In briefMultics: A generative multi-player world model using hierarchical communication for shared video simulation
SummaryThis paper presents Multics, a framework that generates video of shared environments from multiple players' viewpoints. It organizes information exchange across player, group, and global levels, using group tokens to aggregate locally relevant information and limit mixing between unrelated events. It also uses budget-constrained dynamic grouping based on recent camera trajectories. The authors report state-of-the-art results in 4- and 6-player settings and generalization from 4-player training to 10-player inference without retraining.
In briefDeleting Code Without Losing Knowledge: Regenerative Software for Python Libraries
SummaryThe paper proposes regenerative software, in which machine-readable records of a module's architecture, intent, decisions and public contract allow code to be deleted and rebuilt. The authors release spotregen, an open-source package implementing these primitives, and apply it to the safety-critical forecasting library spotforecast2-safe. The report includes executable examples for every primitive.
In briefAgent Ads: On-Policy Attacks That Steer AI Agents' Choices
SummaryThe paper examines how people increasingly delegate option-ranking tasks such as shopping, job search, and travel to AI agents, creating incentives to steer them through "agent ads." The authors introduce on-policy attacks, which inject a truthful negation about a fictitious concept into a product listing to stay within platform rules, then rely on planted websites with fabricated harmful claims to make the agent penalize competitors. Evaluated on four real-world listing datasets, the attack exposes some frontier models, and the authors offer recommendations to model developers and platform providers.
WhyProposes novel on-policy attacks evading alignment, exposing a new agent-ad attack surface; validated on four real datasets with limited evidence breadth.
Tim R. Davidson, Marija \v{S}akota, Joshua Joseph et al.
In briefSize generalization for diffusion routing solvers: the ResizeCO method
SummaryThis paper addresses routing problems such as the Traveling Salesman Problem and Capacitated Vehicle Routing, which are hard to solve at large scale. It proposes ResizeCO, a label-free method that uses reinforcement learning on large instances to adapt a diffusion model trained on small ones, and formulates k-opt as a tree search solved by beam search on a GPU. In experiments on instances with up to 10,000 nodes, ResizeCO's Pareto front of route quality versus time outperforms learned solvers and, at short run times, other solvers.
Barak Itzchak Davidovitch, Eli Meirom, Ethan Fetaya et al.
In briefFreeSolo: Fast Sparse Attention That Generalizes to Longer Inputs Without Retraining
SummaryThis paper introduces FreeSolo, a sparse attention mechanism combining a short-range sliding window with a long-range branch that selects keypoint cells for direct token attention. Its linear-time variant FS-L learns to prune the prefix through short continued pretraining, while FS-B searches balanced trees at O(N log² N) cost including radix-sort construction. The authors report that FreeSolo achieves better zero-shot length extrapolation than baselines without long-context fine-tuning.
In briefCartoVec: General-purpose location embeddings built from open map data
SummaryCartoVec is a coordinate encoder trained, without image inputs or task labels, to predict land cover, land use, transportation, place and connectivity features derived from Overture open map data at multiple spatial scales. The frozen embeddings were tested on US Northeast Census outcomes, Airbnb prices and Philadelphia crime using spatial block cross-validation. CartoVec achieved the highest mean performance on most Census outcomes and the best average across the full suite, though several comparisons were close. The open-weight model is available on Hugging Face.
In briefSoft Teacher Forcing: a training method using soft tokens to improve supervised fine-tuning
SummaryThis paper introduces Soft Teacher Forcing (Soft-TF), a simple training algorithm that mixes teacher-forced embeddings with soft tokens through random convex combinations. The soft tokens come from a single parallel forward pass over the ground-truth sequence. Soft-TF adds no parameters, external models, or curriculum, and inference is unchanged. On GSM8K, it improves over SFT by +2.1 to +4.1 points across seven Gemma, Llama and Mistral models, and it is also evaluated on MATH, MathQA and AQuA.
In briefDuoDiscover: A system that scales autonomous discovery past costly evaluation
SummaryThis paper presents DuoDiscover, a search system in which coding agents jointly find an optimization target and a fast proxy evaluator, reducing reliance on expensive true evaluations. Fidelity tests recheck each proxy's ranking and discard those that drift, while a lightweight orchestrator shares confirmed results among agents. The authors report gains over human experts and prior methods on SWE-fficiency, SOL-Exec, vLLM, SQLite, GPU/TPU kernel suites, and a CPU microarchitecture design task.
WhyProposes a search framework using proxy evaluators with continual fidelity checks, showing large gains across benchmarks and CPU microarchitecture design.
In briefChemExplain: A Multimodal Benchmark for Explaining Chemical Effects
SummaryThis paper introduces ChemExplain, a benchmark of 1,180 multiple-choice chemistry questions drawn from 383 peer-reviewed papers. Each question pairs a molecular or reaction image with an observed effect and four candidate explanations tied to different structural features. The strongest evaluated model reaches 78.6% accuracy, and on a 600-question expert-reviewed subset, 71.4% of its errors are classified as attribution failures.
In briefModRes: A cleaned version of the ResPlan dataset for floor plan analysis and graph machine learning
SummaryThis paper presents ModRes, a cleaned and enriched version of the ResPlan residential floor plan dataset, containing 8,431 valid and unique plans. Each plan includes 2D geometry, images, 3D building representations and three dual graphs with graph-, node- and edge-level features in CSV and JSON formats. The authors characterise the dataset with graph-based statistics and demonstrate its use through several case studies.
In briefFunDM: Cheap, label-free decision models built on frozen language models
SummaryFunDM (Fun Decision Model) adds a small decision head to a frozen language model and trains it label-free by matching the full model's answer probabilities via KL divergence. Using 24 of Qwen3-8B's 36 layers, it reaches 75.0% mean accuracy across eight benchmarks, versus 75.7% for the full model. On MMLU-Pro it achieves about 1.56 to 1.57 times batched vLLM throughput with roughly one-third lower peak GPU memory.
WhyDistills a frozen LM's answer probabilities into a label-free partial-layer head, with broad benchmark comparisons and efficiency measurements.
In briefA generative flow method that discovers robot-specific capabilities using symbolic task abstractions
SummaryThis paper proposes FOSA, a framework that uses a Generative Flow Network to build symbolic task skeletons, each compiled into a simulation environment. The low-level policy learning performance in each environment serves as the reward signal for the abstract policy. Across heterogeneous robot platforms, FOSA discovers diverse, executable tasks near the robot's empirical capability frontier. The discovered task policies are distilled into a single language-conditioned policy, which the authors report improves performance on unseen complex, long-horizon tasks.
In briefA Workflow for Uncertainty-Aware Ionic Liquid Solvent Discovery
SummaryThe paper presents a workflow that couples data processing, deep learning, and physics-based calculations to predict ionic liquid properties and select candidates. It links property predictions with uncertainty estimates and physics-based filtering to enable more reliable ranking of untested molecules when data is scarce and noisy. The authors demonstrate the approach by designing novel ionic liquids for CO2 capture from flue gas and selecting Pareto-optimal candidates based on CO2 solubility, viscosity, and working capacity.
In briefLearning How Natural Odor Mixtures Are Perceived Using LLM-Generated Descriptions
SummaryThe study examines how molecular composition relates to perceived similarity in natural odor mixtures, using a database of over 1,200 food mixtures containing more than 8,000 molecular components. Large language models generate mixture-level perceptual descriptions that serve as scalable supervision for learning low-rank molecular embeddings. These embeddings predict similarities of held-out mixtures better than component-overlap metrics, including pairs sharing no molecules, and align significantly with an independent human odor-descriptor space.
In briefForce-Modulated Dexterity: Unified Compliance for Contact-Rich Manipulation
SummaryThis paper introduces FOMO-Dex, a unified compliance method for contact-rich robot manipulation that expresses actions in task-defined coordinates and applies force-driven motion corrections along selected directions. Across card picking, paper cutting, screwdriving, and nut-and-bolt assembly, it reduces excessive contact forces compared with stiff control. The authors also report that policies trained on the corrected actions achieve higher card-picking success and more rotation in nut-and-bolt and screwdriving tasks.
In briefA generalized McDiarmid inequality for sample reuse and unbounded, non-uniform differences
SummaryThis paper develops a generalization of McDiarmid's concentration inequality and uses it to derive a uniform law of large numbers and error bounds for stochastic optimization. The results account for reuse of training samples and allow mean-difference bounding functions that are unbounded and non-uniform in one component. The authors apply the tools to score-based generative modeling, where integrating out auxiliary Gaussian variables yields tighter error bounds. The approach also covers unbounded score models and data distributions with unbounded support.
WhyGeneralizes McDiarmid's inequality to handle sample reuse and non-uniform unbounded differences, applied to score-based generative modeling; evidence is mainly theoretical.
In briefHow the scoring format limits what an LLM judge can distinguish
SummaryThis paper studies how the scoring format (binary verdict, 1–5 rating, 0–100 score, or comparison) used by an LLM judge affects instruction-tuning data filtering. With the candidate pool, prompt, and selection budget held fixed, the authors scored 15,406 UltraFeedback candidates under five formats and fine-tuned 66 models. They find the format largely determines score resolution and ties, almost independently of the judge, and that Kendall's τb can hide these differences.
Lovedeep Gondara, Jekaterina Novikova, Sagarika Suresh et al.
In briefNeuromorphic Decoding of Retinal Ganglion Cell Spikes to Reconstruct and Track Moving Images
SummaryThis paper decodes moving visual scenes from spikes recorded in up to 1,836 retinal ganglion cells from mouse retina, using an end-to-end spiking neural network. The network reconstructs images with a structural similarity of up to 0.93 and tracks a moving digit with a mean error of 1.33 pixels on a 32×32 grid. Its estimated compute energy is about 70 times lower than a dense artificial network of the same architecture, suggesting potential for implantable and wearable visual prostheses.
WhySpike-domain end-to-end reconstruction and tracking of dynamic scenes, benchmarked against 17 larger decoders with energy estimates, offering substantive advances for visual prostheses.
In briefTwo ways to reparameterize folded Normal and Rice distributions for variational inference
SummaryThis paper derives implicit and explicit reparameterized gradient estimators for the folded Normal and Rice distributions, implemented in PyTorch, and compares their variances. The implicit estimator has lower variance for the location derivative and is proven never to have larger variance than the explicit one. The authors apply both estimators to French-Wilson scaling in crystallography, recovering the exact posterior amplitude to better than 1%, and release the code in the rs-distributions package.
In briefSCOPE: Using Conformal Scores to Guide Fast, Calibrated Bayesian Prediction Intervals
SummaryThis paper proposes SCOPE, a method for Bayesian predictive regions that stay well calibrated even under model misspecification. It uses score gradients on a held-out guide set to identify the few parameter directions that move the region's boundary, explores that subspace with MCMC, and handles remaining uncertainty with a projected variational approximation. On benchmarks, SCOPE's median test NLL is within 0.002 nats of full HMC, and it runs 12 to 15 times faster on the real datasets.
In briefOpenDGO: Benchmarking Multi-Agent Organizations Across Application Domains
SummaryThis paper presents OpenDGO, a controlled benchmark and analysis framework covering six real-world domains and three collaboration settings: sequential dependency, parallel decomposition, and information integration. Under matched tasks, agents, tools, and resource limits, it compares fixed and learned organizations and separately examines communication graphs, role and input assignments, resource budgets, and component failures. The results show no single organization works best everywhere, so suitable organization design depends on the application.
In briefMARS: Multi-level Speedup Method for World Action Models
SummaryWorld Action Models (WAMs) predict future world states and robot actions together, but their visual encoding and iterative generation cause high latency that limits high-frequency closed-loop control. This paper profiles WAM inference, identifies redundancy, and introduces MARS, a training-free, plug-and-play framework that accelerates attention computation, diffusion denoising, and consecutive policy queries. Experiments on LIBERO, LIBERO-PRO, and RoboCasa show end-to-end speedups of up to 4.00x, 3.94x, and 2.10x, with task success rates close to the original policy.
WhyCombines training-free attention, diffusion, and visual-reuse acceleration for WAMs, with broad benchmark experiments showing up to 4x speedup, useful for the subfield.
Zengyuan Zuo, Yongchang Zhang, Guangyu Zhao et al.
In briefMasked Attention Pooling Invariant to Padding and Permutation: Lean Proof
SummaryThis paper studies attention-based multiple instance learning, where batched bags are padded to equal length and the pooled output should not depend on the amount of padding or the order of instances. The authors define masked attention pooling with a guarded normalizer and prove in Lean 4 with mathlib that it is invariant to bijective row reindexing and to appending masked rows, and that an all-masked bag maps to zero. They connect these results to the JAX kernel's Jaxpr using the jax-lean pipeline, but the guarantees cover ideal real arithmetic only and exclude floating-point execution, XLA, and the JAX tracer and importer.
In briefFeature scaling outperforms model architecture in offline respiratory sound screening for low-resource clinics.
SummaryThe study compares classical and deep classifiers on a 1,211-clip respiratory sound dataset covering five classes, using patient-grouped cross-validation. Standardizing SVM features raised macro recall from 0.51 to 0.78, a larger effect than any difference between model families. Augmentation methods showed no measurable benefit, and the classical models avoid a 449 MB PyTorch runtime, suiting GPU-free clinics.
Abubakar Chilala, Cleophas Kadima, Innocent Anyaele et al.
In briefSonicWorld: A Model That Tracks Sound Sources to Reason About Spatial Audio
SummaryThe paper introduces Sonic-Bench, a benchmark of 2.8K first-order Ambisonics clips and 28.8K question-answer pairs from 480 simulated indoor scenes, covering localization, distance, and motion reasoning. It also proposes SonicWorld, which maintains persistent source states through learned transitions and recurrent updates before passing them to an Omni LLM. On the Sonic-Bench test split, SonicWorld improves direction accuracy from 31.98% to 34.69% and reduces azimuth error from 57.22° to 48.01° compared with Spatial-Omni.
In briefEgoProbe: A Benchmark for Long-Range, Exact Memory in Streaming First-Person Videos
SummaryEgoProbe is a benchmark for testing how well AI models remember events in streaming first-person videos, covering both long-range retention and exact recall amid similar events. It includes two subsets, EgoProbe-Base for long-range recall and EgoProbe-Echo for recall with related distractor events, using coarse-to-fine multi-turn questions. Evaluations across various models show clear limits in long-range recall, and models often fail to recall specific details of a designated event even when answering coarse questions correctly.
In briefToolHACE: A Dataset for Spotting Hallucinations in Tool-Based LLM Replies
SummaryTool-augmented language models ground their answers in external tool outputs, yet they can still contradict those outputs, omit key details, or make unsupported claims. The paper introduces ToolHACE, a span-level benchmark covering five fine-grained hallucination categories in tool-augmented LLM responses. Experiments with modern hallucination detectors show that post-tool-call hallucinations are widespread and hard to detect. The authors find that effective detection requires training on specialized tool-augmented corpora, as transfer from tool-agnostic datasets performs poorly.
Nikolay Ivanov, Sergey Pletenev, Elisei Rykov et al.
In briefD2: Turning Half-Duplex Speech Models into Full-Duplex Ones
SummaryThis paper introduces D2, a framework that converts pretrained half-duplex spoken language models into full-duplex models that can listen and speak simultaneously. It adapts the audio encoder for streaming using decoder-free representation distillation and learns full-duplex interaction through supervised fine-tuning. Applied to Qwen3-Omni at a 12.5Hz update rate, the resulting model reports 99.2% turn-taking success, 521 ms latency, 99.5% interruption success, and a VoiceBench score of 70.1.
Cheng-Kuang Chang, Alexander H. Liu, Brian Kingsbury et al.
In briefWhen LLM agents see a conflict but still proceed in network slicing
SummaryThis paper studies how LLM-based autonomous agents handle cross-domain intent conflicts in network management, using a smart-grid case where an idle-looking backup slice carries protection traffic. In a tool-based environment tested across ten models from four providers over 600 runs, most models removed the protection slice at least once after observing the dependency. The authors conclude that dependency verification should be enforced by the management architecture before irreversible actions, not left to the LLM's judgment.
In briefGluing Local Contexts into Global Meaning: A Transformer Interpretability Method
SummaryThe paper proposes a training-free method that uses paraphrase pairs to test whether a transformer's representation keeps the same content when only the wording changes. It builds a cellular sheaf over the paraphrase graph and treats its kernel H^0 as a content subspace, which outperforms the LEACE concept eraser on CounterFact fact retrieval. Ablating the most paraphrase-variable coordinates changes model outputs far more than ablating coordinates of matched variability.
WhyApplies cellular sheaf theory to paraphrase-consistent decomposition of transformer representations, validated with causal ablations and baselines across many models.
In briefD-World: A Robot World Model That Corrects Itself from Measurements
SummaryThe paper presents D-World, which represents a robot's environment as an executable digital twin that a perception agent builds and a planning agent uses to choose actions. After each executed action chunk, the planner revises the twin's object states and physical parameters based on what the robot measured, while the language model stays frozen. On a dual-arm robot performing towel and toaster tasks, local revisions allowed execution to continue without full reconstruction.
In briefCalibrated model for typed entity, relation, and event extraction
SummaryThe paper presents CalOn, a unified 0.6B-parameter encoder that extracts entities, relations, and events across datasets and schemas. It jointly encodes text and type definitions and directly identifies spans, span-pair relations, and event triggers and arguments. Across 40 benchmarks it outperforms the evaluated baselines on most, and a training-bias correction improves confidence estimates with little change in F1.
In briefAIPOCH Open-Science: An open-source AI research workbench that packages work into reproducible .science files
SummaryThe paper presents AIPOCH Open-Science, an open-source research workbench that saves AI-assisted research instructions, computations, revisions and checks into portable .science session packages, and specifies how these records are exchanged. Controlled tests compared complete and partial exports and checked that tested versions and review links were preserved on import. In a registry case, offline re-execution of a retained script with fixed input reproduced two saved CSV tables byte for byte.
In briefMasked Diffusion Model for Synthesizing Travel Diaries
SummaryThis paper introduces DiaryBridge, a masked diffusion model that treats each household's daily travel record as one whole token sequence instead of splitting it across models or cropping it to fixed-length vectors. The model fills in hidden positions in any order, so one trained model can both synthesize complete diaries and complete partially observed ones. Experiments on data from the 2017/2018 Regional Travel Survey of the Baltimore region show that it generates travel reflecting vehicle availability and employment status and better reproduces aggregate demand than the learned baselines.
Pascal Jutras-Dub\'e, Raziq Zaman, Fabian Bastin et al.
In briefLocalTrain: integer-only local learning for low-energy CNN training on edge FPGAs
SummaryThe paper presents an FPGA system that trains a CNN and its regional auxiliary classifiers on a shared integer matrix engine, reducing cross-layer training dependencies. On a KV260 prototype at 200 MHz, it reports 76.57% accuracy on CIFAR-10 and 91.85% on SVHN at about 4.20 W mean module power. Compared with the same integer method on Jetson Orin Nano, measured training energy is 40.4% and 42.1% lower, but runtime is 1.86 and 1.82 times longer.
In briefContrastive learning for detecting reused video footage
SummaryThis paper proposes a framework using a Siamese network fine-tuned with contrastive learning to identify possible source footage for key frames in a video. The authors show that reducing a video to key visuals yields sets under 10% of the full size. These reduced sets enable efficient identification of component footage. Tested on videos from a stock footage library, the method achieves an average accuracy above 0.95.
In briefUniversal Calibration for In-Context Classification: Introducing CALIBFORMER
SummaryIn-context learning improves classification accuracy without fine-tuning, but its effect on confidence calibration has been little studied. This paper introduces CALIBFORMER, a transformer-based calibrator that uses model predictions on in-context examples to estimate confidence for test inputs. Trained on three models and ten tasks, it generalizes to new tasks and four unseen LLMs, including Gemini. The authors report over 50% reduction in calibration error.
In briefPost-training fine-tuning leaves a reproducible shift in model internals; reruns calibrate the measurement
SummaryThe paper argues that comparing a base model with one fine-tuned descendant can mistake run-to-run noise for a training effect. Using repeated LoRA-SFT reruns on Qwen3-1.7B, the authors score each contrast against its own recipe's variation, and replicate the main estimate on Llama-3.1-8B. Supervised fine-tuning produces a large, reproducible structural displacement, while DPO and light GRPO results do not reach a verdict and the causal sweep supports only calibrated displacement, not a changed mechanism.
WhyCalibrates post-training structural displacement against same-recipe rerun noise, with a pre-registered causal sweep; novel framing, but limited to one backbone and recipe.
In briefWhat the Machine Invented: Mining New Math Tools from OpenAI's AI-Generated Proofs
SummaryIn October 2026 OpenAI released 722 AI-written mathematical manuscripts grouped into 372 problem families, including claimed solutions to many prominent open problems. This paper extracts new lemmas, constructions, inequalities and proof mechanisms from them, which it calls mathematical instruments, and distinguishes them from established tools. It examines 38 families across several fields, and its novelty assessments are provisional. No Lean formalization or kernel checking was performed.
In briefInteractive Navigation Beyond Reachability: A Benchmark and Agent Framework
SummaryThe paper notes that existing embodied navigation tasks usually assume fixed traversability and object states, while reaching a goal may require opening doors or containers. It introduces MultINav-Bench, a 3K-episode ObjectGoal benchmark with channel, container, and mixed tasks, annotated for whether interaction is necessary. It also proposes MultINav-Agent, a three-stage framework built on a persistent, state-conditioned interaction graph, which reportedly outperforms adapted baselines across task types.
In briefSuperintelligence and AI Race Risks: A paper on dangers of a US-China race to build AGI first
SummaryThe paper argues that a race between the United States and China to reach artificial general intelligence first poses three dangers: being overtaken by a rival, triggering war, and losing control to AI far more intelligent than humans, endangering survival. It warns that racing may create a race to the bottom on safety practices and regulation. The authors contend the US should prioritise both unilateral and cooperative efforts to improve AI safety and avoid provoking such a race.
Matthew Rendall, Max Hellrigel-Holderbaum, Leonard Dung
In briefExtending unified multimodal models to long interleaved text-image generation
SummaryThe paper studies how unified multimodal models that alternate text and image generation can handle much longer generation trajectories. It introduces a visual-memory framework that keeps persistent visual references, selectively retrieves scene-relevant past states, and uses few-step self-rollout to adapt to imperfect generated histories. The authors also release LongMIG-Bench for evaluating long-horizon multi-image generation, and report strong character and scene consistency.
In briefRethinking LLM-Guided Visual Question Answering: Does Deeper Knowledge Integration Really Help?
SummarySome image questions need knowledge beyond what the pixels show. The study prompts a large language model to describe relevant visual concepts, distractors and external facts, then injects this guidance into a frozen vision model in two ways. On the OK-VQA benchmark, shallow late fusion outperforms deeper attention injection, reaching 39.4% exact match versus 31.8% under data-scarce fine-tuning. The authors conclude that integration depth should match the data regime.
Aneerban Saha, Vijay Kumar, Sameer Madhav D. et al.
In briefUZHCL_for_the_win at Touché: Fine-Tuning Encoders for Reddit Fallacy Detection
SummaryThis paper reports the authors' participation in the Touché 2026 shared task at CLEF, which covers fallacy detection and argument-scheme classification on a Reddit-derived corpus. Their main systems fine-tune RoBERTa-large with task-specific heads and add LLM-generated synthetic contrastive pairs. A broader comparison with other encoders and Qwen LLM variants showed no validated advantage, though the authors judge this comparison underpowered. The paper also reports official test-set rankings and discusses limitations on rare classes.</summary>
In briefNearest-neighbor curve completion for updating renewable generation forecasts within the day
SummaryThis paper proposes a nonparametric method that combines a day-ahead forecast with the morning's realized generation curve to produce probabilistic forecasts for the rest of the day. It selects nearest-neighbor historical curves based on similarity in the observed partial curve, the day-ahead forecast, the date, and location, and derives functional envelopes and depth-based predictive bands. Three case studies using synthetic wind and solar assets and real regional wind data show excellent calibration.
In briefA unified three-branch deep learning framework for detecting and counting wildlife in aerial images
SummaryThis paper presents a unified three-branch deep learning framework covering counting strip delineation (WildSeg), overlap matching between successive images (WildMatch), and animal detection and counting (WildDet). It was evaluated on oblique aerial imagery of four antelope species from the Great Nile Migration Landscape in South Sudan. WildDet improved detection across all four species and reduced detection-based count MAE, RMSE and mean bias by 21.0%, 4.6% and 27.8% relative to YOLO11.
In briefPerturbBridge: Direct Phenotypic Supervision for Generating Perturbed Cell Images
SummaryThis paper presents PerturbBridge, a model that generates images of cells after perturbations by directly supervising perturbation-specific phenotypes during training. Biological response gradients pass through a frozen image encoder. Across three microscopy benchmarks, it reports higher biological fidelity, image quality, and sampling efficiency than CellFlux, with gains in classification and retrieval metrics.
WhyDirectly adds differentiable biological response losses to bridge-model training, a substantive improvement over CellFlux, validated across three microscopy benchmarks with multiple comparisons.
Yucheng Yang, Xiaoxiao Sun, Huangqingbo Sun et al.
In briefTILE: Linking LLMs and Table Foundation Models for Learning from Tables with Text Columns
SummaryReal-world tables often mix numeric columns with free-text string columns, which table foundation models struggle with unless strings are converted to numbers. The paper proposes TILE (Table-Interfaced LLM Embeddings), which uses row serialization, intermediate-layer extraction, columnar pooling, and bypass features to extract better LLM representations for tabular learning. Paired with table foundation models, TILE beats the strongest baseline on 89% of splits across 100 datasets and ranks first on the 108-dataset STRABLE benchmark.
WhyProposes interface design (row serialization, intermediate-layer extraction) letting LLM features scale; strong results across 100 datasets and a 108-dataset benchmark.
F\'elix Lefebvre, David Holzm\"uller, Gioia Blayer et al.
In briefGradient routing to localize and remove reward-hacking behavior in RL
SummaryThe paper addresses reward hacking in reinforcement learning, where a misspecified reward leads models to unintended behavior that an imperfect monitor can only partly detect. It introduces GRAFT (Gradient Routing Adapter Fine Tuning), which confines gradient updates on monitor-flagged samples to a set of sacrificial parameters instead of penalizing or filtering them. Experiments on models from 135M to 8B parameters, plus distillation of frontier LLM trajectories into Qwen3-32B, show that ablating these parameters reduces or stops the unintended behavior while preserving target-task performance.
WhyProposes GRAFT, a novel gradient-routing mechanism, validated across 135M to 8B models and distillation into Qwen3-32B.
In briefCalibration and noise in linear attention for in-context classification
SummaryThis paper studies in-context binary classification in a simple attention model trained with logistic loss. Each prompt has a latent logistic teacher whose direction the learner must infer from Bernoulli-labeled context examples. The authors derive an asymptotic high-dimensional prediction for the test logit and separate the effects of calibration, test-time context noise, and non-isotropic estimation noise. Predictions closely match simulations and fully trained linear-attention models, and standard softmax attention shows qualitatively related behaviour.
In briefPatientWM: A Clinical World Model Built on Unified Patient States
SummaryThis paper proposes PatientWM, a clinical world-model framework that centers on a unified patient state and an ordered latent history. It integrates the observations available at each decision time into the patient state and predicts the next state from the current state, the supplied action, and preceding patient states and clinical actions. Experiments on MIMIC-III, MIMIC-IV, and three external clinical datasets report better performance in clinical prediction, multi-step forecasting, and action planning, and transfer experiments show the patient state generalizes.
In briefVL-DocIR: A Benchmark for Long Visual Document Retrieval
SummaryThis paper introduces VL-DocIR, a page-level benchmark for retrieving evidence from long documents using visual inputs. It contains 29,641 HTML-born documents rendered into 388,548 page images and 271,760 queries spanning single-page, within-document multi-page, and cross-document evidence. The strongest evaluated configuration reaches 69.3 nDCGAll@10 and 80.6 RecallAll@10, and the results show that cross-document retrieval is much harder than single-page retrieval, while document-context scoring helps most for within-document multi-page queries.
In briefMeasuring variable importance with proportional inclusive marginal effects
SummaryShapley-based measures are common for variable importance, but they can assign positive importance to inactive variables correlated with active ones. This paper proposes proportional inclusive marginal effects (PRIME), which overcomes two limits of the earlier proportional marginal effects (PME) measure: it is developed for deterministic models and is computationally costly. PRIME uses a new model-free variable selection technique to identify active variables and includes only them in the importance calculation. Analyses of test functions and real datasets show that PRIME quantifies variable importance accurately and reliably from noisy input-output data.
In briefNeos: Learning to Alternate Search and Visual Reasoning
SummaryThis paper presents Neos, a multimodal deep search framework that trains agents to both gather visual evidence and use it to guide later search steps. It introduces a trainable Multimodal Evidence Interface that turns retrieved content into goal-conditioned observations, and alternates policy reinforcement learning with supervised fine-tuning of this interface. The authors report that Neos-9B and Neos-35B show leading performance among specialized multimodal deep-search agents across ten benchmarks.
In briefAutomatic digital garment assembly from 2D patterns using learning-based sewing and panel arrangement
SummaryThis paper presents two learning-based methods that take a 2D vector pattern as sole input to automatically define sewing operations and arrange pattern panels for digital garment drape simulation. The sewing model uses a Transformer encoder at the segment level, while the panel-level model predicts body-region assignments, layers, and placement coordinates. Evaluated on 4,035 professionally assembled garments, the sewing model exactly recovers all seam and hem topologies for 63.3% of test garments, and the arrangement model meets placement criteria for all panels in 66.2%.
In briefAD4EDU: A Benchmark for University Lecture Audio Description and Its Evaluation
SummaryThis paper introduces AD4Edu, which includes a benchmark (AD4Edu-Bench) and a rule-diagnosable metric (AD4Edu-Eval) for audio description of slide-based university lectures. The authors combined accessibility guidelines such as WCAG 2.2, collected reference descriptions and preference judgments from four blind and low-vision raters, and trained models with direct preference optimization. They find the metric reliably detects standard violations, but descriptions that comply with the standard do not always match blind and low-vision users' subjective preferences, especially on length.
In briefBeyond Model Bias: Rankings of Text Unpredictability Stay Stable Across Small Language Models
SummaryThis paper tests whether rankings of text predictability, based on cross-entropy mean and variance, stay consistent across different language models. Four small transformer models were compared on a corpus of contemporary poetry and a corpus of academic papers, with rank correlations above 0.90 for most measures. Agreement among language models was substantially higher than among parameter-free compression baselines.
Ankit Bhattacharjee, Indrasabarna Hari, Krityapriya Bhaumik et al.
In briefMENTOR: Keeping Tool-Using AI Agents on Company Procedure Manuals
SummaryThis paper addresses tool-using LLM agents that automate workflows governed by natural-language procedural manuals, such as travel booking or hardware procurement rules. It introduces MENTOR, which builds a history-dependent runtime shield from the manual through a multi-step autoformalization pipeline, guiding agents toward compliant progress while remaining context-aware. The authors report that shielded lightweight open-source LLMs achieve higher compliance than unshielded proprietary models. The shield is synthesized once per manual and reused across tasks.
In briefHoneycomb electron-hole crystal in a mass-asymmetric bilayer, studied with neural-network variational Monte Carlo
SummaryThis paper studies a strongly mass-asymmetric electron-hole bilayer, where localized heavy holes form a nonuniform positive background that supports an electron crystal. Using real-space neural-network variational Monte Carlo in the Born-Oppenheimer limit with fixed holes, the authors find that increasing density drives a transition from triangular to honeycomb order, which later melts into itinerant electrons. They also propose a BLG/hBN/WS2-WSe2 heterostructure as a possible platform, at a density of about 10^11 cm^-2.
In briefSPLICE: Discovering and combining physics operators from unlabeled mixed data
SummaryThis paper introduces SPLICE, a framework that discovers neural operators for physical mechanisms from unlabeled trajectories where several mechanisms act at once. It infers each candidate operator's contribution with a convex optimization and uses a calibrated test to assign each operator its mechanism. On one- and two-dimensional systems, the discovered operators recover mechanisms and compose into zero-shot simulators of unseen equations.
In briefTesting quantum kernel SVMs on credit card fraud detection under near-term hardware limits
SummaryThe paper benchmarks three quantum feature encodings (Angle, ZZFeatureMap, IQP) against classical models for credit card fraud detection, adding SMOTE oversampling to the quantum Gram matrix workflow. Classical baselines reached a PR-AUC of 0.983 on the full 30-feature dataset, but dropped to 0.750 when restricted to the quantum models' scale, while the best quantum result was 0.730. Performance fell further to 0.62 when run on physical quantum hardware.
In briefUnderstanding When Personalizing Whisper Speech Recognition Helps
SummaryThis paper studies when lightweight low-rank adaptation (LoRA) improves the Whisper speech recognition model and whether the gains need separate adapters for each group. Comparing zero-shot Whisper, pooled LoRA, and group-specific LoRA on Common Voice, LibriSpeech, and VoxPopuli, it finds gains are concentrated in difficult Common Voice groups and largely vanish on low-error LibriSpeech. Pooled LoRA often matches, and sometimes clearly beats, group-specific LoRA.
In briefLLM Evaluation Is a Forecasting Problem: Time-Series Prediction of Future Model Performance
SummaryThe paper argues that standard random train-test splits assume identical data distributions, yet models are trained on past data and deployed in the future, and web text distributions shift over time. The authors find that model performance across diverse benchmarks also changes continuously over time. They report that future performance is better predicted by recent evaluation data, and best by time series forecasters applied to historic performance, and that selecting models by forecasted accuracy improves deployment outcomes. The authors release their code and data.
In briefProbabilityBench: A Benchmark for Checking Model Uncertainty Against Exact True Probabilities
SummaryReal-world data gives only one observed label per instance, so true class probabilities are unknown and model uncertainty is hard to evaluate. This paper introduces ProbabilityBench, which provides exact instance-level class probabilities across three simulated text domains: peer review, covert information transfer, and credit risk. It measures the quality of reported probabilities using KL divergence and uses the known ground truth to test standard calibration metrics and Bayes-error estimators.
In briefDIME: Scaling Facial Representation Learning via Differential Masked Autoencoding
SummaryThis paper introduces DIME, a differential masked autoencoder for large-scale facial representation learning. It reconstructs identity-matched image pairs and uses an EDDS objective to align the direction of the reconstructed difference with the true difference, capturing subtle facial changes without dense annotations. Pretrained on 20 million face images, DIME's BP4D F1 for AU detection rises by 5.9 points as data grows from 6M to 20M, and it outperforms specialized facial methods on several tasks.
WhyIntroduces identity-paired differential direction supervision for masked autoencoding, validated across many facial tasks and 20M-image pretraining with clear scaling gains.
In briefMAGNET: A manifold-aware diffusion model for generating brain connectomes
SummaryThis paper proposes MAGNET, a diffusion network that generates functional brain connectomes, which are correlation matrices of connections between brain regions. It uses a normalized-Cholesky representation to guarantee valid matrices and compares against prior manifold-aware methods on the ABIDE, ADNI, and OASIS-3 datasets. The authors report 7–21% gains in class-conditional fidelity and over 30% better subclass covariance alignment in a zero-shot setting, and argue that geometric validity alone is not sufficient for clinical utility.
In briefMaskSense: Addressing the Visual Exploration Trap in Masked Image Generation
SummaryThis paper studies how to apply reinforcement learning to masked generative models (MGMs), noting that directly adapting GRPO conflicts with their exploitative decoding. It names this conflict the "visual exploration trap," where confidence-based sampling fixes low-uncertainty background regions early and narrows the generation space. The authors propose MaskSense, which addresses the problem at both sampling and optimization stages, and report improved results over the base model on text-to-image benchmarks, including GenEval accuracy rising from 54% to 86% and HPS from 28.89 to 36.39.
In briefMixture of GFlowNets: a team of independent samplers that finds more diverse high-reward objects
SummaryThe paper proposes Mixture of GFlowNets (MoG) and its exchange variant MoGX, which use a community of K asynchronously trained GFlowNets whose state space is divided by a jointly trained soft classifier. On HyperGrid across six reward landscapes with up to K=64, the methods keep finding new modes as K grows, while data-parallel training saturates. On a genomic ideotype-design task, MoGX covers 39% of 1024 optima at K=64.|
Omar G. Younis, Joseph D. Viviano, Chirayu Anant Haryan et al.
In briefMulti-agent trajectory planning that keeps the true target identified while deferring decisions among moving candidates
SummaryThe paper studies multi-agent interception when one of several moving candidates is the real target and both its identity and terminal locations are uncertain. The authors propose a receding-horizon deferred-decision trajectory optimization controller that combines conformal terminal prediction, sequential candidate rejection, capacity-constrained assignment, and candidate-contingent planning. Under exchangeability assumptions, it establishes fixed-time terminal coverage and anytime-valid target retention. Across 100 paired scenarios, the controller captured 79% of targets under a strict control-effort budget and 85% when the budget was relaxed.
In briefResponsible copyright governance for generative AI training on authors' works
SummaryThe paper examines how copyright rules protect authors whose scholarly works may be used to train large language models, often without their knowledge. It codes 40 legislative, judicial and soft-law instruments from 2019 to 2026 against a framework based on Responsible Research and Innovation theory. It finds that the EU and US frame protection procedurally rather than as an entitlement, and that no instrument creates a general attribution duty for AI training use. The authors derive a six-instrument governance design and a capacity-based allocation of responsibility among stakeholders.
In briefGlasseek: A transparent open framework for training and fairly evaluating deep search agents
SummaryThe paper presents Glasseek, an open deep search framework that releases its data, model weights, training recipes, and code. It introduces a verifiable trajectory curation pipeline, a training method called CM-GSPO that handles context compaction, and an anti-leakage plugin for auditing test-set contamination in open-web evaluations. The authors report that their 9B agent outperforms frontier proprietary agents on almost all deep research benchmarks.
In briefPopularity Trajectories and Mixing Patterns of Urban Venues in Two Cities
SummaryThis study analyses 5.5 years of Foursquare data for about 10,000 venues in London and Barcelona to examine how popularity of restaurants, shops, bars and parks changes over time. It identifies six trajectory shapes recurring in both cities and finds that deprivation and population density do not determine which shape a venue follows. Venues sharing a shape tend to cluster within roughly 200 to 600 metres, and the authors conclude urban vibrancy emerges at block and street level.
Wang Ngai Yeung, Lakmal Meegahapola, Daniele Quercia et al.
In briefPhysics-coupled graph neural operator predicting six-component stress tensor fields in concrete penetration
SummaryThis paper proposes a physics-coupled graph neural operator (PC-GNO) that learns in function space how the six-component stress tensor field on the mid-plane of concrete evolves during penetration. Training uses 359 aggregate-resolved full-scale finite element cases, verified against laboratory measurements. Under the single-step setting the model reaches 80.16% mean accuracy with 20,664 parameters, and it outputs the complete six-component tensor field.
In briefIntrinsic Geometry Matching: learning the true geometry of a process despite how it is observed
SummaryReal data are often indirect observations of an unknown process, and the observation can distort the process's geometry. The paper introduces Intrinsic Geometry Matching (IGM), which learns the carré du champ operator of the underlying process to recover its diffusion geometry, using natural data perturbations to stay equivariant to changes in observation. Because this equivariance links different spaces, the model can be trained with data from other spaces. On molecular dynamics, image, and video data, IGM is reported to outperform existing neural and graph-based geometric methods and to scale better.]
In briefConstant-depth rational-weight GNNs and their link to Schanuel's conjecture
SummaryThe paper asks whether a constant-depth graph neural network using only rational weights can match the expressive power of the 1-dimensional Weisfeiler-Leman (1-WL) test. It proves this is possible assuming Schanuel's conjecture, with a construction uniform across all graph sizes. The authors also give a partial converse linking GNN expressivity to Schanuel's conjecture, and experiments confirm that the constructed GNNs match the 1-WL test.
WhyProves constant-depth rational-weight GNNs match 1-WL under Schanuel's conjecture, with a partial converse and experimental confirmation.
In briefGenerative models as adaptive trust regions for discrete Bayesian optimisation
SummaryThis paper introduces GenTuRBO, which replaces the edit-distance trust region used in TuRBO with an adaptively scaled proposal distribution defined by a pretrained generative model. The generator provides a learned prior over discrete transformations that captures coordinated changes without enumerating edit-distance neighbourhoods. The authors report strong performance across three domains, with the highest mean final scores on Ehrlich sequence optimisation and Sokoban coordination.
Michael Dodds, Erik Bodin, Sebastian W. Ober et al.
In briefTabGift: A Context-Efficient Foundation Model for Tabular Data Generation
SummaryThis paper proposes TabGift, a foundation model for generating tabular data that works from only a few context samples. A pretrained tabular encoder maps heterogeneous tables into a shared latent space, where an in-context diffusion model denoises latent samples conditioned on a small set of clean examples. Experiments on 7 classification and 11 regression datasets show better synthetic data quality than representative baselines, especially with limited context, along with faster adaptation and generation.
In briefSelf-supervised learning of view-consistent visual features using a simple 3D prior
SummaryThe paper notes that features from vision models such as DINOv2 and Stable Diffusion can struggle with semantic correspondence, for example failing to distinguish left from right on symmetric or repeated object parts. It proposes a self-supervised method that uses a simple prior, combining contrastive learning and dataset-level distribution matching to learn a category's coarse shape and appearance. With added geometric losses, the learned features match the global and local statistics of features from real images. The authors present quantitative results on several challenging object categories.
In briefAUGUstus: A Driving VLA Model Unifying Action Generation and Reasoning
SummaryAUGUstus is a vision-language-action model for autonomous driving that combines diffusion-based action generation with autoregressive chain-of-thought understanding in one model, with understanding serving to support generation. Inference needs only the action output, which the authors say enables low latency. The paper reports 94.6 PDMS on NAVSIMv1, 90.7 EPDMS on NAVSIMv2, and average L2 error of 0.27 m with 0.05% collision rate on nuScenes. The authors also report a preliminary real-vehicle deployment running at 12 Hz.
In briefFrom Suboptimal to Optimal: One-Step Latent Flow Policies for Offline RL
SummaryThe paper proposes MeanFlow2Better (MF2B), which reframes offline reinforcement learning as optimizing generative decision chunks in a high-level latent space, moving from current performance toward better performance. An encoder maps trajectory chunks to latent decision representations, and a shared conditional mean velocity field is learned to refine decisions via one-step ODE integration without multi-step sampling. On the D4RL Gym-MuJoCo benchmark, MF2B reports an average score of 90.91, matching or surpassing strong baselines.
In briefPredicting When Reddit Users Seek Mental Health Help Using Language and Behaviour
SummaryThe study frames mental health help-seeking as a time-to-event problem, using only Reddit posts written before any clinical activity. It compares text, MentalBERT embeddings and behavioural features with Cox models, Random Survival Forests and an LSTM. Behavioural features alone predict who will seek help better than text (C-index 0.797 vs. 0.763), and combining both performs best (0.823 cross-validated, 0.856 on the holdout). For predicting when a transition will occur, text is the more stable predictor (C-index 0.698) while behavioural features are unstable.
Nisar Ahmed, Ammara Kazmi, Irtaza Bashir Raja et al.
In briefPRAISE: A reproducible workflow helping non-programmer scientists implement published AI models
SummaryThis paper presents PRAISE, a step-by-step workflow that helps researchers without programming or data science skills implement published AI models. It anticipates common implementation failures and provides a fallback route for each. As a proof of principle, three BSc chemistry students and their tutor, with no prior expertise, reproduced, validated and extended a model predicting enzyme-substrate interactions. The authors argue that this makes implementing literature models a learnable skill.
Max Achterweust, Stefan Groot Nibbelink, Vincent J. de Beer et al.
In briefWeGR: Unified generative retrieval using behavior-aware tokenization and preference alignment for industrial search
SummaryThis paper proposes WeGR, a unified generative retrieval framework for industrial search that generates candidates within an autoregressive model. Its CoCo-VAE tokenizer injects query-to-item and item-to-item collaborative signals before quantization and balances codebook utilization, while DHPO progressively extends reward and policy optimization horizons to alleviate reward sparsity. According to the paper, a six-week online A/B test on a large e-commerce search platform showed relative gains of 2.539% in orders, 1.508% in CTR, and 1.381% in CVR, with CPU and GPU usage reduced by 25.5% and 26.1%.
WhyCombines behavior-aware tokenization with horizon-expanding policy optimization for generative retrieval, a substantive improvement; a six-week large-scale online A/B test gives strong evidence.
In briefThe Value of Being Wrong: Mining a Model's Own Errors as Examples for Visual In-Context Learning (SMILE)
SummaryThis paper proposes SMILE, a method that uses a vision-language model's own prediction mistakes as counterfactual demonstrations for visual in-context learning. Past errors are recorded offline to form a pool of hard negatives, and at test time the model's prediction is used to retrieve the mistakes most likely to recur for the current query. According to the paper, across four benchmarks and four VLMs, SMILE improves average task score by 3.09 points over the prior state-of-the-art baseline at 5x lower latency.
In briefMADDIE: a multi-agent AI engine that coordinates specialized agents to automate the drug discovery pipeline
SummaryMADDIE is a drug discovery system built on large language models, in which a central Orchestrator agent coordinates four specialized worker agents for hit identification, lead optimization, synthesis, and evaluation. Across 50 protein targets, the authors report that it produces molecules meeting key property requirements along with synthesis routes. They state it outperforms existing agentic systems, generative models, and general-purpose LLMs by a wide margin.
In briefEvent Memory and Narrative Dynamics: An LLM Framework for Multidimensional State Spaces
SummaryThe paper presents a large language model framework that builds event representations from autobiographical narratives and models their structure and dynamics, with each event traceable to source excerpts. As a proof of concept on interviews from five participants, it constructed 256 event-states embedded in feature-based spaces. Consecutive event-states were significantly closer than expected under randomized narrative order, and these results were linked to routing profiles and a successor representation model.
In briefUniLAP: A Unified Latent Action Policy for Robot Manipulation Across Different Bodies
SummaryUniLAP is a continuous latent-action policy that models its latent space with flow matching and does not need future frames at inference time. It learns manipulation priors from action-free videos by distilling a latent action model teacher, and it uses a label-free motion target to focus on kinematic cues rather than visual appearance. The authors report that the 1.1B-parameter model outperforms latent-action and Vision-Language-Action baselines on hand pose forecasting, single-arm and humanoid manipulation, and real-world deployment.
In briefProactive physical AI that continuously observes and intervenes at the right moment
SummaryThis paper studies Proactive Physical AI, where a system monitors the physical world without an explicit instruction and decides when to act. The authors present ProAgent, which combines StreamingJev, a timing model that determines whether and when intervention is needed, with a Response Model that generates replies or invokes device actions. An automated pipeline converts videos into training supervision. Experiments report more accurate intervention timing with fewer false triggers than the baseline, and real-robot tests show the system identifies emerging needs and carries out appropriate physical actions.
In briefLearning to Design SR Solvers: Harness Optimization for Symbolic Regression
SummaryThis paper proposes HarnessSR, a framework that jointly optimizes prompts and workflows to build reusable agents for symbolic regression, which finds explicit equations explaining data. It supports either PySR or an LLM-based equation generator during optimization and execution. The method refines role instructions by reflecting on residual patterns and search outcomes. On LSR-Synth in LLM-SRBench, it reports a mean ID Acc0.01 of 61.33% across three backbones, a 45.0% relative improvement over the next-best baseline.
In briefDiTSG: Generating Air Traffic Scenes with Multi-Agent Diffusion Transformers
SummaryThis paper proposes DiTSG, a model that learns from recorded multi-aircraft trajectories to generate air traffic scenes. It uses a diffusion transformer with stacked multi-agent inverted Transformer layers to jointly generate the paths of several aircraft, conditioned on each aircraft's required time of arrival. According to the authors, the generated scenes are diverse, reproduce the statistical properties of recorded traffic, and preserve realistic interactions between aircraft.
In briefSarcasTraps: How Sarcasm Affects Turkish Hate Speech Detection
SummaryThis paper introduces SarcasTraps, a 479-entry external test set for Turkish hate speech detection, annotated for both hate speech and sarcasm. Classifiers built on BERTurk and XLM-RoBERTa were trained on two existing Turkish datasets and evaluated on it, and GPT-4.1 and Llama-3.3-70B were tested zero-shot. Sarcasm substantially affects detection: fine-tuned encoders lose recall on sarcastic hate, while zero-shot LLMs keep higher recall but show large precision drops on sarcastic examples.
In briefDDVR-Net: Restoring Damaged Characters in Historical Document Recognition
SummaryThis paper presents DDVR-Net, a framework for recognizing text in historical documents damaged by ink erosion and structural wear. It separates glyph structure from physical corruption at the feature level and predicts how visible each character is, activating a contextual decoder only for characters that are hard to read. The authors also build a low-resource benchmark of 187,000 document pairs covering Tibetan, Mongolian, and Uyghur scripts. They report that the method outperforms existing approaches across the three languages and different reading orders.
In briefWord-meaning redefinition shows language models follow instructions in a context-sensitive way
SummaryThis paper examines how language models handle conflicting context within a prompt. Using 870 word pairs, it gives an explicit instruction redefining a word, then a later sentence showing the word's usage. The redefinitions were not applied as fixed rules; the instruction could be preserved, weakened, or completely reversed depending on its relationship to the following sentence.
In briefLearning to Design Convex Relaxations for Nonconvex Optimization
SummaryThis paper proposes AutoConvexRelax, which represents a nonconvex problem's symbolic formulation as a tripartite graph and learns a policy that composes symbolic rewrites into problem-specific convex relaxations, replacing manual design. On 241 benchmark quadratic instances, the learned policy strengthens McCormick lower bounds by 13.52%, with a mean lower bound 3.70% below strengthened semidefinite relaxations on the normalized metric while using much smaller formulations.
WhyFrames convex relaxation design as a learned symbolic rewrite policy with validity proofs and error bounds, validated on 241 quadratic instances; impact is confined to a subfield.
In briefSocial Priors: How Descriptions of AI Users Shape Model Trust and Cooperation
SummaryThe study finetunes Qwen3-32B on four corpora built from the same 1,625 usage scenarios, differing only in the motive attributed to users, such as helping others or exploiting them. Finetuning changes the models' beliefs about users and their judgments of world safety, and these shifts appear in behavior. Antisocial models expect less reciprocity and send less money in trust games, and antisocial training also raises dark-triad scores and lowers alignment ratings.
WhyContrastive synthetic finetuning shows user-motive descriptions shift trust and cooperation, supported by trust game and resource experiments with controls.
In briefUniMetric: Recovering Real-World Scale for Visual Geometry Models
SummaryThis paper presents UniMetric, a training-free method that gives real physical scale to reconstructions from visual geometry foundation models, without extra sensors. It uses height ranges of familiar objects estimated by a vision-language model, compares them with reconstructed object heights, and solves a convex problem to find one scale for the whole sequence. The authors also introduce MetricBench, a benchmark with sensor-based ground truth, where the method improves depth, reconstruction, and trajectory errors across nine models.
In briefUA-Bench: A Benchmark for Large Audio-Language Models on Underwater Acoustic Recognition
SummaryThis paper introduces UA-Bench, a large-scale benchmark of 375,051 synthetic underwater audio records for testing how well large audio-language models recognize underwater acoustic signals. The records are generated from parameterized sources and modeled propagation, each with a three-level label hierarchy, dialogue, and metadata. Nine open models without task-specific adaptation reach only 2.0%–13.1% leaf accuracy, below the 7.7% random reference. The audio waveform helps source routing much more than fine-grained leaf recognition.
In briefImplementation details can dramatically change KNN performance: a Cover Tree case study
SummaryThis paper argues that implementation details of nearest neighbor search algorithms are rarely reviewed, yet they can materially change performance and conclusions about which algorithm is superior. Using the Cover Tree as an example, the authors built RustKNN, a unified Rust implementation of three construction variants and three query modes. They find the best query mode depends on k and the dataset, and choosing the wrong one can cost up to a 7.5× runtime factor.
In briefMaking tensor decoupling more robust and less parametric
SummaryTensor decoupling methods represent multivariate functions compactly by combining linear factor matrices with univariate nonlinearities. The authors note that current methods are sensitive to mismatched output or gradient magnitudes and require manual tuning of spline degrees of freedom. They propose rescaling the input Jacobian tensor as preprocessing and replacing B-splines with penalized splines tuned by generalized cross-validation, and release an open-source JAX implementation.
Maxime Rochkoulets, Joppe De Jonghe, Mariya Ishteva
In briefEstimating Uncertainty in Omnimodal LLMs Using External Embeddings
SummaryThis paper proposes a model-agnostic framework for quantifying uncertainty in omni-modal large language models, which accept multiple input modalities. It scores chain-of-thought responses using only external omni-modal embeddings, measuring how well the reasoning is grounded in the multimodal input and its internal coherence. On UNO-Bench, the method yields small gains over self-consistency and is competitive with other baselines, especially on open-ended omni-modal questions.
