In briefScalable random features for hyperbolic latent variable models
SummaryThe paper introduces Random Helgason-Fourier Features (RHFF), an explicit finite-dimensional feature map for kernels on hyperbolic space. It characterizes the spectral density used to draw RHFF frequencies, so kernels such as the Matérn family can be specified in the frequency domain, and proves a uniform approximation error bound. Building on this, the authors develop RHFF-LVM, a hyperbolic latent variable model that needs no quadrature, variational bounds or auxiliary variables, and report that it recovers a synthetic tree's metric more faithfully than all baselines and fits 5×10^5 points in 505 seconds.
Koshi Watanabe, Daichi Mochihashi, Hideitsu Hino et al.
In briefFine-tuned small language model that generates synthetic clinical data to improve vital-sign prediction
SummaryThis paper proposes a clinical data augmentation method that uses a fine-tuned Microsoft Phi-2 small language model to generate synthetic samples. The model was adapted to the medical domain using QLoRA-based supervised fine-tuning on the MedQuad dataset and few-shot prompting. When the generated data augmented a small vital-sign dataset, an LSTM predictive model's classification accuracy rose from 89.0% to 93.7%, with consistent gains in precision, recall, and F1-score.
Mihnea Sebastian Burlacu, Ionut Anghel, Tudor Cioara
In briefRational alignment of superintelligences through debate
SummaryThe paper studies a setup where superintelligent AI agents assist humans in question answering, knowing the correct answer while the human does not. It gives example mechanisms that control how the agent and human principal interact, without changing the agent's utility. It shows that a debate framework in which agents reason and act in a subgame perfect equilibrium yields more predictable outcomes and is sometimes favorable.
In briefCan vision-language models spot video event boundaries the way humans do?
SummaryThis paper examines whether Large Vision-Language Models (LVLMs) perceive event boundaries in videos consistently with humans. On the Kinetics-GEBD dataset, LVLMs refine candidate boundaries proposed by an unsupervised discovery method, achieving state-of-the-art unsupervised performance. The approach is extended to hour-long egocentric video using coarse-to-fine discovery and localized LVLM judgments, and the authors build the EgoBoundary benchmark, on which representative existing methods degrade substantially.
In briefHIFC-IQA: Train-free cross-domain image quality assessment using dual-process cognition
SummaryThis paper addresses the domain shift between synthetic distortions and real-world image degradations in no-reference image quality assessment. It proposes HIFC, a completely train-free framework inspired by human dual-process cognition, which uses asymmetric perceptual decomposition to separate semantic content from distortion-sensitive features. The final quality score combines a memory consensus branch that retrieves a stable Fuzzy Quality Centroid from a synthetic reference gallery with an intuitive branch that captures aesthetic polarization via zero-shot vision-language alignment. The authors report state-of-the-art results across multiple benchmarks, with a notable margin in the synthetic-to-authentic setting.
In briefA unified framework for training physical action models
SummaryThe paper studies how choices about robot model capacity, demonstration budgets, data composition, representation, and optimization are connected through a common conditional-distribution target. It uses nested learning oracles for exact risk accounting, together with local information geometry, conditional mutual information, and empirical convergence, to derive training prescriptions. Numerical checks with known targets reveal capacity-data coupling and representation-induced mixture mismatch. An analysis of 206 public PushT trajectories evaluates fixed-feature information-design criteria without fitting a policy.
In briefFPSRG: First-person post-training and response grids for simulating diverse users
SummaryThis paper studies how to build language-model configurations that simulate individuals with different characteristics, retrievable by multidimensional target scores. The method, First-Person Semantic Response Grids (FPSRG), first post-trains the model to answer as the participant, then refines dimension-specific control vectors using self-report and behavioral targets. Joint scores are measured with the model frozen and stored in a searchable catalog. Reported results include mean five-axis control span rising from 30.75 to 40.45 after refinement, and a five-level vector catalog covering 92.2% of fixed targets in the five-dimensional region.
In briefDesignTaste: a benchmark comparing AI-generated visual designs with award-winning works
SummaryThis paper introduces DesignTaste, a benchmark that compares images generated by text-to-image AI models with award-winning design works. It builds 403 design briefs from Beste Plakate 2025 poster works and ADC 2026 works across 20 categories, and a single model judge scores each output and award-winning work on five dimensions. Only ArtArk achieves a higher mean overall score than the award-winning works on each dataset, and all six models score lower in Creative Originality. The authors conclude that producing finished, visually appealing images does not mean a model can develop good design solutions.
In briefdaVinci-Prover: Agent-Native Mid-Training for Formal Theorem Proving
SummaryThis paper presents a data and training recipe for improving formal theorem-proving models. The authors synthesize a 33.7B-token mid-training corpus covering decomposition, search, self-verification, and revision, and introduce the daVinci-Prover harness, which lets models interact with a Lean verifier in an isolated workspace. Using it, they collect about 36K agent trajectories for supervised fine-tuning. The method improves on a matched Qwen3-8B SFT-only baseline by an average of 13.7 points across MiniF2F, FATE-M, and ProofNet, and performs strongly on the Qwen3.5 series.
In briefSoScale: Second-order scale search for 4-bit floating-point model compression
SummaryThe paper proposes SoScale, a method for choosing hardware-valid scale codes in hierarchical FP4 quantization of large language model weights. It searches scales using a Hessian-aware layer reconstruction objective and combines this with GPTQ-style error recovery, so scales are selected by their effect on layer outputs rather than element-wise error. The authors report that across Qwen3 models, SoScale consistently reduces downstream quantization degradation under MXFP4, NVFP4, and HiF4.
In briefRMIT-ADM+S's TREC 2026 RAG entry: research agents, retriever ablations, and open-weight replication
SummaryThis paper reports the RMIT-ADM+S team's participation in the TREC 2026 Retrieval-Augmented Generation track, using the ClimbMix corpus. The team submitted seven research-agent runs built on proprietary OpenAI and open-weight models. Pre-submission ablations found that writer selection and adjacent-chunk expansion were the main factors affecting answer quality. In the official evaluation, six of seven runs exceeded the track median, and the best run, brief-hybrid, tied the track maximum for Batch-1 vital coverage. Human preferences differed from automated coverage measures across batches.
In briefPhysCapBench: A cross-domain benchmark for video captions that describe physical processes
SummaryThis paper introduces PhysCapBench, a benchmark of 1,500 human-verified video sequences spanning third-person scenes, egocentric manipulation, and handheld-gripper demonstrations. It scores captions for accuracy and coverage across seven physical dimensions. Compared with metadata captions, physical captions reduce video reconstruction LPIPS by 8.6-9.5% on third-person videos and 26.9-30.9% on egocentric videos. Enriched training instructions also improve benchmark scores and manipulation-policy success.
In briefBlinkMimic: Training a Humanoid Motion Tracker in Minutes on One GPU
SummaryThis paper presents BlinkMimic, a method that speeds up training of whole-body motion tracking policies for humanoid robots. On a simulated Unitree G1, it trains one policy for all 40 LAFAN1 motion clips in 3.4 minutes on a single RTX 4090, measured until three consecutive checkpoints reach a normalized return of at least 0.82 with no failed clip (mean of three seeds). A leave-one-out ablation with three seeds per component shows which part of the training time each component affects.
In briefPredicting a Robot Operator's Layered Intent from Video of Teleoperation
SummaryThis paper examines whether vision-language models can infer, from a short visual history of robot teleoperation, both the atomic action a human is performing and the broader mission it serves. The authors introduce VIAMO, a dataset of 6,600 episodes, and fine-tune a model called VIM on it. VIM outperforms GPT-4o and GPT-5.6 Luna on both action and mission prediction, and its predicted mission lets a vision-language-action policy take over from the operator in a real-world setting.
In briefDistilling image-text datasets so they generalize across model training stages
SummaryThis paper studies multimodal dataset distillation, which compresses image-text data into a small synthetic set to cut the cost of vision-language pre-training. It proposes MDA-DD, which uses supervision from a dynamic staggered queue of models at different training stages, along with asymmetric dual projection and an uncentered second-order moment objective. Experiments on Flickr30K and MS-COCO report consistent retrieval improvements over existing methods across synthetic-data budgets.
In briefVBot Exo2Ego Policy: Lightweight Data Collection and Third-to-First-Person Video Generation
SummaryThis paper presents VBot Exo2Ego Policy, a video generation model that converts third-person videos into first-person views of a selected actor, aiming to expand first-person operation data from real-world footage. It also introduces a lightweight mobile app and data collection pipeline that record first- and third-person views synchronously along with camera geometry. The model builds on a pretrained video diffusion transformer, using geometric guidance and auxiliary supervision from semantic masks and hand keypoints, and it runs inference without target-camera calibration.
In briefSmall Language Models for Routing LLM Capabilities: How Model Size and Prompt Tuning Affect Quality and Efficiency
SummaryThis paper studies using small language models to decide which optional capabilities, such as reasoning, retrieval, web search, and tool use, a large language model should activate for each query. The authors compare Qwen2.5 routers from 1.5B to 14B parameters under five prompting strategies, evaluating downstream answers on 127 questions with a fixed GPT-OSS-120B model. They find that higher router capacity does not consistently improve classification, and similar F1 scores can lead to different downstream results. The 14B router keeps answer quality close to enabling all capabilities while substantially reducing downstream token use and median latency, excluding router overhead.
In briefRICapBench: A Benchmark for Fine-Grained Vision-Language Understanding of Remote Sensing Images
SummaryThis paper introduces RICapBench, a dataset and benchmark that combines instance localization with fine-grained captioning for remote sensing imagery. It contains 245,466 instance-level annotations and 498,005 spatial relation tuples across general, UAV, and dense small-object scenarios. The authors also propose the RelCap-F1 metric and a composite RICScore for evaluation. Testing 15 vision-language models, they find instance localization to be a major bottleneck, with LocateAnything achieving the highest RICScore of 31.68.
In briefReinforcement learning that passes experience to future agent runs
SummaryThe paper notes that training LLM agents with scalar rewards discards fine-grained environment feedback such as error logs. The authors fix the Agentic Context Engineering harness and train one model to act, reflect, and curate task-specific playbooks. Each experience-handoff artifact is scored by its effect on future actor rollouts, and the role behaviors are optimized together. The method reports higher held-out performance than baselines across agentic benchmarks.
In briefGravity and Flight: a method teaching AI models to reason about fine details in satellite imagery
SummaryThe paper proposes Gravity and Flight (G&F), a framework that adapts vision-language models to remote sensing scenes using limited supervision. It combines Evidence-of-Thought supervised fine-tuning (EoT-SFT) with Relational Transfer Policy Optimization (RTPO), which verifies visual fidelity and evidence use across paired images. Using 1,008 demonstrations and 500 training pairs for harbor specialization, G&F achieves a mean score of 69.12 across six evaluation sets, outperforming the backbone and supervised initialization. The authors also report gains when applying G&F to urban scenes.
In briefTeaching with Blinks: Fixing Student Error Copying in Fine-Grained Visual Distillation
SummaryThe paper proposes BlinkOPD, an on-policy distillation method for improving fine-grained visual perception in multimodal large language models. Each token is processed by the teacher twice, once with eyes open and once with eyes closed, to counter the teacher following the student's erroneous prefixes. A first-glance gain further strengthens correction on rollouts that conflict with the teacher. The authors report state-of-the-art average results on six fine-grained benchmarks using only 500 training images, and zero-shot gains on ten held-out benchmarks.
In briefFuGIBench: Testing When Streaming Video Models Should Wait, Prepare, or Respond
SummaryThis paper introduces FuGIBench, a benchmark for streaming video in which a model decides whether to wait, make reversible preparations, or respond. It contains 19,893 synthetic tool-assistance videos across ten tool tasks and three tracks: handover, ordinary activity, and cancelled need. A single primary score rewards sustained preparation and alignment with the response interval, and penalizes unnecessary activity when waiting is prescribed.
In briefSaidzoda-Kazakh-EmbeddingGemma-300M: A small Kazakh text-embedding model trained on one epoch of generated questions
SummaryThe paper fine-tunes the small open embedder EmbeddingGemma-300M on Kazakh for one epoch, using 208,511 training pairs, mostly Kazakh questions generated by the authors' language model from encyclopedia, Wikipedia, news and book passages and checked by a second model. On the Kazakh Belebele retrieval test, recall@1 rises from 60.8% to 78.0% and recall@10 from 76.3% to 91.4%. The model is compared with other models that have published MTEB scores and with several Kazakh-specific embedders on the same test.
Tohir Saidzoda, Madina Abdusalomzoda, Abubakr Abdullozoda et al.
In briefTransXplainRRG: A Radiologist-Evaluated Vision-Language Model for Chest X-ray Report Writing
SummaryThis paper presents TransXplainRRG, a large vision-language model that automatically generates radiology reports from chest X-rays. It combines a Swin Transformer image encoder with a text encoder that incorporates patient medical history, and extracts only abnormal findings to refine reports. The model was trained on MIMIC-CXR and tested on the unseen IU X-Ray dataset, scoring 0.327 and 0.605 on the GREEN metric, both above MAIRA-2 under the same setting. A preliminary evaluation by radiologists on 100 generated reports was also conducted.
In briefBeyond Compressing Everything: Contrastive Memory Learning for Selective Context Compression
SummaryThe paper proposes CoMem, a query-guided framework that compresses long inputs selectively by learning memory representations conditioned on both the document and the query. It combines contrastive compressor learning, same-context hard negative mining, and bidirectional attention among memory slots. Experiments on question answering and RULER-NIAH benchmarks report that CoMem outperforms existing context compression methods across compression ratios.
In briefRADIANT: A Redundancy-Aware Method for Universal Multi-Modal Fusion
SummaryThis paper introduces RADIANT (Redundancy-Aware Dynamic Information Allocation Network), a task-general framework for fusing multiple data modalities. It organizes fusion around suppressing redundant cross-modal overlap, preserving modality-unique information, and composing task-conditioned synergy through cross-modal interaction. The method uses stable matrix-exponential projections to reduce interference between shared and unique streams, and a task memory with hard positive updates to choose which synergistic interactions to activate. The authors report consistent improvements across six heterogeneous fusion tasks, including raising downstream detection mAP@0.5 on FLIR from 91.2% to 96.8%.
In briefGoverning Adoption of Self-Generated Updates in Autonomous AI Agents
SummaryThe paper introduces Evidence-Governed Evolution (EGE), a framework that controls when autonomous agents adopt changes they generate to their own prompts, memories, tools, and policies. It separates candidate generation from confirmatory adoption, tests frozen candidates on fresh paired audits, and uses non-inferiority tests to protect declared capabilities. The authors give statistical bounds for adaptively generated candidates, cumulative degradation along reference lineages, and a Monte Carlo stress test of evaluator optimism. The guarantees depend on bounded scores, fresh sampling and declared distributions, and do not establish general agent safety.
In briefRecursive State-Augmented Language Models: a persistent latent state that adapts pretrained language models over time
SummaryThis paper proposes Recursive SALM, a framework in which a persistent latent state updated with each observation generates adaptations of a pretrained language model. Experiments with a frozen GPT-2 small show that on a four-turn hypothesis-inference task, a latent-gated LoRA using only a 64-dimensional state reaches 89.1% accuracy, close to the 88.0% of a full-context LoRA baseline. The authors conclude that persistent latent-conditioned adaptation can support sequential inference when earlier evidence is absent from the decoder context.
In briefCLIP adaptation for few-shot plant disease recognition using descriptions and local evidence
SummaryThis paper proposes a two-stage method that adapts the CLIP vision-language model for recognizing plant diseases from few labeled examples. It enriches class names with symptom and comparative descriptions, then uses these to calibrate global image-text matching with class-relevant image patch evidence. The encoders are first updated with lightweight parameters, then frozen while a linear classifier is trained, with text prototypes used for cross-domain prediction. Experiments on PlantVillage, Cassava, Pear, and PlantDoc report higher accuracy than compared methods in most settings.
In briefMutAtlas: a PDB-wide, energy-guided map of how protein mutations affect function
SummaryThe paper addresses the scarcity of experimentally measured protein mutation data by exhaustively enumerating single-site substitutions on experimentally resolved PDB structures. It aligns mutation signals from physics-based energy models, protein language models, and inverse folding models into a large augmentation dataset, showing that these models often disagree in their mutation distributions. The authors propose an unsupervised multi-source mutation preference distillation framework that models this disagreement without using experimental labels, and report the best overall performance among evaluated zero-shot baselines and naive fusion methods on ProteinGym. The dataset and evaluation pipeline are released.
In briefIsharaBART: An Efficient BART-Based Framework for Low-Resource Bangla Sign Language Translation
SummaryThis paper introduces IsharaBART, a bidirectional generative model for translating between gloss notation and sentences in Bangla Sign Language (BdSL). It is built on BART with denoising training objectives and was trained on a dataset of 1,000 annotated gloss-sentence pairs, expanded to 4,831 pairs through augmentation. Compared with four larger baseline models under identical conditions, it achieves comparable or better bidirectional performance while using fewer than 10% of their parameters.
In briefDIAGNO: A Spherical Neural Operator for Forecasting Uneven Earth Dynamics
SummaryThis paper introduces DIAGNO, a spherical neural operator for forecasting Earth system dynamics. Existing spherical neural operators rely on a rotation equivariance assumption that treats the Earth as a homogeneous system. DIAGNO drops this assumption, explicitly modeling interactions between spectral modes and using the diagonal structure of the spherical spectrum to capture zonal and meridional dynamics. Experiments on simulated fluid dynamics and atmosphere and ocean reanalysis data report state-of-the-art multi-step forecasting performance.
In briefHyBridge: Bridge-Guided Self-Supervised Hyperbolic Hypergraph Representation Learning
SummaryThis paper studies how a node's structural role relates to its radius in hyperbolic embeddings, finding that bridge nodes sit closer to the center in ordinary graphs but not consistently in hypergraphs. It proposes HyBridge, a self-supervised framework that estimates bridge scores using the external connectivity of hyperedges, positions nodes radially according to these scores, and applies contrastive learning to keep nodes distinguishable. Experiments on 10 real-world hypergraphs show it outperforms 22 competing methods, with node classification gains of up to 9.82%.
In briefDRAMA: Speeding Up Multimodal Diffusion Language Models by Analyzing Attention Redundancy
SummaryThis paper addresses the slow inference of multimodal diffusion large language models, which process long sequences dominated by visual tokens over many denoising steps. By measuring attention tensors across spatial and temporal axes, the authors find that attention concentrates on a sparse set of visual tokens and stabilizes quickly across denoising steps. They propose DRAMA, a training-free framework with spatial-redundancy-aware visual token pruning and sufficiency-guided adaptive stopping. The authors report an average 18.67 times inference speedup across six vision-language benchmarks while maintaining competitive accuracy.
In briefSUMNO: A Spatially Unrolled Multigrid Neural Operator for Fluid Flow Surrogates
SummaryThis paper introduces SUMNO, a neural operator for predicting fluid flow fields on regular grids, built by unrolling a multigrid V-cycle into a feed-forward U-shaped network. It uses parameter-free up- and downsampling with additive skip connections, convolutions at finer resolutions, and shifted-window attention at the coarsest level. On four unsteady and steady flow benchmarks against five baselines, SUMNO is reported to be best in 5 of 7 settings and second in the other two. The authors state it cuts 10-step rollout MSE by about 30% versus the strongest baseline while using 9.4 million parameters, fewer than three baselines exceeding 130 million.
In briefReconstructing Hands from Silhouettes: Multi-View Hand Modeling Trained on Synthetic Data
SummaryThis paper presents SilHand, a calibrated multi-view hand reconstruction method that uses only binary hand masks, which can be obtained cheaply, to narrow the gap between synthetic and real images. It is trained on a large synthetic dataset of 46.8k grasping poses and 60k free-hand samples and tested on five real-world benchmarks. Using masks rendered from reference hands, relative reconstruction accuracy rises from 49.2% to 63.4%, and accuracy improves as the synthetic training set grows. Under mask corruption, reconstruction error increases by at most 1.63 mm.
In briefFrom More to Better: Causal Inference and Digital Experimentation in the AI Era
SummaryThis paper argues that AI is rapidly expanding organizations' capacity to produce and analyze data, shifting data scientists' focus toward assessing findings' credibility and importance and judging when evidence generalizes across contexts. Based on a roundtable of academic researchers and industry practitioners, it examines three ways causal inference and experimentation practitioners can turn analytical capacity into better decisions. These are preserving the learning value of experiments beyond individual tests, combining experimental, observational, and synthetic evidence, and keeping evaluation in step with AI's evolving capabilities and uses.
Winston Chou, M{\aa}rten Schultzberg, Lo-Hua Yuan et al.
In briefDoes deep knowledge tracing learn meaning-based links between knowledge components?
SummaryThis paper examines whether relations between knowledge components (KCs) learned by Deep Knowledge Tracing (DKT) models are semantic. The authors build text-derived KC representations with nine pipelines and compare DKT-derived relation graphs with semantic similarity structures, finding only weak and inconsistent alignment. Adding semantic information to DKT did not consistently outperform the standard baseline. The authors conclude that DKT-derived relations are not primarily semantic, and that behavioral, temporal, structural, or psychometric signals may also shape them.
In briefAdaptive Mamba Network for Long-Term Climate Temperature Forecasting
SummaryThis paper presents Deep DualMamba, a deep learning framework for forecasting temperature over time, built on an enhanced version of the Mamba network with selective structured state space models. It uses an adaptive bidirectional Mamba encoder and a hardware-aware parallel recurrent algorithm, and it avoids attention and MLP layers, giving linear complexity with sequence length. The authors also provide a theoretical analysis of the model's approximation ability and convergence. The framework is applied to historical meteorological data from UK monitoring stations.
In briefTurning Large Language Models into Efficient Fast-Decision Models via Adaptive Computation
SummaryThe paper studies how to convert large language models into fast decision-making models, which usually keeps the full backbone and swaps in a decision head, so every decision runs through all layers. The authors find decision information concentrates in middle layers and the final six layers can be removed without loss. They propose training a full model jointly with prefix sub-models that share its final layers and decision head, using a lightweight gate to decide whether extra computation is needed. Across seven decision benchmarks, the approach matches or outperforms independently trained models while offering better accuracy-efficiency trade-offs.
In briefBenchmarking AI risk behaviors: how language models weigh risk
SummaryThe paper evaluates how risk-averse or risk-tolerant large language models are across various contexts and task demands, and how stable and consistent their behavior is. It presents a toolkit for building model risk profiles that can predict behavior and be extended with new risk evaluations. The authors report that 2026 state-of-the-art models are more stable, consistent, and risk neutral than older models. They suggest these trends are related, since newer models more often use expected value maximization as a decision procedure.
Luhan Mikaelson, Robert Adragna, Emmett Brosowsky et al.
In briefWanda-guided sparse perturbations for faster forward-only fine-tuning of large language models
SummaryThis paper proposes WIS-MeZO, which uses Wanda weight-activation saliency to build task-specific sparse perturbation masks while keeping the pretrained model dense. The authors analyze mask quality through gradient energy capture and relate it to gradient estimation and convergence. Across 16 benchmarks with LLaMA and Qwen models from 3B to 32B parameters, WIS-MeZO reports higher aggregate scores than MeZO and reaches comparable validation losses in fewer steps. Comparisons with random and magnitude-based sparse perturbations emphasize the value of selecting informative perturbation coordinates.
In briefShared Item-Level Error Limits Small Language Model Aggregation in Fermi Estimation
SummaryThis paper tests whether teams of small language models can reduce error on Fermi estimation questions by averaging their answers. Using 2.51 million answers from 15 open-weight models given before any communication, the authors split each answer's log10 error into a per-item component shared across a model's draws and a residual. They find the item-level component accounts for a median 0.906 of error variance, so teams of thirty reduce a single agent's error by a median of only 1.7%.
In briefUncertainty-aware diffusion for selective generative video deraining
SummaryThis paper addresses video deraining, where rain streaks and occlusions degrade frames unevenly across regions. The authors propose Selective Generative Restoration, which activates generative priors only where restoration uncertainty is high, and implement it as SGVD, a one-step diffusion framework. Region-level uncertainty is estimated from stochastic differences between two one-step diffusion predictions, and an anchor-guided trajectory consistency mechanism reduces temporal drift across frames. The authors report superior restoration quality, perceptual quality, and temporal consistency compared with state-of-the-art video deraining methods.
In briefORCHID-SAGE: Fully reproducible CALPHAD assessment from evidence to phase diagrams
SummaryThis paper introduces ORCHID-SAGE, a standard that encodes the expert judgments behind thermodynamic database assessments (CALPHAD) into formal, reproducible and automatable protocols. In a first application to Cr-Ni, the database was regenerated from its record by a script, by an AI agent using the Model Context Protocol, and by a refit using only public sources, yielding the same predictions. The authors argue that databases recorded this way can be updated by AI agents as new evidence appears, while human experts retain the judgment and it remains documented.
In briefDEDCA: Test-Time Adaptation for Detecting AI-Generated Images from Unseen Generators
SummaryThis paper addresses the problem that AI-generated image detectors perform poorly on images from unseen generators. It proposes DEDCA, a framework combining CLIP-based semantic evidence with trace evidence from shuffled entropy and residual statistics, fused per sample by a reliability-aware gate. The disagreement between the two evidence streams serves as an unlabeled signal for adaptation to the target domain. The authors report 92.0% average accuracy on the GenImage benchmark.
In briefFixing spurious correlations in conditional diffusion models by doubly robust target correction
SummaryConditional diffusion models can memorize spurious correlations in their training data, causing outputs to drift from the intended target distribution. This paper proposes DR-Diff, a framework that efficiently fine-tunes pre-trained conditional diffusion models to adapt to the target distribution without retraining from scratch. The authors prove that DR-Diff reduces target risk at the population level and has a pseudo-doubly robust property. Experiments on synthetic datasets show noise prediction accuracy matching models trained from scratch, and reduced inter-group disparities in downstream classification.
In briefStateAlign: Reinforcement learning that teaches tool-using AI agents to decide based on task state
SummaryThe paper introduces StateAlign, which combines task-state construction with state-contrastive reinforcement learning to train language model agents for multi-turn tool use. The method pairs states with different action requirements and uses token-level contrasts to reward outputs that fit the current state. Across five base models, it improves results on BFCL v4 multi-turn, ToolSandbox, and When2Call. On Qwen3-8B, BFCL v4 multi-turn accuracy rises from 45.5% to 55.2%, and no-tool hallucination on SimpleToolHalluBench falls from 62.2% to 56.1% compared with outcome-only training.
Zhexi Lu, Subhajit Chaudhury, Tejaswini Pedapati et al.
In briefLoCoJAX: Host-local coordination for training small AI models with reinforcement learning on Google TPUs
SummaryThe paper studies how coordination overhead in reinforcement learning on TPUs changes with model size, finding host orchestration takes 36% of step time for a 0.6B model while network communication dominates for large models. It proposes LoCoJAX, which runs rollouts and reward processing locally on each TPU host while keeping global training synchronous. Compared with a tuned Pathways baseline on Qwen3-0.6B, its colocated mode is 1.3 to 3.6 times faster per step. A Ray-managed disaggregated mode helps only when rollout dominates the step and receives most of the hosts.
In briefSAGE-QA: A Semantic-Aware, Grounded Way to Score Question Answering on SQuAD 2.0
SummaryThe paper introduces SAGE-QA, an evaluation framework for closed-context question answering that combines reference-based Semantic Equivalence Scoring with passage-based Re-Validation Scoring. Tested on 10,000 SQuAD 2.0 validation examples, it compares zero-shot frontier models with open-weight models, untuned and fine-tuned. Fine-tuned open-weight models perform best, with a fine-tuned Llama 3.3 70B surpassing the previous best known Exact Match score.
Chandra N Sekharan, Vishakan Umapathy, and Abhishek Phadke
In briefFinding and reducing coding agents' habit of accepting choices that skip some requirements
SummaryThis paper studies specification satisficing, where coding agents pick an implementation that fails a task requirement for reasons they find acceptable. The authors propose SatisFix, which makes the agent report its choices and reasons and then checks each choice against the task description. Tested on DeepSWE and SWE-bench Pro with five LLMs, it raises pass@1 over the base agent and outperforms three self-correction baselines.
In briefA framework for learning risk-sensitive individualized treatment rules
SummaryThis paper proposes a distribution-first framework for learning individualized treatment rules that account for risk. It first estimates treatment-specific conditional outcome distributions, then builds the treatment rule by plug-in evaluation of the chosen criterion, allowing quantiles, CVaR, and cumulative prospect theory to be compared in one pipeline. The authors give theoretical results on error propagation and regret bounds, and report that the framework outperforms criterion-specific methods in several settings. It is also evaluated on the ACTG175 HIV clinical trial.
In briefSMIC: Shared multiscale input-reference conditioning across Transformer layers for language inference
SummaryThis paper proposes SMIC, a method that builds one input-dependent multiscale reference and reuses it across all Transformer layers, with each layer injecting its own transformed conditioning feature before attention. Trained from scratch, the TCT-SMIC variant achieved the highest ten-seed mean accuracy on all three ANLI rounds, and ablations support the method's contribution. The authors also report lower attention redistribution and smaller accuracy changes under perturbation tests, and derive a finite-sample bound on subset-mass redistribution.
In briefFlowIR: Depth-Driven Rectified Flow for Infrared Super-Resolution
SummaryThis paper addresses the problem that multiple heat sources in low-resolution infrared images often blur together into overlapping responses. It proposes FlowIR, which injects pseudo-3D depth cues into the generation process using depth-guided 3D convolutions and cluster-wise inter-attention to separate objects and reduce cross-source feature mixing. A Thermal-Geometry Consistency Loss enforces geometric alignment during refinement. Built on a lightweight rectified flow backbone, FlowIR restores infrared images in only a few steps, producing sharply separated and geometrically faithful high-resolution results.
In briefCausals: An Open Knowledge Graph for Causal Reasoning Research
SummaryThe paper presents Causals, an open knowledge graph designed to support research on causal reasoning. Its initial release, built from news reports, contains 369,184 nodes and 1,272,418 edges linking events, organizations, commodities, and financial instruments. Structural relationships, observations, and attributed causal claims are kept distinct and linked to their source records. The resource is intended for studying how language models retrieve relevant context and apply mechanisms to specific events, and it is available under CC BY 4.0.
Miaozhe Lin, Qiongzheng Lin, Zhongwei Liang et al.
In briefRectified Looped Transformer: a feedback method that stabilizes repeated-block language models
SummaryThis paper proposes the Rectified Looped Transformer (RLT), which repeatedly applies a shared Transformer block and forms each loop's next input by linearly interpolating between the current state and its prediction. The authors argue that this interpolation keeps hidden-state magnitudes bounded across loops and reduces worst-case perturbation amplification. In language modeling experiments at matched training and inference depths, RLT achieves lower perplexity than a standard looped Transformer and is more robust to input perturbations and inference-depth variance. It does so without extra trainable parameters or additional recurrent block evaluations.
In briefLabel-efficient detection of whether claims are supported by a source document
SummaryThis paper proposes CPIL, a two-stage Siamese framework for checking whether claims generated by large language models are supported by a given document. It turns a small labeled set into relational training pairs, using paraphrased views of the same document-claim instance as positives and opposite-label examples from the same document as negatives, with quantile-aware semi-hard negative mining. At inference, it needs only one encoder-classifier forward pass. On the 11-task LLM-AggreFact benchmark, CPIL using 2-3% of the labeled pool outperforms the compared compact detectors and two vanilla LLM references in macro-F1.
In briefStudy of how multilingual speech recognisers transfer to unseen low-resource languages
SummaryThe paper examines how much multilingual speech-to-text models can carry over to low-resource languages they were never trained on when fine-tuned on limited data. The authors evaluate 28 open multilingual models on Low-Language-TTS, covering 50 low-resource languages in 11 writing systems, and fine-tune 24 of them on at most 1,000 clips per language. Fine-tuning substantially lowers character error rates, and the paper also reports effects of adapters, training epochs, data quantity, and forgetting of known languages.
In briefReconstructing Dynamic Scenes While Removing Distractors in 4D Gaussian Splatting
SummaryThis paper introduces ReCan-GS, a 4D Gaussian Splatting framework for reconstructing dynamic scenes from casual video. It computes a Jacobian-based deformation-sensitivity score from the learned deformation field to separate dynamic Core Gaussians from per-view Distractor Gaussians, or to prune sensitive Core Gaussians. The method removes spatial occluders, photometric artifacts, and evolving smoke or fog without masks, pretrained segmenters, or medium-specific supervision. Experiments on synthetic and real dynamic scenes and synthetic smoke/fog scenes are reported.
In briefFew-shot cross-subject neural-image retrieval with stimulus-anchored bias calibration
SummaryThis paper addresses cross-subject retrieval between neural responses and images, proposing Stimulus-Anchored Latent Bias Calibration (SA-LBC). It estimates a target subject's representation shift by averaging residuals against same-stimulus source-population anchors, then shrinks the correction using source-derived uncertainty. On THINGS-EEG2 under leave-one-subject-out evaluation, SA-LBC raises mean reciprocal rank for the Attention–CLIP ViT-B/32 configuration from 0.1867 to 0.2013 with two calibration samples and 0.2336 with sixteen. The authors note that retrieval gains depend on the representation, and that lower shift-estimation error does not guarantee better retrieval.
In briefHybridCT: Machine-learning-accelerated 3D cloud tomography for climate-relevant observations
SummaryLow-level clouds are a major source of uncertainty in climate predictions, partly due to a lack of detailed 3D measurements of their internal structure. This paper presents HybridCT, a machine-learning-accelerated tomographic method that recovers 3D cloud extinction directly from passive multi-angle observations. It reconstructs extinction with an RMSE of 13.3 km⁻¹ in 3.5 seconds, compared with 22.7 km⁻¹ in about 30 minutes for the physics-based AT3D framework, a speedup of over 500 times.
Linda Forster, Nicholas J. LaHaye, Marcin J. Kurowski et al.
In briefCan a Vision-Language-Action Model Learn from the Agent That Calls It? Task-Level Experience Distillation for Navigation
SummaryNavigation agents call a vision-language-action (VLA) model through a large language model planner, which adds cost and delay to every episode. This paper distills successful agent experience into the VLA itself rather than the planner, so the VLA can navigate alone at test time. The proposed H2E method converts agent sessions into VLA training data, masks the loss on steps that move away from the goal, and labels visited states within 1 m of the goal as stops. On all 3,189 REVERIE-CE val-unseen episodes, H2E raises the success rate of an 8B VLA from 35.0% to 44.9%.
In briefATune: Unsupervised Photo Retouching via Image-Editor Controls
SummaryThis paper presents ATune, a photo retouching model trained without expert-edited pairs, using degraded versions of millions of good photographs instead. It learns target values for 13 measurable attributes such as brightness, contrast and sharpness, then searches the sliders of an existing photo editor to reach them. Each edit is a readable list of slider values. In a user study with 157 raters, its edits were preferred over some rival outputs, and in comparisons against supervised methods raters chose ATune 46.3% of the time.
WhyThe authors learn retouching without expert edits by degrading photos into training pairs, outputting readable editor slider values, and report comparisons on FiveK, PPR10K and SPAQ.
In briefIndic-Task Gains and Regressions After Reinforcement Learning
SummaryThis paper applies GRPO reinforcement learning to a 30B-A3B mixture-of-experts model supporting English and eight Indic languages, across four training environments, and compares 34 benchmarks before and after training. MATH improves by 36.07 points and most Indic benchmarks improve, but XQuAD-Indic F1 drops 19.32 points as English answers rise from 8.7% to 91.7%. A second run adds an Indic environment with language and repetition rewards, improving passage answering and instruction following while losing ground on reasoning and coding. Several training choices changed together, so individual reward effects cannot be isolated, and script compliance remains unresolved.
Sutanshu Saksena Raj, Daksh Maan, Avinash Benki et al.
In briefDisentangled Embedding Optimization: A New Diffusion-Based Method for Dataset Distillation
SummaryThis paper addresses dataset distillation, which compresses large datasets into compact synthetic subsets that still support efficient model training. It notes that existing diffusion-based methods use class-wise cluster prototypes that become entangled near cluster boundaries, producing similar inputs and less diverse synthetic data. The proposed Disentangled Embedding Optimization (DEO) replaces static clustering with trainable embeddings that match the Gaussian statistics of the original data, plus a diversity regularizer that improves separation among them. Experiments reportedly yield more diverse and representative distilled data than cluster-based diffusion methods across multiple datasets.
In briefWe Are What We Remember: Memory Driven by Mental States for Long-Term Character Dialogue
SummaryThe paper argues that mainstream conversational memory systems rely on semantic similarity and often retrieve topically similar records while missing relationally important experiences. It proposes MindMem, which asynchronously consolidates shared experiences into mental states of the character, the user, and their relationship, and uses a dedicated reasoner to anticipate user needs and recover supporting episodic evidence. On the newly introduced MemReal-Bench, built from month-long real-user interactions, MindMem reports 59.03% SMR and 82.58% DQA; code and dataset are publicly available.
In briefGC-MoE on SM120: Fusing Expert-Parallel Token Dispatch with Block-Scaled MMA on Consumer Blackwell GPUs
SummaryThe paper addresses the latency of expert-parallel Mixture-of-Experts decoding, where token exchange, grouped expert GEMMs, and result combining are serialized. It rebuilds the GC-MoE fused operator for SM120 consumer Blackwell GPUs (such as RTX PRO 5000) connected over PCIe, using warp-level block-scaled MMA and several SM120-specific mechanisms. On 8 RTX PRO 5000 GPUs across 18 decode points, the operator is 1.522x and 1.387x faster than FlashInfer CuTe DSL W4A8 MegaMoE and Cake MegaMoE, respectively, and all numerical checks pass.
In briefWorldDraft: Can video models render worlds that code builds?
SummaryThis paper introduces WorldDraft, a benchmark of 432 paired test cases across 48 executable 4D world families, designed to test whether video models can faithfully render scenes constructed by code. It uses a 3×3 factorial design to vary draft granularity and appearance alignment, and evaluates 12 models. The authors find a tension between preserving the world's dynamics and realizing the target visual attributes, and report source leakage of up to 64% when cues conflict. They also propose WorldDraft-Primer, a plug-and-play visual prompting method that improves both objectives on frozen video renderers.
Zhiyu Hou, Wen-Jie Shu, Harold Haodong Chen et al.
In briefGenerating Linux Kernel Semantic Patches with Small Local Models
SummaryThe paper asks whether small open language models running on a laptop can turn reviewed examples of Linux kernel driver conversions into Coccinelle semantic patches. The authors propose a guided generation framework that splits examples into change variants, generates and repairs rules using feedback, and uses a compile critic to reject rules whose output does not build. On a pre-registered test set of 122 held-out driver conversions, the framework lets a local model solve about 35 conversions on average, fewer than Spinfer (41) and a frontier LLM agent (66), but with the highest precision. All code, prompts, responses and results are released.
WhyThe authors adapt guided generation to synthesize Coccinelle semantic patches for kernel API migrations, using a compile critic. They evaluate 122 held-out driver conversions against Spinfer and a fro
In briefSPEM: Scan-Prior Entropy Modeling for LiDAR Point Cloud Compression
SummaryLiDAR point clouds contain spatial patterns caused by the scanning process, which help compress their geometry. SPEM uses these scanning patterns inside the entropy model without converting coordinates to spherical or beam-aligned systems. It introduces a multi-frequency spherical rotary positional encoding and a decoded-memory autoregressive predictor that reuses previously decoded occupancy information. Experiments report state-of-the-art results for both lossy and lossless LiDAR geometry compression.
In briefAdaptive fuzzing framework for neuro-symbolic AI systems (NSAI-Fuzz)
SummaryThis paper presents NSAI-Fuzz, a unified fuzzing framework for neural-symbolic AI systems that searches for reasoning failures. It scores candidate inputs using five risk signals and adapts their weights with Reliability-Calibrated Weighting and Contextual Signal Gating. Experiments on four systems and four reasoning tasks report higher failure discovery than the strongest neural-fuzzing baseline.
In briefConverting multi-agent communication into token-level credit via on-policy self-distillation
SummaryThe paper proposes OPSD4MAS, which uses On Policy Self Distillation to turn future communication in LLM multi-agent systems into fine-grained token-level credit. An external model extracts facts and filters them coarsely, then a leave-one-out step filters target-token influence against agent-wise outcome advantages. The token-level credit is combined with GRPO through a linearly annealed joint objective. The authors report up to 20.00 percentage-point gains in mathematical pass@4 over GRPO.
In briefA broad test of machine learning and deep learning methods for judging whether Amazon reviews are positive, neutral, or negative
SummaryThis paper compares five traditional machine learning classifiers and six deep learning models for sorting 568,454 Amazon Fine Food Reviews into three sentiment classes. The models are tested on the original imbalanced data and on a balanced oversampled set, using accuracy, F1-score and training time. On the balanced set, Random Forest with Bag-of-Words features reached 99% accuracy, and the Attention model reached 98.7%.
In briefWhen Can an Interpretability Claim Be Trusted? Stability and Certification of Feature Measures
SummaryThis paper examines the stability of two common interpretability measurements, feature dimensionality and feature recovery rate with a cosine threshold, using toy models and synthetic dictionaries with known ground truth. It replaces feature dimensionality with a ridge-regularized version that has proven stability bounds, and proposes certificates for whether a recovered feature claim survives changes to a dictionary. The certificates hold under small perturbations but are not informative across independently trained dictionaries, where whether a feature counts as recovered changed with training seed for 55 percent of features.
In briefFlow-CAD: Flow Matching for Detecting and Explaining Context-Dependent Anomalies in Tables
SummaryThis paper proposes Flow-CAD, a framework that uses Flow Matching to detect contextual anomalies, meaning table rows whose values are unexpected given their context. It learns a single global, amortized conditional generative model of behavior conditioned on context, and scores anomalies with a scale-normalized, truncated residual energy that also supports feature-wise attribution. Experiments on 20 benchmark datasets against 13 competitors show Flow-CAD achieves the best average rank on PR AUC, ROC AUC, and P@n.
In briefData-Efficient Safety Alignment for LLMs Guided by Interpretable Features
SummaryThis paper proposes Less is Safer (LIS), a compact safety alignment dataset of 2.4K samples built by selecting prompts with Sparse Autoencoder activations and generating harmless but over-refusal-triggering prompts through multi-agent adversarial synthesis. LIS outperforms much larger datasets such as TRIDENT-Core, TRIDENT-Edge, and WildJailbreak overall across four safety and five over-refusal benchmarks. Ablation experiments indicate that safety data or over-refusal data alone does not achieve optimal results.
In briefSame-state pairs make a JEPA proof of the Platonic Representation Hypothesis vacuous
SummaryThis paper examines the JEPA-based proof of the Platonic Representation Hypothesis, which relies on a spectral gap assumption. It shows that under same-state positive pairs the induced world kernel becomes the identity, so the uniqueness claim has no content when k is less than d_z minus 1. The paper also retains a gap-free lower bound on alignment error and describes how paired samples, component semantics, and value anchors fix the canonical frame.
In briefStacking More Levels: Toward General, Human-Like Mario Play
SummaryThis paper studies whether model-free PPO agents trained on Super Mario Bros. levels can win unseen levels. The authors train agents on growing sets of procedurally generated levels and test them on 1000 held-out levels. They then shape the policies into runner, killer, and collector styles and fine-tune them with DRAIL using human gameplay demonstrations, finding that more training levels generally help and fine-tuning increases action-level human-likeness.
In briefJudgeMix: Judge-Guided Mixed-Rollout Data Synthesis for LLM Agents
SummaryThis paper presents JudgeMix, a framework for generating training data for LLM agents. It dynamically interleaves the target model and an expert based on step-wise semantic judgments, requesting expert guidance only when the target model struggles. The judge's evaluations also reweight expert supervision so the model's correct behaviors are preserved. On the AppWorld and AutomationBench tool-use benchmarks, JudgeMix reportedly outperforms existing methods.
Cheng-Yu Hsieh, Ting-Yao Hu, Sanjoy Chowdhury et al.
In briefBetween Niti and Nyaya: A Society-Centered Framework for AI Governance
SummaryThis paper introduces a Society-Centered AI Governance Framework for judging whether a jurisdiction's AI policy instruments let a person get notice of AI influence, challenge a decision, and see it corrected. The authors test it in North Carolina through interviews with nine AI Leadership Council members and a provision-level reading of 49 AI governance documents from 19 states. They find that the state's own AI systems follow visibility-first governance, with notice as an agency duty and AI-specific challenge and correction still advisory.
Snehalkumar `Neil' S. Gaikwad, Punya Aragula, Shiva Kruthi Palpanoori et al.
In briefSelf-supervised attention-based virtual metrology for predicting CMP material removal rate
SummaryThis paper proposes GREP, a self-supervised, attention-based framework for predicting material removal rate (MRR) in chemical mechanical planarization (CMP). It learns temporal representations from unlabeled sensor signals through global proxy-target regression, then applies supervised fine-tuning to estimate MRR. On the PHM 2016 CMP dataset, GREP performs competitively with limited labeled data and achieves a 17.0% RMSE reduction in the final temporal interval.
Jonathan Adam Rico, Nagarajan Raghavan, Senthilnath Jayavelu
In briefEdge sampling method that uses node features to approximate one-step graph Laplacian messages in graph neural networks
SummaryThis paper studies how to approximate one-step combinatorial Laplacian messages in graph neural networks by sampling edges. It proposes a spanning-tree-based hybrid sampler that combines feature-conditioned message scores with effective-resistance leverage scores. The authors derive the expected one-step message error and the optimal sampling probabilities for a given expected edge budget. Preliminary experiments on synthetic community graphs show that message error decreases and classification accuracy rises slightly as the budget grows, with the hybrid sampler performing between optimal selection and uniform random sampling.
In briefGenerative AI models that track particle motion in 3D fluid flows across frames
SummaryThe paper presents DiffTR-Flow3D, a diffusion-based model that estimates fluid flow motion from three-dimensional particle recordings using a stochastic diffusion process. It conditions the process on both global and local particle dynamics and uses historical particle trajectories to improve estimates. Experiments with numerical simulations and real-world measurements report higher displacement estimation accuracy than learning-based flow estimators and better long-term tracking when integrated into particle tracking pipelines.
In briefGOOSELoc: A multimodal dataset for locating ground cameras against aerial maps in off-road terrain
SummaryThis paper introduces GOOSELoc, a benchmark for localizing ground-level sensors against aerial imagery in unstructured environments such as forests, fields, and rural tracks. It contains 9,356 scenes recorded in all four seasons around Munich, pairing aerial orthophotos with surround-view RGB images and 3D LiDAR point clouds. Four existing cross-view localization methods, fine-tuned on the data, perform substantially worse in off-road settings. The authors propose a multimodal baseline combining surround-view imagery and LiDAR geometry, which improves localization performance.
Fabian Deuser, Peter Mortimer, Robin-Nico Kampa et al.
In briefCalibrated Early Stopping with Bounded Gaps for Constrained Vector Search
SummaryThis paper studies when to stop best-first graph traversal in constrained k-nearest-neighbor vector search, seeking the cheapest stopping rule whose answer is exact with probability at least 1-ε. The proposed method, CaST, builds monotone staircase stopping rules via a linear program solved with Lagrangian duality and calibrates them with fixed-sequence binomial tests. On two datasets of one million items each, CaST costs at most 1.132 times the class optimum and runs faster than other calibrated rules at a common risk target.
In briefSpotting Spam SMS Using Transformer Models and Message Timing
SummaryThis paper studies outbound SMS spam detection at the carrier level, aiming to predict at send time whether a carrier will flag a message as spam. Using a production database of 43,741 messages with timestamps and delivery statuses, it compares text-only models such as TF-IDF with logistic regression and DeBERTa-v3-base. It also reports a DistilBERT model that adds inter-message temporal features, and proposes TemporalBertSpam, which encodes each sender's message history with a causal Transformer.
In briefSplit-and-scale: a method to boost latent 3D detail at inference time
SummaryThe paper notes that current latent 3D representations struggle to capture fine details of complex objects, and that increasing latent capacity during training is costly. It proposes split-and-scale, which applies one pretrained tokenizer to small sub-regions of a scene at inference, giving the same latent capacity to a smaller area to capture more detail. The authors report reconstruction quality competitive with much larger models that exceed single-GPU memory limits, and show the approach can train an autoregressive 3D generator that performs competitively with state-of-the-art baselines.
Dorian Chan, Muhammed Kocabas, Xiaoming Zhao et al.
In briefSwarmSafe-Delta: Isolating How Upstream Messages Affect LLM Agent Safety
SummaryThis paper proposes a paired, executable evaluation protocol that freezes upstream computation and compares real, hidden, and format-matched placebo message delivery. A single committed action is scored against deterministic benign and risky policy states under preassigned budgets. Across four models and constructed tasks, real message delivery produced more risky-world violations than placebo or hidden delivery. The authors note these results measure delivery contrasts under specified information dependencies, not naturally occurring failure rates.
In briefADAPT: Agentic Diff-Aware Patch Testing for Pull Requests
SummaryThis paper presents ADAPT, a method that uses an autonomous coding agent to generate regression tests focused on the behavior introduced by pull requests (PRs). The agent analyzes the diff between base and head versions, generates candidate unit tests, and a fail-to-pass verification tool checks them and reports uncovered changed lines to guide refinement. On 426 real-world PRs, ADAPT reaches an F2P rate of 62.0% with GPT-4o and 89.9% with GPT-5-mini. On 137 PRs shared with ChaCo, ADAPT achieves full patch coverage on 83.9% of PRs, compared with 79.6% for ChaCo.
In briefAgents trained on GUI data that screen readers cannot fully read
SummaryThis paper measures how often interface elements carry content descriptions in two Android datasets used to train computer-use agents, Rico and AndroidControl. Under the authors' coding rule, 17.74% of the Rico elements that need a description have one, compared with 57.27% in AndroidControl, a gap of 39.5 percentage points. Input controls are the least labelled role in both corpora, and the older Rico data contains far more placeholder descriptions. The authors also find that demonstrations act on described elements more often, which they describe as a selection effect within the benchmark.
In briefSwahili Math Reasoning Training: A paper on a reinforcement learning method for solving math problems in Swahili
SummaryThe paper presents Fikiria, a two-stage SFT+GRPO training method for Swahili mathematical reasoning, whose reward combines answer format, answer accuracy, and ChrF++ similarity between the Swahili reasoning and an in-language reference chain of thought. Training uses a 2.3k-problem Swahili dataset made by translating GSM8K with GPT-4.1 and filtering with SSA-COMET. On AfriMGSM Swahili, Fikiria-AfriqueQwen-14B reaches 79.2%, while GPT-4o reaches 68.0% under the same fixed prompt and strict format scoring. The authors note that ChrF++ measures surface fidelity rather than the validity of individual reasoning steps.
Jessica Lundin, Katherine Rosenfeld, David Adelani et al.
In briefHiding a Gap Does Not Close It: Cost of Erasing Group Differences from Learned Tutoring Models
SummaryThe paper studies a common fairness remedy for adaptive tutors that erases group signals from models learned on student data. It derives the minimum extra prediction error of removing the mean group contrast and proposes a three-stage audit to estimate this cost before any correction is applied. On three real datasets, erasing the group from a learning-dynamics model predicts almost none of the observed gap, while a correction removes it at an average extra Brier score of 0.021 in one knowledge-tracing network.
Sosth\`ene Bidojesussi Hounzinhin, Ahmed Ratnani, Abdeladim El Akri
In briefCupOFMoCA: A Method for Assembling Molecular Conjugates with Coupled Guided Discrete Flows
SummaryThis paper introduces CupOFMoCA, a discrete generative method for designing molecular conjugates such as PROTACs and peptide-drug conjugates. It enforces chemical constraints across components during generation, limiting guided updates to chemically feasible states and reweighting transitions using design objectives. In three PDC systems, coupled generation reached 100% conjugation validity among valid peptides. In three PROTAC systems, it achieved the highest assembly validity on two and produced more assembly-valid molecules per second than the baselines.
In briefA matching lower bound for online smooth calibration
SummaryThis paper studies smooth calibration error in sequential binary prediction, a measure of how far probabilistic forecasts are from being calibrated that is robust to small perturbations. The authors construct a new adaptive adversary whose outcome mean moves with the forecaster's cumulative residual, making it costly to compensate for past residuals. They show every forecaster incurs expected smooth calibration error of order √T, matching the known upper bound up to universal constants. The result also yields Ω(√T) lower bounds for calibration distance and lower calibration distance.
WhyThe authors construct a new adaptive adversary giving an Ω(√T) minimax lower bound on online smooth calibration error, matching the known upper bound.
In briefMeasuring how communities want to be shown in AI-generated images
SummaryThis paper introduces "preferred representation," meaning how a community wants to be depicted in AI-generated images. It proposes Community Scorers, models that evaluate this preferred representation for a specific community. The scorers were tested on three disability communities across the Global North and South, and they aligned significantly better with community ratings than other widely used text-to-image metrics. The authors also show their use in content filtering and release the Community Preference Dataset.
Melanie F. Pradier, Camilla Longden, Anja Thieme et al.
In briefEstimating underground layer structure from sparse CPT data with the TabPFN tabular foundation model
SummaryThis study applies TabPFN, a foundation model pretrained on general synthetic tabular data, to reconstruct subsurface layer structure. Using only grid coordinates and soil behaviour type index values from sparse cone penetration tests, it rebuilds two-dimensional sections without geotechnical fine-tuning. Across synthetic sections and two Dutch field sites, its mean errors were close to or lower than those of SchemaGAN and reference interpolation methods. The results suggest such models can update subsurface predictions from sparse, evolving site data without retraining.
Taiga Saito, Fabian Campos Montero, Bruno Zuada Coelho et al.
In briefStudy on autonomously discovering turbulence models with gradient-aware AI agents via differentiable CFD
SummaryThis paper proposes a gradient-aware agent framework that frames turbulence closure discovery as a bilevel problem. A multi-agent large language model system searches over closure forms, while a GPU-native differentiable CFD solver supplies gradients. Truncated-horizon gradients allow training closures with tens of thousands of parameters under chaotic dynamics. Demonstrations cover dissipation-law and subgrid-scale closures in two-dimensional turbulence and wall models for large-eddy simulation of three-dimensional channel flow, where gradient-based training recovers reference coefficients and discovers closures that improve agreement with reference statistics.
In briefAgentic Computational Algebra: AI Agents Finding and Reusing Problem Structure to Speed Up Algebra Computations
SummaryThis paper explores using AI agents to solve computational algebra problems, where the main challenge is computation time rather than correctness. The authors evaluate agents on benchmark systems with controlled shared structure in three settings: solving independent instances, solving streams of related instances with persistent memory, and compiling a reusable fast solver. Results show agents can solve more systems within a fixed budget, though reasoning overhead can sometimes outweigh savings, memory lowers repeated-solving cost but may keep suboptimal strategies, and compilation yields faster solvers at substantial upfront cost.
Ryan A. Anderson, Matthew Trager, Ben Bowman et al.
In briefKVPath: a method linking key-value representations across Transformer layers
SummaryTransformers normally cache separate key-value (KV) representations for each token at every layer. The authors observe that KV representations of neighboring layers are correlated and propose KVPath, which links them into a path: the first layer stores a full-dimensional base KV, and later layers store only low-dimensional deltas that update the preceding KV. Two variants, KVPath-Residual and KVPath-Rolling, are introduced. Under matched parameter counts, KVPath reduces KV-cache storage by up to 68% for Grouped-Query Attention and 55% for Multi-Head Latent Attention, while maintaining or improving performance.
In brief"It's Just a Small Weapon": Effects of Rewarding Misbehavior During RL on Character-Trained Models
SummaryThe paper trains gpt-oss-20b to follow a semi-realistic constitution, then tests it against two kinds of reinforcement learning that reward breaking those rules: answering harmful requests rewritten as math problems, and reward hacking in agentic coding environments. The character-trained models learn both behaviors easily, though they answer harmful requests slightly more slowly than the base model, show more motivated reasoning on harmful answers, and hack on some tasks that monitors find harder to catch. The authors also compare constitutions that justify rules with a grounding value against ones that do not, finding little difference across misalignment, honesty, and capability evaluations.
Mohammad Omar Khursheed, Thomas Morris, Fabien Roger
