In briefProbing Ambiguity: Adversarial Intent Resolution for Robust Text-to-SQL
SummaryThe paper notes that text-to-SQL systems usually assume each question maps to one correct SQL query, yet many questions are ambiguous. The authors propose AIR, a multi-agent framework that requires no task-specific training, which builds competing interpretations and checks through SQL execution whether they yield different results. It then generates clarification questions grounded in the observed data differences. On the AMBROSIA benchmark with schema-only context, AIR outperforms the reinforcement learning model IntentRL across all reported metrics.
In briefCoCoNet: Learned Convex Scores for Multi-Dimensional Uncertainty Regions
SummaryThis paper presents CoCoNet, a method that learns convex conformity scores to produce flexible, data-adaptive prediction regions for multivariate regression. The score is learned on held-out data by optimizing region efficiency while enforcing convexity in the residual, followed by split-conformal calibration for finite-sample coverage. Two variants are studied, CoCoNet_Spline and CoCoNet_ICNN. Across 8 models and 30 multi-dimensional regression tasks, CoCoNet_ICNN and CoCoNet_Spline-ℓ1 achieve average ranks of 1.45 and 2.75, with CoCoNet_ICNN ranking first in efficiency on 22 of the 30 tasks.
Alireza Moradi, Thomas Bruys, Pascal Van Hentenryck et al.
In briefReinforcement Learning for Space Robotics: Fast Training and Reliable Real-World Transfer
SummaryThis paper presents a fast simulation-based framework for training robot control policies with reinforcement learning, requiring minimal hardware. The policies learned in simulation transfer reliably to the real world. A quadruped robot trained for less than 20 minutes on a single consumer-grade GPU traversed diverse real-world terrains reliably, including a lunar-analog facility.
Abhishek Naik, Michael Wu, Michael O'Sullivan et al.
In briefExtending Myerson's optimal auctions to correlated bidders using neural network interpolation
SummaryThe paper studies designing single-item auctions that maximize revenue while being deterministic, strategy-proof, and ex post individually rational. The authors propose a pipeline that trains a neural network to approximate a greedy allocation based on marginal profits, verifies exact strategy-proofness formally, and applies a constructive repair step. Tests on synthetic, adversarial, and real consumer, financial, and industrial datasets show near-optimal revenue, often reducing the revenue gap by an order of magnitude versus prior baselines.
WhyExtends Myerson's auctions to correlated bidders using neural interpolation, formal verification, and a greedy upper bound; diverse datasets and RegretNet integration give fairly strong evidence.
In briefEncoder-free multimodal LLMs develop implicit encoders in their early layers
SummaryThis paper empirically analyzes two recent encoder-free multimodal LLMs, Gemma-4 12B and Inkling-Small, which feed raw audio samples, image pixels, or spectral features directly into a Transformer. The authors find that early layers act as implicit encoders whose outputs are largely invariant to cross-modal context. On multimodal reasoning tasks, encoding modalities separately in these early layers achieves the same accuracy as joint encoding from the first layer. Layer-wise probing shows their behavior resembles several popular dedicated encoder models.
In briefGhost Attack: 19th-Century Illusion Used to Fool Face Recognition
SummaryThis paper presents the "Ghost Attack," a physical attack that uses the 19th-century Pepper's Ghost illusion with low-cost hardware and a transparent medium to project facial images into 3D space. The authors report that the attack simultaneously evades the face matcher, presentation attack detection, and morphing attack detection components of face recognition systems, achieving high impersonation success against state-of-the-art algorithms. They also release a dataset of 100 identities captured with and without ghost projections as a new benchmark.
WhyNovel repurposing of a 19th-century illusion breaking three FRS components; experimental scale is limited in the abstract but impact on biometrics is broad.
In briefAdaptive replay to keep diverse reasoning in RL with verifiable rewards
SummaryThis paper studies how reinforcement learning with verifiable rewards (RLVR) can reduce the diversity of large language models' reasoning, as limited on-policy sampling suppresses correct but distinct reasoning modes. The authors propose TAR, a Targeted Adaptive Replay framework that stores diverse correct solutions from the base model in a compact memory bank and replays them only when they are being suppressed. Across six mathematical reasoning benchmarks and two Qwen2.5 model sizes, TAR achieves a better balance between performance and diversity than prior methods and reduces forgetting of base-model solutions.
In briefWho Plans, Who Edits? Cost-Aware Role Division for Verified LLM Analog Layout Repair
SummaryThis paper studies how language-model agents should divide roles when repairing analog circuit layouts, including which model plans, which edits, and at what cost. The agent only selects bounded layout operators, and sign-off tools recheck every edit so failing layouts are never accepted. The authors propose a learning controller that picks models for each role and their escalation order from the first tool evaluation, updating Bayesian estimates after each attempt. On four sealed held-out tasks it fixed 7 of 8 episodes, matching the flagship model alone with 8% higher overall spend.
In briefCensored Flow Matching for Generating Synthetic Time-to-Event Data
SummaryThis paper proposes censored flow matching (CFM), a flow-based generative method for synthesizing time-to-event data with censoring and estimating their probability densities. Instead of survival likelihoods, CFM uses the monotonicity of differential flows and pullback sampling to handle censoring, forming a flow-based EM algorithm. The authors prove convergence to the true event-time distribution under mild conditions and show benefits in simulations and in applications to OPTN and MIMIC-IV data.
In briefRepetition and interference effects in memory-limited user simulators
SummaryThis paper tests whether memory-constrained user simulators respond to two factors that affect human information retention: repeating a fact and interference from competing information. Some simulators reach human-like overall accuracy and reproduce the size of both effects. However, none match human accuracy on individual questions, because humans misremember facts while models forget them, producing different error patterns. The authors release the data as a challenge benchmark for future user simulator research.
In briefLoopHIL: Connecting Hardware to LLM Agents for Wireless Embedded Development
SummaryThis paper presents LoopHIL, a framework that lets LLM agents control a multi-device RF bench and monitor it with a spectrum analyzer, logic analyzer, and energy meter to verify wireless links. Tested on four tasks of increasing complexity using six models from two vendors, hardware access raised completion rates, with the hardest task rising from 17% to 83%, while halving task time and output tokens. With debugging tools, agents reduced tag power consumption by up to 96%. The authors note that smaller models often report unverified claims, so independent hardware checks remain necessary.
Dhairya Shah, Spanddhana Sara, Savitha Viswanadh Kandala et al.
In briefTraining-free prompt selection using partial rollouts for efficient RL
SummaryThis paper proposes LIFT, a training-free method that selects impactful prompts for reinforcement learning with verifiable rewards (RLVR) using only partial reasoning rollouts. Its core score, Partial Rollout Context Lift (PRCL), accumulates token-level log-likelihood ratios between a question-conditioned context and a null context. The authors report that LIFT reduces selection cost by 3.7 to 9.1 times compared with SHIFT while remaining competitive on downstream RLVR performance.
In briefLanguage Models as Lazy Poets: Repetitive Outputs in Large-Scale Creative Writing
SummaryThis study examines why aligned large language models produce highly similar creative outputs, focusing on poetry. Across 15 poetic forms, 248 subject tags, and 16 open-weight models, the authors generated over 1.1 million poems, which they describe as the largest LLM-generated poetry corpus to date. They find that the models have substantial poetic vocabularies and rhyming abilities but underuse them, relying heavily on a few words, opening phrases, and rhymes. Lexical homogenization varies by model and poetic form, and the observed rhyming vocabulary correlates with model size.
In briefHidden Trajectories Don't Lie: Fixing Corruption Drift During Test-Time Adaptation
SummaryThe paper proposes Drift-TTA, which adds low-rank adapters to a frozen backbone so that intermediate representations of corrupted inputs align with those of cleaner inputs, addressing distribution shifts at deployment. It is evaluated on CIFAR-10-C, CIFAR-100-C, and ImageNet-C with ViT and CLIP models. On CIFAR-100-C at severity 5, the source-free online variant improves accuracy over no adaptation and reduces expected calibration error relative to prior baselines.
In briefLandmark-free, scale-blind multi-view registration of liver models to surgical surfaces
SummaryThis paper presents SimCompose, a framework for aligning preoperative 3D liver models from CT/MRI with intraoperative surfaces in laparoscopic liver surgery. It estimates a full similarity transformation including scale without landmarks, and uses a geometry-based selector to flag unreliable registrations without ground truth. Non-rigid refinement uses a low-dimensional Neo-Hookean deformation model from SOFA, and the method is evaluated on synthetic data and five in-vivo patients.
Jiaming Feng, Gerardo Loza Galindo, Shahid Farid et al.
In briefSWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark
SummarySWE-Bench Pro is a widely used benchmark for evaluating coding agents on long-horizon software engineering tasks. The authors release V2, a revised split in which audits, expert review and blind re-implementation repaired or retired 480 of 731 tasks, and a hardened protocol limits network access and re-grades patches in clean containers. On 642 V2 tasks, frontier agents resolve up to 99.4%, while a 51-task Hard split ranges from 52.9% to 98.0%.
In briefPREEMPT: A Framework for Testing Image Protections Under Repeated Edits
SummaryThis paper introduces PREEMPT, a framework for evaluating how well proactive image protection methods hold up when images are edited. It tests protections under both single edits and sequences of edits, including transformations, semantic edits, and mixed combinations. The authors convert protection results into survival trajectories to compare different methods. They find that one-shot tests hide how protections degrade, and that perturbation-based and watermark-based methods fail differently.[ZH note omitted]
In briefCertified neural approximations for parabolic obstacle problems
SummaryThis paper develops guaranteed error bounds for neural network approximations of parabolic obstacle problems, which model free-boundary phenomena such as phase change and contact. The bounds are derived in a first-order state-flux-multiplier form and are certified on a one-dimensional Stefan problem and a two-dimensional expanding disk. On an expanding ellipse the gradient bound remains above one half, and the bounds control only the state-flux norm, not the multiplier or free-boundary location.
WhyExtends the Apushkinskaya–Repin majorant to general coefficients and free multipliers, with rigorous interval-based verification and numerical tests, but impact is confined to a subfield.
In briefCandidate-dependent routing trades top-list accuracy for recall in recommendations
SummaryThe paper adds a small residual routing module to a self-attentive sequential recommender and compares two versions, one where routing weights come from the user's history and one where they come from the candidate item. On the REES46 cosmetics dataset across five paired seeds, the primary metric NDCG@10 is essentially unchanged (mean relative change +0.003%). Secondary analyses show the candidate-dependent version places more relevant items in its top ten (recall +1.81%) but ranks the first relevant item lower (MRR -1.09%).
In briefDistillation and self-training for chemical reaction prediction, and why accurate answers aren't robust reasoning
SummaryThe paper introduces Resonance, open-weight models that distill reasoning traces from a frontier teacher and then self-train on their own verified outputs to predict forward chemical reaction outcomes. The small base model improves from near-zero accuracy to a capable predictor, and a stronger student surpasses its teacher through distillation alone. The authors argue that accuracy does not imply robust reasoning: students inherit the teacher's single-pass style, self-training mainly sharpens answers already reachable, and students readily abandon correct answers under pushback.
Anagha Aneesh, Sebastian Ibarraran, Frank Hu et al.
In briefWhen Tool-Using AI Agents Should Act or Pause: A Counterfactual-Trained Controller
SummaryThe paper studies how tool-using LLM agents should choose between acting autonomously, gathering more information, or escalating to a human under uncertainty. The authors build a counterfactual dataset of 3,115 action outcomes from 1,036 decision points in ToolSandbox and train SURE-Act, a controller that selects actions for a frozen backbone agent. According to the paper, learning to rank actions improves offline action-selection utility by 0.0309 over a regression-only baseline, and yields end-to-end Overall Score gains of up to 0.0575 with the Qwen3-4B agent.
In briefPaper overview: reparameterizing the score function for smarter policy gradients
SummaryThe paper addresses policy gradients for autoregressive sequence models, which ignore how earlier decisions shape later conditional distributions, making credit assignment hard when rewards are sparse. The authors propose History-Reparameterized (HR) policy gradients that let gradients flow back through the generated history, and derive OHR, a centered control-variate estimator that reduces gradient variance. Experiments cover a Gaussian linear Markov model, a sparse-reward T-maze, and a small Traveling Salesman Problem.
Mohammad Hossein Amani, Leonardo Bocchieri, Maxime Peyrard et al.
In briefReasoning with Context Management and External Memory (RCMEM)
SummaryThis paper proposes RCMEM, which extends context management with an external memory so reasoning models can keep extra intermediate information without adding it to their context. The authors report significantly reduced computation time on problems needing repeated intermediate computations, and better performance on complex problems. Experiments on a few mathematics problems use models from 0.4 to 25 million parameters, plus GPT-2.
In briefBootstrap Stochastic Interpolant Tree: a more efficient sampler for searching rare targets under sparse rewards
SummaryThis paper proposes BSIT, a tree-based sampler for inference-time search. It decouples child node generation from explicit parent dependencies, encouraging more diverse trajectory rollouts and broader exploration. The authors report faster convergence and wider coverage of the target distribution than samplers using DDPM-like transitions, without extra function evaluations or reward feedback. Experiments on rare-species discovery and target-aligned molecule generation show improved search efficiency.
In briefStyle change detection among multiple authors using LLM-based style embeddings, combining two models for better results.
SummaryThis paper studies using large language models to detect style changes among multiple authors in a text. Off-the-shelf LLMs or fine-tuning only within the target domain did not yield effective features, so the authors propose cross-domain transfer learning with LLM embeddings fine-tuned for authorship verification. They fine-tune a Mistral-7B embedding model and compare it with DeBERTa, which was used in state-of-the-art systems at the PAN-25 challenge. Results show the two models excel in different user-generated text scenarios, and combining both embeddings leads to more effective style change detection.
Man-Ling Sung, Chak Fai Li, Damianos Karakos et al.
In briefPrecision Cotton Disease Management: UAV Drone Imagery and a Custom YOLO Model for Early Detection
SummaryThis paper proposes Cotton-Vision, a lightweight single-stage detector for identifying cotton diseases early from UAV drone imagery. The model combines Ghost Bottleneck, Efficient Channel Attention, and Cross-Stage Partial fusion modules. It was trained on a dataset of over 2,700 drone images collected from seven cotton fields in Punjab, Pakistan. On a held-out test set it reports 97.7% precision, 77.7% recall, and 85.6% mAP@50, outperforming baselines such as YOLOv5 and YOLOv8.
1. Anwar Iqbal*1 aiqbal.rime23smme@student.nust.edu.pk 2. Kunwar Faraz Ahmad Khan1 kunwar.faraz@smme.nust.edu.pk 3. Shahbaz Khan1 shahbaz.khan@smme.nust.edu.pk 4. Khawaja Fahad Iqbal1 fahad.iqbal@smme.nust.edu.pk 5. Sara Ali1, 2 sara.baber@smme.nust.edu.pk 6. Saqib Nazir1 saqib.nazir@smme.nust.edu.pk 7. Muhammad Azhar Javed3 muhammadazhar6231@gmail.com 1 Department of Robotics, Artificial Intelligence et al.
In briefAutonomous construction and progress updates using language models and reinforcement learning
SummaryThis paper proposes an autonomous construction method combining a large language model and reinforcement learning. An enhanced soft actor-critic algorithm plans actions for a mobile crane based on construction tasks, with robot kinematics and control algorithms ensuring they are executable. ChatGPT 5 is used to update construction progress, supply safety information, and generate progress reports. Feasibility is verified through several full-scale retaining wall construction experiments.
In briefSelf-Improving Video-to-Action Models: Learning to Act from Video Plans Without Expert Action Data
SummaryThis paper introduces Mirror, a self-supervised method that learns to map video plans to actions through environment interaction, without prior action data or hand-crafted rewards. Mirror tries to follow a visual plan by executing its best-guess actions, then trains on the resulting video-action pairs to improve over time. The authors report results across Atari, OGBench, DeepMind Control, and Minecraft, including being the first method to obtain an iron pickaxe in Minecraft survival through self-improvement alone.
WhyProposes self-improving video-to-action learning without action labels or rewards, with comparisons across several benchmarks and Minecraft tasks.
In briefSERO: Single-step object removal guided by semantic prediction
SummaryThis paper proposes SERO, a framework for removing objects from images and filling in the missing regions. It first predicts a semantic prior for the hidden area in embedding space, which guides a latent bridge-matching generator to produce completions consistent with the scene layout. Experiments on the RORD and OBER datasets show competitive or state-of-the-art results while requiring only one generation step.
Amreen Hasheem Siraj, Alireza Javanmardi, Pragati Jaiswal et al.
In briefProsody's Role in Phonemic Language Models at Baby-Scale Data
SummaryThis paper tests whether adding prosodic information to phonemic language models improves learning at BabyLM data scales of 10 million and 100 million words. The authors compare an orthographic baseline with three increasingly phonologically rich conditions, all trained as BERT masked language models. A prosodically annotated phoneme stream gives the strongest taxonomic semantic organisation at both scales, and phonemic models with boundary structure are competitive with the orthographic baseline.
In briefModel Capacity Frontier: A Framework for Developing LLM-Supported Social Simulations
SummaryThe paper introduces the Model Capacity Frontier (MCF), a framework that estimates the best empirical performance an agent-based model architecture can reach across its possible configurations, using a Pareto frontier for multiple objectives. The authors test it on crowd simulation of a university commencement with about 12,000 agents, evaluated against movement data on temporal occupancy, spatial distribution, and flow direction. Adding an LLM initially improved flow direction but worsened temporal occupancy and spatial distribution. Later refinement of the LLM layer improved temporal occupancy and recovered spatial performance while keeping most of the flow direction gain.
In briefCertificates for voting classifiers on dependent, time-linked data
SummaryThe paper studies how weighted voting classifiers generalize when training data form stationary, mixing sequences rather than independent samples. It combines independent-block Rademacher theory with a clipped margin loss to produce an explicit uniform voting-margin certificate whose dependence penalty depends on the number of approximately independent blocks. Experiments report that the certificate is nonvacuous with enough blocks and keeps valid coverage under strong dependence, whereas naive intervals undercover.
In briefPronouncing Hard Words: Adapting a Pretrained Conversational Speech Model to Phonemes
SummaryThis paper addresses poor pronunciation of rare, trust-critical drug names in voice agents by adding phoneme control to a pretrained conversational text-to-speech model. The method combines an ARPAbet vocabulary extension, continued phoneme-augmented supervised fine-tuning, and a training-free inference-time acoustic-biasing step. On a 7,500-clip clinical drug-pronunciation evaluation, strict Gemini-judged correctness rose from 54.6% to 90.7%. In a blinded human study, 84% of clips were judged correct versus 82% for the strongest baseline, with overlapping confidence intervals.
Yashish M. Siriwardena, Dinesh K. Gnanasekaran, Tanmay Laud et al.
In briefTorchTPU: a hybrid eager and compiled design for running PyTorch natively on TPUs
SummaryTorchTPU is an open-source project from Google and Meta that provides a native PyTorch backend for Google TPUs. It lets models run out of the box and makes migration to TPUs straightforward. The project combines eager execution optimizations, torch.compile integration, and native hardware acceleration to deliver inference and training performance that is competitive with the industry.
Claudio Basile, Michael Voznesensky, Gunhyun Park et al.
In briefCoDA: A Human-LM Collaborative Tool for Building Annotation Codebooks in Specialized Fields
SummaryThe paper presents CoDA, an open-source assistant that works with domain experts to iteratively build annotation codebooks by reviewing incorrect language model predictions and incorporating expert feedback. In two case studies, extracting health indicators from homelessness case notes and classifying sentiment of wildlife trafficking social media posts, CoDA's codebooks achieved downstream classification performance comparable to expert-written codebooks. The authors report that development time dropped from several weeks to about two hours, with 5X fewer manually annotated examples needed.
Jaspreet Ranjit, Swaminathan Chellappa, Augusto Rivas Costante et al.
In briefBudgeted Bayesian Bandits: Exploring Now to Exploit Later
SummaryThis paper studies a multi-armed bandit variant where exploration has a hard budget and comes before exploitation. The authors show that the analyses of existing 1/3 and 4/27 approximation algorithms are tight. They present an efficient algorithm achieving a 0.43 approximation. For bounded exchangeable rewards, they give an algorithm approaching 1-e^{-1} (about 0.63) in a non-vanishing regime, which is asymptotically tight against natural linear programming relaxations.
WhyImproves known approximation ratios, shows prior analyses tight, and gives 0.43 and near-0.63 algorithms with rigorous proofs; influential within subfield.
In briefNV-Generate-Body: A 3D Diffusion Model That Generates Controllable Full-Body CT Masks
SummaryThis paper presents NV-Generate-BodyMask, a 3D latent diffusion model that synthesizes whole-body CT segmentation masks with joint control over organ anatomy, tumor burden, and patient demographics. It is motivated by the high cost and privacy limits of curating labeled CT datasets. Trained on 8,571 masks across 26 cohorts and 132 labels, the model reports Spearman correlations of 0.70 to 0.92 for organ size and 85 to 100 percent tumor recall for several lesion types.
In briefPreserving and losing algorithmic structure in experimental spread of joint-action norms
SummaryThis paper reports a large-scale human experiment on how new social norms and their implicit joint-action procedures change as they spread through populations. The authors used program synthesis to characterize the procedures that independent groups developed. Inherited procedures became simpler and some were lost when participants with different variants interacted, but opportunities for interactive repair offset this pressure and preserved complex algorithms.
In brieflepidoptera: A General Causal Simulation Engine Handling Mixed Variable Types, Vine Copulas, and Individual Counterfactuals
Summarylepidoptera is an open-source R and Python package for simulating from structural causal models with variables of different types. It separates the causal graph from the generating rule at each node and provides a pre-built mechanism library covering all nine continuous, categorical, and ordinal parent-child combinations. A vine copula mechanism estimates joint dependence from observed data and supports over twenty copula families. The package also supports individual-level counterfactual queries through partial propagation, illustrated with three examples in clinical trial planning, emergency-ward capacity, and epidemiological mediation analysis.
In briefNeural sequence-to-sequence surrogate models predicting high-speed link transient responses under variable design parameters
SummaryThis paper proposes sequence-to-sequence neural surrogates for modeling transient responses and signal integrity of high-speed links. For fixed design parameters, feedforward networks match recurrent models at lower cost, while under 14 variable parameters, non-autoregressive TX-only inference remains stable. A recurrent Early Fusion LSTM, injecting design parameters at the initial encoding stage, reports R2 values of 99.51%, 99.57% and 99.94% for waveform, eye-height and eye-width prediction.
Julian With\"oft, Shayan Ghanbari, Werner John et al.
In briefEGGHead: Fast, generalizable Gaussian head modeling from one image
SummaryThis paper presents EGGHead, a method for creating photorealistic human heads from a single image. It introduces a Gaussian allocation algorithm that assigns a fixed budget of Gaussians to semantic regions of the head, reducing the number needed while limiting loss in representation power. The method runs in a single forward pass. The authors report that it outperforms previous single-image face modeling methods in reconstruction quality, novel view synthesis, and rendering speed.
In briefAudio-frontend design for detecting audio-visual deepfakes with low latency
SummaryThis paper studies how to design the audio front end of an audio-visual deepfake detection pipeline. It compares spectrogram-style log-filterbank features with frozen wav2vec 2.0 and WavLM representations, along with dual-SSL fusion, execution depth, and shallow-layer aggregation. On the generator-disjoint MMDF benchmark, full wav2vec 2.0 raises mean macro AUROC from 84.19% to 87.19%, and six-layer execution with learned layer aggregation reaches 90.61%. Shallow execution also reduces measured inference latency relative to full-depth wav2vec 2.0.
In briefContinual test-time learning via evolution strategies
SummaryThis paper proposes ESRL, a framework that helps large language models keep adapting in open-ended settings without external supervision. Instead of relying on the single policy being trained, it perturbs a fixed anchor policy into a population of nearby policies and aggregates their answers into label-free, group-relative learning signals for the student. The anchor is refreshed from the student at block boundaries. Across single-task adaptation in math reasoning, general knowledge, and medical domains, the authors report accuracy gains of up to 31.7% over the base model and 16.6% over the strongest baseline, with substantially reduced catastrophic forgetting.
In briefLooped transformers that reuse layers to solve multi-step tasks
SummaryThis paper studies looped transformers, which reuse one block of layers across many iterations to extend effective depth without adding parameters. The authors use a directed permutation traversal task where the model follows several hops from a start node and outputs the final node. They find that a looped model with the same effective depth solves the task reliably, while an untied model trained the same way does not, and that the looped model advances through intermediate nodes step by step.
In briefAI surrogate model speeds up national wildfire hazard mapping in Canada
SummaryThis paper presents a deep learning surrogate model that approximates outputs of process-based wildfire simulators such as BurnP3+, including burn probability, mean fire intensity, and mean rate of spread. Trained on national-scale data, it produces results in seconds rather than the days to weeks that regional simulations can take. On a representative region, the authors report 806-fold end-to-end CPU acceleration and show zero-shot transfer and lightweight fine-tuning.
WhyA multi-task deep surrogate replaces Monte Carlo wildfire simulation, validated via counterfactuals and zero-shot regional transfer on national data, with substantial impact for Canadian hazard mappin
Hager Radi Abdelwahed, Toluwaleke Olutayo, Chirag Goel et al.
In briefA conditional no-hair theorem: how transformers forget their prompt
SummaryThe paper studies how much an autoregressive transformer retains of its prompt after generating t tokens. Using different prompts followed by the same continuation, the distance between hidden states decays approximately as a power law in t. The authors prove an upper bound on this offset under stated conditions, attributing the forgetting to dilution of the prompt's direct influence as context grows.
In briefDoes team communication help LLMs find scientific laws? A study of symbolic regression teams.
SummaryThe paper studies teams of large language models searching for scientific laws via symbolic regression, under a fixed call budget. The teams differ in how visible peers are and when they update, and both frozen candidate archives and the final selected expressions are evaluated. Spreading the budget across more independent trajectories gave the largest measured gain, while under a shared selector the Ring policy scored 0.33 lower than Independent, and Star and Complete remained unresolved. The authors conclude that candidate generation and final selection should be reported separately, since communication structure alone does not determine an architecture ranking.
In briefThree-way count splitting to reduce double-dipping in single-cell topic models
SummaryThis paper addresses the problem that topic models and gene tests reuse the same single-cell RNA-seq counts. The authors encode topic usage on a separate third of counts and test associations on the remaining counts. On a 104,400-cell lupus cohort with permuted null genes, the false rejection rate at a nominal 5% level fell from 36% for naive testing to 7.7%.
WhyApplies three-way count splitting to single-cell topic models to curb double-dipping; permutation-null tests on a large cohort give solid evidence for a subfield.
In briefA data-driven method for combining resources to simplify automatic digital twin and simulation model generation
SummaryThis paper addresses the difficulty of manually updating simulation models for frequently reconfigured manufacturing systems. It proposes automatically selecting and aggregating resources from production event logs, using ideas from inverse reinforcement learning to learn each resource's behavioral value. Resources with similar functional contributions are grouped together. Experiments on throughput and system time show that this approach substantially reduces complexity while preserving performance accuracy.
In briefTesting how well generative drug-design models avoid off-target binding
SummaryThis paper notes that most structure-based drug design models optimize binding to a single target and ignore selectivity against other proteins. The authors introduce SELECTBENCH, a benchmark with three evaluation tasks and a set of selectivity metrics. They systematically evaluate nine representative models and find that even selectivity-trained models discriminate off-targets only modestly, and inference-time guidance gives limited gains.
WhyIntroduces a benchmark with three task types and systematically compares nine SBDD models on selectivity, offering useful, well-supported findings.
Ella Miray Rajaonson, Jungyoon Lee, William Chau et al.
In briefPrefix-tuning to produce arbitrary output sequences from pretrained Transformers
SummaryThis paper studies the theoretical expressiveness of prefix-tuning, a parameter-efficient method that prepends learned prefixes to inputs to steer language models. The authors prove that, if the target output lies in the range of the pretrained Transformer, a prefix exists that makes the model generate that output, and it can be computed by solving several linear systems. They also show the assumption is necessary, since such a prefix may not exist otherwise. Experiments on GPT-2 and Qwen3-4B show the resulting prefixes give strong control over model outputs.
In briefHAVANA: A Hierarchical Video-Action Model for Robot Navigation Guided by Language
SummaryThis paper presents HAVANA, a dual-system video-action model that separates predictive reasoning from reactive local planning for vision-language navigation. A slow system builds a compact language-conditioned latent plan from observation history, while a fast system uses that plan and the current observation to produce local trajectories. The two run asynchronously, and no future video is generated during inference. The authors report state-of-the-art results on OmniNavBench and robust generalization to unseen environments across different robot embodiments in real-world tests.
In briefDo Language Models Understand Variable Roles in Code?
SummaryThis paper uses Sajaniemi's taxonomy of variable roles, such as accumulation, traversal, and tracking the best candidate, to test whether large language models can recognize these roles and whether sparse autoencoders (SAEs) preserve them. The authors introduce annotation rules for eleven roles and a corpus of 126,926 code files across five languages. Residual probes outperform single SAE features and lexical baselines, but probes on SAE reconstructions are less sensitive to role-changing edits. Interventions during generation often alter role annotations at the cost of correctness.
C\'elian Vasson, Daniil Ognev, Kentaro Inui et al.
In briefGraph-regularized synthetic control for stable counterfactuals with redundant donors
SummarySynthetic control estimates what would have happened to a treated unit by weighting untreated donor units that match it before treatment. When many donors are highly correlated and pre-treatment data is short, many near-equivalent weightings exist and the donor weights become unstable. The paper proposes Graph-SC, a convex estimator that adds a graph Laplacian penalty so that similar donors receive coherent weights, and provides bias-variance analysis and convergence guarantees. Experiments on simulations, the California Proposition 99 benchmark, and Waymo Open Motion Dataset analyses report improved weight stability and reduced sensitivity to redundant donor pools.
In briefLooped GNNs beyond fixed depth: size generalization and cheap adaptation
SummaryThe paper studies how graph neural networks trained on small graphs can generalize to larger ones by repeatedly applying a learned local update beyond the training depth. It proposes Loop-Freeze-Steer (LFS), which freezes the pretrained processor and trains a small controller to steer its repeated use, with far fewer trainable parameters than full fine-tuning. Experiments on synthetic tasks and chip placement evaluate depth extension and the accuracy versus trainable-parameter trade-off.
In briefExplaining AI weather models by decoding their intermediate layers
SummaryAI weather models are often treated as black boxes, with little insight into how they turn input data into forecasts. This paper proposes intermediate decoding, which maps latent states from intermediate depths back into meteorological fields, and applies it to GraphCast, Aurora and ArchesWeather. The authors find that forecast skill improves gradually through the processor layers, that the encoder-decoder is not a simple identity mapping, and that all three models refine forecasts from coarse to finer spatial scales.
Matthias Beylich, Kirsten I. Tempest, Tobias Selz et al.
In briefBLANC: Adaptive Timing Scheduling for Generating Sign Language Motion
SummaryThis paper addresses generating sign language motion from written text. It introduces a Multi-Rate Vector Quantizer that represents the body and both hands as separate discrete streams at different temporal resolutions. Building on this, the BLANC framework generates variable-duration motion blocks causally, with a learned clock predicting block lengths and sequence endings. The authors report state-of-the-art results on PHOENIX-2014T and CSL-Daily.
In briefJoint, Decoupled, or Agent: Which Multimodal Alignment Method Works Best, and When
SummaryThis paper compares three ways to combine multiple data types: a jointly pretrained encoder with a fusion head (JOINT), independently pretrained encoders with a fusion head (DECOUPLED), and a frozen language model given instructions and examples (AGENT). Tested on 16 paired datasets across five domains and several label budgets, no single approach wins everywhere. The authors also propose a two-step selection method using representation and task diagnostics, which picks the winning approach more often than metadata alone.
In briefCache2LoRA: a method that improves how language models internalize document context
SummaryThis paper examines context-to-LoRA methods, which encode documents as low-rank parameter updates so models can answer without keeping the source text. It proposes Cache2LoRA, which aggregates evidence from a reusable frozen document cache for each query before generating LoRA updates. Across 17 tasks it reaches a macro F1 of 0.663, close to full-context in-context learning at 0.671, and decodes 5.9 times faster at 32K documents and 256-token outputs with batch size 16.
In briefA training-free method that prunes visual tokens in high-resolution vision-language models using query guidance
SummaryHigh-resolution vision-language models encode a global thumbnail and several local crops, which produces thousands of visual tokens that slow LLM inference. The paper proposes a training-free framework that uses the query before any LLM layers run, allocating and selecting tokens hierarchically at the instance, crop, and token levels. On LLaVA-NeXT-7B, it retains 98.0% of unpruned performance at 25% token retention with a 2.0x throughput speedup, and surpasses the unpruned baseline on POPE by 0.7 points.
In briefBiNar: A Restorer That Repairs Images Directly from Corrupted Compressed Data
SummaryThis paper tackles bitstream-corrupted image restoration, which aims to rebuild images from damaged compressed bitstreams. The authors propose BiNar, a generative framework that uses fine-grained semantics taken directly from the bitstream, combining bitstream semantic modeling, corruption-adaptive multimodal aggregation, and blind corruption localization. These outputs serve as language and visual prompts for a diffusion model. The authors report that BiNar outperforms state-of-the-art restoration models in both perceptual and objective quality across diverse corruption conditions.
In briefDoes AI-assisted coding make software more alike? A study of code similarity, quality, and growth
SummaryThis paper examines 50 long-lived Java and Python open-source projects across three periods: before LLMs, possibly LLM-assisted, and detected agentic development. It measures code similarity semantically and syntactically, and assesses quality with SonarQube metrics and evolution through growth rates and commit activity. The authors find no consistent increase in similarity; in Java, similarity decreased significantly in the later two eras, while Python showed no significant differences.
Luciano Marchezan, Romain Lefeuvre, Kevin Delcourt et al.
In briefMemory-equipped AI agent learns to draw Chinese characters
SummaryThis paper studies whether an LLM agent that stores memory from past attempts performs better on new programmatic Chinese-character drawing tasks, and whether the memory design matters. Five agent conditions were tested on 668 items. Free-form notes raised first-attempt success by 4.6 percentage points, and a design combining free-form memory with reusable code raised it by 2.8 points. The two designs did not differ significantly overall, though the code bank showed a suggestive edge on later, harder items. A post-hoc comparison found a much larger gain when the code-bank agent received target-specific structural supervision.
In briefHow well human reviewers catch face recognition false positives
SummaryThis study measured how 801 human reviewers detected false positives from four commercial face recognition algorithms across three security thresholds, using a new dataset called DFICT. Human review reduced algorithmic false positives, but performance declined at stricter thresholds and varied by algorithm. The authors suggest these factors should inform algorithm selection and human oversight strategies in operational deployments.
Esha Edakoth, John J. Howard, Isabelle M. Shuggi et al.
In briefFinitely Verifiable Segmentation: Decidable Checks for Tokenizers
SummaryThe paper studies tokenization policies that give each byte string at most one segmentation. It defines a policy as finitely verifiable when a finite automaton over boundary-marked strings accepts exactly that policy's segmentations. Under this assumption, questions such as whether a window certifies a token boundary, whether boundary gaps are bounded, and whether two policies are equivalent become decidable. The authors state these results with executable checks and report statistics produced by their own library.
WhyMakes six segmentation decisions decidable under a finite-state verifier, with polynomial-time inner frontier and closure results; strong theory with executable checks.
In briefOR-License: A Benchmark Testing AI Agents on Real Operations Research Work
SummaryThe paper introduces OR-License, a benchmark of 31 instances across 10 industrial scenarios drawn from paid operations research engagements, using original client materials as inputs. It also presents ORSim, a client simulator that reveals business facts without supplying OR expertise, and a Deployment-Adjusted Quality metric scored against expert solutions. Among 11 frontier LLMs, several exceed the expert reference on conditional objective quality, but none matches it on DAQ; the best configuration scores 0.87 against an expert reference of 1.
In briefCoopshift: Post-training makes LLMs cooperate less as they see more history
SummaryThe paper compares pretrained, supervised fine-tuned, and preference-optimized checkpoints from seven model lineages across seven social dilemmas over 200 repeated rounds. Visible history is limited to the last two or forty rounds. Cooperation declines as visible history grows, and this decline is larger at post-trained stages than at the pretrained stage. The authors argue that post-training evaluation should assess whether cooperation is sustained in repeated play.
In briefGLOW: A Ward-Clustering Method for Finding Brain Regions in Neuroimaging
SummaryThis paper introduces GLOW, a method for finding brain regions whose imaging features correlate with variables such as diagnosis or age. It builds a hierarchy of nested regions by greedily minimizing General Linear Model error, and uses maximum-statistic permutation testing to control family-wise error. A pruning step then seeks whole, pure effect extents. On synthetic effects planted in HCP diffusion images, GLOW often recovers a consistent effect as one region, but its positive predictive value is lower because its regions overrun the effect.
In briefLocation matters: where you ask changes which sources generative engines cite
SummaryThis paper audits how a user's location changes the sources cited by generative search engines. The researchers asked 300 questions to ChatGPT and Gemini from different declared and network locations, collecting 31,200 answers. They found that ChatGPT follows the declared location while Gemini follows the network's country, and the businesses recommended change along with the sources. The authors conclude that audits of generative search must control both location channels.
In briefOrTeN: Jointly reconstructing planetary terrain and satellite camera paths using neural inverse rendering
SummaryThis paper introduces OrTeN, a method that uses neural inverse rendering to jointly optimize a textured height map and orbital camera corrections from line-scan satellite images, producing digital terrain maps. It needs only calibrated line-scan images and initial camera poses, without depth measurements or initial rough DTMs. Evaluated on four planetary imagery datasets, it achieves aligned terrain RMSEs from 0.43 m to 16.18 m.
WhyJointly optimizes height map and line-scan camera trajectory via neural inverse rendering without depth priors, validated on four planetary datasets.
In briefWho Checks the Builder? Independent Assurance for AI-Generated Software
SummaryThe paper argues that when coding agents produce both software and its test evidence, their passing tests are not independent proof of correctness. It proposes separating building from verifying, with a Verification Contract and two verification phases, and leaves the release decision to a human. In a pre-registered study of three stateful services and an EvalPlus pilot, a verifier in a fresh context detected more of 46 seeded faults than the builder's own tests, at the cost of occasional false oracles.
In briefA learned post-hoc model that corrects geometry in generated videos
SummaryGenerated videos often lose geometric consistency under camera motion, with warping structures and inconsistent parallax. The paper attributes this largely to data scarcity and proposes Video Geometry Corrector (VGC), a post-hoc video-to-video model trained on synthetic paired data built from real videos with pose annotations. VGC maps distorted videos to geometry-consistent targets without modifying the source generator, using optical flow to separate camera motion from object motion. Experiments on pose-conditioned and prompt-driven benchmarks report improved geometric consistency across multiple generators.
In briefWorld of SWECraft: Scaling AI Agent Collaboration in Simulated Software Worlds
SummaryThis paper introduces Software World, a framework for simulating teams of AI agents collaborating on realistic ecosystems of interdependent Python packages, called software worlds. Agents maintain these packages with the goal of improving software efficiency, and their collective work is evaluated by speedup in hidden dependent packages. Across three worlds and eight models, totaling over 10,000 agent-hours, the authors find performance scales with inference-time compute, with the best model improving downstream efficiency by up to 31%. Some frontier models, however, stop searching for improvements prematurely and make little progress. The framework and traces are released publicly.<br>
Shannon Zejiang Shen, Yapei Chang, Qiyao Liang et al.
In briefRethinking Pixel Diffusion Transformer Design: MFlow and ConvRefiner
SummaryThis paper studies vanilla diffusion transformers that generate images directly in pixel space, comparing x0-prediction with velocity prediction. The authors propose MFlow, an adaptive parameterization that learns to interpolate between the two across principal data directions during end-to-end training, finding the preferred choice depends on data representation and model capacity. They also introduce ConvRefiner, a lightweight convolutional module trained post-hoc with an adversarial loss while the backbone stays frozen, which restores high-frequency image statistics while preserving FID.
Soumik Mukhopadhyay, Long Mai, Abhinav Shrivastava et al.
In briefBehavioral Couplings: How Fine-Tuning One Behavior in Language Models Triggers Another
SummaryThis paper studies "behavioral couplings," where fine-tuning a language model to show one behavior causes another, unrelated behavior to emerge. The authors find that couplings form when training documents describe two behaviors together, and they can also compose across document sets. This makes unintended misbehavior hard to trace to its source. The authors suggest that preventing such effects may require evaluations that look beyond a model's current behavior.
WhyIntroduces behavioral couplings formed via document training that compose, exposing unintended fine-tuning risks; experimental scope appears limited.
In briefAI Agents Testing Whether Medical Study Findings Hold Up Under Reanalysis
SummaryThis paper examines whether large language model agents can help automate the reproduction and robustness checking of research findings, using studies based on the SEER U.S. cancer registry database. The authors introduce SEER-Repro, a benchmark of 20 published papers, and an agent that reproduces 77.4% of 2,190 atomic claims. They also build an analysis pipeline that identifies analytical choices, tests alternatives, and flags 9.5% of fully reproduced findings as potentially consequential for expert review.
Shicheng Liu, Krista Opsahl-Ong, Andrew Drozdov et al.
In briefStopWatch: A Goal-Grounding Method for Object-Goal Navigation
SummaryThis paper studies object-goal navigation, where fine-tuned vision-language model policies often fail to recognize when the target object is already visible or to estimate its distance. The authors propose explicitly grounding the goal, which improves success and reaches state-of-the-art results on six benchmarks across three scene collections. They also correct annotation defects in these benchmarks to enable more faithful future evaluation.
In briefDiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
SummaryThis paper defines valuable, testable research hypotheses as evidence-based claims about consequential uncertainties, paired with feasible investigations, decision criteria, and alternative explanations. It presents DiscoveryRSI, a framework that builds hypothesis trees through multi-agent evolutionary search and learns recursively via three mechanisms, and introduces DFM-Bench to assess hypothesis quality and research progress. Automated evaluation across four life-science domains reports hypothesis quality rising from 4.86 to 5.39 after two rounds of self-improvement.
In briefGenerative Models and Language-Guided Attention for Recovering Text from Palimpsests
SummaryThis paper proposes an automated method for recovering erased under-text in palimpsests, manuscripts that were overwritten for reuse. The authors compare generative adversarial networks and diffusion models, and design a custom denoising diffusion model that learns structured noise patterns to reconstruct the under-text. The recovered text is then read by a fine-tuned attention-based handwritten text recognition network, which is further guided by text descriptions grounded through CLIP attention maps. The authors also release synthetic palimpsest datasets, text descriptions, and pre-trained models with code.
Robin Hollifeldt, Tianru Zhang, Jessica Str\"om et al.
In briefBeyond the Answer: Student-Conditioned Reasoning Traces for Distillation
SummaryThis paper studies how to build reasoning traces that help a specific student learn more than answer-only supervision. It introduces ReDiST, which revises the student's correct drafts using supplied evidence and repairs incorrect drafts using the verified answer, then uses a frozen student-matched reader to screen and rank candidate traces before distillation. Experiments on question-answering benchmarks report gains over answer-only and original-trace baselines on both source and transfer tasks. The authors also note that rankings of methods can differ between source and transfer benchmarks.
In briefWaiting Buys Only Two Moments: Identifying Timing in Prediction-Driven Allocation
SummaryThe paper studies how a planner with a fixed budget for preventive help should decide whom to help and when, since waiting reveals risk but lets higher-risk people leave first. In the canonical model, each additional observed period identifies two new latent risk moments, which determine average attrition among survivors but not necessarily allocation welfare. Experiments show that longer observation can resolve timing ambiguity, yet the optimal date may already have passed.
Rama AlHamidi, Rasul Khanbayov, Erchin Serpedin et al.
In briefSelf-Building Labs: A System That Designs and Runs Its Own Lab Setup from a Science Goal
SummaryThis paper introduces the self-building laboratory (SBL), a system that designs, configures, and operates its own lab setup given a high-level scientific goal. Starting from whatever exists, even an empty room, it builds the missing layout, digital twin, robot skills, and experimental workflows. The authors show it working across different robots and setups, and validate it end to end in a real quantum-dot synthesis experiment.
In briefENFR: Fast, Amortized Estimation of Absolute Model Evidence Using Flow-Matched References
SummaryThe paper proposes ENFR, a method for quickly and accurately estimating the absolute evidence of a data-generating model, which can be used for model inference and goodness-of-fit tests. It first learns Bayes factors relative to a flow-matched reference and then recovers the absolute evidence from them. The authors report improved estimation over a flow-based baseline across all applications considered, and they introduce an architecture that jointly learns the evidences of multiple models.
Kai Lehman, Jameson Dong, Carolina Cuesta-Lazaro et al.
In briefA neural network that builds in rotational symmetry on the sphere, so it need not learn it from data.
SummaryThe paper presents a spherical convolutional network, equivariant under SO(3) rotations, built for the HEALPix sphere pixelization. It uses steerable filters, parallel transport and spin-weighted features, and the authors state the convolutions are provably exactly equivariant given a band-limit. Reported benchmarks include dense gradient regression and cosmic microwave background foreground removal and denoising. In cosmology lensing simulation-based inference, the authors report the tightest posteriors at a matched simulation count.
WhyProposes a provably SO(3)-equivariant HEALPix CNN, validated on regression, denoising, and cosmology SBI benchmarks, with broad scientific impact.
In briefCocoRNA: Teaching Cooperative AI Agents to Design RNA Sequences That Fold Into Target Shapes
SummaryThis paper presents CocoRNA, a framework that uses cooperative multi-agent reinforcement learning to solve RNA inverse design, finding sequences that fold into a specified secondary structure. It splits the task into smaller sub-problems solved jointly by several agents, guided during training by a centralized critic that uses global structural information. According to the authors, on the Rfam dataset CocoRNA substantially outperforms existing methods in success rate and design speed, and generalizes to unseen structures without extra training.
In briefAcceptance Is Not Speed: Length Inflation in Lossy Speculative Decoding
SummaryThis paper examines relaxed verification in speculative decoding, a method for speeding up text generation with a smaller draft model. Relaxed rules accept more tokens per round, but the authors find they also make completions longer, mainly by extending reasoning traces, which largely offsets the apparent speed gains. Across twelve target-drafter pairs, they conclude that accepted tokens per round alone is an unreliable speed measure and that completion length and verifier rounds should also be reported.
In briefLearning personalized memory gating from delayed one-sided feedback
SummaryThe paper studies how language-model agents should decide which extracted facts to keep in a limited long-term memory. Each fact is labeled by whether it is later retrieved and used in responses, and online gating is framed as contextual apple tasting with delayed one-sided feedback. The authors propose Warm-Start UCB, which combines offline learning from existing data with online adaptation to each user, and give a high-probability regret bound. Experiments reportedly show it outperforms all baselines across different memory budgets and history lengths.
In briefExplainable Machine Learning on Temperature Effects on Fiber-Reinforced High-Performance Concrete Strength
SummaryThis paper uses 1,204 experimental records to compare eight machine-learning models for predicting the compressive strength of fiber-reinforced high-performance concrete after temperature exposure. Gradient boosting and extreme gradient boosting both reach an R² of 0.903 on unseen data. Shapley and partial dependence analyses identify temperature as the primary factor, with strength decreasing significantly as temperature rises, though the degree depends on the mix design.
Jagaran Chakma, Zhiguang Zhou, Badhan Chakma et al.
In briefSteering Beyond Language: Improving Multilingual Training-Free Text Embeddings
SummaryThis paper presents the first systematic study of multilingual training-free text embeddings derived from large language models. The authors find that semantically equivalent sentences remain misaligned across languages, and propose Paired-Residual Subspace Removal (PRSR), an inference-time method that removes language-specific low-rank components from hidden states. Across more than 200 monolingual evaluations in 18 languages and cross-lingual bitext mining over 200 languages, PRSR consistently improves existing training-free baselines.
In briefBeyond SAEs: Consistent Concept Discovery with Sparse Coding
SummarySparse autoencoders explain foundation model representations by breaking activations into sparse combinations of learned dictionary atoms, but dictionaries can differ across training runs. This paper proposes Consistent Sparse Dictionary Learning (ConSDL), which replaces the linear encoder with per-sample sparse coding. ConSDL achieves near-perfect dictionary and coefficient recovery across 16 synthetic settings. In four vision and language settings, it improves dictionary and coefficient consistency by 17.6% and 35.9% over six SAE baselines on average.
In briefFuTransformer: FP8 prefill and INT4 decode for Llama-style inference on NVIDIA Hopper GPUs
SummaryThis paper presents FuTransformer, a CUDA implementation of a Llama-style Transformer block for NVIDIA Hopper GPUs. It uses FP8 for prefill and INT4 weights and KV-cache for single-token decode to speed up inference. On one H100, prefill is up to 1.48x faster than compiled PyTorch FP8, and decode is 1.3 to 2.4x faster than the faster of PyTorch's BF16 and INT4 blocks across tested batch sizes. On WikiText-2, the INT4 weight path raises perplexity from 6.24 to 7.23, or to 6.99 with GPTQ.
In briefPPMem: A Video World Model Using Planar Primitives as Compact Persistent 3D Memory
SummaryThis paper introduces PPMem, a video world model that stores scene structure as a compact set of oriented planes, such as walls and floors, rather than per-pixel point clouds. It adds a planar prediction head and a Depth Token Fusion module to a video diffusion transformer so each generated chunk produces both video and plane geometry that conditions later chunks. According to the paper, it outperforms prior 3D-grounded world models in video quality on DL3DV and RealEstate10K while outputting metric depth, normals, and exportable planar geometry.
In briefDrugHarness: An Evolvable Agent Framework for Computational Drug Discovery
SummaryThis paper presents DrugHarness, an evolvable agent framework that integrates over 80 scientific operations and biomedical resources. It induces skills with applicability conditions from execution trajectories and outcome feedback, which an independent review agent verifies before reuse. Evaluations on MolBench and SMDD-Bench report gains in molecular screening and optimization. In a prospective HPK1 study, a synthesized compound reached a cellular IC50 of 14.19 nM, a 2.64-fold improvement over the best input reference.
WhyProposes an evolvable drug-discovery agent with conditional skill induction and independent review, validated on benchmarks and a prospective HPK1 study; impact is mainly within computational drug des
In briefBuilding a Large Video Dataset for Automated Evaluation of Central Line Dressing Change Skills
SummaryThis paper describes how a large-scale video dataset was collected to support automated skill evaluation of central line dressing change procedures. It contains 365 recordings of registered nurses from the 2026 annual competency exam at Houston Methodist Hospital, each paired with temporal annotations and performance assessments across 17 subtasks. Benchmark experiments with spatiotemporal deep learning models show that video-based skill assessment is feasible.
Ummey Tanin, Shannan K. Hamlin, Hsin-Mei Chen et al.
In briefParametric Resonance in Primal-Dual Learning for Constrained Decision Processes
SummaryThis paper analyzes primal-dual methods for constrained Markov decision processes under a two-timescale assumption, distinguishing additive and parametric channels of non-stationarity. It shows that a mismatch between the constraint estimate used by the primal update and the measurement used by the dual update can produce instability through parametric resonance. The results hold on a tangent linear-quadratic reduction around an active saddle, for one or several constraints. The authors also report experiments with a Lagrangian soft actor-critic agent.
In briefDLite: A History-Free Speculative Decoding Method for Agentic AI
SummaryThis paper addresses speculative decoding for agentic AI, where contexts grow long and multi-turn interactions are frequent. It introduces DLite, which reuses contextual hidden states from the target model and makes local predictions without maintaining historical context, plus a lightweight sequential head to restore dependencies within blocks. Experiments on general, long-context, and agentic workloads report acceptance lengths of 2 to 6, and up to 70% improvement over the state-of-the-art method DSpark.
In briefSampling the XY model with an attention-based neural network
SummaryThis paper presents a transformer-based variational autoregressive network for sampling lattice systems with continuous degrees of freedom. It is applied to the two-dimensional XY model near the Berezinskii-Kosterlitz-Thouless transition. The model adds physics-informed elements such as locality-informed attention and approximate probabilities, and significantly outperforms existing neural samplers. It reproduces energy, magnetization and helicity modulus from traditional Markov Chain Monte Carlo.
Piotr Bia{\l}as, Christian D. Morales, Tomasz Stebel et al.
In briefMesoscopic description of learning dynamics in wide neural networks
SummaryThe paper studies how two-layer neural networks, with width and input dimension growing proportionally, learn internal features of a target network via stochastic gradient descent. It describes the learning through the distribution of overlaps between learned and target features, showing that early learning reduces to a few collective variables tracking linear and quadratic components. It also analyzes when individual features specialize and how a quadratic shielding mechanism accelerates this specialization.]
WhyProposes a mesoscopic Gaussian-closure description reducing wide-network learning to few collective variables, analyzing specialization and quadratic shielding.
In briefVideo Diffusion Avatars: Turning Video Diffusion Models into Controllable 3D Avatars
SummaryThe paper presents Video Diffusion Avatars (VDAs), which fine-tune a video diffusion model to create photo-realistic, multi-view consistent avatars from a single image. Users can control camera viewpoint, facial expression, and upper-body pose, using a learned expression space instead of 3D morphable models. The authors identify viewpoint bias in learned expression codes and show how disentanglement removes it. On BecomingLit, the reported results show a 29% reduction in face LPIPS and a 66% improvement in JOD over prior single-image 3D avatar methods.
WhyFine-tunes video diffusion for controllable single-image avatars, removes expression viewpoint bias, and substantially beats prior methods on BecomingLit.
Simon Giebenhain, Tobias Kirschstein, Liam Schoneveld et al.
In briefLLM-Assisted Bi-Approximate CHC Solving via Certified Derivation Under-approximations
SummaryConstrained Horn clause solvers often struggle to find deep counterexamples needing long rule unfolding. The paper introduces Derivation Power Under-approximations (DPUs), solver-certified under-approximations of multi-step CHC derivations, proposed by an LLM with symbolic intermediate-state witnesses for SMT validation. Its framework HornCraft combines invariant generation with DPU generation and returns certified verdicts for 228 of 260 unresolved integer-arithmetic CHC-COMP benchmarks. It certifies 42 instances as unsafe, versus 22 without DPU construction and reuse.
WhyProposes solver-validated DPUs for under-approximating CHC derivations with LLMs; certifies 228 of 260 unresolved benchmarks, raising certified unsafe verdicts from 22 to 42.
