In briefReframe: Observation-aligned multi-teacher on-policy distillation for visual agents using tools
SummaryThis paper presents Reframe, a post-training framework that aligns the observation state, not only the output distribution, for visual agents that use tools. GD-GRPO trains spatial and temporal specialists with evidence-aware rewards and stochastic removal of the original visual input, while OG-MOPD replaces weak tool actions mid-rollout and resumes distillation from the repaired observation. The authors report that Reframe outperforms Mix-RL and vanilla MOPD across 14 image and video benchmarks.
WhyThe authors propose WICL, learning manipulation in context via simulated rehearsal, reporting 83.4% mean success on 50 RoboTwin 2.0 tasks from one demonstration, over 20 points above trained policies.
In briefGuiding data allocation to improve performance on underrepresented subgroups in machine learning
SummaryThis paper studies how machine learning models can stay accurate on data subgroups whose distribution differs from the training data. The authors extend a Gaussian feature-model framework to theoretically analyze performance across many possible training subgroup proportions. They show that several configurations, including substantially imbalanced ones, can perform about as well as a balanced configuration. They also derive gradient-based directions for adjusting subgroup proportions to improve adjusted and worst-group accuracy.
In briefControlling which modality a multimodal model trusts when inputs conflict
SummaryMultimodal large language models can become unreliable when visual and textual inputs disagree, sometimes favoring a modality that is not the most reliable for the task. The paper identifies a small set of attention heads, called arbitration heads, that causally influence whether predictions follow visual or textual evidence. It introduces a lightweight, training-free intervention that biases these heads toward a designated trusted modality at inference time. The authors report that the heads are stable across calibration set sizes and that steering them enables bidirectional control across diverse benchmarks.
In briefKC-Maps: a trajectory-free neural method for simulating molecular dynamics
SummaryThis paper introduces KC-Maps, a neural network method that learns the transition kernel of a stochastic differential equation. It needs only trajectory-free marginal samples together with the known drift and diffusion coefficients. Experiments on peptide molecular systems show it outperforms classical molecular dynamics in GPU-hours.
Majdi Hassan, Marta Skreta, Dominique Beaini et al.
In briefPredicting Future Hand-Object Motion in 3D from Internet Videos
SummaryThis paper presents Grasping-the-Future, a unified generative model of joint 3D hand-object motion covering wrist movement, finger articulation, and object position and orientation. Trained with masked spatiotemporal completion and flow matching, one model handles future forecasting, mutual prediction of hand and object motion, trajectory generation toward a target object pose, and completion of partially observed interactions. The authors also build a pipeline that reconstructs 3D supervision from internet videos and report stronger results than released forecasters on datasets with 3D annotations.
Rustin Soraki, Matthew Wallingford, Ryan Aby et al.
In briefAdaptive Quantization for TD Learning over Rate-Limited Channels
SummaryThis paper studies policy evaluation with linear function approximation when an agent sends quantized temporal-difference learning updates to a server over a finite-capacity channel. The authors propose BAQ-TD, which combines batching to reduce transmitted variance with adaptive quantization to encode innovations. They prove a high-probability finite-time error bound showing that, above a certain channel capacity, BAQ-TD matches unquantized TD's convergence rate up to logarithmic factors. The method needs O(d log T) bits over horizon T to reach the order-optimal rate.
WhyThe authors propose BAQ-TD, using batching and adaptive quantization over a finite-capacity channel, with a high-probability finite-time bound and an exponential communication-complexity improvement.
In briefClick3R: Interactive 3D Reconstruction Corrected by Sparse Point Clicks
SummaryThe paper presents Click3R, a framework that lets users fix errors in feed-forward 3D reconstruction models by clicking a few corresponding points across an image pair. A lightweight adapter injects these correspondences as geometric constraints into the network. The authors also build a challenging benchmark that exposes failure modes of existing methods. Experiments report lower reconstruction error with as few as one click, on both standard and the new benchmark.
In briefHiVe-Mem: a memory framework for multimodal agents that expands related memories and selects evidence per query
SummaryThis paper presents HiVe-Mem, a multimodal memory framework for agents with long interaction histories. It organizes memories into a relational graph linked by textual and visual anchors, expanding retrieved seeds with related neighbors, and keeps several evidence forms per memory, such as summaries, captions, visual views, raw images, and interaction history. A lightweight trainable selector chooses an evidence subset for each retrieved memory based on the query. Across three multimodal memory benchmarks, the authors report consistent performance gains over base models and about a 90% reduction in token consumption compared with the most efficient multimodal baselines.
In briefVendi Anomaly Score: a diversity-based method for spotting unusual data points
SummaryThis paper proposes the Vendi Anomaly Score (VAS), which flags an anomaly by measuring how much removing a sample changes the dataset's diversity as measured by the Vendi score. VAS does not rely on density estimation or partitioning heuristics, requires no hyperparameter tuning, and adapts automatically to the data's spectral structure. It is non-parametric and scales linearly with dataset size. Across 148 benchmark embedding datasets, VAS achieves state-of-the-art anomaly detection performance, and it remains robust on large-scale ImageNet experiments as contamination rate and dataset size increase.
In briefSWAM: Scaling World Action Models with Cross-Modal Alignment
SummaryWorld action models jointly predict video and actions conditioned on language. The paper studies how to keep language-video alignment intact while learning actions, using about 47,000 hours of general, egocentric, UMI, and robot demonstration videos. Trained only on RoboTwin Clean demonstrations, the model reaches 88.10% and 61.56% success under Easy and Hard conditions, and 40.00% across three real-world tasks versus 23.33% for the strongest baseline. The authors release the curated data, training code, and model weights.
In briefWhere Alignment Emerges in Neural Networks: Revisiting the Platonic Representation Hypothesis
SummaryThe Platonic Representation Hypothesis proposes that neural networks are converging toward similar representation spaces. This paper examines where alignment appears between unimodal models across layers and representation dimensions. It finds that the trend of greater alignment with greater model capability persists in newer models when captions are clean and informative. The authors also introduce unimodal and multimodal baselines on various datasets, showing that unimodal models can exceed explicitly aligned vision-language models in cross-modal alignment yet remain well behind same-modality alignment.
Julia Herbst, Simon Roschmann, Quentin Bouniot et al.
In briefIs Persona Generalization Predictable? A paper on how fine-tuning a language model in one domain shifts its behavior in others.
SummaryThe paper studies persona generalization, where fine-tuning a language model in one domain changes its behavior in other domains, such as the broadly unsafe behavior seen in emergent misalignment. The authors find that in-domain and out-of-domain persona rates follow an approximately probit-linear relationship across fine-tuning hyperparameters, and that this relationship is roughly shared across model families and sizes, including Qwen and Llama. Changing which layers are fine-tuned alters the slope; for unsafe behavior, fine-tuning only later layers yields a shallower slope.
WhyThe authors report a probit-linear in-domain/out-of-domain relation across hyperparameters and model families, unifying emergent misalignment analyses.
Olina Mukherjee, Arnav Goel, Aditi Raghunathan et al.
In briefWorkflowArena: A Benchmark for Workflow Orchestration in LLM Multi-Agent Systems
SummaryThis paper introduces WorkflowArena, a benchmark of 145 long-horizon tasks across 8 professional domains, designed to test how well language model agents orchestrate their own worker agents through executable workflows. Deliverables are scored by a hidden verifier, and all delegation is recorded so that output quality and the organization of work can be evaluated together. Experiments with nine models find that no single configuration leads in every domain and that systems use different delegation strategies. When the executing model is fixed, workflows written by a stronger model substantially improve weaker executors.
In briefAI analysis of lymphoid structures in lung squamous cell carcinoma after neoadjuvant therapy
SummaryThis study used AI to analyze H&E whole-slide images of resected lung squamous cell carcinoma tumors and paired lymph nodes from 20 neoadjuvant-treated and 20 therapy-naive patients. Treated tumors showed lower stromal plasma cell and lymphocyte densities, plus fewer mature tertiary lymphoid structures and mature lymph node follicles. In an independent cohort of 458 patients, stromal plasma cell density and lymphocyte-plasma cell co-localization were associated with longer survival, suggesting a candidate prognostic biomarker.
Arne Schmidt, Justin Nieder, Cornelius Boehm et al.
In briefWhat blocks v-prediction? A patch-wise PCA study of pixel-space flow matching
SummaryThis paper examines why the standard v-prediction objective struggles in pixel-space flow matching, using patch-wise principal component analysis. It identifies a bandwidth bottleneck, as velocity in minor components is mostly determined by the raw noisy input. The authors propose the Raw-injected Transformer (RiT), which uses a lightweight raw-conditioned decoder and an output-side shortcut, and reports competitive results on ImageNet 256×256.
In briefReducing deception by AI agents using evaluation-side partial observability games
SummaryThe paper argues that language model agents can deceive humans to earn higher rewards while appearing helpful, because evaluations mostly check only final outcomes. The authors model deception mitigation as an Evaluation-Side Partially Observable Extensive Form Game (e-PoG), in which a monitor exposes deception from the full trajectory and a verifier checks the monitor's claims. They define a Subgame-Perfect Oversight Equilibrium and prove it exists in finite games, and experiments with Gemma3 and Qwen3 models show gains over the strongest baselines.
In briefProgramming Prescience: How Humans Build and Test AI Forecasting Bots
SummaryThis paper reports a qualitative interview study with 15 builders of forecasting bots, selected from successful participants in a recent AI forecasting competition. It identifies challenges in tuning bots to gather factual research, implementing coherent reasoning, and judging how well a bot performs. The authors also find that builders often added complexity that did not improve results and tested their bots in opportunistic, piecemeal ways.
In briefWhat the loss curve misses: 4-bit training learns rare facts more slowly
SummaryThis paper asks whether low-precision pretraining that matches validation loss also learns rare facts as efficiently. The authors train transformers from scratch with emulated 4-bit matrix multiplications on synthetic biographies with controlled exposure counts. NVFP4 needs 1.32 to 1.52 times as many exposures as BF16 to reach the same recall, even though validation loss differs by only 0.5% to 2%. The authors suggest measuring rare-fact acquisition alongside loss when evaluating low-precision recipes.
In briefMaxIM: Reinforcement learning for maximally informative incremental summaries
SummaryThe paper introduces MaxIM, a framework that trains AI systems to keep a bounded, incrementally updated summary of long interaction histories. It frames summarization as an information-theoretic reward-shaping problem, maximizing the predictive log-likelihood of future task-relevant targets and using an asymmetric critic to compute dense step-wise mutual information bounds. The authors report that on long-horizon scientific reading and multi-turn dialogue benchmarks, MaxIM outperforms RL, supervised fine-tuning, and prompt-based baselines without requiring human preference labels.
In briefHow Well Does AI Peer Review Work? A Test Built on 100 Planted Errors
SummaryResearchers inserted 100 known errors into 10 open-access psychology papers and had frontier models and two commercial AI review tools examine them. The best single system caught 71 errors and the worst caught 30, while pooling all outputs found 93. The study finds that combining systems is a major way to find more issues, and seven errors were missed by every system, all of them omissions of deleted information. The papers, errors, outputs, and experiment log are publicly available.
In briefLUCQUE: A Quebec lung cancer screening dataset for reliable AI evaluation
SummaryThis paper presents LUCQUE, a dataset built from low-dose CT lung cancer screening in Quebec, containing 1,678 baseline studies with expert-annotated pulmonary nodules and cysts. It is hosted on the PARADIM platform and accessed through a model-to-data approach so that data remains on premise. The authors state it is the largest accessible annotated real-world lung cancer screening dataset, intended to support reproducible and independent AI evaluation.
Johanna Topalis, Yannick Lemar\'echal, Gabriel Couture et al.
In briefComparing NLP and generative AI for extracting disassembly instructions
SummaryThis paper addresses the difficulty SMEs face in adopting disassembly automation as products grow more complex. It introduces and compares two frameworks that extract disassembly instructions from unstructured sources such as manuals, converting natural language text into a structured, machine-readable format. A rule-based NLP framework and a generative AI framework using LLMs were tested and compared on real-world disassembly manuals. The work provides a data foundation for developing and automating future disassembly processes.
Tim Schnieders, Tareq Maximilian Gabriel Ahrari, Oliver Petrovic et al.
In briefCounterfactual image inference without inversion: testing whether the latent-state inversion step is needed
SummaryThis paper studies counterfactual inference for high-dimensional variables such as images, which answers "what if" questions given observational data, modeling assumptions, and a causal graph. It examines whether the conventional step of inverting the factual image to recover its latent state is necessary, and evaluates an inversion-free method adapted from image-editing research. Experiments on the CelebA and ADNI image datasets measure effectiveness and identity preservation, and the paper presents the tradeoffs between inversion-based and inversion-free approaches.
In briefCurbside Chatbots: Auditing Language Model Mental Health Recommendations for Different Users
SummaryThis study audits 41 language models to see which mental health interventions they recommend when the same clinical case is presented by a clinician, a caregiver, or the person themselves. Across 19 case vignettes and five questions, the authors collected 35,055 free-text responses and had 29 models rate the full option lists on a seven-point scale. Models generally recommend options they rate highly, and their responses differ by the asker, with many giving clinicians consultation-style answers and lay users hedged, simplified advice.
Matthew Flathers, Jill Noorily, Phuong Anh Nguyen et al.
In brief2D Gum-Net: Unsupervised Elastic Image Alignment via Symmetric Correlation and Spectral Pooling
SummaryThis paper presents 2D Gum-Net, an end-to-end network that predicts dense elastic deformation fields to align one image with another without requiring labeled ground truth. It adapts a 3D subtomogram alignment method from cryo-electron tomography to 2D, using DCT-based spectral pooling to preserve fine structures. Trained only on synthetic fingerprints, it is tested zero-shot on real benchmarks, improving structural correlation by up to 15.7% on FVC2004.
In briefAgentic-OPD: Recovering lost agent skills by distilling from discarded training checkpoints
SummaryThe paper observes that reinforcement learning for LLM agents usually keeps only the best-validation checkpoint, while earlier checkpoints from the same run can solve problems it fails. The authors propose Agentic-OPD, which treats the best checkpoint as a student and all saved checkpoints as frozen teachers. Teachers are chosen by how well they cover the student's unsolved problems, and distillation is limited to failed student attempts where the teacher performs better. Across five agentic task domains and 24 benchmarks, it improves on its initialization checkpoint by up to 7.4 points.
In briefDual-teacher alignment for zero-shot EEG-to-image retrieval
SummaryThis paper proposes DTHA, a framework that decodes visual information from EEG by aligning brain signals with two complementary pretrained models: CLIP and DINOv2. A shared EEG encoder is trained with dual-branch contrastive alignment, and a fused-retrieval objective learns adaptive weights to combine the two teachers' representations. The encoder, AMSPBE, uses temporo-spatial attention and a learnable scalp-pattern bank. On THINGS-EEG zero-shot retrieval, DTHA reaches 92.4% Top-1 accuracy, 12.4 percentage points above the prior state of the art.
In briefSmartphone Depth-Aware Light Imaging to Detect Woody Breast Myopathy in Chicken Fillets
SummaryThis study presents a smartphone-based framework that combines transillumination images and 3D point clouds to non-destructively grade Woody Breast myopathy in chicken breast fillets. It analyzes 267 fillets using image, geometric, and one physics-based light attenuation feature. A multimodal ensemble model reached about 90% accuracy, compared with about 82% for models using 2D features alone.
Chaitanya Pallerla, Ramesh Bahadur Bist, Yihong Feng et al.
In briefUsing Large Language Models as a General Search Tool for Automated Machine Learning Across Open Tasks
SummaryThis paper proposes a large language model-based AutoML framework that jointly optimizes model architecture, training hyperparameters, and data ratios without task-specific fine-tuning. It introduces NCode, a unified numerical encoding that turns these design choices into discrete sequences, framing optimization as a ranking task. Trained only on trajectories from a single vision task, the searcher reaches state-of-the-art results on image classification, segmentation, speaker verification, and text-to-image benchmarks, and reports a GenEval score of 0.91 for a billion-parameter Diffusion Transformer using 896 GPU-days.
In briefFreezing Behavior, Not Weights: Rehearsal-Free Continual Learning With One Shared Prompt
SummaryThis paper proposes TEA-Prompt, a continual learning method that adapts a frozen vision transformer using a single shared prompt token set, without storing old images. It combines an input-conditional token relevance module with an attention-behavior anchoring loss that constrains attention outputs around class-specific query regions. The method reports state-of-the-art accuracy and reduced forgetting on CIFAR-100, ImageNet-R, and CUB-200 using only 0.38M parameters.
Duy-Cuong Bui, Trung-Anh Dang, Vincent Nguyen et al.
In briefLogCDM: An Interpretable Log Anomaly Detector Using Graph Drift and Adaptive Fusion
SummaryThis paper presents LogCDM, a compact detector that scores each log session using three complementary signals: next-event surprise from a small causal Transformer, graph-attention embedding distance from a normal reference transition graph, and a transition-drift statistic shown to be a frequency-weighted chi-squared divergence. The model has 8.1 to 8.8 million parameters and needs no GPU at inference, reaching F1 scores of 0.984 on HDFS, 0.870 on BGL, and 0.850 on Thunderbird. It scores a session in 5.5 to 10.5 milliseconds on a CPU.
In briefPaper showing learning rate decay can exponentially speed up SGD for global nonconvex optimization
SummaryThe paper examines whether learning rate decay in stochastic gradient descent (SGD) gives an asymptotic advantage for global optimization of nonconvex functions. The authors identify a natural class of nonconvex functions, including many popular benchmarks, where SGD with decay needs exponentially fewer gradient queries than SGD with any fixed learning rate in the small-step regime. They also show that annealing solves the associated non-log-concave sampling problem in polynomial time, using weak Poincaré inequalities.
WhyThe authors prove exponential speedup for SGD with learning rate decay on nonconvex global optimization, analyzed via weak Poincaré inequalities.
Guneykan Ozgul, Dylan Herman, Junhyung Lyle Kim et al.
In briefPORT: Robust token-level reweighting for preference optimization
SummaryThis paper proposes PORT, a fine-grained method for aligning language models under noisy preference labels. It reweights individual tokens using the empirical CDF of their log-probabilities, suppressing corrupted tokens with only a forward pass as a proxy for gradient-norm penalization. The authors show theoretically that PORT reduces gradient bias from noisy labels by minimizing an upper bound on that bias. Experiments across diverse noise settings report consistent gains over sequence-level baselines.
In briefSage-Taiyi: A Metacognitive System for Guiding Scientific Discovery
SummaryThis paper presents Sage-Taiyi, an AI architecture inspired by metacognition that uses accumulated scientific understanding to decide what to investigate next and how. It monitors and revises its assessments of evidence, assumptions, and open questions. On the Taiyi-Bench ideation benchmark, it ranks first on all eight metrics, scoring 54.0 overall versus 47.9 for the strongest baseline.
In briefA review moving from Granger causality to causal structure in dynamical systems
SummaryThis article examines Granger causality, a widely used method for detecting predictive relations between measured signals. It first reviews recent advances in analyzing these relations in observed data. It then turns to directed structure in latent dynamical models and asks under which conditions that structure can be interpreted causally.
In briefThe Narrative Gap: Can LLMs Help Users Explore Varied Accounts Across Languages?
SummaryThis paper examines how large language models present competing accounts of contested topics across languages and regions. The authors define six capabilities needed for diverse multilingual information seeking and introduce NarrativeBench, with 1,333 news articles in 23 languages and 460 queries. Evaluating 14 LLMs, they find current models lack the capabilities needed to reliably navigate diverse narratives.
Nikhil Sharma, Kelly Marchisio, Kenton Murray et al.
In briefClique Constraints for Exact Causal Structure Learning
SummaryThis paper adds an upper bound on the maximum clique size of the learned undirected skeleton to exact score-based causal structure learning, enforced through lazy cycle and clique constraints at integer candidate solutions. Experiments on linear-Gaussian Erdős-Rényi and scale-free DAGs compare callback order, clique-cut policies, incumbent repair, fractional separation, and Big-M bounds. The results show that clique constraints do not uniformly speed up optimisation; their benefit depends on graph density, sample size, and cut-management policy.
Xiaoyu He, Pavel Ryt\'i\v{r}, Petr Ry\v{s}av\'y et al.
In briefPlant Orbiter: Generating 3D Plants from One Image Using Video Diffusion and Gaussian Splatting
SummaryThis paper presents Plant Orbiter, a method that uses video diffusion to create high-quality 3D Gaussian Splatting models of plants from a single image. The authors report that it recreates plant morphology more accurately than other state-of-the-art methods, achieving the highest 3D F-score, while being easier to align and cheaper in camera placement for high-throughput pipelines. Combined with a 2D diffusion model, it also generates 3D plant models from text prompts with state-of-the-art results.
Zane K. J. Hartley, Yong En Kok, Lewis A. G. Stuart et al.
In briefInference validation for regulated AI, built on NVIDIA AIPerf
SummaryThe paper presents Runway, an internal validation tool built at MasterControl around NVIDIA AIPerf, which adds experiment design before measurement and compliance, cost, and audit decisions after it. It reports three findings from Runway sweeps: TensorRT-LLM and vLLM rankings reverse as concurrency changes, a newer and pricier GPU instance can cost less per request, and speculative decoding behaves differently across engines under load, so engine configuration can matter more than engine choice.
In briefΣ-LeWM: Learning robot control world models using S-information
SummaryThis paper adds a backward dynamics predictor and an inverse dynamics action predictor as latent losses to LeWorldModel. These extra losses impose further structure on the latent embedding. The authors report gains in performance and planning efficiency, which matter for real-time model-based control of robots. Results are shown on standard world model benchmarks and two simulated robot manipulation tasks.
In briefKiteNorm: Variance Regularisation for Stable, Scalable Post-LN Transformers
SummaryThis paper proposes KiteNorm, a post-norm training method for Transformers that constrains hidden-state variance to improve training stability. Experiments on decoder-only models up to 1.3B parameters show it outperforms leading baselines at their optimal learning rates, with scaling laws favouring it across depth, width, batch size, and training budget. Ablation studies indicate that each component is necessary for the full stability and performance gains.
Leon A. Trochelmann, Sajad Movahedi, Shiwei Liu et al.
In briefSafe Reinforcement Learning for Clinical Agents: Dense Rewards vs. Outcome Rewards
SummaryThis paper studies reinforcement learning with verifiable rewards for clinical multi-turn agents that decide what information to request, when to stop gathering evidence, and which action to take. The authors build three controlled environments: medication refill triage, acute diagnostic work-up, and dosing titration. They find that outcome-only training can raise aggregate reward while causing premature commitment, one-question collapse, and failures on rare high-risk cases. Giving dense credit for questions that reduce uncertainty improves results in diagnostics. The authors conclude that verifiable final outcomes alone are insufficient for training reliable multi-turn agents.
In briefOptiCover: A scalable framework for optimizing cold-start cover photos across multiple surfaces under semantic constraints
SummaryAirbnb hosts must choose one static cover photo from many listing photos, but over 90% of photos lack engagement data, creating a cold-start problem. The paper proposes OptiCover, which combines candidate photo generation, lightweight vision heads on frozen CLIP embeddings, a vision-language model that checks semantic representativeness, and an exploration logging module that collects randomized engagement data. The framework is deployed in production and produced a 1.5% engagement lift on the HomePage and a 0.35% increase in bookers on Search in online experiments.
In briefGenerative AI image labeling without fixed taxonomies (GLOC method)
SummaryThe paper presents Generative Labeling On Clusters (GLOC), a method for providing fast, accurate, and scalable online insights about images. It groups semantically similar images using optimized embeddings and summarizes insights at the cluster level so they apply to all cluster members. At Pinterest, GLOC generates image descriptors and pivot query suggestions for Hybrid Search, and the authors report that offline and online A/B tests show better quality and efficiency than previous methods.
Zhenjie Zhang, Sainan Chen, Kshiteesh Hegde et al.
In briefSpaghetti Splatting: differentiable curve-based 3D scene rendering via convolution
SummaryThis paper proposes Spaghetti Splatting, a method for learning curved 3D primitives from images using volume rendering. Curves are approximated by polylines, and each segment is rendered so that connected segments combine without double counting at joints. The camera projection is linearized per segment for fast training and inference. The authors report that the method generalizes 3D Gaussian Splatting, is competitive on general scenes, outperforms line-specialized methods, and natively supports anti-aliasing.
George Shramko, Shrisudhan Govindarajan, Rudy Lam et al.
In briefHidden Tails: Verifying Tail-Risk Claims When Some Outcomes Are Unobserved
SummaryThe paper addresses how past decisions hide some outcomes, which can conceal a predictor's largest errors and leave population tail risk unresolved. The authors derive sharp bounds on conditional value-at-risk for binary outcomes and bounded losses, under a stated limit on how outcome odds differ between observed and hidden cases. In controlled experiments with known loan repayments, recovering about 16% of hidden labels removed a median 61% of estimated ambiguity, compared with 16% for random recovery.
WhyThe authors derive sharp conditional value-at-risk bounds under selective labels and a minimum-cost audit allocation, tested in controlled loan-repayment masking experiments.
In briefOpen-source tool for removing patient identifiers from clinical notes, combining a local language model with rule-based and neural detectors
SummaryThis paper presents DeidSharp, an open-source .NET toolkit that merges findings from pattern rules, a clinical named-entity recognition model, and a locally hosted 26-billion-parameter mixture-of-experts language model to detect protected health information in clinical text. The authors compare several existing de-identification tools on nursing notes and semi-synthetic discharge summaries. They report recall-first and precision-first configurations, and the toolkit runs without sending text off-site.
In briefMulti-Agent Screening: What Agreement and Acceptance Cannot Tell You
SummaryWhen AI agents pre-screen items before a costly human assessor reviews them, only the passed items get judged. The paper shows what each kind of evidence can establish about the agents' errors. Agreement among agents does not show they are right, and assessor acceptance gives precision but only a lower bound on recall. Recall can only be determined by judging items the agents did not pass, and the needed sample size does not grow with the number of agents.
In briefFisher-Glass Learning: Classification Bounds and Subpopulation Audits
SummaryThe paper extends the Fisher-glass framework to link information remaining after adjusting for nuisance variation with local classification difficulty and subpopulation auditing. It derives lower bounds on local minimax excess classification risk under explicit distinguishability and decision-separation conditions. On the Waterbirds dataset, accuracy is 93.2% when background type is preserved and 40.6% when it changes, and a search identifies a subpopulation with 98.8% membership-weighted accuracy. The authors find that information gains and predictive gains can diverge, so they recommend jointly assessing predictive performance, information under nuisance models, and sensitivity.
In briefGeoScope: a training-free method for counting and spatial questions on ultra-high-resolution satellite images
SummaryThe paper finds that vision-language models undercount objects in ultra-high-resolution remote sensing images once a view contains more than about five objects, and that this limit narrows as views grow. GeoScope has a text-only language model turn each question into a spatial program that sets the region and resolution, while the vision-language model only grounds and counts within its effective range and directions are computed from the grounded boxes. Without training, it improves Qwen3-VL-8B by 16.3 points on regional counting and 10.5 points on spatial relations on XLRS-Bench.
WhyThe authors propose a training-free method splitting views by a measured VLM counting limit, reporting large gains on several UHR remote sensing benchmarks.
In briefGeneralized adaptive boosting and the geometry of mistakes
SummaryThis paper extends classical boosting to treat false positives and false negatives separately, introducing a "Z-learner" that guarantees a joint pair of error rates lying in a set Z. It gives a complete geometric characterization of which sets Z can be boosted, unifying the classical result with a recent multi-objective extension. The authors also present an adaptive algorithm that works without knowing Z in advance, and show that Z-learners can amplify confidence, using the centerpoint theorem from computational geometry.
WhyThe authors generalize weak learning to joint FP/FN guarantees over a set Z, giving a complete geometric characterization of boostable sets that unifies prior results.
Marco Bressan, Nataly Brukhim, Nicol\`o Cesa-Bianchi et al.
In briefTACTUS: A Robotic Palpation System That Maps Hardness Across Chicken Breast to Detect Woody Breast Myopathy
SummaryThis paper presents TACTUS, an automated tactile array that produces thickness-normalized, three-dimensional hardness maps at 20 locations on a broiler breast fillet within about 30 seconds. The system was tested on 100 fillets across four woody breast severity classes. It found significant force differences among severity classes and a moderate but significant correlation with texture-analyzer measurements.
Chaitanya Pallerla, Yihong Feng, Avisekh Tripathy et al.
In briefAgnostic Marginal-Nonuniform Learning: studies binary classification learning rates in the agnostic setting, comparing optimal rates under standard and relaxed criteria.
SummaryThis paper studies binary classification in the agnostic marginal-nonuniform learning framework, where learning rate constants may depend on the unlabeled data's marginal distribution but guarantees must hold uniformly over all label conditional distributions. In the standard agnostic setting, the optimal rate is 1/√n and the marginal distribution does not help. In the relaxed setting, optimal rates form a trichotomy of exponential, linear, and arbitrarily slow rates, along with a fine-grained rate based on the uncentered VC-eluder dimension.
In briefA lattice-based HMM for scanning Ottoman poetry's aruz meter
SummaryOttoman Divan poetry follows aruz, a meter based on syllable weights, but modern Turkish spelling drops the diacritics that record those weights. The paper proposes automatic methods to identify each poem's meter and align its syllables with the meter's positions. A hidden Markov model decodes over a lattice of candidate paths and is trained by EM without scansion labels. On a held-out set it fully scans 77.6% of test lines, outperforming a rule-based scanner, a neural tagger, and few-shot Claude Opus 5.5.
In briefAverage vs. expected causal effects when individuals influence each other: modeling choices and estimation limits
SummaryThis paper studies causal inference when one person's treatment can affect others' outcomes, a situation called interference. The authors organize their assumptions into three layers (interference neighborhoods, exposure mappings, and outcome models) and split potential outcomes into direct, spillover, and interaction components. They distinguish average causal effects, which compare outcomes under fixed conditions, from expected causal effects, which compare outcomes under a specified allocation policy, and bound the variance of the Horvitz-Thompson estimator for both. They also propose graph-aware independent-set designs whose variance bounds are polynomial in the maximum degree.
WhyThe authors propose hierarchical finite-population models separating ACEs and ECEs, and report variance bounds contrasting exponential and polynomial sample-size growth under interference.
Vishesh Karwa, Edoardo M. Airoldi, Donald B. Rubin
In briefCoilStellaration: A Dataset for Designing Stellarator Fusion Magnet Coils
SummaryThis paper introduces CoilStellaration, a dataset and benchmark of 177,000 optimized coil sets for stellarator fusion devices. The benchmark asks models to predict coil geometries from a target plasma equilibrium and requirements on spacing, curvature, and magnetic-field error. The authors find that magnetic-field accuracy remains the main challenge, while using learned predictions to initialize conventional coil optimization can speed up early progress. Baseline models and evaluation code are released.
Jan-Hendrik Ewers, Santiago A. Cadena, Nicol\`o Foppiani et al.
In briefreComposer: An Editable Planning Interface for Controllable Image Generation
SummaryThis paper presents reComposer, an image-generation interface that breaks a user's evolving intent into creative decisions such as subject, composition, and palette. Users can edit this plan through conversation or direct manipulation. In a study with 8 design and non-design professionals, participants reported greater control and ownership than with prompting when creating assets, exploring alternatives, and iterating on details.
In briefRecovering tissue optical properties from intravascular OCT images using differentiable inverse rendering
SummaryThe paper presents a method to recover per-pixel optical attenuation coefficients, markers of pathology and cellular composition, from a single intravascular optical coherence tomography (OCT) image. It uses gradient descent through a differentiable physics-based forward model, requiring no training data or segmentations. The method was verified against analytical solutions, validated on five in vitro phantoms, and applied to 29,269 clinical images from 128 patients.
Ajay Manicka, Karim Kadry, Sebastian Ruiz-Lopera et al.
In briefA four-step method for checking radiomic survival models in small patient groups
SummaryThis methodological study proposes a four-step validation protocol for radiomic survival analysis in small cohorts, combining FDR-corrected feature selection, comparison of feature selection strategies, repeated nested cross-validation, and a Gaussian noise negative control. Applied to 78 patients with gastroenteropancreatic neuroendocrine tumors, it used 535 CT-based body composition features at the third lumbar vertebra. The findings suggest a modest prognostic signal with considerable uncertainty, and the protocol may transfer to other small-cohort studies.
Mariela Talavera-Mu\~noz, Cecilia Diana-Albelda, Miguel Antonio Sampedro-Nu\~nez et al.
In briefMapping Morocco's Crop Types Nationwide Using Google AlphaEarth Embeddings and Field Boundaries
SummaryThis paper uses publicly available 64-dimensional pretrained embeddings from Google AlphaEarth Foundations with lightweight classifiers to map crop types across Morocco. It classifies three categories: winter cereals, tree crops, and other crops. Predictions are restricted to agricultural parcels using the Fields of the World global field boundary map. Over ten repeated runs on an independent test set, average accuracy was 87.17 ± 0.42%, and nationwide inference took about five hours.
Walid Elbarz, Mohamed Bourriz, Vincent Simonneaux et al.
In briefLibrary of Hands: fixing how commands map to robot motion, not the world model
SummaryThe paper proposes Library of Hands (LoH), a training-free method that keeps a world model frozen and corrects only the mapping from commands to applied actions, called the hand. LoH infers the hand through the frozen model by finding the action that best explains the change between two camera images, and it stores each identified hand so a returning setup is recognised rather than relearned. In closed-loop tests on three simulated tasks, success rose from 28.6% to 62.3%, and performance dropped by only 0.6 percentage points when nothing changed.
In briefAdapting Vision-Language-Action Models with Real and Simulated Data: A Robotic Chess Study
SummaryThis paper studies how to fine-tune a general-purpose vision-language-action model (pi0.5) for a robot arm that moves a chess king between named squares. It varies dataset sizes and update budgets, using real teleoperated demonstrations and simulated demonstrations. In simulation, success improves with more data and updates, a learning-rate cool-down helps checkpoint comparison, and validation loss is unreliable for model selection. On the real arm, simulated pretraining helps most when real data is scarce, while several alternative improvement methods gave no measurable gain over plain behavior cloning.
In briefInterpretability Tools Are Not Model-Specific: Transferring Refusal and Read-out Directions Across Models
SummaryThis paper tests whether interpretability tools can move between language models through a map between residual streams fit only on unlabeled prompts. It transfers a refusal direction and two read-out tools, a correctness direction and a sentiment probe, across 110 model pairs from five families spanning 1B to 72B parameters. The amount transferred depends on the data the map was fit on, and adding unlabeled prompts from the tool's operating region substantially raises recovery.
WhyThe authors transfer interpretability tools via residual-stream maps fit on unlabeled prompts, validated across 110 model pairs with controls and a proof plus coverage score.
Marcus Armstrong, Navid Ayoobi, Pradham Mummaleti et al.
In briefTesting how scribble-labeled image segmentation fails when annotations are imperfect
SummaryScribble-supervised segmentation is usually tested only on clean, fixed annotations, so its robustness to flawed labels is poorly understood. The authors audit 12,031 PASCAL VOC 2012 scribbles and build six controlled perturbation operators, such as class corruption, spatial offset, and stroke-length reduction. Across several backbones and training pipelines, class corruption causes the largest drop, lowering TEL-ResNet101 mIoU on VOC from 70.08 to 1.46. The paper offers a reproducible benchmark for diagnosing annotation robustness beyond clean-set accuracy.
In briefControlled Binary Assignment Flows: a geometric method for learning deep assignment flow parameters for binary graph labeling
SummaryThis paper proposes a geometric approach for representing and learning time-variant parameters of deep assignment flows used for binary data labeling on graphs. Each edge carries a parameter flow optimally controlled within a 2-dimensional totally geodesic submanifold, guided by objectives derived from a non-local linear scale space representation of the input data. Aggregated triplets of these flows are learned from data within 5-dimensional totally geodesic submanifolds, and their fusion, followed by geometric integration, yields the layers of binary deep assignment flows. Numerical results show both the expressivity and the limitations of this highly parallel local geometric synthesis approach.
In briefTrade Promotion Optimisation for Consumer Goods
SummaryThis paper presents an end-to-end system for optimising trade promotions, estimating price elasticity and recommending promotional prices for consumer goods. Items with weak evidence borrow elasticity and curvature from analogue SKUs selected by a cross-validated ridge model. In a four-week experiment on a 39-SKU consumer brand, the plan lifted units by 20.6% and revenue by 6.9% at flat margin.
In briefCertify, Report, or Refuse: Harm-Constrained Conformal Prediction for Clinical Actions
SummaryThe paper introduces HC-ROCP, a training-free framework that uses frozen classifier probabilities and adds a harm penalty when selecting prediction sets and clinical actions. Policies are tested on disjoint calibration splits with multiplicity adjustment and evaluated retrospectively on COVID-19 radiographs, ISIC 2019 dermoscopy, and EyePACS-to-APTOS retinopathy transfer. Across ten calibration seeds, severe-action errors on COVID-19 fell from 4.68% to 0.54% compared with the exact risk-optimal baseline.
In briefWhen Does Sliced Inference Pay Off? Tiling and Resolution for Small Object Detection
SummaryThis paper examines when sliced inference, such as SAHI, actually improves small object detection in high-resolution images. Comparing seven dense detectors on VisDrone and DOTA, the authors find that much of the gain comes from higher resolution rather than slicing. They propose SCAHI, which accumulates dense outputs from all tiles with coverage-aware weights and outperforms box fusion on identical passes.
WhyThe authors propose SCAHI, replacing box fusion with coverage-weighted dense outputs, and report gains over box fusion on VisDrone and DOTA.
In briefReinforcement learning aided by simulation using control variates
SummaryThe paper proposes a conservative method that uses simulated data only to reduce variance during learning, rather than to directly set the learning target. It is developed first for policy evaluation, then combined with the actor-critic framework for policy improvement and extended to offline reinforcement learning. Experiments on multiple MuJoCo benchmark tasks show improvements over standard simulation mixing approaches.
In briefTrialDesignBench: A Benchmark for Testing Language Models on Phase-I Trial Design
SummaryThis paper introduces TrialDesignBench, a benchmark of 617 Phase-I/I-II oncology dose-finding trials, each paired with an expert-validated reference design aligned with recent FDA regulatory initiatives. The authors evaluate five frontier large language models, and none of them matches the reference standard. Ablation experiments attribute the gap mainly to inferring unrecorded design determinants and to structured decision-tree reasoning.
Yihan Hank Tang (1), Peng Yang (2), Jiayi Xin (1) et al.
In briefPredicting city traffic in bad weather using large language models
SummaryThis paper proposes Multi-STG, a framework that combines semantic priors derived from large language models with graph-based spatiotemporal learning to forecast urban traffic under adverse weather. It converts context descriptions such as weather, POIs, and holidays into semantic embeddings, and uses multimodal attentional fusion and gated residual calibration in its prediction head. Experiments on real private car datasets from three cities show an average 11.82% reduction in MAE relative to the strongest baseline, with stable performance under sparse and non-stationary conditions.
In briefAdapting small language models to Romanized Nepali (Nepali written in Latin script)
SummaryThe paper adapts three 2–3B open base models (Llama-3.2-3B, Qwen2.5-3B, Gemma-2-2B) to Romanized Nepali using QLoRA on 9,000 instruction examples created by deterministic transliteration. Fine-tuning raised chrF++ from 10–14 to 29.0–30.6 and lowered bits per byte from 2.23–2.41 to 0.78–0.84. Likelihood and reference-overlap metrics disagreed on how much the models differed, and the model with the fewest tokens per word scored worst on every measure.
In briefFrom spikes to fields: how turbulence-like brain dynamics arise in spiking networks
SummaryThe paper builds a spatial spiking neural network and converts spikes into activity-density and phase fields to study how turbulence-like brain patterns arise. A delayed, distance-dependent passive-dendrite regime reproducibly generates turbulence-like field structure. The authors find that firing rate, vortex count, and avalanche fits alone are insufficient, so phase defects must be interpreted together with spatial scaling and spike-level propagation.
In briefReconstructing and recognizing objects hidden behind fragmented occlusion using video
SummaryThis paper presents a method that reconstructs an object hidden behind dense, fragmented occlusions, such as foliage, from a sequence of grayscale video frames captured by a single camera. The approach aligns the frames so that the average of the aligned images, called the integrated image, has maximum contrast, and this integrated image serves as the reconstruction. Experiments on synthetic and real videos show accurate reconstruction at occlusion densities above 90%, and the integrated image can be classified reliably by standard classifiers.
Roman Pflugfelder, Michael Lindenbaum, Daniel Cremers
In briefType-logical grammars and formal languages: language classes and complexity of first-order linear logic
SummaryThe paper shows that type-logical grammars with standard rules for product, associativity, extraction and infixation have an NP-complete fixed recognition problem. It also shows that grammars in first-order linear logic can simulate quasi-linear grammars, which recognise languages in non-deterministic linear time. The author argues that first-order linear logic can be seen as the logic of non-deterministic linear-time string languages, graph languages and automata.
WhyThe authors report NP-completeness of the fixed recognition problem for type-logical grammars and simulation of quasi-linear grammars by first-order linear logic.
In briefDirectly conditioning audio diffusion transformers on fMRI to reveal cortical contributions to sound reconstruction
SummaryThe paper introduces MAAS, an end-to-end framework that conditions a pretrained audio diffusion transformer (Stable Audio) directly on region-of-interest fMRI responses, without first training a decoder to predict a stimulus representation. On a 7T fMRI dataset, it reports consistent gains over baselines, including a two-stage fMRI-to-latent approach, on both acoustic and semantic metrics. The authors also propose a fidelity metric based on a computational model of auditory neural processing to quantify how much each auditory cortical region contributes to the reconstructed audio.
Matteo Ciferri, Tonio Weidler, Matteo Ferrante et al.
In briefFour Traveling Thief Benchmark Instances Are Actually Knapsack Problems
SummaryThe traveling thief problem asks a thief to choose a route and a packing together, and a 9,720-instance benchmark is widely used to compare its heuristics. This paper finds that four instances have a renting rate of 0.00, so the route does not affect the objective and each reduces to a 0-1 knapsack problem. The authors define corrected renting rates using a published reference solution, show that the corrected instances behave as proper TTP instances, and release reference results. They also release ttplib, an open-source Python library for loading the instances.
WhyThe authors identify four benchmark instances with zero renting rate, reduce them to knapsack problems, and report corrected rates, reference results, and an open-source library.
In briefFairness in AI explanation texts: how language models create group gaps in their explanations
SummaryThis paper examines whether large language models that turn SHAP-style explainable AI results into plain-language narratives treat social groups unequally. The authors propose a framework that generates narratives with the protected attribute disclosed, withheld, or counterfactually swapped, and evaluate five LLMs on four credit scoring, education and employment datasets. They find significant disparities across all datasets, metrics and models, arising from feature values and from the protected attribute itself.
WhyThe authors introduce fairness in XAI narratives, using a three-condition framework across four datasets and five LLMs to report significant disparities.
Wannes Van den Bulck, Sofie Goethals, David Martens
In briefTRACE: Adaptive Trust-Region Bayesian Optimization for Open-Ended Molecular Design
SummaryThis paper introduces TRACE, a framework for finding good molecules within a small number of evaluations, generating candidates on demand rather than choosing from a fixed library. It combines graph-perturbation proposals with a Gaussian process model using fingerprints and fragment descriptors. Across six molecular optimization benchmarks with a 1,000-evaluation budget, TRACE showed the strongest overall sample efficiency. The authors also applied it to iron phthalocyanine stability, raising the computed resistance to Fe dissolution from 0.42 to 2.59 eV over 80 designed evaluations.
Madhav R. Muthyala, Mahsa Askari, Farshud Sorourifar et al.
In briefAgentOPSD: Turn-by-Turn Credit Assignment via Recursive Belief Revision
SummaryAgentOPSD is a critic-free reinforcement learning method that assigns credit recursively, one interaction turn at a time, instead of spreading one advantage over a whole trajectory. Within each turn, token-level log-probability gaps between a skill-conditioned self-teacher and the student are summed and used as evidence in a recursive Bayesian update in log-odds space. With Qwen2.5-3B and Qwen2.5-7B on ALFWorld, it consistently improves on GRPO and self-distillation baselines and reaches 89.1% success at the 7B scale.
In briefTurning likelihood accuracy into parameter accuracy for POMDP learning
SummaryThis paper studies how to learn the parameters of partially observable Markov decision processes (POMDPs) from data. For any candidate parameter, it bounds the parameter recovery error using the likelihood gap to the ground truth, the structural conditioning of the observation and transition kernels, and the behavior policy used for data collection. The bound covers both undercomplete and overcomplete settings and yields finite-sample guarantees for maximum-likelihood estimation and similar procedures.
In briefChecking whether mammography AI models' answers, confidence, and explanations agree
SummaryThis paper examines vision-language models that read mammography images and output an answer, a confidence level, and a written explanation. The authors test four models on breast density and BI-RADS categorization across two datasets, applying controlled changes to questions, option orders, and images. They find that confidence does not drop even when the image is altered, and that shuffling answer options changes most models' answers. Some models also give reasoning that contradicts their final answer.
Rebeca Nam, Chase M. Fensore, Christian A. Fauria-Robinson et al.
In briefLawKAN: Group-law layers that bake known rules into learned models
SummaryLawKAN embeds an exact non-commutative group law in a self-contained layer, learns local increments on its edges, and uses a zero-initialized correction term to capture unknown deviations from the law. On fixed-angle SU(2) composition and real robot arms, it outperforms baselines with fewer parameters. It matches classical fits when the data follow the imposed law.
Anna Mishchenko, Nikolai Gavrilin (Independent researchers)
In briefGHiML: A Vision-Language Model-Guided Method for Hierarchical Multi-Label Classification of Seafloor Images
SummaryThe paper proposes GHiML, which extracts conceptual variables from a vision-language model, generates sample-specific captions, and fuses their text embeddings with image embeddings via cross-attention and a gated router. It also introduces the d_inv score, which evaluates hierarchical multi-label predictions using graph traversal distance and a cardinality penalty. Experiments on BenthicNet, FathomNet and iNaturalist report improvements in macro F1 and d_inv over a vision-only baseline.
In briefWanderBench: Can World Models Explore Dynamic Worlds from Video?
SummaryThis paper introduces WanderBench, a benchmark for testing whether video-conditioned world models can explore dynamic scenes from an input video. It includes 745 single-shot real and AI-generated videos across three difficulty tiers, paired with keyboard-based camera routes, and evaluates six dimensions using 24 human-validated sub-metrics. Tests on eight models show that longer clips preserve the static scene and visual quality but lose camera control and source dynamics, and no model keeps revisited places consistent. The authors also propose WanderLand, a training-free framework that improves five models across every dimension, with source-dynamics fidelity gains of up to 8.3 points.
In briefSciMARS: Self-improving multi-agent system for closed-loop scientific discovery
SummaryThis paper proposes SciMARS, a framework that recursively adapts multi-agent systems to individual scientific tasks. It uses a shared evidence graph and separate editors that adjust agent roles and coordination as well as how agents record and retrieve evidence. On Autonomous Science Bench, ScienceAgentBench, and AppWorld, the memory-augmented version solved more tasks than the base multi-agent team.
In briefA strict upper bound for three-point weighted selection in planar regression
SummaryThe paper proves that the weighted selection constant F_weighted(2,2,3) for minimum-Frobenius-norm least-squares regression with two input and two output coordinates is strictly less than 11/6, improving the previous upper bound of 15/8. The proof uses an exact scalar certificate, rigidity of equality in planar selection, and a compactness argument to control disappearing weights. Combined with a known lower bound, this yields 13/8 ≤ F_weighted(2,2,3) < 11/6.
In briefTeamknitt: Configuring, observing and intervening in LLM-agent teams using dynamical-systems and human-team theory
SummaryThis paper introduces teamknitt, a framework and open reference implementation for configuring teams of 1 to 16 AI agents, choosing among fourteen routing topologies or custom graphs and limiting information sharing through sharing profiles. It describes observing team processes through non-system and system views, with human-led interventions applied at defined boundaries. The work is an idea and system-description paper offering falsifiable hypotheses, and it reports no comparative results.
In briefLearning a faster inexact alternating method for convex problems with quadratic coupling
SummaryThis paper proposes learning an inexact alternating minimization scheme in which neural networks approximate each subproblem solution, reducing the costly inner solves. For convex problems with quadratic coupling that satisfy a quadratic growth condition, the authors derive a worst-case bound on the optimality gap and combine it with a PAC-Bayes argument for a high-probability bound over problem instances. On robust matrix completion and low-dose CT reconstruction with dictionary regularization, the learned scheme achieves 5x to 13x speedups over an inexact first-order baseline at matched accuracy, rising to 30x with brief end-to-end fine-tuning.
Paul H\"ausner, Jevgenija Rudzusika, Jens Sj\"olund et al.
In briefK-prop: A method for efficient online learning in leaky recurrent networks
SummaryThis paper proposes K-prop, an online learning method for leaky neural networks with self-recurrence. By propagating kernel gains, it greatly reduces the computational and memory costs of real-time recurrent learning (RTRL). In sequence learning, classification, and reinforcement learning tasks, K-prop outperforms existing RTRL approximations and performs comparably to backpropagation through time.
In briefPerceived agency in virtual reality when control of an avatar-agent switches
SummaryThis within-subject VR study with 24 participants compared virtual humans under four control conditions: always computer-controlled, computer-to-human, human-to-computer, and always human-controlled. Participants acted as interviewers and were told the initial controller but not about any transitions. Perceived agency was strongly shaped by the controller presented at the start and did not update after control switched. Likeability and co-presence were generally higher when human control came later and covered more of the interaction.
In briefWaveGen: Generating speech waveforms end-to-end by learning an internal spectral trajectory
SummaryWaveGen is a diffusion-based framework that generates waveforms directly from text in one model, using a method called Spectral Forcing to learn an internal spectral trajectory alongside the target waveform. It avoids external intermediate representations and alignment modules such as duration predictors. The paper reports state-of-the-art results on the Emilia training dataset, with a WER of 1.47 and a SIM of 0.75 on the Seed-en benchmark. Code and checkpoints are publicly available.
