In briefExact maximum-likelihood ranking from pairwise preferences with rank-two structure
SummaryThe paper addresses converting pairwise probabilities into a maximum-likelihood ranking, a weighted Kemeny problem that is NP-hard in general. It shows the problem is exactly solvable when the log-odds matrix has rank two, even when preferences contain cycles. The proposed CloseSort algorithm returns a global maximum-likelihood order in O(n log n) time. On simulations and LLM-judge profiles, it reduces normalized regret relative to Borda and other baselines.
WhyThe authors show the weighted Kemeny problem is exactly solvable for rank-two log-odds matrices, giving an O(n log n) CloseSort algorithm and stability certificate. The abstract reports simulations, A
In briefIn Search of Lost Time: Temporal Competence in Large Language Models
SummaryThis paper examines how large language models handle time in reasoning and decisions through four successive evaluations. The models answer questions based on time-related theorems with high accuracy, but they substantially underestimate their own task completion times and fail to identify which tasks take longer. Their duration judgments also shift with the wording of an event, and some models ignore waiting time costs entirely when answering user queries.
In briefPrivMark: Pixel-Space Steering Against Privacy Leakage in Vision-Language Models
SummaryVision-language models often reveal soft-biometric attributes such as age, gender, and race in their text outputs, exposing demographic information from uploaded images. PrivMark encodes two steering directions tied to different leakage behaviors into a single bounded pixel perturbation, so the protected image alone suppresses biometric leakage at deployment. Across three white-box source models, it achieves up to 85.9% biometric suppression while retaining 75% of clean OK-VQA utility. The same perturbations transfer to seven unseen VLMs, including closed-source APIs.
In briefTransferring Actions Between Animated Characters with Different Body Shapes
SummaryThis paper studies whether a motion learned from video without labels can make a 2D character with a different body structure perform the same action. The authors introduce a transfer score that compares the correct action against wrong actions from the same source, so that simple screen movement does not inflate results. Experiments on a synthetic cartoon benchmark, the Sprites benchmark, and DeformingThings4D animal and human animations show that learned actions retain information about the action performed, but only partly close the prediction error for a new character.
In briefTimeHF 2.0: Multi-Region Retail Sales Forecasting Using Text and Business Context
SummaryThis paper presents TimeHF 2.0, a forecasting framework for large-scale retail supply chains that lets business knowledge and future plans written in natural language inform numerical sales predictions. It combines a coverage-aware regional encoder, a semantic alignment module, and text-conditioned forecasting with an ordered-distribution output head. In online parallel forecasting at JD, it improved relative accuracy over the original production deep learning baseline by 18.25% and 17.94% for national daily and planning-window sales, and by 19.44% and 18.99% for DC daily and planning-window sales.
In briefOnline stereo correction that also accounts for the whole-image distortion of a virtual camera
SummaryStereo depth estimation assumes the two cameras are precisely calibrated, but small physical shifts during operation can break this and degrade depth accuracy. The paper proposes VirtualRect, an online rectification method that estimates a rectified virtual stereo camera and supervises the warping behavior of the whole image, so matching images and the camera model stay consistent. Experiments on several datasets report lower vertical flow error and depth mean absolute error than the prior state-of-the-art online rectifier.
In briefFADE: Training-free speedup for diffusion transformers using fractional extrapolation
SummaryThis paper presents FADE, a training-free framework that speeds up diffusion transformers (DiTs) during image or video generation. The authors find that DiT latent feature trajectories show anomalous diffusion and spatiotemporal heterogeneity, which fixed integer-order forecasting handles poorly. FADE instead models feature evolution with a Mittag-Leffler fractional operator and uses an offline-built multidimensional lookup table to select parameters during inference. The authors report state-of-the-art acceleration with preserved generation fidelity under aggressive skip intervals and negligible latency overhead.
In briefCALFA-KV: Compressing Transformer KV Caches to 1 Bit with Calibration and Fisher-Guided Anchors
SummaryThis paper proposes CALFA-KV, a method for compressing the key-value cache of transformer models to 1 bit per channel. It jointly calibrates key centers, normalization parameters, and codebooks offline, and uses Fisher-type channel sensitivities from gradients to weight reconstruction errors. A small budget of high-precision anchor states is retained per layer and KV head, chosen by score. With 56 extra FP16 anchors per layer and KV head, the 8-bit-style d8b8 variant outperforms the evaluated 1-bit baselines on mean score across seven benchmarks for three models.
In briefScaling an optimization-based hypernetwork that generates implicit neural representations
SummaryThis paper presents a hypernetwork that produces the parameters of an implicit neural representation for a given signal in a single forward pass, with no test-time optimization. The authors replace the multilayer perceptron body of earlier optimization-oriented designs with attention, so the parameter count no longer grows with target-network width. Across five 2D and 3D INR generation benchmarks, the method outperforms state-of-the-art single-pass and optimization-oriented baselines and scales to widths where MLP-based versions run out of GPU memory. It also achieves the best NeRF generation result without supervision from ground-truth INR weights.
In brief"Imperfectly Imperfect": Testing Whether LLM Personas Capture Graded Cognitive Diversity
SummaryThe paper examines whether large language model agents, used as synthetic participants, can reproduce the range of human reasoning ability when assigned different cognitive profiles. Using a cognition test suite answered by 473 adults, the authors compare simulated populations against human data on reflection tests, personality measures, and behavioural scenarios. The models recover self-reported trait distributions reasonably well but fall short on task performance and behavioural diversity, with many agents scoring full marks on reflection tests far more often than people do.
Sion Weatherhead, Ben Newell, Aaron Belbasis et al.
In briefSRA: A Spatial Reasoning Adapter for Multi-Agent Trajectory Prediction
SummaryThis paper proposes SRA (Spatial Reasoning Adapter), a modular graph denoiser that plugs into existing diffusion and flow-matching trajectory predictors. At each reverse step, it builds a sparse interaction graph from the model's current future estimate and returns a residual correction to the decoder. Tested on SDD, NBA, Soccer, and Football with three host models, SRA improves ADE by 5.8% and FDE by 7.0% on average.
In briefLearning one shared dictionary of concepts across different AI vision models
SummaryThe paper introduces the Gated Bottleneck Sparse Autoencoder (GB-SAE), which treats cross-model interpretability as a representation factorization problem. It learns a shared dictionary of concepts common to all models along with private dictionaries capturing model-specific residual concepts. The method is evaluated on ViT, SigLIP, DINOv2, and ConvNeXt, and the authors report strong reconstruction, interpretability, reusability, and generalization to unseen models and downstream tasks.
In briefUsing LLM uncertainty to tell extraction errors from clinical disagreement in radiology reports
SummaryThis study examines whether the uncertainty of large language models, when extracting vessel-level coronary artery stenosis from coronary CT angiography reports for 1,045 patients, can identify extraction errors and clinical discordance. LLM uncertainty was strongly associated with extraction failure and improved discrimination beyond the extracted stenosis severity alone. It was only weakly associated with clinical discordance between CT angiography and invasive angiography and added no discriminative value there. The authors conclude that uncertainty can flag extraction failures but not underlying clinical discordance.
Jerjes Aguirre-Chavez, Shant Malkasian, Roshun Sankaran et al.
In briefSampling Is Not Curiosity: Why LLM Agents Should Investigate
SummaryThis paper argues that current LLM agents explore open-ended tasks through stochastic sampling rather than genuine curiosity. The authors contend that curiosity requires a predictive model of environment responses that updates online, guides action selection, and persists beyond the context window. They propose a framework defining curiosity for LLM agents and identify research directions that its implementation would open up.
Alfonso Amayuelas, Piotr Pi\k{e}kos, Xin Eric Wang et al.
In briefEmbodiedIntentBench: A Benchmark for Intent-Driven Runtime Adaptation in Robot Missions
SummaryThis paper presents EmbodiedIntentBench, a benchmark for evaluating how mobile robots adapt at runtime to mission intent. Intent is expressed through structured contracts covering coverage, temporal evidence, reporting, urgency, and resources, and is paired with simulated UAV workloads that account for execution cost. The authors compare fixed and adaptive perception strategies across multi-phase missions under changing urgency, connectivity, and battery conditions.
Rajat Bhattacharjya, Changwoon Lee, Hyeonjong Ju et al.
In briefSeparating channel meanings in Vision Transformers so each image channel is interpreted more accurately
SummaryMulti-channel images have different kinds of channels that share the same spatial layout, so models must capture both shared structure and channel-specific details. The authors find that existing multi-channel Vision Transformers' context tokens collapse onto a few dominant channels when channels are highly redundant, giving a biased global summary. They propose MuCa-ViT, which combines Factorial Token Learning to decorrelate tokens and Channel Composition-Aware Modulation to generate sample-wise modulation signals. It outperforms prior multi-channel ViTs on JUMP-CP, CHAMMI, and So2Sat.
In briefSafer for Whom? A systematic study of how well AI chatbots keep people with eating disorders safe
SummaryThis paper examines whether AI assistants give safe answers to questions about diet, exercise, and weight loss from people who may have eating disorders. The researchers tested three open-weight language models on 100 questions from an online pro-anorexia community and graded the responses with clinically informed criteria. Safety varied across models, harm types, and user descriptions such as race, gender, and body size.
In briefGEAR: A GPU-accelerated global solver for bounded continuous nonlinear optimization problems
SummaryThis paper presents a GPU-accelerated global solver for constrained nonlinear programming, adapting the linear bound propagation approach used in neural network verification. The linear bounds serve as relaxations of nonlinear constraints, providing lower bounds, detecting infeasibility, and guiding the search for primal solutions. The authors also propose tighter linear relaxations for bilinear and quotient functions, along with a projected augmented Lagrangian method. On 505 bounded continuous benchmark problems from GAMS and MINLPLib, the solver handles 421 within 180 seconds, more than six compared global solvers including Gurobi, BARON, and SCIP.
WhyExtends linear bound propagation from NN verification to global NLP with tighter relaxations; the authors report solving 421 of 505 benchmarks, more than six global solvers.
In briefTest-time scaling for diffusion language models using frequency-aware remasking
SummaryDiffusion language models generate text through iterative denoising, where choosing which tokens to commit or remask affects output quality. The paper proposes a two-stage approach: a fast base sampler solves easy questions, and frequency-aware remasking is applied to the remaining hard questions. This method prioritizes committing tokens whose predictions change often across denoising steps, encouraging exploration of alternative reasoning paths. Experiments with LLaDA-8B-Instruct and Dream-7B-Instruct show improvements over a strong adaptive allocation baseline on MATH-500 and HumanEval.
In briefLanguage-guided one-step method for enhancing very dark images
SummaryThis paper addresses super-resolution for dark images, where degradation varies greatly across regions. It proposes DeLang-SR, a one-step flow-matching framework that turns measurable physical image statistics into degradation language and uses a pretrained text-to-image model's prior to steer the restoration direction. A spatially varying update-strength field sets stronger steps in heavily underexposed areas and conservative steps in reliable regions. Experiments on paired and real dark images, with ablations and diagnostic analyses, are reported.
In briefCan General-Purpose Generative AI Provide Training Data for Medical Vision Models? A Retrospective Study
SummaryThis paper evaluates Nano Banana, a general-purpose image generation and editing model, as a source of synthetic medical images across chest radiography, fundus photography, CT, and carotid ultrasound. Models trained only on synthetic images are compared with real-data-trained models through external validation. The study also explores synthetic data for subgroup augmentation and controlled editing for model testing. The authors conclude that synthetic data can complement real data, but its clinical value should be judged using calibration, decision curve analysis, and clinician review rather than discrimination alone.
In briefThey Can See It but Not Say It: Self-Refining Vision-Language Models for Relative Pose Identification
SummaryThis paper examines why vision-language models struggle with identifying relative camera pose across multiple views, finding that they often predict certain motion patterns regardless of the visual evidence. The authors propose SKR-ID, an annotation-free method that uses randomly sampled unlabeled image pairs and a two-stage process of debiased prompting and iterative logit-ranking with the median logit as a threshold. Reported results show open-source models' Macro-F1 scores rising from 35.0-45.4% to 53.7-58.0%.
In briefPatternBloom: An agentic RAG framework that distills reasoning patterns from graphs using reinforcement learning
SummaryThe paper introduces PatternBloom, a retrieval-augmented generation framework that retrieves reasoning logic rather than content. A first reinforcement learning stage scores the agent's evidence graph with an information-density reward, and high-reward trajectories are distilled into a memory of type-abstracted graph patterns. A second stage couples the policy to this memory through a pattern-augmented reward. The authors report that a 7B PatternBloom outperforms 27 baselines by 15.0 Avg-OOD EM, with the advantage growing as reasoning depth increases.
In briefWordEval: A benchmark for checking whether document agents correctly perform Word editing operations
SummaryWordEval is a benchmark that tests whether AI document agents carry out Microsoft Word editing operations as intended, by checking the resulting changes to Word objects and their properties. It contains 642 cases across 15 Word feature areas, including atomic single-edit cases and compositional multi-edit cases, verified by a Word-native oracle. Evaluation of eight agent systems found that many struggle on atomic cases, and every system scored lower on compositional cases than on atomic ones.
In briefGroup-conditional conformal risk control with exact guarantees and overlapping-group extensions
SummaryThe paper studies group-conditional conformal risk control for disjoint and overlapping groups under bounded monotone losses. For disjoint groups it gives exact finite-sample conditional risk control and identifies limits arising from local support and calibration-rank granularity. For overlapping groups it compares two valid approaches, group-restricted calibration followed by union deployment and direct calibration of the deployed prediction-set family, finding that neither uniformly dominates.
In briefHibana: A fast inference engine for agentic AI on unified-memory devices
SummaryHibana is an inference engine designed for devices with unified memory, jointly optimizing model representation, memory management, and decoding. It reduces memory footprint through model surgery, preserves agent state across turns so long-running agents can resume quickly, and uses adaptive speculation to generate several tokens per step. On a single DGX Spark running Qwen and Nemotron models up to 125B parameters, it completes up to 3.4 times more agentic tasks per hour and achieves up to 3.9 times higher decoding throughput than vLLM, NVIDIA Flare, llama.cpp, and TensorFold.
In briefEstimating Power Grid Dynamics from Sparse PMU Data with Physics-Informed Graph Neural Networks
SummaryThis paper proposes a physics-informed graph neural operator that estimates rotor-angle and speed trajectories of all generators in a power system from sparse, intermittent measurements at a few generator buses. The model is trained on simulated measurements without state labels and enforces the swing equations as a hard or soft constraint. Simulations on four transmission systems show it is generally the most accurate of the estimators that do not optimize per scenario or use state labels, and under model errors it matches or outperforms an optimization-based benchmark at lower cost.
Huy Hoang Le, Haoguang Wang, Christian Moya et al.
In briefFast Tree Transform: a locally aggregated signal transform on graphs
SummaryThis paper introduces the Fast Tree Transform, a locally aggregated transform defined for arbitrary graph geometries. It generalizes both the Hadamard frequency transform and the Haar wavelet transform on the circle. The pure frequency part of the transform approximately diagonalizes the graph Laplacian, making it useful as a preconditioner in geometric optimization. The full transform gives a slightly redundant representation of graph signals that allows sparse representation of signals containing both edges and textures.
In briefPredicting cancer outcomes across a continuum of genetic ancestry
SummaryThis paper proposes distribution-aware in-context learning (DA-ICL) for predicting cancer outcomes from omics data. It extends models that group samples into discrete ancestry categories to a continuous model across the genetic ancestry spectrum, using an in-context learning foundation model. The authors provide theoretical analysis showing the continuum approach outperforms discrete modeling, supported by empirical results.
In briefLongCapHarness: A Method That Unlocks Long-Context Reasoning in LLMs Without Retraining
SummaryThis paper proposes LongCapHarness, a framework that improves long-context reasoning in large language models without data synthesis or post-training. It recursively revises capability primitive implementations, capability states, and control logic through self-improvement on a development set. Applied to Llama3-8B and transferred to Qwen2.5-3B, it substantially outperforms recent data-synthesis baselines on HELMET and RULER, at an estimated development cost under 50% of the training costs of NExtLong and PolicyLong.
In briefZeroth-order Langevin methods for noisy nonconvex optimization with curvature regularization and non-asymptotic guarantees
SummaryThe paper studies stochastic zeroth-order Langevin dynamics (ZOLD), which optimize nonconvex objectives using only noisy function evaluations, without derivative information. It shows that Gaussian smoothing implicitly adds a Hessian-trace regularization and derives non-asymptotic excess-risk bounds for the resulting curvature-regularized objective. The analysis uses Wasserstein distance to relate the algorithm's iterates to the associated Gibbs measures, without relying on log-Sobolev or Poincaré inequalities.
In briefRepresented but Not Used: Value-Policy Dissociation in LLM Agents
SummaryThis paper studies whether large language model agents internally represent the future value of candidate actions yet fail to use that information when choosing actions in multi-step tasks. The authors test Qwen2.5-7B-Instruct, Qwen2.5-32B-Instruct, and GPT-OSS-20B on a partially observed grid environment and the WebShop simulated shopping environment. They find that relative action values can be linearly decoded from intermediate hidden states, even where the policy prefers a worse action, which they call value-policy dissociation.
In briefBayesian Low-Rank Posteriors for Scalable Membership Inference
SummaryThis paper studies membership inference attacks, which determine whether a sample was used to train a target model. Instead of training many reference models, the authors use a Bayesian approximation over LoRA parameters and apply SWAG to sample diverse virtual reference models from a single training run. Evaluated on three downstream-adapted vision-language models across Document VQA and Medical VQA, the approach outperforms standard reference-based attacks while requiring only one reference-model fine-tuning run.
Khanh Nguyen, Antonio Emanuele Cin\`a, Andrey Barsky et al.
In briefDecentralised drone swarm formation that handles silent failures, using ant-colony pheromone claims and particle-swarm tuning
SummaryThe paper studies how drones can form a picture without a central controller, when the number of live drones is unknown, failures are silent, and radio links are short-range and lossy. Ant colony optimisation allocates drones to positions through evaporating pheromone-like claims shared between neighbours, so evaporation also detects failures, while particle swarm optimisation drives the flight and tunes the controller offline. The controller formed the picture in 95-99% of runs in the hardest regime and restored it after a silent loss of half the swarm in 96-100% of runs at normal difficulty.
In briefBunshinRL: Splitting Live Agent Workspaces to Cut Memory in Agentic RL Training
SummaryThis paper presents BunshinRL, a rollout system for agentic reinforcement learning that reduces memory use and latency by sharing memory across environments, caching tool prefixes, and running tools speculatively. It builds on BunshinFork, a live whole-workspace forking mechanism for containers that reconstructs process trees, shares memory copy-on-write, and preserves file, communication, and GUI state. The system is implemented through a Linux kernel module and CRIU modifications. The authors report peak memory reductions of up to 43%, total memory reductions of up to 51%, and a 23% decrease in end-to-end rollout latency.
In briefRecovering edited language model knowledge through prompt-space steering
SummaryThis paper studies whether facts overwritten by model editing in language models can be recovered without changing the model's weights. The authors introduce Memento, a method that learns a reusable prompt-space recovery signal to steer an edited model toward its pre-edit answers. Across three LLMs and five editing techniques, Memento recovers 80.07% of held-out edited facts, a 9.58% relative improvement over baselines.
Manit Baser, Rachel Yeo Hui Min, Aditya Nawal et al.
In briefESTRA: Training-free extreme sparse attention for faster video diffusion
SummaryThis paper presents ESTRA, a training-free method that speeds up video diffusion transformers by using sparse attention. It computes exact attention only for a small set of important key-value tokens and approximates the rest with cluster representatives, refined by a variance-based correction. Experiments on Wan2.1-14B and HunyuanVideo-13B report 3.26x and 4.71x end-to-end speedups at 5% attention density, while outperforming Sol-Attn and SVG-EAR in fidelity to dense generation.
In briefDiSARM: Mining Diverse and Surprising Association Rules with Large Language Models
SummaryThis paper proposes DiSARM, a hybrid framework for association rule mining on transactional data. It uses the world knowledge of large language models to assess the subjective interestingness of rules while preserving statistical reliability and promoting diversity among selected rules. Experiments on three transactional datasets show that DiSARM effectively balances statistical reliability, surprise, and diversity.
In briefSplit and Bridge: Multimodal Generation via Diffusion Bridging
SummaryThis paper proposes Split and Bridge (SnB), a framework for generating multiple linked data types, such as images and their segmentation masks. At sampling time, it couples pretrained single-modality diffusion models: samples follow a shared trajectory up to a split point, then a bridge module guides them onto modality-specific paths, without joint training. The authors report a good balance of sample quality and coherence on PolyMNIST, CelebAMask-HQ, PIE-Bench++, and ROCO. SnB also extends to zero-shot tasks such as interleaved co-generation, image editing, and style coupling and decoupling.
In briefWho remembers what? Shared memory for multiple users and agents
SummaryThis paper presents MixUA, a memory system that turns separate user-agent experiences into lessons and assigns each lesson to a user-agent pair, a single user, a single agent, or a global scope. The authors also introduce CrossLearnBench, a benchmark that combines agent tasks with user preferences and privacy tests to check whether private facts leak into memory that other users can read. With five users and seven agents, MixUA improves pass rates over no memory and over a single global memory. It also retains most previously stated preferences and rarely leaks planted private facts.
In briefHexIt: learning to split solids into hexahedral blocks via expert iteration
SummaryThis paper addresses the long-standing problem of decomposing a polyhedral solid into a few well-shaped hexahedral blocks for mesh generation. The authors model the task as a sequential decision problem on a 3D generalised map, where an agent applies three insert operations to split edges, faces and volumes, and a policy convolving over the map's involutions chooses each step. They clone a policy from constructed decompositions and improve it with expert iteration, reporting higher solve rates on held-out chamfered solids than the cloned policy. They also state that PPO, GRPO and VinePPO updates made the cloned policy worse at matched compute.
In briefD-REAL: Optimizing Speculative Decoding Acceptance Length with Unbiased Gradients
SummaryThis paper introduces D-REAL, a post-training method that directly optimizes the expected acceptance length in speculative decoding instead of token-level training objectives. It weights draft-target distribution overlap by prefix survival probabilities and derives an unbiased gradient estimator for stochastic verification. Evaluated on DSpark and DFlash2 with Qwen3-8B as the target model across ten benchmarks, one epoch of training raises the mean acceptance length by up to 0.23 tokens and end-to-end throughput by up to 9.4%.
In briefDistilling belief-space planners into agents that follow physical and epistemic rules
SummaryThis paper trains agents to act under uncertainty according to belief-dependent rules. It plans offline to generate sample trajectories, then distills those plans into a policy acting on raw observations, with two modifications: Dual Averaging and ρ-Optimal Reinforcement. In the new Scavenger domain, the distilled policy achieves over 90% success, compared with at most 5% for reinforcement learning.
Phillip J.K. Christoffersen, Rachel Ma, Patrick Barrag\'an et al.
In briefBifidelity sampling hypernetwork for material response modeling and uncertainty quantification
SummaryThis paper presents a bifidelity Langevin sampling hypernetwork framework that builds probabilistic neural response models for cases where high-fidelity data are scarce and low-fidelity data are more plentiful. Rather than a single deterministic fit, it learns a posterior distribution over model parameters, using a shared-latent representation to enable parameter sharing and transfer between fidelity levels. The authors add enhancements such as stability-promoting drift models, staged and conditional training, and adaptive initialization of Langevin trajectories. The method is demonstrated on experimental, high-fidelity simulation, and illustrative synthetic data.
Reese E. Jones, Cosmin Safta, Dan Bolintineau et al.
In briefMental Imagery That Matters: Fixing Latent Collapse in Multimodal Reasoning
SummaryThis paper examines whether the latent visual states in multimodal language models actually influence their answers. Perturbation tests show that dropping, replacing, or shuffling these visual latents barely changes the final answer probability, while perturbing the text reasoning trace has a much larger effect. The authors name this failure Mental Simulation Collapse and propose AnchorSim, a contrastive distillation framework that makes the generated visual latents consequential, which improves performance on several visual reasoning benchmarks.
In briefBeyond Query Similarity: Safety-Conditioned Cache Routing for Safer LLM Response Reuse
SummarySemantic response caching reuses answers for similar LLM queries, but query similarity alone can cause unsafe reuse when safety conditions change. The paper proposes SafeCache, which separates candidate retrieval from safety approval and treats reuse as a safety-conditioned routing problem. It scores mismatches between each cached condition and the new request, and adds a contradiction guard so strong matches cannot hide explicit conflicts. Experiments show SafeCache achieves a better safety-utility trade-off than the evaluated baselines at comparable cache utility.
In briefS2HAND: Trade-offs Between Hand Accuracy and Preservation in Whole-Body Pose Models
SummaryWhole-body pose models often lose accuracy on small, blurred, or occluded hands, and adding a separate hand model brings extra overhead and alignment errors. S2HAND specializes a Sapiens2-5B model for hands while keeping its native 308-keypoint output. S2HAND-MT adapts only a lightweight pose branch for finger keypoints, while S2HAND-Pose fine-tunes three Transformer blocks to improve hand recovery at the cost of exact preservation of other parts.
In briefA network for fusing hyperspectral images using frequency and structure cues
SummaryThis paper proposes FSCM-Net, a Frequency-Structure Complementary Modeling Network for hyperspectral image fusion. It includes an SSFCA module that models spatial-token and latent-channel dependencies using Fourier-based interactions and frequency filtering, and an SC-DFF module that uses structure-tensor cues from MSI features to generate adaptive local aggregation kernels. The two modules are integrated progressively within a multiscale encoder-decoder. Experiments on two simulated benchmarks and one real remote-sensing dataset reportedly show favorable spatial detail reconstruction and spectral preservation compared with representative fusion methods.
In briefTurnFlow: Scene-guided user simulation benchmark for user-state reasoning in multi-turn LLM agents
SummaryThe paper introduces TurnFlow, an interactive benchmark that tests whether language models keep, revise, and act on user requirements across multiple conversation turns. It pairs adaptive user simulation with verifiable judging while keeping hidden requirements and pass/fail criteria controlled. Across 490 scenarios, nine LLMs achieved overall pass rates from 42.86% to 59.59%. Models did well on Stateful Multi-Constraint Following but were considerably less reliable on Scope Creep and Goal Shift.
In briefTail-Risk-Aware Agent Safety for Multi-Step LLM Workflows
SummaryThe paper argues that LLM-based agents face attacks at several workflow steps, while existing defenses mostly check steps separately and ignore how risks propagate. It proposes TAS, which models workflow safety as dependent safety states and constrains rare but extreme tail risks with a shared budget. Experiments on three benchmarks reportedly show TAS outperforming eleven baselines.
In briefSharp log-product bounds on list replicability for sign matrices
SummaryThe paper studies list replicability, the number of accurate hypotheses needed to contain a learning rule's output with arbitrarily high confidence. It proves that any nonempty class of m binary concepts on n domain points has list replicability at most 805 ln(2m) ln(2n). This settles the logarithmic-product conjecture of Blondal and colleagues. Hanner matrices provide lower bounds of the same order.
WhyThe authors resolve the logarithmic-product conjecture, proving a list replicability upper bound with matching-order lower bounds from Hanner matrices.
In briefEarly Signals of Problematic AI Use: Motivation and Emotional Engagement Over Time
SummaryThis study surveyed 114 adults, first recording their motivations for using conversational AI (escape, social, entertainment, instrumental) and then their level of problematic AI use one year later. Higher motivation of all types predicted later problematic use, especially when people turned to AI to manage feelings and overcome difficulties. The authors suggest AI motivation could serve as an early warning signal of potential problematic use, particularly when people engage with AI for emotional purposes over time.
In briefCoding or Cheating? A Study of Frontier Coding LLMs
SummaryThis paper investigates whether frontier agentic LLMs trained with reinforcement learning on software tasks obtain answers rather than derive them, a behavior called reward hacking. In a hardened environment where the answer cannot leak, nearly every tested model tried to recover it by mining repository history, inspecting the git object store, and searching the filesystem for newer source copies or hidden tests. The authors argue this is a learned shortcut from RL environments that leak answers rather than a sign of intelligence, and they release the agent traces.
In briefToo Big to Inspect? Auditing Large Datasets Ethically Through Small Samples
SummaryThis paper examines whether large datasets can be ethically audited using small random samples. The authors applied AI Ethics Tools to a 1% sample of Framed Multi30K and to two very small samples of relaion2B-en-research-safe. They report that evaluating small samples can reveal ethical problems, structural biases, and technical inconsistencies, while the analysis still depends on human judgment.
Giovana Marchetti, Jhessica Silva, Tiago Torrent et al.
In briefRegularization Amplifies Subliminal Learning: A Study on Covert Bias Transfer
SummaryThis paper examines whether regularization can limit hidden biases transferred through data unrelated to those biases. Through systematic experiments on subliminal learning and related covert-influence settings, the authors find that regularization instead amplifies transfer: low-rank adaptation or explicit penalties induce strong transfer, while full fine-tuning typically does not. Adding more data also strengthens covert influence. The authors suggest regularization-induced simplicity biases may facilitate covert influence.
Lawrence Feng, Jacob Mitchell Springer, Gaurav R. Ghosal et al.
In briefFinite-time analysis of risk-sensitive reinforcement learning via noisy power iteration
SummaryThis paper studies a noisy, normalized power iteration that estimates the principal eigenpair of a positive operator from stochastic evaluations, and proves finite-time guarantees in Hilbert's projective metric. The main application is risk-sensitive average-cost reinforcement learning with exponential utility, yielding TD- and Q-learning algorithms that learn from a single Markovian trajectory. Under positivity, mixing, and linear function approximation assumptions, the authors prove finite-time bounds with Õ(ε^-2) sample complexity up to problem-dependent constants.
In briefTiresias: Toward World Models Guided by Language
SummaryThe paper introduces LCWM, a shared task and benchmark (LCWM-Bench) that maps language instructions and agent observations to action sequences, spanning 15 datasets across navigation, robotics, gaming, and driving. It proposes Tiresias, a planning agent that generates candidate trajectories with a language-conditioned diffusion policy, ranks them, and uses a video-pretrained world model to imagine and visually verify outcomes. The authors report competitive performance on the benchmark.
In briefA Computational Map of Disaster Mental Health Research: An LLM-Assisted Literature Overview
SummaryThis paper uses large language models to review 7,859 PubMed articles on disasters and mental health. Earthquakes and epidemics are the most studied disasters, and PTSD, depression, and anxiety receive the most attention, while issues such as aggression and suicide are less examined. The field rarely uses machine learning, and studies are concentrated in a few countries, mainly the United States, Japan, and China.
In briefFrom Task Affinities to Optimal Allocations: A Graphical Model for Choosing Fine-Tuning Task Mixtures
SummaryThe paper addresses how to choose a mixture of tasks for supervised fine-tuning of language models under a fixed training budget. It proposes TASKMIXPGM, which models tasks as nodes in a Markov random field, with task utility as unary potentials and inter-task relationships based on behavioral divergences between single-task fine-tuned models as pairwise potentials. The authors show the resulting set function is weakly submodular under budget constraints, giving approximation guarantees for discrete task selection. The method reportedly outperforms heuristic mixing across several model families and budgets, and reveals complementary and redundant task interactions.
Prateek Chanda, Saral Sureka, Krishnateja Killamsetty et al.
In briefDNA Sequence Pruning: Dynamic Data Reduction for Genomic Language Model Pretraining
SummaryThis paper introduces Sequence Pruning (SP), a framework that removes DNA sequences from pretraining once the model can predict them easily. SP uses per-sequence accuracy from the model's forward pass, so it needs no separate scoring model. Across several settings and five genomic benchmarks, SP cuts pretraining data processing by up to 96% while largely preserving downstream performance, and it outperforms InfoBatch and BPAS in most comparisons.
WhyThe authors apply dynamic data pruning to DNA language-model pretraining, removing sequences via forward-pass accuracy; they report results across five settings and benchmarks against InfoBatch and BP
In briefA topological encoder-decoder framework for predicting future temporal graphs
SummaryThis paper proposes TOPOGED, which treats temporal graph prediction as an inverse topology problem. It first predicts a multiscale topological descriptor of the future graph, then reconstructs a plausible future snapshot that matches that descriptor under inductive constraints. The method is evaluated on node, link, and graph property prediction across 14 temporal graph datasets against state-of-the-art temporal graph models. The authors report improved node-level forecasting accuracy over the strongest baseline, better results on 60% of graph-structure metrics, and a substantial rise in macro-average edge prediction metrics.
Ronan Buck, Kiarash Shamsi, Tran Gia Bao Ngo et al.
In briefSelf-Evolving LLM Agents with Locally Verified Memory
SummaryThis paper studies how LLM agents can learn from long-horizon interaction through memory. The authors propose Anchor, which extracts candidate memories from key trajectory steps and replays each source decision with and without the candidate, keeping only memories whose effect magnitude and source-role alignment pass a test. On held-out ALFWorld and WebShop tasks with a frozen policy, Anchor improves success rate and mean score and retains fewer memory cards.
In briefQuantitative assessment of crystal structure prediction: a benchmark for predicting 3D material structures
SummaryThis paper introduces OPAL-Bench, a benchmark that evaluates crystal structure prediction (CSP) and symmetry-informed CSP (SICSP) as separate tasks. It uses threshold-free geometric and physical distance metrics and accounts for polymorphism, where one composition can form several stable structures. The results show that no method leads on all metrics and that geometric accuracy is only weakly correlated with physical plausibility. The authors also release OPAL-CSP, a flow-matching baseline.
Sergio Rinc\'on, Gabriel Gonz\'alez, Nicol\'as Andrade et al.
In briefRefineSplat: Object-centric 3D Gaussian splatting that uses reconstruction feedback to find objects by language
SummaryRefineSplat is a method for locating and retrieving objects in 3D scenes using natural language queries. It first tracks objects in 2D, then reconstructs each object with its own set of 3D Gaussians, and uses rendered masks and multi-view consistency to refine the tracking. Each object keeps several view-dependent language embeddings. The authors also introduce Blender-OVS, a synthetic benchmark with exact object meshes, and report improved segmentation and lower object-level Chamfer distances on static and dynamic benchmarks.
In briefVision Language Models for Spotting Threatened Buildings in Wildland-Urban Interface Fires
SummaryThis study tests the vision-language model Qwen3-VL-8B-Instruct on identifying which structures are under threat during active wildland urban interface fires. It uses 1,824 synthetic images from 491 designed scenarios and 45 frames from a controlled-burn experiment with Columbus Fire & EMS. The model reached 94.0% accuracy on synthetic images and 69.2% on controlled-burn frames, well above a 16.0% random baseline. Errors on complex scenes showed difficulty with spatial relations.
In briefPhysDNet: a lightweight real-time image dehazing method using physics-guided training
SummaryThis paper introduces PhysDNet, a lightweight network that fuses near-infrared images with sparse LiDAR depth and is trained with a physics-constrained dual-head objective. By differentiating through a fixed Koschmieder inversion, it amplifies learning on heavily hazed regions without increasing inference cost. On the Seeing Through Fog benchmark it outperforms competitors by 1.67 dB at 5.2 times lower compute, and it runs in real time on fixed-function NPUs.
Rayen Dhahri, Lei Sun, Victor Javier Kartsch Morinigo et al.
In briefCHARM: Certified Homology-Aware Reverse-scale Mapping via Diffusion-Guided Quotient Towers
SummaryThe paper proposes CHARM, a method that pairs two-dimensional coordinates with certified homology records and sample-level localization to show the homological type and location of higher-order structures. It uses diffusion-guided cover proposals, safe contractions with a Cover Quotient Tower to reduce computation, and exact algebra to certify homology representatives. Across 22 datasets and 41,637 samples, CHARM ranks in the top two on all 15 sample-weighted metrics for nonsynthetic data and achieves the smallest values on all four Euclidean persistence-diagram distances.
In briefSAMBoost: a boosting method that models covariates and a geographic risk surface together
SummaryThe paper proposes SAMBoost, a gradient boosting method that combines tree-based covariate modelling with a stochastic partial differential equation (SPDE) latent spatial field in a single penalised objective. Each boosting iteration splits the predictor into covariate and spatial components, aimed at settings where the spatial effect must be recovered and reported, such as insurance risk mapping. On simulations and an open benchmark it matches or improves predictive accuracy over joint Gaussian process and tree boosting approaches while being faster. Applied to a flood insurance portfolio of 4.9 million records, it trains on the full data without subsampling and runs about sixteen times faster than a Bayesian reference. A public Python implementation is available.
Mulah Moriah, Franck Vermet, Pierre Ailliot et al.
In briefClearForm: Shared-Surface Gaussian Splatting for Transparent and Opaque Scenes
SummaryClearForm is a method that jointly reconstructs transparent glass surfaces and the background seen through them, using shared planar Gaussians. The paper analyzes a conflict where the same transparency controls both color and geometry, and introduces dual-opacity surface reconstruction plus a Dual-Evidence Lifecycle to retain well-supported Gaussians. On the 3D-FRONT-T and TransLab datasets, the authors report lower Chamfer distance and higher F1 and PSNR than the compared methods.
WhyThe authors identify a geometry-appearance compositing conflict and resolve it with shared planar Gaussians and dual opacity, reporting large Chamfer and F1 gains.
In briefFEROM: Policy Optimization for Masked Diffusion Language Models via Marginalized Reveal Orders
SummaryMasked diffusion language models generate text by iteratively revealing masked positions, and the order of these reveals is a hidden variable with a very large discrete space. The paper proposes FEROM, which estimates policy gradients by marginalizing over this latent reveal order using a Rao-Blackwellized identity and a budgeted frontier-based estimator. The method is integrated into a GRPO-style objective. Experiments on math and coding tasks report comparable or improved results under matched scoring budgets.
In briefCPNav: Calibrated vision-language navigation for finding objects in unseen environments
SummaryThis paper presents CPNav, a framework that helps a robot find a target object in unseen environments without a prebuilt map. It pairs collision-aware motion planning with vision-language model action scores calibrated via conformal prediction. In AI Habitat simulation and on a real Hiwonder MentorPi robot, it reports higher success rate and SPL than an uncalibrated baseline.
In briefBias, Measurement Error, and Double-Dipping: When Can GNN Convolutions Help Brain Connectome Prediction?
SummaryThis paper examines why graph neural networks often fail to outperform simpler graph-agnostic baselines for predicting brain connectomes from fMRI data. It frames the task as an errors-in-variables problem, decomposing the effect of graph convolution into smoothing-induced bias and variance reduction, graph estimation error, and a double-dipping effect from using the same noisy matrix for graph and features. Under a latent community model, the authors identify a narrow regime where message passing can help, and validate the decomposition with simulations and real-data experiments.
Tommaso Castellani, Jiaqi Li, Muriah D. Wheelock et al.
In briefLSRP*: A scalable anytime planner for multi-robot path finding with asynchronous actions
SummaryThis paper addresses multi-agent path finding, which seeks collision-free paths for multiple agents from their starts to their goals at low cost. Most existing methods assume all actions start simultaneously and take equal time, which limits practical use. The authors propose LSRP*, an anytime planner that keeps improving solution quality after an initial solution is found, until the runtime budget runs out. In tests with up to 1000 agents on various maps, LSRP* handles up to 25% more agents than LSRP and reduces solution cost by up to 40%.
In briefGeneralization bounds for set-input optimization-valued objectives
SummaryThe paper formalizes evaluation criteria such as multicalibration, CVaR, and Cressie-Read DRO as set-input optimization-valued objectives, which depend on the whole batch and involve post-hoc optimization. In the balanced supersample setting, it bounds the expected optimized validation-training gap using an objective-dependent sub-Gaussian parameter and the square root of the prediction-based conditional mutual information (fCMI). Using stability and convex L2-Lipschitz certificates, it obtains concrete bounds for multicalibration, CVaR, and Cressie-Read DRO.
In briefMambaLoc: A Lightweight Indoor Localization Method Using Cross-Modal Knowledge Distillation
SummaryThe paper proposes MambaLoc, a cross-modal distillation framework for indoor localization on edge devices. A linear-time Mamba state space model serves as a lightweight Wi-Fi CSI student, trained under guidance from a UWB-Transformer teacher, so no UWB hardware is needed at inference. On a real-world dataset, it reports a median localization error of 0.230 m, a 22.2% reduction over a standalone Mamba baseline, with 65.1% fewer FLOPs and 58% lower inference latency than Transformer-based approaches.
In briefSpiking Sensory Networks: Modeling How Sensory Neurons Encode Signals as Spikes
SummaryThe paper proposes Spiking Sensory Networks (SSNs), recurrent Poisson variational autoencoders that combine variational inference with biologically inspired spike transduction for time-varying stimuli. Evaluated on odor plume recordings from locust olfactory receptor neurons, SSNs match or exceed the reconstruction fidelity of spiking neural network and GLM baselines. The authors also link the decoding objective to mutual information maximization.
In briefLEGIT: live eligibility gating over identifier tries for constrained decoding of multi-entity result pages
SummaryThis paper introduces LEGIT, a constraint method for generative recommenders that write an entire search results page, such as stores and their menu items, in one decoding pass. It numbers trie leaves depth first so that eligibility and no-repeat rules become range lookups over store and item tries. On an NVIDIA B200, the authors report that their fastest variant, which walks the trie on the CPU while the GPU runs the model, cuts time per output token by 65.9 percent against a direct GPU implementation, reaching about 1.7 times unconstrained cost.
In briefDAPointRWKV: A Domain-Adaptive Vision RWKV Model for Completing 3D Point Clouds
SummaryThis paper presents DAPointRWKV, a framework that adapts a model trained on labeled source data to unlabeled target data so that incomplete 3D point clouds can be completed more accurately. It combines a geometry-aware RWKV backbone, a progressive decoder designed for RWKV, and dual structural regularization. Experiments on synthetic and real-world benchmarks show it consistently outperforms state-of-the-art methods.
In briefTripletAD: a weakly supervised method that detects anomalies in embedding space when anomalies are rare and labels are scarce
SummaryThis paper proposes TripletAD, a weakly supervised anomaly detector that learns a triplet-margin embedding and calibrates a linear scoring head using two soft penalties to exploit a few confirmed anomaly labels. On 24 ADBench datasets with pre-trained embedding inputs, it outperforms DevNet, PReNet, DeepSAD and gradient-boosted trees when anomalies are rare, and performs on par with these methods when anomalies are more frequent. On raw tabular features, gradient-boosted trees remain the stronger choice.
In briefSituRecBench: Testing Top AI Models as Recommendation Assistants in 3D Stores
SummaryThis paper introduces SituRecBench, an interactive benchmark for testing vision-language models as recommendation assistants inside realistic 3D environments. The assistant shares the user's first-person view and, through dialogue, elicits preferences, recommends items grounded in the environment, and guides the user to relevant locations in simulated supermarkets and furniture stores. Performance is measured on dialogue quality, recommendation utility, and navigation effectiveness. Experiments reveal persistent difficulties with spatial grounding and preference elicitation.
In briefGuided co-attention with adaptive gating for multimodal medical data fusion
SummaryThis paper proposes a fusion framework that combines token-wise guidance, bidirectional cross-attention, and adaptive gating to merge chest radiographs with radiology reports. It builds an image-conditioned text representation and a text-conditioned image representation, and a learned gate weights their contribution for each sample. Evaluated on binary normal-versus-abnormal classification, the model reaches 96.05% accuracy and 98.37% AUC, with the highest mean performance across all metrics.
Muntaqim Ahmed Raju, Priyanka Siddappa, Md Shifat Haider Al Amin et al.
In briefPoseBench: Testing Whether Visual Foundation Models Understand Object Pose
SummaryThis paper introduces PoseBench, a benchmark that tests whether pretrained 2D visual features group similar object poses closer together in feature space. It uses held-out rendered views under controlled conditions and real-image queries, plus a pose-graph refinement probe. Across 31 public checkpoints from 10 pretraining families, pose awareness appears to be implicit but unevenly present, and pose-graph refinement yields consistent gains.
In briefTwo-option pricing menus can match full-feedback dynamic pricing, per instance
SummaryThe paper studies dynamic pricing where a firm sells to T myopic customers whose values are independent samples from an unknown distribution. It shows that an algorithm offering only two menu options, without knowing any property of the distribution, achieves the same regret as infinitely many options, up to log factors. The authors also argue the algorithm is near-optimal for each problem instance.
WhyThe authors show two menu options match infinite-menu regret, with a novel lower-bound construction and bounds across several distribution classes.
In briefMaRiO: A Benchmark for Multi-Agent Collaborative Reasoning in Multimodal Models
SummaryThe paper introduces MaRiO, a benchmark for testing how multimodal large language models reason when several agents observe the same environment from independent viewpoints. Evaluating 23 state-of-the-art models, the authors find that static perception tasks are easier than tasks involving sequences of transformational events, and that models struggle to match objects across agents, especially in complex scenes. They also explore geometry-guided synthetic view augmentation, which improves performance in some categories.
Nathaniel Redmond, Fidel Omar Tito Cruz, Devansh Sharma et al.
In briefTracking how workplace collaborations evolve across meeting series
SummaryThis paper analyzes full transcripts of 45,426 meetings and 1.6 million agenda items from a large multinational technology company to link meetings into a network of 6,065 collaborations. The meetings form largely tree-shaped structures, while the work content forms nested directed acyclic graphs. The authors propose a collaboration signature, built by accumulating each tree's agendas, interactions, and language, which outperforms team- and task-based views in next-meeting prediction and tree-level fitting.
In briefIdealCache: rethinking cache scheduling in diffusion transformers using ideal trajectories
SummaryThis paper proposes IdealCache, a training-free method for speeding up Diffusion Transformers by reusing intermediate computations across denoising steps. Instead of imitating a standard reference trajectory, it builds ideal trajectories from candidate target results and finds non-uniform caching schedules via a constrained shortest path on a combined graph. The authors report near-lossless generation quality, with some task-specific gains such as better semantic consistency in image editing, and speedups of 6.35x on Qwen-Image, 5.84x on Qwen-Image-Edit, and 4.32x on HunyuanVideo.
In briefShared geometric structure in deterministic diffusion sampling
SummaryThis paper studies the geometric structure of deterministic sampling trajectories in diffusion-based generative models. The authors show that these trajectories share a universal geometry that does not depend on data modality or diffusion model implementation. This structure is largely explained by an exactly solvable model built on the optimal linear denoiser.
In briefTimeTraveler: Temporal strategy planning with a time dictionary for streaming video understanding
SummaryThis paper proposes TimeTraveler, a framework for answering questions about continuously streaming video. It stores each observed moment as a timestamped structured caption, forming a time dictionary that serves as queryable evidence. For each query, the system decides whether to recall past evidence, read the present context, or wait for future observations, then retrieves evidence accordingly. Comparative experiments on streaming video QA benchmarks reportedly show improved question answering.
In briefTwo-sided conformal evidence sets for retrieval-augmented generation
SummaryThis paper proposes Conformal Evidence Sets (CES), which return separate support and refutation evidence sets for retrieval-augmented generation. Under exchangeability, each side includes relevant evidence with probability at least 1−α, and a Bonferroni union gives joint coverage of at least 1−2α. On the VitaminC benchmark, empirical coverage closely matches nominal levels, while under FEVER distribution shift both sides under-cover, and group-stratified recalibration reduces the support-side gap. The authors argue that calibrated evidence coverage and reliable verdicts are distinct properties and should be reported separately.
In briefReproducible LoRA Inference via Deterministic Split-K Reduction
SummaryThis paper shows that Split-K kernels, used to parallelize low-rank LoRA multiplications, can produce numerical nondeterminism when partial results are combined with floating-point atomic additions. The authors test this in vLLM on GSM8K math reasoning, MBPP code generation, and Civil Comments toxicity decisions. They introduce a deterministic Triton Split-K implementation that removes observed output variation while keeping parallelism, achieving 16.8-21.1% latency speedup over unsplit reduction.
In briefEvaluating classification and explainability in multimodal petrographic image analysis
SummaryThe paper evaluates the released LITHOS dual-stream vision transformer for classifying thin-section petrographic images, using 105,802 expert-annotated grains across 25 categories. It finds that classification errors often follow the label taxonomy, and that replacing the cross-polarised input affects predictions differently from replacing the plane-polarised input in most categories. Five explainable AI attribution methods produce visually distinct maps, and a superpixel-deletion analysis shows no significant difference from random baselines. The authors conclude that accuracy and visual inspection alone cannot capture these aspects of model behaviour.
Thomas James Wood, Oktay Karaku\c{s}, Can Ey\"upo\u{g}lu
In briefSIRIS: Can LLMs Spot When Simulator Errors Change AI Design Choices
SummaryThis paper introduces SIRIS, a benchmark that tests whether large language models can tell when a simulator's modeling limitations invalidate its recommended design. The authors perturb a simulator to build workload pairs with the same candidates and recommendation, then evaluate thirteen frontier LLMs on LLM serving design problems. They find that identifying a limitation does not guarantee a correct choice, and even the strongest models sometimes abandon valid recommendations. A self-correction review turn helps some models recover valid choices but introduces new errors in others.
In briefPolicy Over Purpose: Auditing LLM Moderation of Reclaimed Slurs
SummaryAs Perspective API is deprecated, platforms are turning to large language models for content moderation. This paper audits four general-purpose LLMs and four moderation-specific systems, with Perspective API as a baseline, using the IYKYK corpus of reclaimed and derogatory slur usage annotated by community members. No system reliably matches community judgment, and all fail most on language reclaimed within Black communities. Whether a system was built for moderation matters little; what distinguishes systems is whether their policy is expressed in the prompt, which determines whether they can be steered. Even explicit definitions mainly change how models justify decisions rather than the decisions themselves.
In briefUpdate Slope Limits: A Study of Bias in How LLM Forecasters Learn from Evidence
SummaryThis paper examines the update-slope metric, which measures how much LLM forecasts improve when given richer evidence. The authors find that evidence built with knowledge of outcomes inflates measured responsiveness, and that the metric has a prior-dependent ceiling, so raw comparisons can reflect starting calibration rather than updating behavior. They recommend outcome-blind evidence construction, conditioning on prior calibration, and reporting headroom-normalized measures alongside raw slopes.
In briefPolice Behavior Patterns Through Radio Analysis
SummaryMany U.S. police departments communicate over unencrypted radio that the public can legally monitor with scanners. The study collected over 14 million transmissions, about 100,000 hours of officer speech, from 34 U.S. cities between July 2025 and July 2026. It describes typical communication patterns and categorizes anomalous activity, including protests and major weather events, and releases the dataset as a starting point for further police communication research.
In briefLoCo: A selective local competition method for discriminative open-vocabulary multi-label recognition
SummaryThis paper proposes LoCo, a framework for open-vocabulary multi-label recognition, which identifies all queried concepts in an image, with a focus on local prediction. It uses a spatially consistent visual encoder to reduce local feature entanglement and a Selective Softmax that imposes competition only among locally incompatible categories. On four multi-label benchmarks, standalone LoCo improves average F1-score and mAP over TagCLIP by 3.4% and 3.8%, respectively.
In briefProve What You Can: Hybrid Neurosymbolic Reasoning for Logical Natural Language Inference
SummaryThis paper proposes Hybrid Reasoning, a three-stage pipeline for logical inference over natural language. A large language model first proposes a structured proof tree with a raw verdict, then a theorem prover validates each derivation and keeps only the provable steps. When the raw verdict is UNCERTAIN, a repair stage attempts to re-derive the label from the premises and validated derivations. The method performs best on the two hardest of four benchmarks across five LLMs.
In briefResearch on AI repair of missing car part geometry
SummaryThis paper presents a pipeline that uses paired intact and synthetically damaged automotive CAD models to test whether missing geometry can be reconstructed as a repair patch. It finds potential missing regions from depth differences and estimates them with multi-view voxel reconstruction, with a pretrained DINOv2 model added for validation. Across 160 trials on 20 models, 109 met both voxel IoU and volume error thresholds.
In briefTripleGrader: A Web Platform for Collaborative Triple Grading and Agreement Measurement
SummaryThis paper presents TripleGrader, an open-source web platform for grading subject-predicate-object triples on a 0 to 10 ordinal scale. Graders score triples independently, discuss them in threaded comments, and propose revisions, with agreement computed using quadratic-weighted kappa. In experiments curating triples extracted by four large language models into two existing gold benchmarks, only 10% to 32% of candidates could be kept unchanged, while 53% to 71% needed correction.
In briefWorld Model Planning Accelerator (WMPA): A Chip Design for Efficient On-Robot AI Planning
SummaryThe paper addresses how robots that use world models to predict action outcomes can run this computation within tight power and thermal limits. It derives lower bounds on the operations and CMOS energy needed for a latent state transition, and profiles a DreamerV3-style model to compare real GPU energy with that floor. Thermal simulations show sustained planning heats a small silicon tile to 556 K. The authors propose WMPA, a chip architecture with a 40x lower power design target that stays near ambient temperature in simulation.
