the directions of greatest variation in the model's activation space, labels them, and shows them as sliders with pole previews, diff comparison, and reset. In a 16-person within-subjects study against a prompt-only baseline, DimSteer reduced mental demand, effort, and frustration while keeping perceived success comparable. Fifteen of 16 participants disagreed that they would have thought to request the same changes in a prompt.
In briefRisk-calibrated proposal transport for stabilizing diffusion steering with finite particles
SummaryThe paper studies inference-time steering of pretrained diffusion models using Feynman-Kac correction and sequential Monte Carlo, where the proposal's finite-particle behavior matters. It shows that variance-controlling guidance can nearly eliminate its fitting residual yet greatly increase residual risk on new states, potentially degrading generation or causing particle collapse. The authors propose Risk-Calibrated Proposal Transport (RCPT), which uses deletion leave-one-out residuals to calibrate how much of the VCG update is retained, with no extra model calls. Experiments on 2D checker distributions, scaffold decoration, molecular optimization, and class-conditional CIFAR-10 report improved terminal metrics over uncalibrated VCG.
In briefAgentic Cognitive Depth: Operational Criteria for Judging LLM Agents
SummaryThis paper argues that LLM agents should be judged beyond end-to-end task success, and proposes "agentic cognitive depth," a trajectory-level profile built on five criteria: context sensitivity, temporal continuity, multimodal coordination, adaptive interaction, and metacognitive monitoring. The first four assess how well control flow, memory, tools, and planning are used across a run, while the fifth assesses whether the system monitors and regulates the whole run. For each criterion the authors give operational proxies and a perturbation procedure, and they outline how to add per-criterion diagnostics to benchmarks such as GAIA, SWE-bench, WebArena, and TRIP-Bench.
In briefA principled training method for top-k selection in language models using hybrid gradients
SummaryThe paper addresses selecting k items from m candidates in language model systems, such as document selection in retrieval-augmented generation and expert routing in Mixture-of-Experts models. It proposes a principled training objective whose gradient combines a supervised component and a policy-gradient component. The algorithm is shown to converge at rate O(1/√T), and experiments on synthetic regression, RAG, and MoE tasks report better next-token perplexity and QA accuracy than baselines.
In briefDiagnosis-conditioned spatial gating and decoder-level contrastive learning for radiology report generation
SummaryThe paper addresses finding-level errors in radiology report generation, where models produce fluent text that may still be inaccurate. It proposes a Position-Aware Gate that uses predicted finding representations to modulate visual patches spatially without region supervision, and a Decoder-level Supervised Contrastive Loss that structures decoder representations by shared positive findings. On MIMIC-CXR, PAG+DSCL raises clinical efficacy F1 from 0.484 to 0.502 over a matched global-gate reference, and on IU X-Ray it reaches 0.226 without additional fine-tuning, compared with 0.211 reported by PromptMRG.
In briefHow reinforcement learning changes LLM reasoning: transfer, coverage, and scaling
SummaryThis paper examines how reinforcement learning (RL) affects the reasoning of large language models, finding both cross-domain gains and forgetting, with coverage at large sampling budgets rising on some tasks and falling on others. By analyzing solution traces before and after RL, the authors propose a two-stage autoregressive policy model that separates strategy selection from problem-specific execution. The framework is used to theoretically explain cross-domain transfer, coverage changes, and log-sigmoid and log-linear scaling laws in RL compute.
WhyProposes a two-stage strategy-selection framework, proves RL's implicit bias reshapes strategy preferences, and validates across Qwen and Gemma with scaling-law theory.
In briefEasing over-squashing in GNNs by changing the sheaf, not the graph
SummaryGraph Neural Networks often fail to capture long-range dependencies because repeated compression of node embeddings causes over-squashing. The paper introduces sheaf effective resistance, a generalization of effective resistance based on the sheaf attached to the graph, and proves an upper bound on over-squashing sensitivity for flat vector bundles. It presents FlatNSD, a simple message-passing variant of Neural Sheaf Diffusion, which performs well on over-squashing benchmarks without changing the original graph topology.
Andr\'e Ribeiro, Germano Barcelos, Amauri H. Souza et al.
In briefTrajectory-Based Confidence for Reliable, Resource-Efficient Clinical Text-to-SQL Agents
SummaryThe paper addresses the problem that LLM agents for clinical text-to-SQL cannot judge whether their own reasoning or outputs are trustworthy, which matters in high-stakes healthcare. It introduces Sentinel, a classifier-based confidence layer derived from the agent's reasoning, code and database outputs, which decides whether to refuse, halt, or withhold answers. On the EHRSQL benchmark, Sentinel raises the Reliability Score from +0.08 to +0.24 at low error weighting and reduces agent steps by 13-28%.
Mincheol Daniel Song, Joshua Ward, Jake Jung et al.
In briefOne-cycle fault classification and faulted-line identification benchmark on PROTECT-90
SummaryThis paper provides an initial application benchmark on the PROTECT-90 electromagnetic-transient dataset for fault-type classification and faulted-line identification. A compact 1D convolutional neural network is evaluated on post-inception windows, with a multilayer perceptron as an architecture-control baseline. Under full observability both tasks are nearly saturated, reaching 99.84% fault-type and 100.00% line-identification accuracy in one cycle. Performance drops under voltage-only inputs, indicating that measurement information rather than network architecture is the limiting factor.
Emad Abukhousa, Abdulaziz Qwbaiban, Saman Zonouz et al.
In briefKepler4D: controllable future video generation by evolving a 4D scene state
SummaryKepler4D is a framework that predicts future video from a single input video by building a shared 3D scene representation. A vision-language model selects object speed and heading, which are rolled out into future object trajectories that can be inspected and edited. These trajectories are rendered as geometric controls for a pretrained video generator. Experiments on real-world videos show controllable object motion while preserving scene consistency.
In briefWatermarks and fingerprints as soft bindings: an open-licence benchmark for image, audio and video provenance
SummaryThe paper benchmarks invisible watermarks and registry-based fingerprints, two ways to recover C2PA provenance manifests after metadata is stripped, using openly licensed models and public media. It evaluates many image, audio and video watermark configurations and fingerprint methods on registries of up to 98,985 images, measuring quality, robustness, false matches and attacks. The results show the two families fail differently, and that embedding a watermark can move ISCC codes past their match threshold while platform colour conversion alters watermark bits.
Seyedmahdi Kazempourradi (1), Ramtin Mojtahedi (1), Behrang Mohseni (1) ((1) Original Pictures Technologies et al.
In briefChecking Which Inception Crop Implementation You Use
SummaryInception crop is a widely used crop-based data augmentation for training deep vision models. The paper finds that the standard TensorFlow/JAX implementation samples crop scale with a different density than the PyTorch version. Training 522 ViT-S/16 models on ImageNet-1k, the authors report findings on training budget, the lower tail of the crop scale distribution, and saliency, and revisit Beta crop.
In briefConsideration Circuits: Depth Separation Beyond a Single Softmax
SummaryThis paper introduces consideration circuits (CC), feature-based multi-stage choice models built from directed acyclic graphs of multinomial logit units. On a three-item compromise task, the authors show that increasing depth from 2 to 3 reduces the required maximum taste-vector norm from order log(1/ε)/ε to order log(1/ε). They also characterize geometric conditions for approximating arbitrary choice tables, with depth 3 sufficing under them. In experiments, small standalone tree circuits achieve the lowest mean test negative log-likelihood among evaluated models on several benchmarks.
In briefIdeal paths approximating logistic gradient descent under large initialization
SummaryThis paper studies full-batch logistic gradient descent on strictly linearly separable data, starting from a large initialization of scale R that reflects prior training. It builds a geometric ideal path made of finitely many linear segments, with a negative-margin correction stage followed by minimum-margin growth. After an explicit two-stage time reparameterization, the GD trajectory divided by R converges uniformly to this path as R tends to infinity, yielding asymptotic formulas for peak evaluation loss and cumulative training loss.
In briefExpertMuon-Compass: Alignment-Guided Step Sizes for Training Mixture-of-Experts Models
SummaryThis paper proposes ExpertMuon-Compass (Compass), a method that adjusts the Muon update step size for each expert in a mixture-of-experts language model based on how well the expert's update aligns with its gradient. It keeps Muon's update direction and momentum buffer while multiplying each step by two alignment-based factors. In FineWeb-Edu pretraining, Compass performs as well as or better than Muon, NorMuon, and other optimizers, and it is most effective when the data seen by each expert changes during training. The authors also prove a perturbation bound for the two factors.
Omatharv Bharat Vaidya, Ashwin Vinod, Pedram Akbarian et al.
In briefFrom Sight to Foresight: Predicting Spatial Changes in Vision-Language Models
SummaryThis paper introduces SpatialMind, a vision-language model that reasons about real-world-scale space and predicts future spatial states. Its metric depth adapter ties spatial reasoning to real-world scale, and a progressive state chain uses current states and observed motion to predict future frames. The authors also release the SpatialMind-30K dataset and the SpatialMind-2K benchmark, covering driving and everyday egocentric scenes. Experiments report that SpatialMind substantially outperforms general and spatially specialized models on its benchmark and performs competitively zero-shot on VSI-Bench, OSI-Bench, and VLM4D.
In briefDual-scale graph transformer for fraud detection that considers the wider ecosystem
SummaryThis paper introduces HERMES, a model for detecting account takeover fraud in digital banking. It combines local relational structure around each session with global context about the evolving fraud ecosystem, rather than scoring sessions in isolation. Evaluated on over 130 million high-risk transaction sessions from a U.S. financial institution, it reports a 44.44% relative reduction in customer friction and a 24.66% relative improvement in fraud recall over the production system.
WhyCombines local heterogeneous graph attention with global ecosystem context; evaluated on over 130M sessions against production and graph baselines with ablations.
Mohsen Nayebi Kerdabadi, Xinrou Li, Yao Xiao et al.
In briefDynamic Harness Search: Predicting the Best Multi-Agent Setup for Each Query
SummaryThe paper introduces SHIFT, which builds a tailored multi-agent harness (roles, instructions, tools, and communication structure) for each query without executing every alternative at inference time. A local LLM architect learns a policy over harness-building actions, while a value function predicts a utility balancing accuracy and execution cost. Across 9,193 tasks in six benchmarks, SHIFT reports the highest mean accuracy, about 80%, outperforming 17 baselines.
In briefAdaptive resource allocation for batch multi-objective Bayesian optimization in autonomous materials discovery
SummaryThis paper develops a multi-objective Bayesian optimization framework for alloy design under resource constraints, comparing an exploitation-focused strategy, a fixed mixed policy, and two LLM-based adaptive strategies. The evaluation covers refractory high-entropy alloys and Fe-Co-Ni soft magnetic alloys, including mid-campaign budget and timeline disruptions. Mixed allocation strategies accumulated substantially more mutual information than the exploitation baseline, at a smaller cost to hypervolume and speed. Adaptive strategies outperformed the fixed mixed policy, reducing predictive uncertainty on Pareto-optimal compositions.
Robert Robinson, Shakti Prasad Padhy, Sushant Sinha et al.
In briefAuditing How Hypothesis Equivalence Is Judged in Multi-Agent AI Systems
SummaryThis paper studies multi-agent systems built on large language models that generate scientific hypotheses. It compares how many rounds of self-critique (0, 1, or 5) change the delivered hypotheses, and how three pairwise equivalence rules (TF-IDF, dense embeddings, and an LLM-as-a-judge) affect measured diversity. The authors find that the chosen equivalence rule substantially changes both the estimated effect of self-critique and the measured diversity of the generated hypotheses.
In briefPB-GRPO: Training Socially Adaptive LLM Agents with Persona Simulation and Preference-Batched GRPO
SummaryThe paper introduces PB-GRPO, a post-training algorithm that teaches large language models to adapt to users' latent preferences over multi-turn conversations. It uses a persona-driven simulation environment with a persona library, LLM-based user simulators, and a user-satisfaction score from 0 to 1. Unlike vanilla GRPO, PB-GRPO estimates advantage normalization within buckets of users with similar preferences to stabilize training across diverse users. The authors report that it improves social behavior over strong reinforcement learning baselines in their simulated environment.
