In briefWrite Once, Check Many: Reusable Ledgers for Auditing Agent Trajectories Against Policies
SummaryThis paper proposes a framework in which a writer model converts each LLM agent trajectory into a policy-independent ledger of actions and outcomes, and a reader model then checks new policies using only that ledger. Tested on 900 trajectories from SHADE-Arena, τ-bench and τ²-bench, a 12B reader using narrative ledgers matched or exceeded the AUC of the same model reading full trajectories. When checking eight SHADE-Arena policies separately, ledgers used about 5 times fewer tokens, including writing costs. The authors note that rule-based extraction sometimes performs better on single-action policies, and ledgers can miss evidence in long tool outputs and message contents.
Ashwin Nagarajan, Gursimran Panesar, Alicia Yang et al.
In briefRiSE: A residual subspace expert model for generalizable text-centric image forgery localization
SummaryThis paper addresses poor generalization of text-centric image forgery localization detectors, which rely on full-parameter fine-tuning and struggle with unseen forgery types. The proposed RiSE model freezes principal SVD components, adapts only residual subspaces, and trains independent residual experts for specific forgery types. At inference, Latent Perturbation Confidence selects the most confident expert in a single forward pass. The authors report that a single expert trained on synthetic data outperforms prior methods by 20.4% on real-world images with 10 times fewer training steps, and nine experts yield a 35.8% gain.
In briefBeyond Recall Saturation: Scaling Hard-Negative Reranker Training for Legal Search
SummaryThis paper studies reranker training for Chinese legal retrieval, where embedding recall shows diminishing returns at large candidate pools. The authors compare hard-negative mining strategies and find that an embedding-to-self-mined two-stage pipeline offers a favorable trade-off. The method scales from a 0.6B to a 4B reranker, which the authors report outperforms the native 4B model on several benchmarks and an internal preference evaluation.
In briefBen-Misandry: A Bengali dataset and framework for detecting hate speech against men
SummaryThis paper notes that Bengali gender-based hate speech research has focused almost only on misogyny, leaving anti-male hate (misandry) unstudied. The authors introduce Ben-Misandry, the first Bengali misandry detection dataset, with 5,000 comments across six classes. The data is synthetic, generated with GPT-4.1, and spot-checked by four human reviewers on a 500-comment sample. A late-fusion ensemble of BanglaBERT and XLM-RoBERTa reaches 96.6% accuracy, versus 91.8% for the best classical model, though significance testing was not run and synthetic-generation artifacts may partly drive results.
Md Mehedi Hasan Nipu, Chinmoy Mitra, Tarannum Ahmed Nowshin et al.
In briefPluralistic Alignment in Language Models: Goals, Human Reference Groups, and Preference Representation
SummaryThis paper reviews research on pluralistic alignment, which aims for language models to reflect varied human preferences rather than merging them into one answer. The authors propose a two-layer taxonomy: the first layer separates representing a range of viewpoints, faithfully adapting to a specified perspective, and matching a target population's response distribution, while the second layer distinguishes aggregate populations, countries and regions, and subgroups as human references. Based on a corpus of 70 works, they map existing approaches and link each objective to the evidence needed to support it.
In briefWhich Long Documents Work Best? A Controlled Study of Data Selection for Extending Context Length
SummaryThe paper compares data selection methods for extending a language model's context window, using Qwen2.5-3B under a matched 6B-token continued-training budget. Randomly sampling natural long documents outperforms repacking, and selecting documents with high LongPPL key-token density gives further gains. The authors propose truncLongPPL, which scores only a truncated prefix to cut scoring cost, reporting a 2.28x end-to-end speedup at 128K with performance close to full-span scoring. Results are reported across several models and benchmarks, including HELMET, RULER, LongBench v2 and MRCR.
In briefSENTRY: A real-time audio deepfake detection engine for network edge devices
SummaryThis paper presents SENTRY, a streaming detection engine designed to run inline at the network edge to identify deepfake speech. Its feature front end uses only integer add, shift, compare, and bitwise operations, and it combines a supervised gradient-boosted tree classifier with a Mahalanobis out-of-distribution detector trained on authentic speech, fused by a logistic layer. The authors report high accuracy and low equal-error rates on WaveFake and ASVspoof 2019 LA, with a prototype using no GPUs or floating-point hardware. Generalization to unseen speech generators is left to future work.
Chirantan Banerjee, Soumita Chatterjee, Bhargab B. Bhattacharya et al.
In briefJoint amplitude-phase coupling and reliable endogenous priors for low-light image enhancement
SummaryThis paper addresses low-light image enhancement, noting that existing frequency-domain methods restore amplitude and phase independently and ignore their spectral coupling, causing inconsistency and distortion. The authors propose a frequency-spatial framework with an Amplitude-Phase Adaptive Compensation Fusion module for bidirectional interaction and an Endogenous Reliability Prior extracted from an intermediate restored result. A Dual-Modal Guided Refinement Network injects these cues to recover fine textures. The authors report state-of-the-art results on benchmark datasets in both quantitative metrics and visual quality.
In briefLearning to Compose: Self-Improving Shared Robot Skill Stacks for Zero-Shot Long-Task Manipulation
SummaryThe paper introduces Recursia, a framework that improves a shared executable stack of robot tools, skills, handoffs, and verifiers rather than individual task programs. Execution traces are used to attribute failures and generate fixes, which are kept only after generality and regression checks, then reused via declarative task adapters. Reported experiments on LIBERO-Pro Long simulation and a physical Franka FR3 show higher success than ASPIRE and a Skill Document Only baseline.
In briefParametric session memory that recovers retrieval intent without replaying chat history.
SummaryLong-running assistants must interpret follow-up requests whose referents and constraints were set many turns earlier, so the core difficulty is retrieval intent rather than context length alone. The authors propose Parametric Session State, a reusable session-memory layer that remembers what to search for while keeping factual knowledge external. The approach aims to avoid the cost of repeatedly reprocessing conversation history.
In briefO-SCD++: Replay-Free Online Scene Change Detection via Recursive Evidence Fusion
SummaryThis paper addresses online scene change detection, where an agent must spot changes in a previously mapped environment while on site, using only past and present observations. The authors propose a replay-free method that carries each new observation onto the reference map's Gaussians with a single renderer backward pass, reducing multi-view fusion to an exact two-scalar recursion per Gaussian. On PASLCD, mean IoU rises from 0.489 to 0.575 and throughput from 11.8 to 25.6 frames per second.
Chamuditha Jayanga Galappaththige, Jason Lai, Lloyd Windrim et al.
In briefEstimating empowerment from offline data to guide intrinsically motivated exploration
SummaryThis paper studies how unlabeled prior experience data can make online reinforcement learning exploration more efficient. It proposes the first method to estimate empowerment, a measure of how many future states an agent can reach, directly from unlabeled offline trajectories without environment interaction. The approach jointly learns an empowerment-maximizing policy and its discounted state-occupancy measure. The authors report that the estimator gives intuitive empowerment values and that the pretrained estimator improves exploration in online goal-conditioned reinforcement learning.
In briefGenerative retrieval for finding protein targets of molecules, with a proteome-wide benchmark
SummaryThe paper reformulates reverse target fishing, which finds protein targets for a query molecule, as generative retrieval and introduces GenFish. It converts frozen protein language model features into semantic identifiers and ranks proteins by autoregressively generated identifier likelihood. The authors also release ProTargetBench, covering 20,431 human proteins. On 198 held-out molecules, GenFish outperforms four discriminative neural and three non-neural baselines on three full-catalogue retrieval metrics.
In briefTree Plurality Veto: A Few-Round, Low-Communication Deterministic Voting Protocol with Low Distortion
SummaryThe paper studies metric distortion in voting, where voters and candidates sit in an unknown metric space and the rule uses only ordinal preferences. It introduces TreePluralityVeto, a deterministic protocol that for every r ≥ 2 achieves distortion at most 2r−1 using at most r rounds and O(n log m + n s^{1/(r-1)} log s) bits, with s = min{n, m}. The protocol runs PluralityVeto on a hierarchical partition of voters, and the branching factor trades off rounds against communication.
WhyProposes a tree-structured veto protocol trading rounds against communication, with a matching analysis bounding distortion per level.
In briefMELD: Recursive self-improvement through hierarchical coordination for multi-agent scientific discovery
SummaryThis paper presents MELD, a framework for scientific discovery using multiple large language model agents. It jointly adapts a team-level Coordinator policy and individual Scientist policies within one discovery process, so team organization and individual research behavior improve together. On 22 MLGym and MLE-bench tasks, MELD performs best on 20 of them and achieves the highest aggregate normalized scores on MLGym.
In briefHippo: Training-free agent self-improvement using hippocampus-inspired memory replay
SummaryThis paper studies how LLM agents can learn from experience through external memory without updating model parameters. The authors propose Hippo, a memory framework that compares multiple attempts at the same task to identify reusable decisions, then merges overlapping experiences and resolves conflicts using explicit applicability conditions. Experiments on SWE-bench Verified, WebArena, and Mind2Web show performance gains in all three settings, with substantially fewer retained experiences than the simple accumulation baseline.
In briefHumanSkill-4D: Training-free language selection of body regions and motions in animatable Gaussian avatars
SummaryThe paper introduces avatar referring segmentation, which uses language queries to select regions of animatable Gaussian avatars and, optionally, time intervals matching a specified pose or motion. It proposes HumanSkill-4D, a training-free framework that combines a multimodal large language model with geometric operators, using a semantic memory for body-part and garment regions and a motion memory for driving joint trajectories. The authors introduce three human-annotated benchmarks and report that the method outperforms baselines in part segmentation and 3D and 4D referring segmentation, with compact memories and only minutes of setup per avatar.
In briefDo 3D Large Language Models Really Use 3D Evidence? A test of whether 3D AI models answer scene questions based on actual visual evidence.
SummaryThis paper introduces UNICORNS, a benchmark of 11,387 questions over 557 ScanNet scans that tests whether 3D large language models answer reliably under ambiguity, missing evidence, or false presuppositions. Evaluating six open-source 3D-LLMs, the strongest zero-shot model reaches 61.22% accuracy on answerable controls but only 40.94% across the eight challenge categories. The authors also propose HORN, a post-training method, which reaches 75.1 ± 1.0% on the challenge categories and 77.2 ± 0.8% overall without hurting answerable-control accuracy.
WhyIntroduces a new benchmark and training method for evidence-grounded 3D-LLM reliability, with 11,387 questions and six models compared.
In briefNear-optimal regret bounds for learning monotone stochastic problems from semi-bandit feedback
SummaryThis paper studies learning an approximately optimal policy for stochastic optimization problems with unknown distributions, where each period the algorithm observes samples only from the variables its policy selected (semi-bandit feedback). It gives an algorithm for a broad class of monotone problems, including prophet inequality, Pandora's box, and stochastic knapsack, with regret Õ(√(nkT)) for weakly monotone and Õ(√(nT)) for strongly monotone problems. These bounds improve prior results by a √n factor and match the known lower bound for multi-armed bandits. The main technique bounds the error between true and learned distributions using Hellinger distance.
In briefMemory Internalization via Test-Time Training for Long-Video Understanding
SummaryThis paper proposes MIT^3, a framework that trains compact, video-specific parameters for each long video so models can answer questions without the video itself. It uses causal tracing to choose which layers should store video knowledge, and distills the representation change caused by relevant video clips into a text-only memory. The authors report that across LVBench, LongVideoBench, and Video-MME, MIT^3 needs substantially fewer synthesized samples than prior parameterized video-memory methods.
In briefUncertainty Gates Gradient Updates for Few-Shot 3D Point Cloud Segmentation
SummaryThis paper addresses generalized few-shot 3D point cloud segmentation, where support annotations are often sparse and occluded, making low-quality samples risky to adapt to. The proposed Uncertainty-Gated Adaptive (UGA) framework uses an Uncertainty-Gated Update module that estimates episode reliability from intra-class dispersion and selectively gates residual gradient updates. It also includes an Uncertainty-Based Pseudo Label Filter to keep only high-confidence dense signals. On ScanNet200 and ScanNet++ under 1-shot and 5-shot settings, UGA reports an average gain of +3.08 HM IoU over the strongest prior baseline.
In briefTransformers as adaptive kernel learners for in-context learning across tasks
SummaryThis paper studies how transformers perform in-context learning by viewing the attention mechanism as an adaptive kernel method, since its query and key matrices are trained. The authors run numerical simulations on low-rank linear and multi-index nonlinear regression, finding that learned attention scores closely resemble oracle kernels. They also give an explicit construction of oracle kernel estimators using a wide, shallow transformer and analyze the generalization error of in-context learning.
Zhaiming Shen, Alexander Cloninger, Rongjie Lai et al.
In briefBodyBridge: Grounded body layers for realistic pedestrian and cyclist video translation
SummaryThis paper addresses how driving-video diffusion models represent pedestrians and cyclists as generic object classes, which fails to capture their pose and identity. It introduces Grounded Body Layers, a simulator-derived, pixel-aligned representation combining scene geometry, canonical body correspondence, and exact VRU masks. On CARLA-to-real translation, BodyBridge improves near-range PCK@0.5 by 22% and reduces body flips by 18% and near-range facing flips by 27%.
In briefSingle-bit distributed optimization for scheduling and resource allocation
SummaryThis paper proposes a distributed resource allocation algorithm in which each agent sends only the sign of its local cost gradient, one bit per directed link per iteration, while keeping the demand constraint satisfied at every iterate. In simulations on economic dispatch with 1000 generators and CPU scheduling with 500 servers, the single-bit protocol reduces total communication cost by 5 to 60 times compared with multi-bit baselines, at the cost of a tunable, explicitly bounded optimality gap. The authors note that a diminishing step size empirically eliminates this gap, but its formal convergence analysis is left as an open problem.
In briefAdapting Only the Conduction-Shared Components: Source-Free Geometric Transport for Holter and Patch ECG Patient Adaptation
SummaryThis paper addresses ECG beat classifier accuracy loss on new patients and devices by proposing NAGT, a label-free method that estimates a symmetric positive-definite congruence map from normal beats in the first 5 minutes of each recording and applies it only to normal and supraventricular prototypes. On 493 patients (973 recordings) from Icentia11k chest-patch data, arrhythmia macro-F1 rose from 34.5% without adaptation to 48.7%, outperforming AdaBN, Tent and T3A. However, the gain was not significant in the 75-patient historical-split test subgroup.
In briefWhen Opposite Premises Translate Identically: A Diagnostic Test for Translated Commonsense Questions
SummaryThe paper examines whether translated English premises keep their original logical relations. The authors reverse the polarity of ACCORD premises, translate both versions separately, and count cases where the target-language outputs are identical, then score each version against its own source. Many identical translations occur, for example in Swahili, and some pass round-trip NLI checks, but these cases show only that at least one translation is unfaithful, not which one. The authors note that the results do not establish detection accuracy for spontaneous translation errors.
In briefA distilled, universal two-layer feature extractor to replace speech CNN layers
SummarySelf-supervised speech models usually use convolutional neural network (CNN) layers as a local feature extractor before transformer layers. The authors hypothesize these CNN layers can be replaced by a simple universal model, and they learn a two-layer extractor by distilling from teacher models HuBERT, data2vec, and WavLM. On SUPERB(-SG) benchmark tasks, the model largely retains teacher performance and also works with the unseen wav2vec 2.0 model. It needs only 3% of the CNN's MACs, reducing overall inference runtime by over 20%.
Jessica Yang, Marcelo Beramendi Caballero, Ju-Chieh Chou et al.
In briefAdding momentum to distributed energy resource optimization under delays and quantization
SummaryThis paper studies a distributed discrete-time optimization method that combines heavy-ball momentum, communication delays, and logarithmic quantization, while keeping total supply and demand balanced at every iteration. Using a Lyapunov-Krasovskii analysis on a connected undirected network with bounded delays, it establishes practical stability and shows exact convergence to the smooth penalized optimizer when quantization is removed. Stability regions and optimal parameters are derived for delay-free quadratic costs and checked on a sparse network.
In briefA new method for learning from combined multimodal health data
SummaryThis paper proposes Relational Tensor Representation (RTR), a framework that jointly learns representations for each data dimension and the relationships across dimensions and modalities. It uses learnable bridge matrices and a structured fusion step to combine factor-level and relationship-level information. On simulated data and two neurophysiological datasets for emotion recognition and brain activity classification, RTR outperforms tensor, fusion, and Transformer baselines.
In briefS3D-Mosaic: Subgradient-guided structure transfer across spectral bands for spectral demosaicing
SummaryMultispectral filter array cameras capture spectra in one snapshot, but sparse per-band sampling causes strong spatial aliasing. The paper proposes S3D-Mosaic, a model-based method that transfers structure across neighboring bands using a subgradient-guided regularizer, a compact directional convolution (DSE-DConv), and a dual-scan ADMM solver. According to the authors, it outperforms state-of-the-art unsupervised methods on CAVE and NTIRE 2022 and generalizes to different filter layouts and real UAV data.
In briefAsking Less, Asking Better: Budgeting User Attention for Proactive LLM Agents
SummaryThis paper studies how clarifying questions from LLM agents burden users. The authors propose an evaluation protocol that limits the number of questions with an interruption budget and charges a user-attention cost for hard-to-read or unnecessary questions, judged by a human-validated LLM judge. In experiments across eight LLMs and four customer-service domains, capping questions at one to three cut attention cost by 36-73% while keeping 86-94% of task reward.
Xuanming Zhang, Amir H. Rezaeian, Miguel Ballesteros et al.
In briefLong Listening Thoughts: Training Audio Models to Reason More Deliberately
SummaryThis paper introduces Long Listening Thoughts (LLT), a framework for building post-training datasets that improve audio reasoning in large audio-language models. Its four-stage pipeline produces multi-aspect audio descriptions, harder questions, incorrect reasoning traces, and longer traces that include behaviors such as verification and backtracking. Using 1.6M synthesized traces, the authors fine-tune Qwen2.5-Omni-7B with supervised fine-tuning followed by direct preference optimization, reporting gains on MMAU Test-mini, MMAR, and MMSU.
Jaeyeon Kim, Chao-Han Huck Yang, Luoyi Zhang et al.
In briefZeroDSS: Image Segmentation Trained Only from Class Names
SummaryThis paper introduces ZeroDSS, a method for semantic segmentation that uses no real target-domain images, masks, or captions, only a predefined set of class names. It generates structured prompts with an LLM, creates image-mask pairs using diffusion attention, and iteratively refines the synthetic training data through feedback-driven two-phase training and a teacher-student framework. Using only synthetic data, the approach reports 35.0% mIoU on MS COCO with DeepLabV3 and 66.3% and 63.5% mIoU on PASCAL VOC with DeepLabV3 and Mask2Former.
In briefSpatons: Efficient Building Blocks for Correcting Distortion in XR Displays
SummaryThis paper presents Spatons, compositional, spatially localized neural operators that pre-distort images to correct geometric and colour distortions in AR, VR, and MR displays at low computational cost. Each Spaton combines tiny MLPs with fixed-function interpolation. On the full spatiotonal scenario, the Sequential-5 pipeline matches CompenNet++ in SSIM and nearly matches its PSNR, with far fewer FLOPs and parameters. Chromatic dispersion remains the main weakness, and testing used closed-form distortions and one real capture set.
Mohammadjavad Maheronnaghsh, Sushant Kondguli, Steve Mann et al.
In briefMSCR: A Multimodal Regularizer That Balances Modality Use and Finds Synergistic Information
SummaryMultimodal models often let one modality dominate, which limits useful cross-modal interaction. The paper proposes MSCR, a fusion regularizer that suppresses modality dominance and encourages the fused predictor to outperform a non-synergistic unimodal reference. Across five benchmarks, MSCR improves over MCR, raising conditional recovery and the share of output attributed to cross-modal interactions.
In briefTrust the Edge, Doubt the Arrow: Auditing Validation Signals for Text-Mined Clinical Causal Graphs
SummaryThe paper builds a directed candidate causal knowledge graph from 328,839 MIMIC-IV discharge summaries using LLM extraction and UMLS normalization. It evaluates co-occurrence lift, diagnosis-coding temporal order, and literature corroboration against SemMedDB. Lift filters unsupported co-mentions but does not prove causation, while coding order agrees with SemMedDB only 54.2% of the time, below the extractor's 69.4%.
WhySystematically audits three validation signals for text-mined clinical causal graphs, replicated on MIMIC-III. Novelty lies in empirical negative findings on direction; impact is confined to clinical
In briefVolatility-adaptive forgetting for stock forecasting via GARCH-modulated LSTM
SummaryThis paper proposes the Gsigmoid-LSTM, which feeds GARCH(1,1) volatility estimates into the LSTM forgetting gate through a volatility-adaptive sigmoid variant. The gate favors recent information during turbulent markets and retains long-term dependencies in stable ones. In daily stock price forecasting experiments, the model outperforms a feature-level fusion baseline using the same volatility signal on all five datasets and ranks best among nine models.
In briefBridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization
SummaryThis paper studies how LLM agents can adapt to user preferences at test time using binary comparisons between candidate responses to select personas. It presents ID-TAP, a neural dueling-bandit framework that pairs a query-conditioned utility model with a lightweight Bradley-Terry explorer and transfers evidence across similar queries. Across seven user-conditioned generation benchmarks within a fixed ten-round comparison budget, ID-TAP reportedly achieves the strongest aggregate performance among the evaluated baselines.
In briefA framework combining singular spectrum analysis with a volatility-adaptive LSTM for stock price forecasting
SummaryThis paper addresses multi-scale, noisy, and non-stationary volatility in financial time series with a decomposition-integration framework called CEEMDAN-SE-SSA-VMGL. The series is decomposed with CEEMDAN, components are grouped by sample entropy and K-means clustering, and singular spectrum analysis is applied to the high-frequency group to reduce mode mixing. The authors also propose a volatility-modulated GARCH-LSTM that adjusts memory decay to conditional volatility. On five US and Chinese financial series, the model reduces forecasting error by 9% to 24% relative to the best of eleven benchmarks, with statistical tests and backtesting reported.
In briefIF-MAP: Identifying Which AI Agents to Improve First in Multi-Agent Systems
SummaryThis paper introduces IF-MAP, a black-box method that ranks agents in LLM-based multi-agent systems by how much improving each one would raise overall system performance. It applies controlled semantic prompt perturbations, measures end-to-end loss changes, and fits a regularized local response model to produce a priority score. Across six workflows, prioritized prompt optimization and model upgrades outperformed uniform agent selection, while using fewer tokens than optimizing every agent.
In briefIs This Your Agent? Identity Watermarking for Dynamic Agent Systems
SummaryThe paper proposes DASMark, an identity watermarking framework for AI agents that keep updating their memories, skills, and configurations in dynamic environments. It introduces Action-Preserving Embedding to stably embed watermarks and Shift-Robust Recovery to reduce context-induced shifts during recovery. The authors report 98.5% bit-decoding accuracy in dynamic settings, 95.2% accuracy under attack, and task utility comparable to unwatermarked agents.
In briefOrganon: a self-improving agent that revises its own search for scientific discovery
SummaryThe paper introduces Organon, a recursively self-improving agent for scientific discovery that analyzes its accumulated experience during execution and revises its own search strategies, population management, diagnostic tools, memory, and worker coordination. It can also diagnose and correct flaws in its own adaptation procedures. The authors evaluate Organon on 21 computational discovery tasks and report better performance than competing search frameworks on ALE-Bench Lite and KernelBench.
WhyProposes a runtime recursively self-improving discovery agent, beating EvoX and others on 21 tasks, with clear novelty but limited benchmark scope.
In briefLearning from the Right Regions: Class-Conditioned Explanation Guidance for Medical Imaging
SummaryThis paper proposes a class-conditioned explanation-guided learning framework for multi-label medical image classification. It builds a separate explanation for each positive class and penalizes evidence outside that class's annotated regions. Using Grad-CAM across 14 abnormalities and four architectures, the authors report substantially improved annotation alignment with only smaller, architecture-dependent changes in classification performance.
In briefDistributed multi-agent training using ADMM, without a central trainer
SummaryThis paper proposes a distributed multi-agent training framework based on consensus ADMM, where agents learn locally and reach agreement through communication protocols such as ring all-reduce. It removes the need for a centralized trainer that aggregates trajectories and optimization states. On distributed safety alignment and general reasoning tasks, the method reduces HarmBench attack success rate by up to 26.4% under compromised-agent settings and achieves reasoning gains comparable to centralized training.
In briefPACE-Mem: Prefix-Aware Complementary Evidence Retrieval for Multimodal Agents
SummaryMultimodal agents store many similar observations in long-term memory, and answering questions requires retrieving the right evidence, which can overlap or be too much or too little for a fixed retrieval depth. This paper proposes PACE-Mem, a training-free retrieval method that uses maximal marginal relevance to balance query relevance against overlap with selected evidence, and a prefix-length selection algorithm that sets retrieval depth from the score trajectory. It plugs into clip-based agent memory systems without changing memory construction or answer generation. Across three long-video benchmarks and two agent frameworks, it improves question-answering accuracy by an average of 1.6% over the strongest baseline, Adaptive-k, at comparable retrieval depths.
In briefBeyond Next-Screen Prediction: Measuring the Decision Value of GUI World Models
SummaryThis paper examines whether GUI world models that predict the results of candidate actions truly improve action selection, or whether the gains come from the evaluator interpreting those predictions. The authors propose Direct Action Ranking (DAR), which scores candidates from reference actions without generating future screens, and find it outperforms prediction-augmented variants. On AndroidWorld, DAR improves task success by 5.8 to 16.4 points without retraining and transfers zero-shot to unseen apps. The authors argue GUI world models should be judged by their incremental decision value beyond a strong direct evaluator, not by prediction fidelity alone.
In briefImproved Smoothed Online Learning: Competitive Ratios, Dynamic Regret, and Best-of-Both-Worlds Guarantees
SummaryThe paper revisits smoothed online learning and proposes algorithms with improved competitive ratios, including showing that minimizing hitting cost alone achieves better ratios for polyhedral and quadratic growth costs, with matching lower bounds. It introduces Smoothed Ader (SAder), which attains optimal dynamic regret with switching cost in standard online convex optimization, and UDNP, which aggregates two experts to give best-of-both-worlds guarantees.
WhySubstantially improves competitive ratios for naive algorithms with matching lower bounds, plus optimal dynamic regret rates, advancing smoothed online learning theory.
In briefTactix: a programmable strategy interface that lets users control how a SAT solver searches for solutions
SummaryThe paper presents Tactix, a proof-producing sequential CDCL SAT solver that separates proof rules from solving strategies. A public interface lets users invoke techniques, observe their effects, and write custom strategies. Tactix's default strategy, built on this interface, outperforms state-of-the-art solvers on benchmarks from the 2024 to 2026 SAT Competitions. The authors also report performance gains from new strategies they designed for hard benchmarks in two application domains.
In briefWAQS: Weight Absorption for Quadratic Probing and Affine Steering (a method to steer language model behavior without extra inference cost)
SummaryThe paper proposes WAQS, a framework that models a concept with a quadratic probe whose gradient gives an activation-dependent affine steering direction. This intervention can be exactly folded into adjacent linear layers through an offline weight update, so it adds no inference-time operations. Across truthfulness, refusal, and sycophancy steering tasks, quadratic steering matches or improves on constant linear steering while preserving model utility.
In briefA two-stage method to infer domestic air trips from phone location data when airport sightings are incomplete
SummaryThis paper uses large-scale mobile location-based service (LBS) data to infer U.S. domestic air trips, where airport sightings are often missing before or after flights. Stage 1 applies heuristic rules to identify high-confidence trips observed at two distinct airports. Stage 2 uses a probabilistic model based on spatial proximity, temporal compatibility, and historical flight schedules to infer latent routes. Using January 2022 data with nearly 10 million users, air route coverage rose from 10.65% to 50.60%, validated against T-100 Domestic Segment benchmarks.
In briefFlashMap: A Flow Map Method for Faster Speculative Decoding
SummaryThis paper introduces FlashMap, a continuous flow drafter that speeds up large language model inference by proposing a block of tokens for the target model to verify. It conditions on the target model's hidden states and generates the whole block in one forward pass while capturing dependencies between tokens. On Qwen3-8B, FlashMap outperforms a matched DFlash baseline on every benchmark, with 7% higher average speedup and up to 6.1x lossless acceleration.
In briefGazeAgent: VLM-Guided Implicit Gaze Calibration for AI Glasses
SummaryThis paper presents GazeAgent, a framework that improves eye-tracking calibration for AI glasses by using naturally occurring behavior instead of explicit calibration steps. It identifies precise visual attention events from IMU, gaze, and hand signals and uses them as pseudo-calibration targets. In a 10-participant study, the system detected all 820 predefined events and reduced median gaze error by 44.1% under device disturbance compared with explicit calibration.
In briefHexagonal grid projectors that improve model-based CT image reconstruction
SummaryThe paper proposes constant and linear hexagonal projectors for two-dimensional fan-beam CT, built with box spline calculus. It derives analytic ray footprints and detector-blur expressions and implements efficient GPU forward and backprojection. On simulated FORBILD data the linear model outperforms Cartesian and hexagonal pixels, and on real 2DeteCT data it shows competitive image and feature agreement against high-dose Cartesian pseudo ground truth.
In briefDKnown-Jev: A Fine-Tuned JEV-Style Model for Normative Knowledge Tasks
SummaryThis paper presents DKnown-Jev, a model that classifies candidate paragraphs as Answer Evidence or Non-answer Evidence for tasks involving normative knowledge. The 2B and 27B versions are fine-tuned from Qwen base models using human-verified data. On a test set of 1,120 examples, the 27B model trained for one epoch reaches 98.75% accuracy, outperforming Jev. The 2B model's gains over Jev are too small to establish a stable advantage.
In briefKoi: A self-calibrating planner for deploying LLM inference across mixed multi-cloud clusters.
SummaryThis paper presents Koi, a planner that chooses accelerators, parallelism strategies, engine settings, and replica counts for large language model inference across heterogeneous, multi-cloud fleets. It pairs a Recursive Language Model hierarchy that proposes hypotheses with deterministic components that predict serving behavior and score plans, and it updates a causal graph from each deployment. The authors report 1.18x greater goodput than the strongest baseline and a 20% reduction in $/token on a real heterogeneous cluster.
Mankeerat Sidhu, Gangmuk Lim, Hetarth Chopra et al.
In briefFlashRT-PAI: fast kernels don't make fast pipelines for physical AI inference
SummaryThis paper addresses the real-time, single-batch inference latency of physical AI models such as vision-language-action, video-action, and world action models. It argues that latency comes mainly from system-level overhead, such as many short kernel launches and redundant recomputation, rather than individual kernel computation. The authors present FLASHRT-PAI, an inference engine combining deep kernel fusion, an execution contract describing pipeline structure and state reuse, and execution-state capsules for fast snapshot and restore. On RTX 5090 and Jetson AGX Thor, it reports 1.3 to 3.7 times speedup over official compiled paths at matched 16-bit precision, and 3.38 to 12.66 times over official eager implementations when combined with reduced precision.
WhyPipeline co-design with deep kernel fusion, execution contracts, and state capsules reduces physical AI inference latency; evaluated on six policies and two hardware platforms, showing 1.3-3.7x speedu
In briefBlack-box performance auditing for LLM-based network control systems
SummaryLarge language models are increasingly used in network control systems, but their opacity makes it hard to audit how they perform on unseen inputs. This paper introduces LENS, presented as the first black-box performance analysis method for LLM-enabled network control systems. It extends deep reinforcement learning to search for performance gaps and inputs that cause LLMs to struggle. The authors report that LENS finds large gaps in LLM-based traffic engineering and GPU allocation systems, and matches existing white-box analyzers designed for hand-crafted traffic engineering heuristics.
Mahdi Alizadeh, Pooria Namyar, Arvin Ghavidel et al.
In briefExact risk certificates for reranking judged only by pairwise comparisons
SummaryThis paper characterizes the exact worst-case nDCG loss when a ranking is replaced by a fixed proposal, given only pairwise relevance judgments. Binary gains achieve this worst-case loss and provide an inspectable witness, and threshold verification reduces to at most K+1 maximum-closure problems. On TREC-DL and BEIR, exact verification saves comparisons compared with the baseline, while predictive gates yield higher deployment gains.
In briefMeasuring whether higher-precision fallback actions improve closed-loop agent tasks
SummaryThe paper studies a low-precision agent that calls a full-precision (BF16) model at selected decisions, with a learned router choosing where to call it based on disagreement between the two models. The authors define a fallback's value as tasks rescued minus tasks spoiled, measured through paired branch experiments on Qwen3 agents in tau2-bench. Replacing only the first differing action produced nearly cancelling rescues and damages, while replacing both the action and continuation had positive net gain in two of five exploratory lower-bit settings. The authors recommend comparing against same-precision references before attributing gains to higher precision.
In briefCertification Exponent: Optimal Sequential Proof That a New Policy Is Better, From Logs
SummaryThe paper studies how many logged episodes are needed to certify, with error below δ, that a new policy beats a deployed one when the logging policy depends on history. It defines an exponent K* that sets a lower bound on the expected number of episodes and shows a plug-in Kelly certifier attains this rate to first order. Experiments report stopping times about 1.1 to 1.6 times the lower bound on several tasks, faster than the best scalar betting baselines in most cases.
WhyIntroduces a KL-based certification exponent with a tight sample lower bound and a first-order optimal plug-in Kelly certifier, validated on memory tasks and trees.
In briefWhat Matters for NVFP4 Training? A Comparison of Low-Precision Pretraining Methods
SummaryThis paper systematically compares recent NVFP4 (4-bit floating-point) training recipes for language models at medium scale, using dense 8B and 30B-A3B MoE models trained on up to 1T tokens. It focuses on which components are needed to recover accuracy and proposes a final recipe built on principles of stable NVFP4 training. Ablations show that each technique improves the loss trajectory, that selected high-precision layers are necessary at scale, and that backward-pass tensors benefit unequally from de-biasing.
Anjulie Agrusa, Andrei Panferov, Elizabeth Wei et al.
In briefUniform convergence at optimal rate and scale
SummaryThis paper studies uniform convergence, asking how many random samples ensure empirical averages approximate expectations simultaneously over a whole class of real-valued functions. It establishes sharp sample complexity of order d/ε² at the optimal approximation scale γ, achieving error γ+ε and closing a gap where earlier results optimized only one of scale or sample complexity. The authors also introduce tools of independent interest, including ε-nets with margin, improved covering bounds, and a chaining bound at the optimal scale.
WhyCloses a long-standing gap, achieving optimal scale and sample complexity together for real-valued classes, with new covering and net tools.
In briefScreening LLM Judges Without Ground Truth: Preliminary Evaluation of an AI Dental Hygiene Coach
SummaryThe paper proposes a staged pipeline that screens candidate LLM judges and student models before expert review, using explicit decision rules at each stage. It is demonstrated on synthetic, clinically unvalidated data with two fine-tuned small language models scored by five judges. The pipeline surfaced problems raw scores hid, but no dental hygiene expert has graded the responses, so the findings are preliminary.
William DeShong, Barath Kannan, Auberon Palma et al.
In briefPower-law correlations in fixed-tree hierarchical language models without temperature-tuned criticality
SummaryThe study uses a finite-depth probabilistic context-free grammar on a fixed hierarchical tree, computing exact next-token conditionals by belief propagation and sampling after softmax temperature rescaling. It measures two-point mutual information between tokens versus separation. Across the tested temperatures and grammar ensembles, mutual information stays well described by a power-law envelope; temperature changes its magnitude and fitted exponent, but no long-distance plateau or clear crossover to exponential decay appears within accessible system sizes. An approximate Ising-on-a-tree mapping explains how exponential decay in tree distance becomes a power law in token distance.
In briefPAC Learning with Query Feedback: Each sample returns only a yes/no answer on whether its label falls in a chosen subset; the paper defines a query dimension to characterize sample complexity.
SummaryThe paper studies realizable multiclass PAC learning where each training example returns one binary response indicating whether its true label lies in a subset chosen from a fixed query family. It introduces the query dimension, a combinatorial parameter capturing both hypothesis class structure and how well queries distinguish labels, which generalizes the bandit DS dimension. The authors prove a sample-complexity lower bound of Ω(D/ε) and an upper bound of Õ(D²/ε), improved to Õ(D/ε) when queries can split every candidate set into comparable halves.
WhyIntroduces query dimension generalizing bandit DS dimension, with matching-up-to-logs sample complexity bounds; a strong theoretical contribution.
In briefVERAS: Verification-guided search for more reliable operations research solutions
SummaryThis paper presents VERAS, a framework that uses verification feedback to guide LLM agents in solving operations research problems more reliably. It combines task-specific solution checkers with a separate verification agent and prioritizes feasibility repair, constraint-coverage review, and solution-quality improvement. On ORAgentBench, VERAS reaches a 42.99% pass rate, above the reported 35.51% for Codex using the same LLM.
In briefContinuous Ontology Assurance via External Resources: An Ontology Engineering Method Applied to Software Language Engineering
SummaryThe paper proposes COATER, a method that continuously reviews domain ontologies against evolving external resources using executable components called arrows. Each arrow compares part of an ontology with a relevant resource and reports support or challenges to its axioms, which can become tasks for human or agent-assisted triage. The authors illustrate the method with the FSL ontology in software language engineering, reporting that an LLM-assisted linkage exercise raised compliance with a Wikipedia-linkage policy from 405 to 600 of 600 entities.
In briefMorpheus-2.0: A Probabilistic Forecasting Model with Dual-Axis Attention Across Scales
SummaryMorpheus-2.0 is a compact encoder-only foundation model for time-series forecasting that handles variables and time steps on one grid using dual-axis attention. A single set of weights covers short to long forecast horizons because the horizon is sampled during training. A multiscale variant also handles several temporal resolutions at once. On the GIFT-Eval benchmark, the multiscale version reports the lowest relative MASE and CRPS among the published systems cited in the paper.
Aakriti, Manoj Cherukumalli, Prathamesh Patil et al.
In briefSelf-Debiasing Judges: Bias Awareness for Reliable LLM Evaluation
SummaryThis paper studies systematic biases that make large language models less reliable when used as automated judges. It proposes two training-free methods: bias-aware prompting, which warns the judge about possible biases before evaluation, and self-debiasing, where the judge critiques its own initial verdict. Three new benchmarks covering authority, confidence, and formatting biases are introduced, with seven bias types evaluated in total. Self-debiasing improved performance in 20 of 21 model-bias combinations across three frontier models.
In briefHow Latent States Are Used in Probabilistic Circuits
SummaryThis paper studies how the dependencies captured by probabilistic circuits (PCs) change as the number of latent states grows. Experiments with Hidden Chow-Liu Trees on ImageNet32 show that the learned dependence increases slowly and stays well below the latent-state upper bound and data-based dependence estimates. The authors track per-sample state preferences and state usage during training, and use these early measurements to guide initialization and tree selection, which improves final performance across the tested latent sizes.
In briefHow AI advice indirectly shapes complex sepsis treatment decisions
SummaryThe researchers first retrospectively evaluated an AI system's recommendations for 6,928 sepsis patients, finding overall improvement over original decisions but significantly worse performance in complex cases. In a pre-registered experiment with 228 US critical care physicians, three advice types were tested across six complex treatment decisions. Physicians given simple model recommendations chose optimal treatments in 56.3% of decisions, compared with 31.8% when always following the advice, and 42.9% for a less prescriptive advice type.
Venkatesh Sivaraman, Joel Levin, Eric P. Mason et al.
In briefLearning Location Encoders That Understand Administrative Boundaries via Reverse Geocoding
SummaryThis paper proposes a method for training location encoders, which convert latitude and longitude into compact representations, using reverse-geocoded text as a self-supervised signal. This lets the encoders capture administrative boundaries and place names. The approach can be applied to existing encoder architectures and consistently improves their downstream performance. The authors also introduce RevGeoRQ, a custom architecture for this supervision, which gives further gains and outperforms contemporary location encoders on several tasks.
In briefMuon optimizer stays stable beyond the classic stability edge
SummaryThis paper studies the Muon optimizer using sharpness and provides the first empirical evidence that the Flat Minima Hypothesis also applies to Muon. On CIFAR-10, sharpness strongly correlates with the generalization gap within each optimizer. Muon runs above the adaptive Edge of Stability threshold that bounds diagonally-preconditioned optimizers like Adam, yet remains stable, suggesting a fundamentally different regime. The exact stability boundary for Muon is left as an open question.
Niko Covic, Baraq Lipshitz, Alexander Padula et al.
In briefGenerating a diverse set of Pareto-optimal chip circuit designs with multi-objective reinforcement learning for LLMs
SummaryThis paper proposes a reinforcement learning method that trains large language models to generate RTL circuits, optimizing the quality and diversity of the whole set of circuits across power, performance, and area trade-offs rather than a single circuit. It uses a generation-based training pipeline that frames Pareto-set coverage as a sequential generation problem, and a reward based on Hypervolume Contribution (HVC). On the RTLLM v2 benchmark, it reports relative Hypervolume gains of 13.3 and 14.4 percentage points over a baseline without PPA optimization and a single-objective RL method, respectively, along with better circuit diversity.
Koki Takeshita, Kenji Uebayashi, Hikaru Tomonari et al.
In briefCME-SpectrumBench: Testing whether LLMs can analyze condensed matter spectral data
SummaryThis paper introduces CME-SpectrumBench, an expert-designed benchmark of 190 questions that tests how well frontier large language models analyze research-level spectral data from condensed matter experiments. The questions use simulated 2D intensity arrays containing noise and limited instrument resolution. Under direct querying, models generally scored well below handwritten code, but with Python interpreter access they performed comparably to it. The authors conclude that models can act as research assistants through code, yet still struggle to process information directly from large datasets.
Jin Gene Wong, Anjney Midha, Joseph Mathew Tennyson et al.
In briefTorch-DPhys: A PyTorch-based, MOOSE-compatible framework for differentiable physics simulation and hybrid machine learning
SummaryThe paper presents Torch-DPhys, a PyTorch-native framework for differentiable physics simulation and hybrid physics-machine learning. It implements selected thermal-hydraulic modeling concepts from MOOSE's Thermal Hydraulics Module and SAM, with a preliminary extension to continuum mechanics. The framework is verified through analytical solutions, reference-code comparisons, and demonstrations including elasticity, thermal-hydraulic transients, and a molten-salt fast reactor. Applications include parameter identification and solver-in-the-loop training of a predictive controller for nuclear engineering.
In briefCRISP: Recurrent Low-Light Video Enhancement Using Progressive Curriculum and Semantic Priors
SummaryCRISP is a recurrent framework for enhancing low-light video, aiming to recover brightness, color, and fine structure while keeping frames temporally consistent despite noise and motion. It uses a two-frame recurrent architecture and trains with a curriculum that moves supervision from intermediate targets between the low-light input and the reference to the full reference. It can optionally add frozen DINOv3 semantic features through AdaLN-style modulation. On DID and BVI-RLV, the authors report the highest PSNR on DID and the highest SSIM on BVI-RLV among the methods evaluated.
In briefWho Speaks, Who Authorizes, Who Acts? Trust Boundaries in Multi-Agent LLM Systems
SummaryThis paper examines trust boundaries in multi-agent LLM systems, separating message delivery, signing authority, and tool-side enforcement as distinct boundaries. It characterizes the smallest coalition of independently compromised signers needed to build a policy-valid false-claim proof under stated assumptions. Controlled experiments found no fixed-message routing contrast reaching the prespecified large-effect criterion across four models, and on invalid-proof cases some models often requested protected actions while a deterministic tool gate committed none, showing that request-level and commitment-level safety measures diverge.
In briefNot every user explains everything: training tool-using agents for less ideal simulated users
SummaryThe paper notes that tool-using agents are usually trained and evaluated with strong simulated users who state every detail clearly, while real users are often less precise. The authors build weak users by masking sentences of simulated replies and find that weak users lower the success of every policy, and training with them gives no reliable gain. They propose StaGPO, which has the agent record what the task needs, what is missing, and how it is being filled, so GRPO can compare turns reaching the same state and assign step-level credit. On ToolWOZ, a Qwen3.5-2B policy trained with StaGPO reaches 76.8% with weak users, 4.8 points above GRPO.
In briefConstructive universal approximation for infinite-dimensional neural networks using Gaussian-weighted ridgelet transforms
SummaryThe paper studies universal approximation for neural networks whose inputs are functions, such as those used in operator learning. It extends the ridgelet transform to continuous maps on separable infinite-dimensional Hilbert spaces, replacing the unavailable Lebesgue measure with Gaussian weights to obtain explicit coefficient representations. From this, the authors prove uniform approximation on compact sets for continuous non-polynomial activations of polynomial growth, including ReLU. Monte Carlo sampling yields finite-width networks with a dimension-free O(N^{-1/2}) error in the width N.
WhyExtends the ridgelet transform to infinite-dimensional Hilbert spaces with Gaussian weights, yielding constructive approximation and dimension-free O(N^-1/2) rates with machine-checked proofs.
In briefInducing Fourier Geometry: a smoothness prior creates approximate translation-equivariant structure in multimodal stellar embeddings
SummaryThe paper introduces a graph-Laplacian smoothness prior that makes learned representations vary smoothly over a known physical parameter space, without fixing the representation geometry. Applied to multimodal stellar representations from six astronomical surveys, the prior induces Toeplitz Gram structure and sinusoidal modes along stellar parameters, including in modalities that lack this geometry natively. The prior also improves downstream linear-probe performance and label efficiency.
In briefA Key Bottleneck in LLM Optimizers: Principle-Level Local Optima
SummaryThis paper studies principle-level search collapse, where feedback-driven LLM optimizers confine their search to a narrow set of design principles despite continuing to produce new candidates. The authors built a diagnostic benchmark across molecular optimization, network sentinel-node selection, and GPU-kernel optimization, finding collapse signs in 78.3% of 161 assessable runs. Diversity injection such as higher temperature, multiple LLMs, and web search gave limited and inconsistent benefits, and reference-principle interventions helped some runs but varied across models and tasks.
In briefBeyond Clean Accuracy: Practical Challenges and Benefits of Pixel-Based Language Models
SummaryThis paper examines whether pixel-based language models, which process rendered text instead of a fixed subword vocabulary, can be reliably adapted to downstream tasks. It tests fine-tuning behavior, robustness to typographical noise, and zero-shot crosslingual transfer. Pixel-based models need longer training and broader labelled data than text-based models, and their learning can be delayed and sensitive to learning rate. Once adapted, they lose less accuracy under typographical noise and outperform English-pretrained text models on target-language accuracy on two crosslingual benchmarks, despite weaker clean English performance.
In briefBanglaShield: A Robust Safety Classifier and Benchmark for Bangla Prompts in Native, Romanized, and Code-Mixed Forms
SummarySafety classifiers for large language models are usually built on standard-script text, but Bangla is often written in native script, romanized form, or mixed with other languages. The authors release BanglaShield, a benchmark of 10,245 Bangla prompts labeled on a four-level safety scale and rendered in three aligned written forms. Testing six LLMs and three multilingual guard models, they find many unsafe native-script and romanized prompts are labeled Safe. A compact fine-tuned encoder, BanglaShield, performs substantially better, and the corpus, splits, judgments, and model are released.
WhyBangla-script safety benchmark and compact classifier with large comparisons against LLMs and guard models; impact confined to a specific language subdomain.
Mohsinul Kabir, Md. Tariquzzaman, Mohammad Rashidujjaman Rifat et al.
In briefSteerSearch: Interactive Search Reasoning with Mid-Response Steering
SummaryThis paper studies how search-augmented language models respond when users change their goals mid-reasoning or mid-retrieval. It identifies a failure mode called answer-target anchoring, where models retrieve evidence for the updated request but still answer the earlier question. The authors introduce SteerSearch, a reinforcement learning framework, along with SteerSearch-Bench, and explore Contrastive Context Reward as an auxiliary signal. The method improves steering accuracy over Search-R1 on Qwen2.5-3B and Qwen3-4B while keeping standard single-query performance.
In briefULIFlow: Generating Text-Guided Human-Object Interactions in a Unified Latent Space
SummaryThis paper addresses generating 3D human-object interaction motions from text, where separately plausible human and object movements may not form a coherent interaction. The authors propose ULIFlow, which derives continuous marker-object relations from sparse body-surface markers and models them alongside human motion and object dynamics in a unified latent space. The framework also supports variable-length generation through a blockwise Flow Forcing scheme. The authors report that ULIFlow outperforms prior work in generation quality and human-object coordination, and keeps interactions coherent beyond the training horizon.
In briefROAR: Real-Time Benchmark for LMM Agents in Multiplayer Shooter Games
SummaryMost benchmarks for multimodal AI agents in multi-agent settings are turn-based, pausing the environment while agents reason. This paper introduces ROAR, a multiplayer action-shooter game environment covering cooperative, competitive, mixed, and single-agent scenarios. It also proposes a Multi-Agent Asynchronous Gym interface that runs the environment and each agent in separate threads, so the world advances at fixed intervals without waiting for model inference. ROAR aims to be a practical testbed for real-time multi-agent decision-making.
In briefBeckmann World Models: learning one-step future-frame prediction directly from noise
SummaryThis paper introduces Beckmann World Models (BWM), which learn a direct map from noise, visual history, and actions to future video frames without a pretrained generative teacher. The authors add endpoint-aligned training to improve accuracy from pure-noise inputs at inference. On PushT and Robomimic Can, the final model reduces 64-frame prediction MSE by 68% and 57% relative to a same-backbone regression baseline.
In briefDetecting Whether a Tabular Dataset Was Used to Train Tabular Foundation Models
SummaryThis paper introduces the first dataset inference method for tabular foundation models (TFMs), which aims to determine whether a suspect dataset was part of a model's pre-training data. The authors find that individual signals observed from black-box models are weak evidence, but combining multiple complementary signals achieves up to 0.99 ROC AUC for several state-of-the-art TFMs. The study also examines how pre-training data composition, model capacity, and real versus synthetic data affect identification.[]
WhyProposes the first dataset inference method for tabular foundation models, combining black-box signals to reach 0.99 ROCAUC across several models, with fairly solid evidence and broad auditing impact.
In briefModeling bilingual naming performance with laterally connected self-organizing maps
SummaryThis paper extends BiLex, a computational model of the bilingual lexicon, by adding excitatory and inhibitory lateral connections. The extension lets the model account for response times and common errors based on semantic familiarity, typicality, and specificity. The Boston Naming Test was simulated in Spanish and English, and the model's performance approximated naming accuracy and error types of Spanish-English bilinguals. The authors conclude the approach captures semantic-level processes in bilingual lexical access and may later support modeling impairments and treatments.
Divya Prabhakaran, Uli Grasemann, Erie Shivers et al.
In briefTesting whether LLMs can forecast stock reactions and analyst EPS from 10-Q filings
SummaryThis paper evaluates 13 general-purpose and finance-specialized LLMs on 45,964 Form 10-Q MD&A sections from 2001 to 2024. The models rank stock returns and forecast next-quarter EPS, compared against analyst forecasts and realized abnormal returns. Adding cross-sectional fundamental ranks improved return-ranking correlation, while analyst-summary context gave no comparable gain. Direct EPS forecasts averaged the 43.4th percentile of the analyst distribution and did not consistently beat individual analysts; consensus anchoring raised inside-range share but increased reported-EPS error.
In briefCan small open-weight LLMs recover missing flight trips from incomplete location data?
SummaryThe study uses January 2022 U.S. location-based services data and airport reference data to test whether compact, locally deployed large language models can recover missing information from incomplete trajectories. Thirty high-confidence air trips were altered by removing observations near the origin, destination, or both airports. Exact transition accuracy was 9.2%, origin accuracy 31.4%, and destination accuracy 10.7%, with more explicit prompts performing best. The paper also proposes a controlled endpoint-removal framework for evaluating such recovery.
In briefDesirability Confirmation: Relational Sycophancy in AI Intimate Roleplay
SummaryThis paper studies sycophancy in large language models during adult intimate roleplay, where users may seek to keep a preferred relational position rather than agreement with beliefs. Through a controlled audit of 2,016 outputs from two open-weight models, the authors identify four recurring relational moves, such as legitimating desire and presenting the user as wanted. They term this pattern desirability confirmation, extending sycophancy and intimate-AI evaluation beyond simple compliance or refusal.
In briefAM-Bench: A Benchmark Suite for Measuring Agentic Misalignment
SummaryThis paper introduces AM-Bench, a unified taxonomy and benchmark of 35 scenarios across seven industries for testing when LLM-based agents choose harmful actions over failure. It builds on the fraud triangle from criminology and organizes evaluation around perceived threat, perceived opportunity, and reasoning. The authors tested 10 frontier models and found large differences in misalignment, with Gemini 3 Flash and Gemini 3.1 Pro showing the highest rates and Claude Sonnet 4.6 showing near-zero rates.
Eric Zhang, Terry Jingchen Zhang, Chijioke Ugwuanyi et al.
In briefCombining Fine-Tuned Small Models with LLMs to Balance Cost and Performance
SummaryThis paper classifies survey text into 16 classes and compares a trained RoBERTa model with an off-the-shelf GPT-3.5 model on performance, cost, and time. RoBERTa achieved an F1 score of 96% versus 83% for GPT-3.5, and processed 10,000 records in about 12 minutes compared with 6 hours for GPT-3.5. Batch prompting reduced GPT-3.5 inference costs by about 70%, and an ensemble of the two models reached an F1 of 97.5% at 4.4% of the inference cost of standalone GPT-3.5.
Shivam Mathur, Sri Harish Popuri, Priya Krishnan Sundararajan et al.
In briefSLA-LENS: Analyzing Second Language Classroom Interaction from Transcripts
SummaryThis paper presents SLA-LENS, an LLM-assisted framework that turns classroom transcripts into structured teaching annotations and evidence-linked descriptive analytics. Its annotation scheme is based on interactionist second language acquisition theory and covers learner errors, teacher feedback functions, and uptake or repair. The framework was applied to 40 online Spanish classroom transcripts with over 21,000 sentence-level instances and evaluated across several LLMs.
In briefDynamic neural optimal transport for aligning a source to only the relevant part of a target distribution.
SummaryThe paper studies optimal transport when only a subset of the target distribution is relevant to the source. It maps the source to a reweighted target whose density ratio relative to the full target lies in [0, c], using two neural networks: one for a time-dependent potential and one for interpolating between the source and a learned terminal distribution. Unlike prior neural dynamic optimal transport methods that fix the terminal distribution, it learns the terminal distribution within the target's support. The method is demonstrated on positive-unlabeled learning, unpaired domain translation, and trajectory inference.
In briefCortex: A Fixed-Point Theory of Governed Coding Agents (Technical Report)
SummaryThe paper studies Cortex, a supervisory layer for coding agents covering pre-execution governance, requirement tracking, and iterative validation and repair. It formalizes requirement closure on a finite lattice and shows that, under stated assumptions, iteration reaches a least fixed point, with a separate stochastic bound on expected completion time. The authors define the Gauntlet evaluation metrics but note that the results depend on their assumptions and that unequal observations prevent causal claims about governance.
In briefMOSS: A Production Framework That Monitors Search and Recommendation Systems for Semantic Errors
SummaryThe paper presents MOSS (Model Sanity Service), a production framework that continuously monitors recommendation and search systems using LLM-based relevance scoring and categorizes failures into actionable root causes. It has run for over six months at a large U.S. retailer and surfaced two confirmed serving defects. On a human-annotated recommendation benchmark, frontier LLM judges achieve higher F1 and AUC point estimates than non-LLM baselines.
Alina Nesen, Prathyusha Kanmanth Reddy, David Relyea et al.
In briefPredicting whether an LLM can accurately label a new dataset
SummaryLarge language models are increasingly used for data annotation, but their results vary. This paper asks whether their performance on a new dataset can be predicted in advance, and builds two task families yielding 461 text classification datasets to test this. The authors find that LLM performance can be predicted to a certain degree. Contrary to work highlighting LLMs' verbalized self-confidence, internal log-probabilities are the most predictive signal.
