In briefAdaptive Entropy-Sparing (AES) for Efficient Reasoning
SummaryLarge reasoning models often produce long, inefficient chains of thought that raise computational cost and risk errors. This paper proposes Adaptive Entropy-Sparing (AES), which spares high-entropy tokens appearing before an efficient in-group reference and penalizes overlong high-entropy suffix tokens tied to superfluous branching. Across eight reasoning benchmarks, AES reduces reasoning length by up to 52.9% while raising mean accuracy by up to 3.8% over base models.
In briefHow Far Must a Random Test Explore? Certifying Read-Write Interfaces Through Their Own Random Edits
SummaryThe paper studies whether random tests of stateful code can reliably show that an interface behaves like a record with hidden data. It defines a distance measure and analyzes how many samples and how much exploration are needed under different start distributions and write-chain depths. From a single start state, law checks suffice to certify the interface, while for the worst start distribution, exploration linear in the number of variables is needed when writes are counted.
In briefDisagreement-Aware Distribution Matching Distillation for Few-Step Diffusion Models
SummaryThis paper proposes Disagreement-Aware Distribution Matching Distillation (DA-DMD), a method that improves few-step diffusion model training. It uses the magnitude of disagreement between the teacher model and the fake model already computed during DMD training, applying updates only in the high-noise regime and scaling their strength accordingly. According to the paper, DA-DMD improves few-step generation across several diffusion backbones at the same inference cost, raising HPSv2 on SD3.5-L from 28.23 to 30.93 and achieving a 64.4% average win rate over DMD2.
In briefShopMind: A mental world model that makes e-commerce user simulations more personal and realistic.
SummaryThe paper proposes ShopMind, a mental world model for e-commerce user simulation that captures how consumers interpret shopping information and how their mental states evolve through interaction. The authors train a Mental State Reconstruction (MSR) model on behavior-mental state pairs synthesized by ShopMind to infer latent mental states from observable behavior. The paper reports that ShopMind improves personalization by 18.0% and behavioral realism by 19.6% over the strongest alternatives, and that MSR improves consumption value measurement and CTR prediction.
In briefLEGION: Testing AI Agents on Finding Rare Attack Clues in Noisy Data
SummaryThe paper introduces LEGION, a benchmark combining a simulated enterprise-telemetry environment, a library of 74 adversarial scenarios, and an evaluation protocol that scores both efficiency and detection accuracy. Five agent architectures were tested on two simulated companies. Recall was high (0.82–0.88) but precision was lower (0.61–0.72), as agents missed some attack indicators or confused genuine indicators with attack-related ones. The authors release the benchmark, scenario library, and pipeline.
Samuel Joseph Amouyal, Ofek Glick, Yaniv Menasherov et al.
In briefFixing a robot's misreading by directly editing its model weights
SummaryThe paper asks whether a robot policy's single-step failure can be repaired by editing its weights directly, without gradients, training data, or changing the instruction. It uses closed-form instruction distillation to copy the per-layer effect of an explicit rewording into the language model's MLP weights. On 128 held-out seeds, success rose from 27 to 106 episodes, while a same-size random update left success unchanged.
WhyThe authors propose gradient-free closed-form weight edits that distil a rewording into an MLP, raising success from 27 to 106 on 128 held-out seeds versus a random-direction control.
In briefA street-view image pipeline for checking which house is assessed as vacant
SummaryStreet view imagery is often used to remotely audit buildings, but an image may not clearly show the target house or may show a neighboring one. This paper presents a reproducible processing pipeline that makes image suitability and target-house identity explicit and checkable. Applied to 10,311 images of 1,609 detached-house parcels in South Korea, the pipeline raised vacancy classification AUROC from 0.684 to 0.719. The authors conclude that image suitability, target identification, and model evidence are separate problems that scores alone do not resolve.
In briefWhether a Chain-of-Thought Answer Is Decodable Does Not Show When the Model Decided
SummaryThe paper examines whether probes on chain-of-thought hidden states read the model's final answer or the candidate it is about to mention. It introduces ECDP, which fits the same low-rank reader separately to the imminent mention and the final answer. On the DecisionDepth task, Qwen3.5-4B decodes the next mention well while final-answer decoding stays near its null, whereas in a separate MMLU setting both targets are decodable and the final answer is read more strongly. The authors conclude that a strong answer readout alone does not establish when a model commits to a decision.
In briefDataset for cross-language patent claim drafting across countries from PCT filings
SummaryThis paper frames cross-lingual patent drafting as generating candidate claim sets for a specified language, jurisdiction, and prosecution stage from a source claim set in another language. The authors build a Japanese-English dataset by linking Japan Patent Office and USPTO records across 884,179 PCT-derived patent families, yielding 4,067,168 directional claim-set instances. Commercial machine translation and prompting capture only part of the task, while supervised adaptation gives the strongest overall baseline.
Trung Phuong Le, Hirofumi Nonaka, Koji Marusaki et al.
In briefTeaching vision transformers to focus, so attention heads specialize spatially
SummaryThe paper addresses a limitation of self-supervised vision transformers such as DINO, whose representations can be ambiguous or overly similar because they lack a focusing mechanism during training. The authors introduce two visual focusing loss functions that link the model's attention to specific image regions, steering its multi-head attention without annotations. This leads to spatial attention heads that are more diverse and spatially selective. Qualitative and quantitative results show improved interpretability and better performance on dense prediction tasks such as zero-shot image segmentation.
In briefHow many candidates does a one-step world model need? Random-codebook view of nearest-of-m training
SummaryThis paper studies one-step stochastic world models trained on frozen DINOv2/v3 patch tokens, comparing nearest-of-m matching (conditional IMLE) with flow matching and energy-score generators. With m=16 it covers enumerated futures better than 50-step flow matching at about 1/30 of its latency. The authors interpret the objective as the expected distortion of a random codebook and state three predictions about memorization, mode flattening, and under-dispersion.
In briefWhen physics beats data: data-efficient traffic state estimation with amortized PINNs
SummaryThis paper studies traffic state estimation, which reconstructs traffic density at unmonitored locations from sparse sensors. The authors impose the Lighthill-Whitham-Richards conservation law as a physics-informed prior and test it on four real corridors: NGSIM, I-80, PeMS04 and PeMS08. Physics-informed estimators generalise to held-out sensors better than data-driven models, including classical and deep kriging, with the clearest gap on the two shock-rich corridors. The paper also proposes a loss-conditional PINN that removes per-corridor loss-weight tuning at about 5x lower cost.
In briefOptimal regret lower bound for online learning in tree-structured games
SummaryThis paper studies the minimax regret of full-information online decision making in extensive-form games. For treeplex strategy spaces, it shows that the logarithmic dependence on the number of reduced normal-form strategies in known efficient algorithms such as KOMWU is minimax optimal. The authors introduce tropical Gaussian formulae and prove a lower bound on their Gaussian width, yielding a regret lower bound of Ω(√(T log|V|)) that matches known upper bounds up to universal constants.
WhyThe authors introduce tropical Gaussian formulae and report a regret lower bound matching known upper bounds for treeplex online learning.
Ashkan Soleymani, Zhiyuan Fan, Lillian J. Ratliff et al.
In briefBeyond Head Importance: Collective Computation in Video Transformers
SummaryThis paper studies how attention heads in video Transformers work together as a collective rather than only by individual importance. Ablation experiments on four video Transformers show that heads important for one task also support other video understanding tasks and help preserve representations that distinguish videos. Sets of heads with shorter spatiotemporal attention distances, when jointly ablated, cause greater performance loss than equally sized sets of high-importance heads.
In briefDiagnosing where frozen vision-language models fail when choosing among detector-proposed regions
SummaryThe paper studies modular visual grounding, where a frozen detector proposes candidate regions and a vision-language model (VLM) picks the referenced one. With candidate coverage held fixed, the authors vary the selection interface and find that what the selector reads matters more than how it is built. Outside the training distribution, such as on Objects365 images, trained selectors degrade sharply while direct Qwen coordinate decoding stays strong.
In briefFading to Red Noise: Scheduled Coloured Exploration Reduces Robot Movement Jerk
SummaryThis paper studies exploration in deep reinforcement learning for continuous robot control. The authors systematically compared 49 linear schedules moving from white to red noise, together with generalised state-dependent exploration (gSDE), on four sparse-reward manipulation tasks. They found that the noise colour at the end of training did not affect learning efficiency or final performance, while fading toward red noise halved exploration jerk in every task. Red noise learned as efficiently as pink noise with almost two-fold less movement jerk, so the authors recommend red-noise-fading exploration schedulers as the default for off-policy robot learning.
In briefCompeting Influence in LLM Populations: Budget Timing and Baseline Drift
SummaryThis paper studies how two competing campaigns affect opinion change in a network of persona-conditioned large language model agents. It compares runs with and without campaigns and tests all combinations of early, sustained, and late spending. The authors find that opinions drift substantially even without intervention and that stance diversity declines. After accounting for this drift, campaign effects align more with the claim's epistemic standing than with the campaigner's assigned role.
In briefCan Hindsight Critics Guide Delayed Actors?
SummarySensory delays cause a mismatch between the environment's state and what an agent observes, making control problems partially observable. The paper proposes Hindsight Critics, an asymmetric extension of Soft Actor-Critic that trains critics on hindsight-aligned state-action transitions. The actor, which sees only delayed observations, learns from the resulting privileged value estimates. On the DelayBench suite, the method is compared against delay-aware and delay-unaware baselines and partially compensates for observation delays without changing the agent's observations at deployment.
In briefAre UFOs Real? Introducing Unified Factored Organizers, networks that learn to model good representations for other networks
SummaryThe paper proposes unified factored organizers (UFOs), networks trained to model regularities in the representations of other networks. A task network can be trained to lower its task loss while aligning its representations with what a UFO has learned. In an image-regression domain, UFO-trained networks learn qualitatively better, more interpretable representations. In a transformer grokking experiment on modular arithmetic, UFO regularization cuts training steps needed to grok by over 10x with weight decay and over 100x without it.
Thomas X. Sheppard, Alon Albalak, Joel Lehman et al.
In briefDeltaForge: A post-hoc mixture-of-experts driving model combining multiple domains
SummaryDeltaForge is a post-hoc mixture-of-experts vision-language-action model for autonomous driving that merges separately finetuned per-domain models into one multi-domain model. It uses only their weights and a small proxy dataset, decomposing weight updates via singular value decomposition into shared and small domain-specific parts. Two cooperating routers are trained on the proxy data to select among the experts. Across six domains including urban, off-road, and highway driving, it approaches per-domain expert accuracy and outperforms merging, routing, and federated baselines.
In briefDASH: A decoupled asynchronous system for scaling vision-language-action models in humanoid robots
SummaryThe paper presents DASH, which splits a vision-language-action policy into a slow vision-language model and a fast flow-matching action head that run concurrently over LCM. This makes the control rate depend only on the action head, not on backbone size. On an asynchronous RoboTwin benchmark, DASH keeps a 20 Hz control rate from 0.8B to 8B backbones, while the synchronous rate falls from 9.7 to 4.1 Hz. On a real humanoid with dexterous hands, DASH achieves clearly higher pick-and-place success than the synchronous pipeline.
In briefIdentifiable but hard to estimate: rotation in overcomplete ICA
SummaryThis paper studies how many observations are needed to estimate a common rotation in two-dimensional overcomplete independent component analysis with equally spaced mixing columns. For a fixed non-Gaussian fourth-order Gram-Charlier source law, the rotation is identifiable modulo the column spacing, yet the required sample size grows exponentially with the number of sources. The authors characterize this sample size via the reciprocal of the squared Hellinger distance between rotations separated by half the column spacing, and give a radially weighted phase estimator that attains this scale.
In briefCausal credit assignment that improves Verilog code generation using counterfactual hardware tests
SummaryThis paper introduces Causal Signal Credit Assignment (CSCA), a training method for generating Verilog hardware code. It checks candidate modules at the signal level, applies counterfactual edits to output-driving assignment statements, and uses the resulting changes to weight each statement's contribution during preference optimization. Evaluated on VerilogEval 2.0 and RTLLM v2.0, CSCA with a Qwen2.5-Coder-7B-Instruct backbone leads the 7B-scale models and is comparable to the much larger GPT-5.4-mini.
In briefSeverity-Controlled Medication Prediction Sets
SummaryThis paper proposes Conformal Severity Control (CSC), a post-hoc method for calibrating medication prediction sets. It uses the Anatomical Therapeutic Chemical (ATC) hierarchy to measure how severe each mismatch is, and controls both average severity and upper-quantile severity for patient-level predictions without retraining the predictor. Experiments on MIMIC-IV show that CSC follows user-specified severity tolerances and shifts prediction sets away from distant mismatches toward exact or close matches.
In briefMOTIF: LLM-generated material groupings for predicting properties of unseen materials
SummaryThis paper addresses out-of-distribution generalization in materials property prediction, where models must handle chemistries unlike their training data. The proposed MOTIF framework uses large language models to generate candidate taxonomies that partition materials into chemically meaningful groups, keeps only those validated on training-only pseudo-OOD splits, and applies invariant feature selection within each accepted group. On 2D halide perovskite bandgap prediction, mean OOD MAE falls from 0.122 eV to 0.111 eV, and the authors report transfer gains on other materials datasets.
In briefCalibration failure from summation in multi-label facial expression recognition label expansion
SummaryThis paper examines why a label-space expansion method (CALE) fails for multi-label facial expression recognition. The authors evaluate it on a self-built dataset and the public RAF-DB and RAF-ML benchmarks across seven backbones. Aggregate scores vary widely at a fixed threshold while threshold-free mAP stays stable, pointing to the summation step as the source of the fault.
In briefBenchmark for Evidence-Based Shared Belief Updating in Connected Robotics
SummaryThe paper introduces the Connected Robotics Evidence Graph (CREG) Benchmark, which evaluates how systems move from evidence to belief across fused observations, rather than scoring only final answers. It contains 350 generated scenarios across seven regimes, with normative posterior references at each evidence prefix. Experiments with four open-weight models and six fusion baselines show that removing semantic relations lowers strict belief-state accuracy on relation-dependent scenarios, while graph serialization gives no consistent advantage, and belief accuracy varies non-monotonically along the evidence sequence.
In briefCorrect Answers from Insufficient Histories: A Controlled Study of Agent Memory Evaluation
SummaryThis paper uses executable device logs in a controlled study to separate three things: whether an answer matches a reference, whether the evidence is sufficient, and whether the behavior fits the context. It tests two hosted reader models on 80 confirmation histories and 203 insufficient-evidence instances. Both models often match hidden references while rarely abstaining. The estimated effects of the compact method remain inconclusive and do not demonstrate gains for a general memory system.
In briefMOOD: A Benchmark for Post-Hoc Out-of-Distribution Detection in Materials Property Prediction
SummaryThis paper introduces MOOD, a benchmark testing whether post-hoc out-of-distribution detectors, which flag unfamiliar inputs without retraining, work for graph neural networks predicting materials properties. It evaluates 15 detectors and 2 supervised probes across 4 GNN architectures, 5 MatBench tasks, and 15 physically motivated shifts. The results show that all 15 detectors perform near chance, with an AUROC around 0.5.
Shirley X. Yu, Amey P. Pasarkar, Adji Bousso Dieng
In briefUsing linear representations of input ambiguity to improve confidence estimates in language models
SummaryConfidence scores from language models are often poor predictors of whether their answers are correct. The authors find directions in model activations that represent input ambiguity and propose DART, a fine-tuning method that restricts parameter updates to the span of these ambiguity directions. DART greatly improves calibration while maintaining accuracy on multiple-choice and generative question answering tasks.
In briefMDA framework: multi-architecture domain adaptation for unsupervised person re-identification
SummaryThis paper proposes a Mamba-enhanced multi-architecture domain adaptation framework (MDA framework) for unsupervised person re-identification, aiming to address domain gaps. It is built on a composite MDA operator that encapsulates Mamba as a component and supports multiple network architectures. Experiments on person re-identification benchmarks show gains with both CNN and Transformer backbones, with mAP improvements of 10.2% and 19.8% respectively on the CUHK03 to Market-1501 adaptation task.
In briefAnswer or Defer? Testing Stated Confidence of Vision-Language Models on Brain MRI
SummaryThis paper examines whether vision-language models reading brain MRI scans can use their stated confidence to decide which answers to accept and which images to refer to a clinician. Eight open-weight models assessed 3,888 brain MRI slices from post-treatment glioma patients (BraTS 2024) and healthy volunteers (IXI) for tumor or abnormality. On the BraTS data, no confidence threshold kept a nonempty set of answers with at most 5% error for any model. The authors caution against using unvalidated stated confidence as a standalone deferral signal and recommend population-specific validation before deployment.
Amir Sabbaghziarani, Parsa Ghazvinian, Sepehr Salem et al.
In briefJointly Training Super-Resolution and Artifact Removal for Industrial Machine Vision
SummaryThe paper addresses how downsampling and compression on edge devices degrade defect detection in industrial machine vision. It proposes an end-to-end framework that integrates super-resolution and compression artifact removal, trained to optimize both human and machine perception. Experiments on PCB Defect, DeepPCB, and NEU-DET show the joint model generally outperforms separate pipelines, raising PCB Defect mAP@50 from 0.461 to 0.887 under severe JPEG compression with a slight PSNR drop.
Dang BuiDinh, Sang NguyenQuang, Anh NguyenPhuong et al.
In briefA communication method for distributed online optimization that prioritizes updates by their effect on KKT conditions
SummaryThis paper studies fully distributed online convex optimization where costs and constraints couple all agents together, so each agent relies on a local belief about joint actions and a shared Lagrange multiplier. Exchanging full-dimensional beliefs every round is costly, so the authors propose a randomized communication mechanism that triggers each primal and dual channel according to its estimated impact on the KKT operator. Under standard regularity assumptions, the algorithm achieves O(T^{2/3}) expected static regret and O(T^{2/3}) expected cumulative constraint violation. Numerical experiments show better learning and constraint performance per transmitted bit than belief consensus baselines.
In briefHow path integral diffusion decides a finite spin glass
SummaryThis paper studies how diffusion-type generative models, specifically Path Integral Diffusion (PID), decide macroscopic alternatives when sampling a frustrated spin glass. The authors show that isotropic PID is equivalent to a time-changed Gaussian stochastic localization. They test the theory on Bethe-lattice and Sherrington–Kirkpatrick glasses with up to 64 spins, finding that trajectories first pick a mirror half and then a glassy class within it.
In briefAcrossVAM3.0: Resource-Bounded 4D Scene Construction with Adaptive Surface Matching
SummaryAcrossVAM3.0 is an offline system that turns text, images, or monocular RGB video into a versioned point representation, running frozen components sequentially on one GPU. It introduces multi-hypothesis adaptive surface chaining (MHAC), which keeps several candidate matches and an explicit missing state to preserve long-range tracks. On held-out videos, MHAC improves Average Jaccard over a position-only baseline under the tested protocol. The results do not establish metric scale, semantic identity, or high-fidelity novel-view rendering.
In briefWhen to Trust Adaptation: Contextual Fusion for Transferable Vision-Language Prompting
SummaryThis paper proposes Contextual Adaptive Fusion (CAF), a method that helps vision-language models decide how much to trust learned prompts when facing new classes or domains. CAF keeps four prediction branches built from frozen and prompted representations, and fuses them using each branch's confidence together with how often the branches favor a class across unlabeled target data. Fusion coefficients are learned once on source data and kept fixed at inference. Trained on ImageNet, CAF reaches 69.96% average accuracy across ten cross-dataset benchmarks.
In briefDirect Trajectory Optimization for Aligning Text-to-Image Diffusion Models at Inference Time
SummaryThis paper proposes Direct Trajectory Optimization (DTO), a training-free method that aligns diffusion image generators with reward functions by directly optimizing their sampling trajectories. It treats intermediate denoising scores and sampled noise as optimization variables, so it does not need gradients through the denoising network. To handle out-of-distribution issues, the authors introduce Reward-Optimized Score Guidance, which interpolates between optimized and pretrained predictions, with a theoretical guarantee of reward improvement under constrained distributions. Experiments across three diffusion models and two benchmarks show better alignment than existing inference-time methods while remaining efficient.
In briefWhen AI Should Overturn a Human Call: 2026 World Cup Controversy
SummaryThis paper studies how to combine a human binary call, such as a referee's offside or onside decision, with an AI numerical measurement, such as VAR's estimate of the attacker's position relative to the offside line. The authors derive an explicit maximum-likelihood rule that decides whether the referee's call should stand or be overturned by a conflicting VAR measurement. The model treats the true offside margin as fixed and unknown, with independent additive Gaussian errors in the human perception and the AI measurement.
In briefCaseLaw: A Training-Free Content Moderation Framework Decoupled from Policy
SummaryThis paper proposes CaseLaw, which reframes content moderation as precedent-based retrieval and thresholded voting over labeled cases. Multimodal content is converted into structured semantics, embedded in a stable space, and classified by the labels of retrieved precedents. Policy changes are handled by updating the case library without retraining models or modifying prompts. The authors introduce the MFO-RedNotes benchmark for male-female opposition risk, and report that CaseLaw outperforms baselines in precision and recall.
In briefA method for choosing camera views under a budget to speed up real-time 3D perception for mobile robots
SummaryThis paper presents BCVS, a pose-free view selection algorithm that runs at over 22 FPS on a CPU for online filtering of redundant images. It scores views using only feature correspondences, iteratively picking the most distinct frames within a user-set budget while suppressing near-duplicates. BCVS outperforms CPU-friendly baselines at tight budgets on benchmarks and, on large aerial and marine robotics data, reduces redundancy by up to 16% while maintaining robust SfM registration and high-quality 3D reconstruction.
Vaishnav Ramesh, Jacob Carter, Shivali J. Patel et al.
In briefEfficient online submodular maximization when only selected-set values are observed
SummaryThis paper studies online monotone submodular maximization under a cardinality constraint, where the learner observes only the value of each selected set. The authors present a polynomial-time algorithm with expected (1-1/e)-regret of O(k√(nT log(n/k))) over T rounds. The authors state it is the first polynomial-time algorithm with √T dependence on the horizon for arbitrary monotone submodular functions in the bandit setting. The approach combines randomized sampling with a carefully constructed potential to reduce the problem to online linear optimization over the probability simplex.
In briefLearning microstructure-based models of inelastic material behavior with neural operators
SummaryThis paper studies how neural operators can learn homogenized constitutive laws for inelastic heterogeneous materials, replacing costly unit-cell simulations. For periodic Kelvin–Voigt viscoelasticity, the authors prove a universal approximation theorem in two and three dimensions and give an explicit bound on the internal-variable dimension. Numerical experiments cover viscoelasticity and crystal plasticity, and the theoretical results are formally verified in Lean 4.
WhyThe authors prove a universal approximation theorem for recurrent neural operators in 2D and 3D periodic Kelvin–Voigt viscoelasticity, with logarithmic internal-variable bounds.
Lianghao Cao, George Stepaniants, Andrew M. Stuart
In briefHiddenPathQA: A Benchmark for Knowledge Graph Q&A When Questions Hide Their Paths
SummaryThis paper points out that knowledge graph question answering is usually tested on questions whose wording reveals the correct graph path, while real users often do not know the graph's structure. The authors build HiddenPathQA, a benchmark of 1,055 manually verified questions spanning 2 to 6 hops, and propose Holistic Path Selection, which compares candidate relation chains as whole units. It outperforms six strong baselines across two LLM backbones and both original and pseudoword settings, and the pseudoword test indicates the gains cannot be explained by entity memorization alone.
In briefUsing papers as benchmarks to audit AI agents on spatial transcriptomics alignment
SummaryThis paper presents a framework that turns scientific publications into auditable benchmarks for AI agents and demonstrates it on spatial transcriptomics alignment tasks. Across 121 tasks, 16 baseline methods, and six AI agents, the best agents matched or outperformed all baselines on 80% of tasks. Building the benchmark also revealed errors in prior work and shortcuts in current evaluation metrics.
In briefChemInsight: A Benchmark for Testing How Well Language Models Understand and Reason About Chemistry
SummaryThis paper introduces ChemInsight, a benchmark of 203 challenging chemistry questions grouped into 50 bundles that require multi-step reasoning. It scores models on two dimensions, scope and depth. Evaluating ten representative large language models, the authors report consistently low results, with scope scores below 30% and depth scores below 11%, well under human expert performance.
In briefComparing Image Models for Explainable Brain Tumor MRI Classification, Tested on External Data
SummaryThis paper benchmarks six deep learning architectures, including convolutional networks, vision transformers, and foundation models, for classifying brain tumor MRI images into four classes. It finds that 36% of test images had near-duplicates in the training data, and accuracy dropped on the duplicate-free subset. On an independent external dataset without duplicates, the architectures differed significantly in accuracy, while internal public-data testing did not distinguish them.
In briefAPO2CS: Advantage-Aware Preference Optimization for Two-Level Constructive Search in Combinatorial Optimization
SummaryThis paper proposes APO2CS, a preference-based training objective for constructive neural solvers used in combinatorial optimization. At the macro level, it adds an advantage term based on objective-value differences to pairwise preferences; at the micro level, it compares partial solution trajectories to supervise the sequential construction process more densely. The objective can train a policy from scratch or fine-tune a pretrained one. Experiments reported in the paper show effectiveness and adaptability across several combinatorial optimization tasks.
In briefFMA-Net: Robust Multimodal Face Attack Detection with Flexible Modalities
SummaryThis paper proposes FMA-Net, a framework for face presentation attack detection (identifying spoofed faces that try to fool recognition systems) that works even when RGB, depth, or infrared inputs are missing. It estimates missing modalities from RGB inputs through instruction-guided latent-space translation, and uses learnable text prompts, a facial feature adapter, and synthetic-to-real domain alignment to strengthen discriminative features. On WMCA with a 70% missing ratio, it reports a mean ACER of 3.22% across four intra-domain protocols, and 5.23% when depth and IR are both unavailable. Experiments are conducted on WMCA and PADISI-Face.
In briefHolidayFactory 1.0: A Framework for Simulating and Evaluating Industrial Robot Tasks
SummaryThis paper presents HolidayFactory 1.0, a framework for simulating, generating, and evaluating industrial robot work. It includes HOLIDAY-SIM, a physics simulator for contact, actuator, and tactile modeling, and an agentic method for building simulation-ready scenes and evaluations. The 24/7-BENCH benchmark defines seven industrial tasks evaluated without scene resets, measuring completed, failed, and unfinished jobs, throughput, latency, and motor energy.
Holiday Robotics: Jeongmin Lee, Taehyeon Kim, Dohyun Jang et al.
In briefPredicting population game dynamics when payoffs change
SummaryThe paper addresses how to predict the population trajectory produced by a new payoff design before data under that design are available. It proposes PA-RmD, a payoff-aware method that keeps payoff information explicit and learns the payoff-independent response structure from data. The method is evaluated on synthetic zero-sum and potential games and a YouTube Trending multi-country dataset, where it is reported to perform well under changing payoff designs.
In briefCABrain: A correlation-aware brain model for decoding EEG brain signals
SummaryThis paper presents CABrain, a foundation model for decoding electroencephalography (EEG) signals. It uses a graph convolutional network to build a conditional positional encoding that explicitly models correlations between EEG channels, and adds an adaptive masking strategy that adjusts how difficult masked patch reconstruction is during pretraining. Experiments on ten benchmark datasets show that these methods, alone or combined, achieve state-of-the-art or competitive performance.
In briefPlanners exploit deterministic error: dithered low-bit world models
SummaryThe paper studies low-bit quantized world models that speed up sampling-based robot planners such as MPPI and CEM. Without calibration, round-to-nearest quantization can collapse planning success, while one-step latent prediction error stays no larger than that of a stochastic rounding scheme that plans well. The authors attribute the failure to rounding error that persists across planner iterations and is shared across candidates. Redrawing shared dither at each iteration removes most of the damage without training.
In briefToward Trustworthy Mobile Voice Interaction: Verifying Speech Authenticity and Transmission Integrity
SummaryThis paper presents MoBiMark, a real-time watermarking system for smartphones that checks whether a voice captured during speech is from the genuine speaker. It emits ultrasonic probes, extracts vocal tract echo features, and combines them with speech content to embed a dynamic watermark containing a user identifier and a per-utterance identifier. The system supports both source authentication and detection of tampering during transmission. Tested with 50 users on four smartphone models, it averaged 99.79% accuracy with negligible impact on speech quality and intelligibility.
In briefWireless AI Networks: Jointly Adapting Communication Resources and Distributed Intelligence
SummaryThis paper proposes Wireless AI Networks, a design approach in which wireless communication and distributed AI adapt to each other. It uses an AI-executable resource budget to characterize candidate AI configurations, and task-level information value to guide their selection and wireless resource allocation. Two experiments show that update selection with statistical correction improves federated learning under the same model-update payload budget, and that sensitivity-aware power allocation reduces accuracy degradation under the same average transmit-power constraint.
In briefARMOR: Adaptive curriculum meta-learning for making retrieval-augmented reasoning robust to noise
SummaryThis paper addresses the problem that retrieval-augmented generation (RAG) systems pick up noisy documents during retrieval. It splits RAG noise robustness into three subtasks (knowledge boundary perception, retrieval verification, and robust reasoning) and proposes ARMOR, which uses an adaptive task scheduler that plans training according to the generator's learning progress. The authors report improved noise robustness and output quality over baselines on 2WikiMultiHopQA, MusiQue, and HotPotQA.
In briefMulti-view pose estimation using only rotations for imaging geometry
SummaryThis paper presents GRAM, which estimates global rotations directly from initial relative rotations and inlier matches, replacing separate pairwise pose refinement and rotation averaging with a single stage. The authors extend rotation-only imaging geometry to multiple views with MORE for nonlinear global rotation refinement, and ART, which recovers global translations via weighted least squares. Synthetic and real-world experiments demonstrate the methods' accuracy, robustness, and efficiency.
In briefLangDepth: Scene-Guided Depth Estimation from Single Images Using Vision-Language Priors
SummaryThis paper presents LangDepth, a monocular depth estimation model that first infers a compact scene descriptor from the image using a frozen CLIP encoder and learnable concept keys. A Scene Prior Module maps this descriptor to a scene-specific depth range, which a per-pixel assignment network then uses to place each pixel within the predicted bins. A confidence gate combines the dense visual estimate with this prior. The model is trained on NYU Depth V2 and KITTI and reports results across several indoor and outdoor benchmarks.
In briefBridgeKV: Using a Large Model for Prefill and a Small Model for Decoding in LLM Inference
SummaryThis paper proposes BridgeKV, which uses a larger language model to process the prompt (prefill) and a smaller model from the same family to generate output tokens (decoding). A lightweight module translates the large model's key-value cache into the small model's format, and only this module is trained while both models stay frozen. The authors report that BridgeKV beats fine-tuned small-model baselines with matched trainable parameters across six model pairs and five datasets. On one A40 GPU, Llama-3.1-8B to Llama-3.2-1B runs requests 3.0X faster than the larger model alone.
Payman Behnam, Debopam Sanyal, Pranav Ponnusamy et al.
In briefSpinSCI: A Mask-Conditioned Spin Transformer for Video Snapshot Imaging
SummaryThis paper addresses video snapshot compressive imaging, which recovers multiple video frames from a single spatiotemporally coded measurement. It proposes SpinSCI, a spatial-temporal Transformer whose mask-conditioned block combines attention and sensing information to refine features. Compared with HiSViT₉, it improves average PSNR on three benchmarks while using fewer parameters, FLOPs, and memory. Code and checkpoints are publicly available.
In briefSpectral rank and update direction in language-model post-training: how weight changes spread across dimensions and align over training.
SummaryThe paper examines how weights change during language-model post-training, reconciling reports of near full-rank updates with low stable-rank changes. It analyzes singular value decompositions of OLMo-2-7B updates and cosine similarities between consecutive checkpoint displacements across reinforcement learning stages. A controlled OLMo-2-1B study compares group-relative advantages with a learned critic using generalized advantage estimation, finding that the advantage method affects directional alignment.
In briefGeoPilot: Diagnosis-Guided Fixes and Geometric Checks for Aerial 3D Mapping
SummaryThe paper presents GeoPilot, a workflow that uses diagnostics to propose program or camera-setting changes for reconstructing 3D meshes from aerial imagery, then checks the results with an external geometric evaluator. On 224 development images, shared camera intrinsics improved all six native-coordinate measures and cut mean surface-to-reference distance by 7.04%. Reference-assisted alignment reversed that ranking, so the gain applies only to native-frame delivery.
In briefReproducible quantum clustering for single-cell transcriptome atlases
SummaryExisting single-cell clustering pipelines often discard most genes and can produce inconsistent clusters across random seeds. The paper introduces KQE-2, a training-free hybrid quantum-classical framework for reproducible whole-transcriptome clustering that maps cells into entanglement-free quantum states. The authors report that KQE-2 gave consistent biological results across runs, scaled to 11.4 million nuclei on one server, and outperformed compared methods on most of the atlases tested.
In briefAdaptive Hamiltonian Monte Carlo that cuts communication cost in federated sampling
SummaryThis paper proposes Adaptive Federated Hamiltonian Monte Carlo (AFHMC) for Bayesian federated learning. It adds a control mechanism for local node shifts and proves a communication cost of at most O({d/ε²}^{1/3}) up to logarithmic terms under log-convex distributions. When client heterogeneity is low, the cost can reach O(log(d/ε²)), claimed as the first logarithmic communication complexity for federated sampling. Numerical experiments support its advantage in the low-heterogeneity regime.
WhyThe authors add a local node shift control, reducing federated sampling communication to O({d/ε²}^{1/3}) and logarithmic under low heterogeneity, with numerical experiments.
In briefSilinx: an open browser-based HDL platform for teaching digital systems on FPGA boards
SummaryMany digital systems courses use FPGA boards whose vendor tool, Xilinx ISE, is discontinued and hard to install on current computers. This paper presents Silinx, a free HDL design platform that runs in a web browser with nothing to install and keeps ISE's Project Navigator workflow. It supports VHDL, Verilog, C/C++ via Silinx HLS, and graphical entry, and programs real boards over WebUSB. Compared with ISE at the same optimization goal, the open flow reports 15% fewer slices and 14% shorter critical paths for area optimization, and 21% fewer slices and 8% shorter critical paths for speed optimization.
In briefCreative Priming: Separating Process from Output to Avoid the Creativity Paradox
SummaryThe paper argues that AI models trained on past creative successes tend to reproduce those specific examples, which hinders genuine creativity. It proposes treating creativity as an explicit process and introduces creative priming, in which a trained model generates a prompt-specific primer that helps a response model produce creative answers. The primer model is trained with GRPO to reward primers that yield novel, high-quality, and diverse responses.
Alon Albalak, Sebastian Gabriel, David B. D'Ambrosio et al.
In briefDense optical flow via adaptive pixel-to-token correspondence (FATE)
SummaryThis paper proposes FATE, a method for estimating optical flow at every pixel of an image. It separates the visual regions used for matching from dense flow reconstruction, with regions expanding within coherent areas and contracting near motion boundaries, trained jointly without grouping labels. Trained only on FlyingChairs and FlyingThings3D, FATE reports the highest zero-shot accuracy on Sintel Final and KITTI, with its largest gains on occluded or out-of-frame pixels.
In briefCorrectness Metric Matters: NWD-Based Calibration Loss for Tiny Person Detection in UAV Infrared Imagery
SummaryThis paper addresses confidence calibration for detecting tiny persons in UAV infrared imagery. The authors argue that the IoU-based correctness target used by existing calibration losses is geometrically ill-posed for very small objects, causing binary prediction calibration to suppress confidences that should be high. They replace IoU with a scale-invariant normalized Wasserstein distance (NWD) as the target, which lowers localization-aware ECE across signal-to-clutter ratio groups on HIT-UAV while keeping mAP essentially unchanged.
In briefA theory of why smarter AI systems do not automatically become easier to control
SummaryThe paper argues that the adaptability that comes from emergence and the steerability needed for accountable deployment are governed by two distinct bifurcations in the same system. It concludes that increased capability does not imply increased controllability. The authors report quantitative results and calibrate their criteria against the two-dimensional Ising transition.
In briefCan missing ratings flip a personalized model's prediction? A worst-case stability check
SummaryThe paper asks whether unobserved ratings could change a personalized rating model's current prediction. For completed models with additive response scores, it computes exact worst-case bounds that check all possible missing-rating assignments. The method is evaluated retrospectively on 108 clinicians rating three simulated psychiatric interviews and on 126 counseling demonstrations annotated by ten experts. The authors report lower log loss for Fisher adaptation based on observed ratings and certified coverage levels, and note the certificate shows agreement only with the completed model, not intermediate correctness or clinical usefulness.
Priyanka Siddappa, Muntaqim Ahmed Raju, Shifat Haider Al Amin et al.
In briefParticle filtering with generative trajectory models for robot state estimation and control
SummaryThis paper examines whether a pretrained generative diffusion model that jointly represents states, actions, and observations can improve robot state estimation and control without being modified. The authors use particle filtering to maintain multiple hypotheses about state histories, each with predicted observations and action proposals. Particles are reweighted by how well their predicted observations match new measurements, and these weights are also used to guide action selection. Simulations in partially observed navigation and vision-based manipulation show gains in state estimation and task performance over action generation without filtering.
In briefState Media at Retrieval Time: An Audit of Russian State-Aligned Sources in AI Chatbots
SummaryThis paper audits 11 leading AI chatbots to see how often they retrieve articles from the Pravda Network, a Russian state-aligned media aggregator with over 110 websites. Using a 10,000-prompt corpus built from sampled headlines, the authors collected 104,997 completed responses. They find Pravda Network articles appear in 7.3% of responses, concentrated in three platforms, while all 11 chatbots retrieved other Russian state-aligned media. The paper also describes “headline washing,” where some headlines credit other outlets for reporting that was republished from state-aligned sources.
WhyThe authors audit eleven AI chatbots on 104,997 responses, reporting Pravda Network retrieval in 7.3% of answers and identifying a headline-washing tactic.
Laura Kurek, Elijah Covert, Rafael I Bonilla et al.
In briefClinical Data's Dark Matter: Limits on Separating Physiology from Care Process in EHRs
SummaryThis paper argues that representations learned from electronic health records mix patient physiology with the clinical process that determines what is measured, when, and for whom. The authors formalize the data-generating process and analyze two strategies for separating these components. They show that a single dataset cannot identify how information from laboratory values and orderings is divided, and that cross-site invariance tests are falsifiable hypotheses rather than identification methods. Identification becomes possible under certain conditions, such as when tests are performed for every patient.
WhyThe authors formalize EHR data generation and prove that physiology and ordering are not identified from a single dataset, with a worked example reporting a 9.6-point cross-site gap.
Mohammad Shahin, Krittaphas Chaisutyakorn, Dimitrios Proios et al.
In briefThe Value of Not Knowing: How Naive LLMs Foster Creativity
SummaryThis paper introduces naive large language models, deliberately made ignorant of a specific topic through training deprivation or intentional forgetting. Freed from existing dogma, these models can make new associations and conjectures. They are paired with a stronger mentor model that refines their ideas into concrete proposals. For 1B and 7B models, naive models produced seed ideas that were more novel and useful than baselines, and mentor refinement improved them further.
David B. D'Ambrosio, Alon Albalak, Joel Lehman et al.
In briefExplaining and stopping group-level misalignment among language model agents
SummaryThis paper examines how language model agents in a group can exchange individually harmless messages that together lead to a policy violation. The authors propose Collective Dependence Analysis (CDA), which uses necessity and sufficiency to identify which messages drove a violation and the fewest removals that break that support. In controlled cases, removing selected messages reduced violations from 93% to 6%.
In briefFaster reinforcement learning for soft robots using a compact learned model
SummaryThis paper proposes training reinforcement learning policies in a low-dimensional polynomial model, learned from pre-collected data, instead of costly hardware interaction or a simulator. On a pneumatically actuated soft arm, a trajectory-tracking policy was trained in 3 hours and deployed without fine-tuning, a process the authors say would take 3 years of hardware interaction. The policy achieved 40% lower tracking error than policies trained on RNN or transformer models.
Roshan S. Kaundinya, Hehui Zheng, Robert K. Katzschmann et al.
In briefContinual learning and knowledge transfer with brain-inspired neural assemblies
SummaryArtificial neural networks tend to overwrite earlier solutions when trained on new tasks, a problem called catastrophic forgetting. This paper proposes Functional Task Networks (FTNs), a continual learning method modeled on properties of the mammalian neocortex, which computes a sparse binary routing mask from a small support sample at inference time. Unrelated tasks route through mostly separate subnetworks while related tasks share components. Experiments show that support-set mask reconfiguration is needed for retention under correlated concept shift, while a distributed network with replay suffices under domain shift, and the approach also yields forward and backward transfer.
In briefSteering Collective Behavior in LLM Agent Groups by Designing Who Observes Whom
SummaryThis paper studies groups of large language model agents where an operator controls which agent observes which other agent, but not their payoffs or messages. Across 1,080 episodes of six repeated group-decision games with ten identical agents, the authors compare how much information to share and how to wire observation links under a fixed budget. They find that the welfare effect of information depends on whether individual incentives align with the collective goal, and they argue against defaulting to full transparency.
WhyThe authors treat the information network as a design object and report that its wiring reverses network-reciprocity intuitions from evolutionary game theory.
In briefA better way to weight data from several logging policies when evaluating bandit decisions offline
SummaryThe paper studies off-policy evaluation in contextual bandits using data from multiple logging policies. It derives an unbiased inverse propensity scoring estimator with sample-dependent weights that minimize variance, obtained in closed form via variational calculus. Experiments on benchmark datasets show it generally improves on existing multi-logger IPS methods, and a doubly robust extension is compared with the estimator of Kallus et al. [2021].
In briefCRUSTPairs: A dataset of matched C and Rust programs for cross-language comparison
SummaryThis paper presents CRUSTPairs, a dataset of 1857 pairs of C and Rust implementations of the same competitive programming solutions. The pairs are normalised, deduplicated, and grouped into difficulty tiers by cross-language embedding similarity, with provenance and licence recorded per item. Experiments show cross-language retrieval is difficult, with a lexical baseline beating a pretrained code model, while finetuning substantially improves accuracy.
Vojtech Hejlek, Leonardo Jesus Almeida, Alneu de Andrade Lopes
In briefCASCoDe: A closed-loop AI agent framework that co-designs analog circuit sizing and layout
SummaryCASCoDe is a closed-loop framework in which a large language model agent turns a circuit netlist into a verified layout. It models the sequential design decisions that experts make, such as sizing, floorplanning, local refinement, and routing choice, as a hierarchical discrete action space searched with feedback-driven tree search. On 20 SKY130 amplifiers, it was the only flow to produce DRC- and LVS-clean routed layouts that meet specifications for every circuit, reducing area by 71% on average against a conventional place-and-route flow and raising mean device density from 0.39 to 0.74.
WhyThe authors formalize expert layout decisions as a hierarchical discrete action space with feedback tree search, reporting about 71% mean area reduction on 20 SKY130 amplifiers.
In briefPrecision control of soft helicoid robots using data-driven models
SummaryThis paper extends a data-driven method based on adiabatic spectral submanifolds (aSSMs) to learn reduced models from data collected on a hardware soft robot. The authors learn a five-dimensional nonlinear ODE that models the dynamics of a soft-helicoid trunk robot and use it in model predictive control for real-time 2D and 3D trajectory tracking. The method is reported to improve accuracy and precision over data-driven linear methods and model-free reinforcement learning.
Roshan S. Kaundinya, Mark Leone, Roberto Sabella et al.
In briefFaster Time Series Counterfactual Explanations via Learned Policies
SummaryThis paper proposes a policy learning framework for generating counterfactual explanations for time series, framed as a one-step contextual-bandit problem. An actor-critic agent is trained in the latent space of a pre-trained autoencoder to learn a transferable policy that captures global structure. Once trained, the policy produces valid and plausible counterfactuals in a single forward pass without per-instance optimization. Experiments on 13 UCR benchmark datasets report high validity and strong plausibility across multiple anomaly detection metrics.
Peiyu Li, Pouya Hosseinzadeh, Soukaina Filali Boubrahimi et al.
In briefSurgTAP-3D: A Large Surgical Video Dataset for 3D Motion Tracking
SummaryThis paper presents SurgTAP-3D, a large-scale pseudo-3D motion tracking dataset and benchmark for minimally invasive surgery, comprising 520K frames, 4,549 action clips, and 1.4M 2D/3D point trajectories from robotic and manual procedures. The authors introduce a framework that converts uncalibrated monocular surgical videos into structured 4D annotations for scalable labeling. They establish benchmarks on SurgPose, SurgT, and STIR covering both instrument and tissue tracking.
In briefDirect GPU energy measurements of batch sharing in mixture-of-experts and recurrent models
SummaryThis paper directly measures GPU energy use for two workloads: recurrent policies in multi-agent reinforcement learning and mixture-of-experts (MoE) language models. It asks whether rows within a batch can share computation to save energy. Recurrent policies showed no measured saving. On Qwen3-30B-A3B, batch-aware routing read 25-31% fewer expert weights within KL 0.05 and used 12-18% less energy per token. The authors state these are pilot measurements on one GPU type.
In briefA study on how language model abilities fail in a predictable order under compression and perturbation
SummaryThis paper examines whether losses in language model abilities caused by compression or perturbation follow a consistent order. The authors applied seven perturbation and compression methods to eight instruction-tuned models and scored 16 tasks. They report that short-answer capabilities fail in nearly the same order across different methods, and that this order can be predicted from factors such as intact answer margin and answer frequency.
In briefTopological Obstacle to Stably Learning First-Order Nash Equilibria
SummaryThis paper studies how to stably learn local Nash equilibria (first-order Nash equilibria, FONE) in constrained non-concave differentiable games. The authors show that no continuous deterministic learning dynamics can make the full FONE set a stable global attractor when the target set has nontrivial positive-degree cohomology and the state space is compact and contractible. The barrier is demonstrated in a simple non-concave quadratic zero-sum game on a box.
WhyThe authors show homological obstructions block stable learning of first-order Nash equilibria, with a six-edge loop quadratic game example extended to higher degrees.
Ashkan Soleymani, Gabriele Farina, Patrick Jaillet et al.
In briefLocateAnyMesh: Using vision-language models to find 3D mesh queries in 2D images
SummaryThis paper introduces LocateAnyMesh, a vision-language model that uses a textured 3D mesh as part of its prompt to locate the matching object in an image. A frozen 3D encoder converts colored surface points into a compact token sequence that is fed alongside image and instruction tokens. New objects need only a single encoder pass and no object-specific fine-tuning or rendered query views.
In briefResolvent ellipsoid method for solving Minty set inclusion problems
SummaryThis paper studies Minty set inclusion on a compact convex set with a set-valued operator, aiming to find a point and an operator value satisfying a specified error bound. In a fixed-scale constrained resolvent oracle model, each oracle response either certifies an approximate solution or provides a valid geometric separating normal. Combining this separation principle with the ellipsoid method, the authors propose RESOLVENT-ELLIPSOID, which computes an approximate solution in polynomially many oracle calls.
Ashkan Soleymani, Gabriele Farina, Patrick Jaillet
In briefPhysical Testbed for Self-Driving Labs Comparing Bayesian and Agentic Optimization
SummaryThis paper presents a physical testbed built from commercial lab automation instruments that runs closed-loop experiments autonomously for 24 cumulative hours. The system formulates aqueous dye mixtures to match a target absorbance spectrum, with AI deciding each next formulation without human input. Traditional Bayesian optimization and an agentic optimizer were compared, and the agentic approach matched targets in fewer iterations, including one target Bayesian optimization did not match within 48 iterations.
In briefStreaming Failure Forecasting: A Failure Monitor That Keeps Learning After Deployment
SummaryFailure monitors for vision-language-action robot policies are normally trained offline and then left frozen, even as conditions and policies change after deployment. This paper proposes Streaming Failure Forecasting (SFF), which refits only the last linear layer after each episode using the success label that every episode provides. Across simulated test suites and 434 recorded runs of four VLAs on real SO-101 arms, SFF ranks failures better than the same monitor kept frozen.
Amit Jadhav, Shaurya Beriwala, David Q. Liu et al.
In briefSelf-Indexing Latent Memories: An Access-Aware Method for Continual Instruction Tuning of LLMs
SummaryThe paper introduces TILA-CT, a replay-free method for continual instruction tuning of LLMs in which each task has its own latent memory and adapter bank within a frozen model. The memory supplies a routing prototype that a shared prompt-only router uses to select the correct task module. On an 8-task stream, the authors report a final macro score of 63.98 and average forgetting of 1.38 points, with final macro routing accuracy of 93.9%.
Dhruv Dixit, Paritosh Pandey, Snehalkumar 'Neil' S. Gaikwad
In briefA sequential Monte Carlo method for jointly estimating layered geological uncertainty in underground flow data assimilation
SummaryThis paper presents an adaptive-tempered sequential Monte Carlo method for jointly estimating higher-level geological metaparameters and cell-by-cell properties in subsurface flow models, using flow surrogates for function evaluations. The approach is extended to successive data assimilation and tested on two 3D systems, a faulted carbon storage system and a channelized oil-water system. For four synthetic true models, the posterior distributions and flow predictions closely match reference MCMC results. The method achieves speedups of 2.5 to 7.0 for single-stage data assimilation, with greater gains on multiple GPUs.
Wenchao Teng, Guido Di Federico, Louis J. Durlofsky
In briefA theoretical reduction showing membership inference leads to data reconstruction attacks on learning algorithms
SummaryThis paper studies reconstruction attacks, which recover training data from a learned model. The authors show that, in their adversarial model, membership inference implies reconstruction for any learning algorithm. For stochastic convex optimization, they make the reduction computationally efficient over exponentially large domains. Combined with an existing result, they prove that low-excess-risk learners admit polynomial-time attacks recovering Ω(n) training points.
WhyThe authors give a general reduction from membership inference to reconstruction, and a polynomial-time attack on stochastic convex optimization recovering Ω(n) training points.
In briefThe Average Patient: Do Simulated Psychiatric Patients Answer Like Real People?
SummaryThis paper examines using language models as simulated psychiatric patients to train clinicians and evaluate AI therapists. The authors propose a framework and compare simulated respondents with item-level data from eight human samples, including clinical and instructed-feigning groups. Simulated patients vary far less than real respondents with the same totals and endorse nearly every symptom at a middling level on most questionnaires. Prompt changes, larger models and reasoning reduce this pattern without removing it.
WhyThe authors propose an average-patient hypothesis and information-budget framework for simulated psychiatric patients, testing it across eight human samples and 17 models.
In briefSurfScaff3D: Linking geometry and meaning for LLM-based 3D object generation
SummaryThis paper introduces the Hierarchical Patch Scaffold (HPS), a 3D representation built from part-level abstractions that both humans and large language models can interpret. Building on HPS, the authors present SurfScaff3D, an interactive system for generating 3D objects from text or images. The hierarchical part labels allow natural-language edits at any level without regenerating the model, and per-patch control over surface detail is supported. Experiments against existing LLM-based methods report strong geometric fidelity, semantic alignment, and flexible editing across diverse object categories.
In briefQuery- and candidate-aware long user sequence modeling for sponsored product ranking in search
SummaryThis paper presents a production framework for ranking sponsored products in search that models long-term user behavior sequences. It retrieves relevant past actions using two channels, candidate-nearest-neighbor and query-nearest-neighbor retrieval, and encodes them with query and candidate context. Offline ablations show the dual retrieval improves major ranking metrics over the production baseline. An online A/B test on sponsored-product search traffic reported gains in CTR, attributed conversion rate, and orders per impression.
Dongyue Xie, Sergey Kolbin, Manavender Malgireddy et al.
In briefBrain-Aligned Latent Actions: Shaping a Robot's Action Codes with Monkey Premotor Cortex
SummaryRobot policies increasingly act through a small set of discrete latent actions learned from unlabeled video, but which motion receives which code is unstable during training. The authors impose the representational geometry of the macaque ventral premotor cortex, recorded while a monkey drew shapes, onto the tokenizer's code assignment using a straight-through similarity loss trained alongside reconstruction on robot video. On five unseen shape categories, the resulting tokenizer reaches a representational similarity of 0.52 with PMv, compared with 0.20 to 0.26 for released and no-prior tokenizers.
In briefNonasymptotic causal structure learning with continuous optimization in generalized linear DAGs
SummaryThis paper studies learning directed acyclic graphs from observational data under nonlinear models, with finite-sample nonasymptotic statistical guarantees. Each node follows a generalized linear model, and the authors analyze a global nonsmooth objective that combines an exact acyclicity constraint with a quasi-minimax concave penalty, which standard gradient-based solvers can optimize. Finite-sample guarantees are established for this setting, including the widely used logistic and Poisson models. Experiments check structure recovery accuracy and computational performance.
