In briefPreparing Base Models for Agent Post-Training with an Open Midtraining Recipe
SummaryThis paper presents an open midtraining recipe, a stage between pretraining and post-training, that prepares base models for agentic post-training. Training runs at 256K context on natural software-engineering data such as packed repositories, pull requests, commit histories, and build logs, plus a small share of agentic traces. The authors report results on two Nemotron models of different sizes, along with ablations and an evaluation protocol based on short fixed SFT runs.
Soumye Singhal, Joseph Jennings, Ka Shun Shum et al.
In briefWhen Models Misread Ink: Auditing Vision-Language Models for Recognizing Arabic Calligraphy
SummaryThe paper evaluates 12 vision-language models on a cleaned six-way benchmark of 1,195 Arabic calligraphy images. The best numeric-prompt accuracy was 60.9% for closed/API models and 47.2% for open-weight models. The main failure is default-label collapse, where most predictions fall on one script; for 8 of 12 models this default is Thuluth, and GPT-4o labels 52.9% of images as Thuluth.
In briefCARA: A Causally Guided Retrieval Agent for Question Answering
SummaryThe paper proposes CARA, a training-free retrieval agent that guides search over text-rich knowledge graphs using causal relevance rather than only textual similarity. It scores retrieved evidence by consequentiality (whether adding it improves answerability) and essentiality (whether removing it weakens answerability), combined with path consistency and reciprocal rank fusion. Experiments on three STaRK datasets, augmented with ground-truth answers, show CARA outperforming existing retrievers, improving answer quality and faithfulness while examining fewer nodes and retrieval operations.
In briefClose Enough Is Nowhere Near: Why Substituting KV Caches Between Frozen LLMs Fails
SummaryThis paper studies sending KV caches instead of text between frozen multi-agent LLMs. Replacing a reader's cache with a sender's cache without training yields performance at the level of an uninformative cache baseline. The authors attribute this to the reader only being able to use caches close to its own, and report that adding the sender's cache as extra attended context while keeping the reader's own cache does beat the reader's baseline.
In briefGraphBEAT: A time-bucketed graph model for early gastrointestinal cancer risk prediction from health records
SummaryThis paper introduces GraphBEAT, a framework that groups irregular, sparse longitudinal electronic health record events into time buckets and converts each patient's history into a temporal directed graph. It is used to stratify early risk for pancreatic, liver, and intestinal cancers at 12, 24, and 36 month lead times. The authors report that GraphBEAT outperforms GNN and sequence baselines such as GraphSAGE, GAT, GCN, Transformer, and LSTM while requiring far fewer operations and parameters.
Leo Y. Li-Han, Chris Varghese, Elizabeth B. Habermann et al.
In briefFrom Natural Language to Executable Control: Feedback-Guided LLM Planning for Cyber-Physical Systems
SummaryThe paper addresses the problem that LLM-generated plans for cyber-physical systems may misread user intent or be infeasible under component dynamics and safety constraints. It proposes a hierarchical feedback framework that generates both a structured task plan and signal temporal logic constraints, detects contradictions between them, and uses diagnostic information from failed control synthesis to guide replanning. Experiments on seven heterogeneous multi-component CPS scenarios report an 87.9% success rate with GPT-5.
SummaryThis paper introduces Visual Sanad, a benchmark of 2,400 images that tests whether vision-language models falsely accept incorrect attributions of Qur'anic verses. The images are built from 100 English verses, with four factual states per verse and six visual layouts. Results show a clear capability split: some proprietary models verify citations nearly perfectly, while many recent open-weight models often default to VALID.
In briefTesting deep learning compilers for bugs caused by retained state
SummaryThe paper studies deep learning compilers that compile models lazily and cache artifacts, where retained state can make a later call fail even if it works alone. It presents Rondo, a stateful testing method that pairs models with invocation sequences, mutates them with an MCMC sampler, and uses a three-way differential oracle. Evaluated on four frameworks and 24 model architectures, Rondo detects 117 bugs, 35 of them stateful.
WhyNovel stateful testing paradigm for compiler cache bugs; finds 117 bugs across four frameworks, 35 stateful, with broad impact.
In briefUsing sparse autoencoders to build local geometric constraints against catastrophic forgetting
SummaryThe paper proposes SAGA, which trains a sparse autoencoder at each task boundary and converts it into local, low-rank constraints on representation drift, with replay examples specifying where constraints apply. On three pretrained ViT-B/16 benchmarks with replay and snapshot history, it reports the highest mean Class-IL accuracy and lowest mean forgetting among the compared baselines. Results from training from scratch are more mixed.
In briefShared-coordinate graph for scalable weight-space learning across network architectures
SummaryThe paper proposes a typed graph that represents neural network parameters through shared coordinate-permutation actions, with interfaces as nodes and modules as hyperedges, derived automatically from supported PyTorch models. It introduces the Universal Weight Encoder (UWE), which combines local equivariant updates with global attention over invariant interface summaries. UWE is reported to be competitive with architecture-specific methods on established benchmarks. On a new zoo of 1,200 transformers across 100 architectures, it outperforms the evaluated baselines at ranking networks within unseen architectures while using substantially less memory.
In briefData-efficient ConvNeXt pretraining for high-resolution Sentinel-1 radar image representation learning
SummaryThe paper proposes SparK-SAR, a sparse convolutional masked pretraining framework built on SparK and ConvNeXt, for learning representations from Sentinel-1 GRD radar imagery. It introduces a domain-adapted ConvNeXt-RS encoder and a lightweight decoder, and reports instability in FP16 mixed-precision pretraining of the standard decoder. On four Copernicus-Bench S1GRD tasks, the finetuned 27.8M-parameter encoder beats a non-pretrained baseline and outperforms a larger frozen Copernicus-FM ViT on three tasks. A data-scaling study finds classification saturates with 10-25% of the pretraining corpus, while dense prediction keeps improving up to 25-50%.
In briefspora: A Unified Multimodal Dataset and Platform for Spatial Proteomics
SummarySpatial proteomics captures the molecular makeup and spatial organization of tissues at single-cell resolution. The paper introduces spora, a dataset of 13,546 standardized multiplex images from 5,477 patients across 35 cohorts and five major technologies, with cell and tissue segmentations, structured clinical metadata, and paired H&E images where available. It also presents spora [io], a unified interface for tiling, sampling, and standardization, and spora [bench], a benchmark suite across twelve cohorts that compares current spatial proteomics foundation models with task-specific baselines.
Benedikt von Querfurth, Johann Wenckstern, Eeshaan Jain et al.
In briefVesperMUSE: A Multi-Modal Vision Transformer for Image Segmentation in Varied Lighting
SummaryThis paper presents VesperMUSE, a framework that improves semantic segmentation, which helps autonomous systems understand scenes, when lighting is poor. It combines a physics-informed synthetic low-light adaptation method, thermal-guided low-light image enhancement, and a fusion of RGB, thermal, and surface normal inputs through a self-attention mechanism. The authors report that the approach recovers missing visual cues and achieves state-of-the-art results across diverse illumination conditions.
In briefGRID-Planner: Graph-Structured Relational Reward Shaping for Diffusion-Based Driving Planning
SummaryThis paper addresses autonomous driving planning, where a vehicle must choose trajectories that are safe, rule-compliant, and comfortable amid complex multi-agent interactions. It proposes GRID-Planner, a framework that integrates diffusion-based planning with structured relational signals through graph-aware importance weighting, a time-indexed relational reward surrogate, and an advantage-weighted diffusion objective with on-policy self-distillation. Experiments on nuPlan and interPlan report that compared with a diffusion-only baseline, it reduces collision rate, lane deviation, and jerk by 35.8%, 60.8%, and 27.4%, and improves the planning score by 3.4%.
In briefOntoAct: Hypothesis-Based Neuro-Symbolic Agents via Ontological Abstraction
SummaryThis paper presents OntoAct, a framework that turns experiences into reusable axioms over entity types, relations, and conditions. An LLM proposes hypotheses and actions, while a symbolic verifier checks their supporting conditions against current observations. The authors also introduce Tacit-FLE, a Factorio-based benchmark for diagnosing and repairing hidden faults, and report leading performance on it, Mars, and ALFWorld.
In briefDeep Learning for Topological Equivalence Classes
SummaryThis paper applies deep learning to decide whether two generating vectors correspond to topologically equivalent finite group actions on Riemann surfaces. It shows that architectures developed for semantic understanding can be useful for pure mathematical problems. The authors discuss how desired algebraic structure shapes model design and examine which structural properties of topological equivalence classes the model successfully learned.
In briefP2TPO: A structure-aligned method assigning credit from full-page to tokens in document parsing
SummaryThis paper addresses local errors and omissions in end-to-end full-page document parsing by proposing P2TPO, a credit-assignment framework for policy optimization. It aligns reference elements with generated spans, uses Unified Anchor Pricing to handle matched and missing elements, and redistributes block credit across correct and erroneous spans. The authors report that P2TPO outperforms all evaluated policy optimization baselines in seed-averaged scores on six document parsing benchmarks.
In briefLearns Only Where Gaussian Processes Fall Short: Structured Corrections to GP Posteriors
SummaryThis paper proposes a framework that starts from a tractable Gaussian process as a reference and uses a diffusion model to learn corrections within a low-dimensional subspace, leaving the remaining components Gaussian. The authors report that across three scientific inference tasks, these corrections recover hidden modes, improve sample structure, and reduce prediction error relative to the Gaussian reference. The work also compares the approach with full-field generative models, noting task-dependent tradeoffs between accuracy and computational cost.
In briefTime to Fake It: Continual Learning Worsens Unfairness in Deepfake Detection
SummaryThis paper studies how deepfake detectors trained progressively on a chronological sequence of 11 real-world generation methods change in fairness across demographic groups. It tracks four fairness dimensions over time and finds that accuracy degradation typically worsens unfairness, while fairness improvements are scarce. The authors argue that inequitable forgetting is not explained by underrepresentation alone and call on continual learning methods to account for social outcomes.
Kajaree Das, Oliver B\"ose, Subarnaduti Paul et al.
In briefUncertainty-aware federated learning for jointly optimizing communication and computing in XR systems
SummaryThe paper studies a 6G network where several base stations serve multiple extended reality (XR) devices at once. It proposes a distributed deep learning method that jointly manages transmit power and computation allocation under channel and computing uncertainties, injecting estimation-error samples during training to minimize worst-case delay. The authors report better robust delay performance and scalability than uncertainty-agnostic methods and state-of-the-art benchmarks, with federated training achieving near-centralized performance.
Robert-Jeron Reifert, Hayssam Dahrouj, Alaa Alameer Ahmad et al.
In briefWhy shallow neural networks learn sinusoidal components first: an orthogonal polynomial dynamics framework
SummaryThis paper theoretically explains spectral bias, the tendency of neural networks to learn low-frequency components before high-frequency ones. The authors expand network output as a general Fourier series in an orthogonal polynomial basis and write gradient-descent dynamics of the coefficients as ordinary differential equations. They analyze eigenvalue decay and eigenvector shapes of the dynamics matrix H to explain the bias.
In briefChronoSheaf: Joint spatio-temporal spectral filtering on sheaf-based graph networks for forecasting signals over time
SummaryThis paper introduces ChronoSheaf, a spatio-temporal sheaf neural network that forecasts signals evolving over a graph. It defines a spectral sheaf Laplacian jointly over space and time, along with Transitional Restriction Maps for asymmetric temporal influence and a Dual Spectral Filter. The authors report that ChronoSheaf outperforms state-of-the-art baselines on diverse spatio-temporal forecasting benchmarks.
In briefSelectPFN: Amortized Bayesian Variable Selection for Sparse Nonlinear Regression
SummarySelectPFN is a prior-data fitted network that performs Bayesian variable selection and prediction jointly in a single forward pass, returning posterior inclusion probabilities for each input variable. It is pretrained on synthetic tasks drawn from a meta-prior built on a group spike-and-slab sparse-input neural network, and uses knockoff statistics to achieve exact finite-sample false discovery rate control. On nonlinear designs with up to 500 variables, it improves the area under the precision-recall curve by 0.23 to 0.28 over several baselines.
In briefTracing How People Delegate Tasks to Multimodal AI to Study Human Oversight
SummaryThis paper reports an online study with 56 AI-literate participants and 392 video question answering trials, examining how users decide to delegate tasks after reviewing multimodal AI responses. Post-interaction trust strongly predicted delegation choices, but trust and delegation remained distinct, and verification was the most common policy. The authors propose enacted delegation traces, a non-intrusive observation approach, as a methodological framework for studying multimodal and agentic AI use.
In briefLearning Speech Attribute Representations via Multi-Teacher Distillation (one shared encoder learning several speech traits at once)
SummaryThe paper proposes a multi-attribute, multi-teacher distillation framework that uses one shared self-supervised speech encoder with attribute-specific branches. It learns semantic, speaker, language, and emotion representations from four frozen teacher models. Experiments on multilingual semantic retrieval, speaker verification, language identification, and speech emotion recognition show performance close to dedicated single-task students. The learned layer aggregation also indicates that different attributes emerge at different encoder depths.
In briefAgentic-Priors: Using LLMs to Solve the Cold Start Problem in Learning User Preferences
SummaryThis paper addresses the cold start problem, where an agent must learn a user's preferences before observing any of their choices. The proposed Agentic-Priors method has a large language model describe a new user's trade-offs, such as willingness to pay, which is mapped to a prior over multinomial logit coefficients and refined by a Bayesian learner as choices arrive. Across seven choice datasets, it improves cold start accuracy over an uninformed prior by 5 to 40 percentage points and outperforms AutoElicit on three datasets while losing on one.
In briefFlow-CLIP: Image-Conditioned Residual Scoring for Compositional Image-Text Matching
SummaryThe paper addresses a weakness of sentence-level image-text alignment: it struggles to tell apart captions that use the same objects and attributes but bind them differently. Flow-CLIP refines the global matching score using changes in token-patch evidence conditioned on the image, and adds a cross-image margin ranking loss to reduce reliance on language priors. On compositional matching and attribute-object binding benchmarks, it reports better binding accuracy than pretrained CLIP and learned global alignment.
In briefTheoretical analysis of a greedy cost-penalized method for choosing which features to acquire
SummaryThis paper studies Active Feature Acquisition, where features are chosen one at a time to improve prediction under a budget. It analyzes a common greedy heuristic, Expected Information Gain minus a cost penalty, and gives a certificate splitting its gap to the optimal hard-budget policy into a budget-slack term and a greedy-suboptimality term. The authors report near-zero greedy gaps on small instances and prove a bound for equal costs. Experiments include synthetic settings and clinical differential diagnosis on MIMIC-IV.
In briefBridgeAfford: Closing the Query-Target Gap for Generalizable 3D Affordance Grounding
SummaryThis paper addresses locating functional regions on 3D objects from language-described interaction intents. Existing methods generalize poorly to unseen objects because the query features and target-region features differ semantically. The proposed BridgeAfford framework reduces this gap through Interaction Region Alignment and Visual Evidence Aggregation. Experiments on the LASO and PIAD datasets show it outperforms prior methods, especially on unseen objects.
In briefCode health and the token cost of AI agent code review: a pre-registered study
SummaryThe study analyzed 465 agent-authored pull requests from 246 repositories, processing each with a frozen five-agent review-and-fix pipeline across three LLMs. Higher baseline CodeHealth was associated with lower token cost, about 3.5% to 6.1% per point, and this was not explained by PR size. Adjusting for the size of the scored files removed the association, which the authors interpret as mediation by file size.
Lo Gullstrand Heander, Markus Borg, Emma S\"oderberg et al.
In briefLearning Rewards in Strategic Games: Active inverse reinforcement learning to infer player rewards from observed behavior
SummaryThe paper studies how to infer agents' rewards from their observed behavior in strategic multi-agent settings. It shows that learning from a fixed policy pair can be insufficient to strategically identify rewards, regardless of the number of expert samples. The authors propose an active inverse reinforcement learning algorithm for two-player repeated Stackelberg games, derive finite-sample guarantees, and evaluate it in an anti-poaching setting.
In briefCoWorld Reasoner: Agentic multimodal reasoning driven by an evolving video world model
SummaryThis paper presents CoWorld Reasoner, an agentic framework that uses a video world model as a unified visual reasoning engine, guided by tools and skills that help a vision-language model decide when to imagine and how to interpret the results. It also introduces VLM-VWM Harness Co-Evolution, which revises tools, skills, and execution code from execution feedback while keeping model weights fixed. Across seven benchmarks, the framework outperforms direct reasoning with the same model, with relative accuracy gains of 11.8% on MMSI and 13.0% on ERQA.
In briefTilted Sampling: Optimism, Posterior Sampling, and Logarithmic Regret
SummaryThis paper studies tilted sampling, an online learning problem where the goal is to sample from a Gibbs distribution induced by an unknown reward function rather than pick one best action. It defines sampling regret as temperature-normalized cumulative KL divergence and proposes sampling analogues of UCB and Thompson sampling (UCS and TS²). Under standard settings at fixed temperature, the authors show sampling regret grows logarithmically in T, unlike the square-root scaling of standard regret.
WhyProposes a sampling-regret framework showing logarithmic regret via squared confidence widths, with UCS, TS², and a Bayesian decomposition.
In briefSeeing Is Believing: Testing and Reducing How Language Habits Distort Vision in Multimodal AI Models
SummaryMultimodal large language models can let their built-in language expectations override what the image shows, weakening visual perception. The authors introduce ViSiP-Bench, a benchmark covering six categories of rare but real prior-driven scenarios, with three metrics that measure resistance to these priors more precisely than accuracy. Evaluating 24 models, they find prior reliance is widespread and does not decrease with scale. They also propose MAPS, which fine-tunes each model on factually grounded knowledge to reduce excessive reliance on its strongest priors.
In briefIn-context learning before specialisation: a learning theory in a solvable topic model
SummaryThis paper studies two ways sequence models can generalise: extrapolating structure seen in training, or using the prompt to adapt to new tasks. The authors build a solvable topic model and derive a learning theory explaining why one solution is favoured over the other. They predict the in-context predictor emerges early, while the specialised predictor appears only after training tokens exceed the parameters of the data-generating process. Trained transformers match both scales and the crossover closely, with no fitted parameters.
WhyProvides a solvable-topic-model theory of in-context versus specialised learning, predicting emergence scales matched by trained transformers without fitted parameters.
In briefTEM: A memory method that tracks changing facts for LLM agents
SummaryThis paper presents TEM, a training-free memory framework for LLM agents that stores facts as versioned records linked to their sources. At answer time it retrieves relevant versions, orders them by source, and labels historical ones. On 420 questions, key-entity hit rates reached 75.71% to 97.62% overall, with fewer evidence tokens than session-aware RAG.
In briefSelf-calibrated robust PCA that removes mixed noise and line stripes from hyperspectral images without manual tuning
SummaryHyperspectral images are often degraded by Gaussian noise, impulses, and stripes whose intensity varies across bands. The proposed SC-RCCTV model adds band-wise dense noise and line-stripe terms to RCCTV and estimates noise parameters adaptively during iterations, so no noise level or stripe budget must be specified. Reported experiments show it beats the parameter-free FastHyMix and, without noise information, exceeds S3TTV given true noise parameters, while running far faster.
In briefSurgStreamer: request-driven monitoring of surgical video with structured state and reusable reasoning
SummaryThe paper formalises request-driven event detection for surgical video, where an assistant alerts users to requested events and answers questions about past events. It presents SurgStreamer, which separates procedure-specific perception, mapping 1-FPS video to structured state over tools, actions and phases, from procedure-agnostic reasoning over a monitoring condition. On 12 sealed laparoscopic cholecystectomy videos it reached a macro event-F1 of 0.598, above the tested streaming vision-language model configurations. The authors identify perception under distribution shift as the main limitation.
WhyFormalises request-driven surgical event monitoring, separating perception from reasoning, with sealed-test and multi-centre evaluation, though scope centres on cholecystectomy.
In briefRank Measures of Fine-Tuning Updates Need a Noise Baseline
SummaryThis paper examines rank measures used to summarise fine-tuning weight updates, comparing them with random matrices, pretrained weights, and fine-tunes without alignment content. Counting-based measures place the updates only slightly below noise, while stable rank places them far below, and this pattern holds across thirteen released model pairs up to 8B parameters. The authors find it is not specific to alignment, and argue that rank claims need a noise baseline and alignment claims need a control without alignment content.
In briefFitting-free causal direction test using graph Dirichlet energy
SummaryThe paper proposes GDECI, a method for deciding which of two variables is the cause, based on the Dirichlet energy of a kernel similarity graph built from one variable with the other treated as a graph signal. The lower-energy direction is inferred as causal, and no regression fitting or independence testing is needed. On 1899 cause-effect pairs it ranked fifth of eleven on synthetic suites and third on real Tuebingen pairs in unweighted accuracy, and was the fastest method up to about 10^4 samples.
In briefMetaShare: A framework where agents learn collectively from both successes and failures.
SummaryThis paper introduces MetaShare, a framework that records both successful and failed agent modifications along with their signed performance changes. Before each generation, it retrieves diverse high-performing strategies and high-confidence failures from other lineages so agents can reuse effective mechanisms and avoid known dead ends. Experiments on paper review, coding, math grading, robotics reward design, and financial audit-risk assessment report that its median matches or exceeds DGM-Hyperagents' final performance while reaching its own best agent in 1.7 to 2.9 times fewer iterations.
Vishal Bhutani, Prasang Gupta, Sumanth Chundru et al.
In briefAnvil: Scaling Embodied Spatial Intelligence via Knowledge Recirculation
SummaryThis paper presents Anvil, a unified framework covering data generation, training, and benchmarking for embodied spatial intelligence. Anvil-Gen produces diverse data across 11 forms and recollects failure cases, while Anvil-MOPD is a 9B student model that distills knowledge from specialist teachers. On average across seven public benchmarks, Anvil-MOPD outperforms Qwen3.5-9B and RynnBrain-8B, reaching 80.67% on VABench-Point.
In briefLearning Lean feedback to speed up proof-candidate checking on GPUs
SummaryThe paper trains neural models on 4.84M labelled Lean 4 programs to predict verifier verdicts and compiler feedback on GPUs, replacing slow proof checks. A two-stage evaluator reaches 87.8% accuracy on a 68,932-program benchmark and processes 3.5 candidates per second on one H100, 4.2x a 64-worker Lean baseline. Using it with Lean verdicts as hybrid rewards cuts GRPO training steps by about 31%, with miniF2F pass@1 nearly unchanged.
Lazar Milikic, Etienne Bamas, Emmanuel Abbe et al.
In briefOffline reinforcement learning that accounts for risk, with proven performance bounds
SummaryThis paper studies offline reinforcement learning when decisions must account for risk, using the optimized certainty equivalent (OCE), a general class of risk measures. It considers two formulations, dynamic-OCE and static-OCE, and proposes pessimistic value-iteration algorithms for the tabular setting with sublinear suboptimality bounds and matching minimax lower bounds. It also extends the approach to large state spaces with linear function approximation, giving guarantees that depend on the feature dimension.
In briefLethe: a benchmark that tests edge unlearning in graph models using link inference attacks
SummaryGraph unlearning removes the effect of specific training data from a graph neural network without full retraining. The paper finds that current edge unlearning evaluations have two limits: dedicated methods can be slower than retraining, and accuracy metrics say little because removing an edge barely changes accuracy. The authors introduce Lethe, the first comprehensive edge unlearning benchmark, with 15 methods and 8 datasets, emphasizing large graphs and using link inference attacks as evaluation measures.
WhyProposes the first comprehensive edge-unlearning benchmark with link-inference-based metrics replacing proxy accuracy, exposing flaws in current evaluations across 15 methods and 8 datasets.
Andrea D'Angelo, Francesco Gullo, Davide Mottin et al.
In briefFrom Sketch to Scene: Rebuilding Relightable 3D Garments from Annotated Sketches
SummaryThe paper presents a system that turns a rough garment sketch, marked with sparse star (highlight) and moon (shadow) symbols plus a short texture prompt, into a textured 3D mesh with an explicit directional light. The result is a scene that can be relit and edited after generation. The system predicts the light direction and a displacement field that makes moon-marked regions fall into shadow, rather than painting them dark. The two symbols were the most frequent proposals in a study with art students.
Ahmet Ne\c{c}irvan Do\u{g}an, Eren Acar Ba\c{s}aran, Metin Sezgin et al.
In briefDynamic generation order for Transformer tagging in low-resource settings
SummaryThis paper proposes TagInsert, a Transformer model that drops the usual left-to-right generation order and lets the model choose the order of tag generation at inference time. It is tested on part-of-speech tagging and CCG supertagging across several languages and low-resource data settings. TagInsert often beats a left-to-right encoder-decoder model, matches encoder-only models when data is scarce, and performs better on rare categories.
In briefCounterfactual 4D Heart Shape Generation via Causal Neural Distance Fields
SummaryThis paper proposes a method that combines a structural causal model of clinical variables with a spatio-temporal neural distance field to generate counterfactual four-dimensional left atrial anatomy over the cardiac cycle. The generated shapes follow dose-response under age interventions, and inverse interventions recover the original anatomy with sub-millimetre precision. The authors state this is the first counterfactual generation of 4D cardiac anatomy and the first integration of an SCM with an implicit neural representation.
Mathias Lowes, Kurt Butler, Kristine Aavild S{\o}rensen et al.
In briefCombining Whole-Object and Part-Level 3D Priors for Part-Decomposed Generation
SummaryThe paper introduces STRUCTPARTS, a framework for generating 3D assets with explicit parts while keeping detail within each part and coherence across the whole object. Its core mechanism, dual-frame coordinate interleaving, processes shape latents under both part-local and object-global frames using a global-local interleaved transformer. The model takes a segmented mesh scaffold as structural input and requires only lightweight finetuning of pretrained weights. On PartObjaverse-Tiny, the authors report it matches or surpasses prior part-decomposed models at a fraction of the training cost, and it supports part-mesh refinement and part-level editing.
In briefTemporal Periodic Image Registration Using Implicit Neural Representations
SummaryThis paper extends implicit neural representations (INRs) to deformable registration of periodic temporal image sequences, proposing Temporal-IDIR. A single INR models deformations across all frames at once, with a temporal consistency loss enforcing agreement between direct and intermediate-frame transformations. On the DIR-LAB dataset, using target registration error, it reports 1.03 mm, outperforming other INR-based methods, and supports smooth interpolation between frames.
Mathias Lowes, Kristine Aavild S{\o}rensen, Maxime Sermesant et al.
In briefPhysics-aware distillation of a power-grid foundation model for real-time edge inference
SummaryThe paper distils a 20.1M-parameter, 12-layer grid foundation model (GridFM) into a 0.62M-parameter, 6-layer student, a 32.5× parameter reduction with half the message-passing depth, for edge inference. The distillation objective combines the masked power-flow loss with output, physics-residual, latent-state and attention transfer. Evaluated on a renewable-rich variant of the IEEE 118-bus grid, distillation lowers student error on all four predicted quantities, by up to 48% on active-power dispatch.
Vlad Alexandru S\u{a}nduleac, Shahariya Rasheed Rayamarakkarveettil Puthanpurayil, Noman Khan et al.
In briefLocal Epochs as Optimal Stopping in Federated Learning
SummaryThe paper models how long each federated learning client should train locally before communicating as a finite-horizon optimal stopping problem, continuing only while expected loss reduction outweighs squared client drift. It proves a one-stage look-ahead threshold rule is optimal when stopping decisions are nested and introduces FedStop. FedStop's mean accuracy is never more than one point below tuned FedAvg, while it raises worst-decile accuracy by 4-8 points where budgets vary.
WhyCasts local training as optimal stopping with closed-form thresholds, backed by federated learning experiments showing worst-decile gains over strong baselines.
In briefCompiling multi-threshold operators for non-monotonic, fine-grained quantized activations in custom DNN accelerators
SummaryThe paper argues that aggressively quantized neural networks make activation functions a hardware bottleneck, and that multi-threshold implementations can help. The authors extend the multi-threshold representation to non-monotonic functions and fine-grained, multi-dimensional thresholds, and automate its conversion using algebraic rules in an ONNX-based compiler. They validate the approach by extending the binary-search multi-threshold operator in the FINN framework and compare it with alternatives and hls4ml.
Christoph Berganski, Felix Jentzsch, Marco Platzner
In briefA closed-form depth-width law for neural collapse in deep linear networks
SummaryThis paper studies how network depth and width shape learned representations, using neural collapse as a measure of representation geometry. In a tractable feature-learning setting, the authors express NC1 to NC3 directly through the learned feature kernel and derive a layerwise law for within-class variability in deep linear networks. The paper states that at large width, the accumulated compression is a function of the depth-to-width ratio L/N.
In briefMeasuring how learning-rate transfer curves locally change shape
SummaryThis paper studies how the learning rate affects trained model performance, which is important for optimizing neural networks. It gives two relational equations explaining the gradient of the learning-rate transfer curve: one theoretical formula holding at the edge of stability, and one approximate expression treating the gradient as a hyper-parameter gradient. Using both, the authors derive a lower bound on the curve's gradient and show conditions where it is zero for normalization layers. The approximate expression also decomposes the learning-rate gradient into layer-wise components to quantify each layer's contribution to the curve shape.
In briefN-gram Memory Compression: Quantizing Lookup Tables in Large Language Models
SummaryThe paper proposes Conditional Program (CP), a compressed representation that reconstructs only the embedding rows an input requests, each stored as a short code and local scale with a shared reconstruction program. At two bits, CP lowers output KL divergence by 41–63% relative to 2-bit round-to-nearest quantization. For DeepSeek, table storage drops from 203GB to 51GB.
In briefTesting Chest X-Ray AI Models on Children After Training on Adults
SummaryThis paper examines whether foundation models trained on adult chest X-rays can classify pediatric chest X-rays accurately. The authors merged VinDr-CXR and PediCXR into a five-label task and compared six foundation models and two CNN baselines, adapting them only with adult data. All models performed substantially worse on pediatric data, with drops above 10 percentage points, and deeper adaptation did not significantly help cross-population generalization.
In briefRoleQ: Role-Aligned Scales for Ternary Language Model Quantization
SummaryThis paper introduces RoleQ, a quantization method that assigns scaling factors based on the functional roles of Transformer components rather than weight matrix layout, requiring no search or calibration. With ternary weights, it reports cutting perplexity by up to half versus layout-based placement across models from 0.6B to 14B parameters. The authors also present a lightweight post-training pipeline that converts existing checkpoints to about 1.59 bits per weight in a few GPU-hours, with gains on perplexity, zero-shot accuracy, and MMLU.
Johannes Knittel, Alec Helbling, Hanspeter Pfister
In briefWhich words link to best answers and engagement in logistics Q&A forums
SummaryThis exploratory study analyzes logistics-related questions and answers from Yahoo! Chiebukuro, a Japanese community question-answering platform. It examines which nouns and semantic noun clusters are associated with best-answer selection and with question engagement. Operational nouns such as distribution and warehouse relate to best-answer selection, while broader nouns such as car, tax, and shinkansen relate to higher engagement.
In briefCIM-Embed: Instruction-Conditioned Multimodal Embeddings for Grouping Items
SummaryThis paper presents CIM-Embed, a framework that maps an object's image, text, and instruction into one embedding, then uses HDBSCAN to form groups without a preset cluster count. The authors build ProductMIC from Amazon Reviews 2023 product metadata, with a benchmark of 5,373 products and 449 grouping tasks. With direct within-task supervised contrastive training, the mean per-task V-measure is 95.01, compared with 53.06 for WeMM 9B, the strongest general-purpose baseline evaluated.
In briefCo2In: A Brain-Inspired Two-Level Memory Framework for Incremental Learning
SummaryThis paper proposes Co2In, an incremental learning framework that lets models learn new knowledge while retaining old knowledge. It uses a two-level memory design, where a working memory learns new data and a consolidation step compresses it into a compact representation integrated into long-term memory. The authors evaluate Co2In on CIFAR-10, CIFAR-100, and Tiny-ImageNet under exemplar-free class- and task-incremental settings, reporting state-of-the-art performance with efficient memory use.
In briefTeaching DINOv3 Partial 3D Geometry via Self-Supervised Learning
SummaryThis paper addresses partial shape matching in 3D vision, where shapes are often only partly observed. The authors refine DINO features with LoRA-based self-supervised learning to produce descriptors robust to partial observations. According to the paper, these features outperform traditional and vision foundation features on partial shape matching, and achieve state-of-the-art results on partial shape matching and left-right prediction benchmarks when integrated into existing pipelines.
Viktoria Ehm, Dongliang Cao, Riccardo Marin et al.
In briefBehavioral forgetting is not privacy: a study of machine unlearning in language models
SummaryThe paper argues that testing machine unlearning only on a designated benchmark is not enough to show privacy-oriented forgetting. The authors propose checking whether forgotten information stays inaccessible across other access paths, such as different languages. In a controlled cross-lingual study across several multilingual model families, forgetting varied substantially when the same information was queried in different languages and scripts.
In briefTransferring Human Video Skills to Robot Manipulation Across Different Bodies
SummaryThis paper studies how knowledge from human manipulation videos can benefit robot learning, despite differences in morphology and action spaces. It proposes Causal-Beam-Guided Subtask Transfer (CBST), which learns manipulation representations from human videos and grounds them in task-level semantic variables using robot demonstrations. The selected semantic information conditions a world-action model so the robot can act from its own observations. Experiments on robot manipulation benchmarks report higher task success than the corresponding world-action model baselines.
In briefSingle-stage joint diffusion of sky images and clear-sky index for short-term solar irradiance forecasting
SummaryThe paper proposes a single diffusion model that co-denoises future sky-image and clear-sky-index sequences, with an irradiance head reading a denoised preview of the future sky, removing the separate regression stage. On 2023 SIRTA data it reports a pooled RMSE of 104.1 W/m² and a CRPS skill score of +0.34 against smart persistence, outperforming twelve baseline configurations. The authors also report that a separately trained regressor on the generated frames performs worse (150.6 W/m²) than the joint head.
In briefMedPAO-VL: An Interpretable, Protocol-Compliant Agent for Structured Radiology Report Generation
SummaryThe paper presents MedPAO-VL, a multimodal agent that generates structured radiology reports from chest X-rays through an iterative plan-act-observe loop. It uses an explainable vision module that captures diagnostic uncertainty and refines reports by detecting and correcting discrepancies with visual findings. The authors report expert-rated scores of 3.5/5 for accuracy and 4.3/5 for structure.
Shrish Shrinath Vaidya, Gowthamaan Palani, Minmini Selvam et al.
In briefRSIFlow: Self-improving AI agent workflows that evolve their own meta-level controller
SummaryThe paper introduces RSIFlow, a framework in which language model agents diagnose performance bottlenecks from task trajectories and feedback and select component-specific improvement workflows. A meta-agent distills workflow results into skills and cross-component principles and evolves its own harness to guide later orchestration. The authors report state-of-the-art average performance across six benchmarks and two backbone models, and ablations isolate the effects of adaptive orchestration and meta-harness evolution.
In briefCF-NODE: Counterfactual Spatial Node Intervention for Test-Time Adaptation in Diabetic Retinopathy Grading
SummaryAutomated diabetic retinopathy grading from fundus images often degrades when models move across cameras, clinics, and patient groups. The paper proposes CF-NODE, which adapts each image independently by intervening on high-ranked internal spatial nodes and calibrating predictions based on how stable the expected grade response is. It uses no target-derived information or model updates. On five external datasets, the authors report it outperforms TENT, EATA, SAR, and T3A, with the highest average AUC and macro-F1.
In briefIn the Guise of Correction: How Framing and Cultural Adaptation Weaken LLM Safety in Real Criminal Cases
SummaryThe paper introduces GUISE, a benchmark of criminal-assistance scenarios drawn from real cases and organised by the UNODC's International Classification of Crime for Statistical Purposes. Each scenario is tested under three linguistic framings and across five cultures. Correction-seeking framings raised the mean unsafe compliance rate from 28.9% to 66.4%, and culturally adapted prompts drew more unsafe responses than literal translations. The authors also report that Anglocentric features in unsafe responses grow with harmfulness, and they release their code and data.
WhyBuilds a real-case multilingual safety benchmark showing correction framing and cultural adaptation significantly erode refusals, with statistically tested, fairly thorough evidence.
Mohsinul Kabir, Md. Tariquzzaman, Karam S. R. Elradie et al.
In briefNICE Toolbox: an open-source Python tool that automatically measures nonverbal communication signals such as posture, gesture, and gaze.
SummaryThe paper presents the NICE Toolbox, a free, open-source Python framework that automatically measures nonverbal interpersonal communication signals. It combines existing computer vision and speech-processing methods into one package, with support for multi-camera systems, interchangeable detectors, 2D and 3D visualization, and export to third-party software. The authors aim to make data exploration of interpersonal communication scalable, reproducible, and accessible for psychological research.
Aleksandr Evgrashin, Carolin Schmitt, Timo L\"ubbing et al.
In briefTeaching vision-language models to explore 3D scenes through action
SummaryThis paper studies how vision-language models can think and actively explore 3D space by viewing one first-person image at a time and moving their viewpoint. After trajectory-supervised finetuning, the authors apply reinforcement learning rewarded mainly by final-answer correctness, so the model learns its own exploration policy. The resulting model, ACT3D, reaches 53.2% average accuracy across six 3D benchmarks while using 50-70% fewer input frames than existing open-source 3D methods.
In briefDINO-FIL: multi-layer transformer fusion for locating AI-generated inpainting edits in images
SummaryThe paper presents DINO-FIL, a transformer-based model that pinpoints pixel-level regions in images altered by generative inpainting. It combines features from multiple layers of a pretrained DINOv3 encoder with a lightweight fusion decoder to capture both semantic structure and subtle texture inconsistencies. Trained on B-Free and tested on four other datasets, it outperforms existing baselines on average and on four of five datasets. It also keeps a high true negative rate on pristine images.
Simon Freitter, David Fischinger, Silvia Poletti et al.
In briefPW-Pruning: Phase-Aware Expert Pruning for Speech-Centric Omni-Modal Models
SummaryThis paper studies how to prune experts in speech-centric omni-modal models with sparse routing to reduce memory use. It proposes PW-Pruning, which builds separate importance-weighted kernels for acoustic-input processing and text-answer generation, then selects one static expert bank. On Qwen3-Omni, keeping 32 of 128 experts per layer improves MMAU-mini and MMSU accuracy over matched pooled selection by 3.20 and 4.38 points after supervised fine-tuning.
In briefFuseAdapt: Parameter-Efficient Multimodal Fusion for Segmentation with Missing Sensor Data
SummaryThis paper presents FuseAdapt, a framework that fuses multiple sensor modalities for semantic segmentation inside a frozen vision foundation model using lightweight adaptation branches. It also introduces Predictive Modality Modeling (PMM), a training-only objective that predicts a masked modality's embedding from the visible ones. Across three benchmarks (MCubeS, DELIVER, MUSES), the method reports gains over prior state-of-the-art of up to +3.01 mIoU with full modalities and +12.56 mIoU with missing modalities.
In briefCF-AVSS: A Coarse-to-Fine Audio-Visual Speech Separation Framework for the Real-World AVSE Challenge
SummaryThis paper proposes a two-stage coarse-to-fine audio-visual speech enhancement framework that uses visual cues to separate target speech from mixtures affected by reverberation, noise, and overlapping speakers. The first stage produces coarse estimates with a dual-visual-stream progressive network, and the second stage refines the target speech using the original mixture and these estimates. The two results are fused using development-set calibration. The system tied for first on Track 1 and ranked second on Track 2 (OVRL) in the ISCSLP 2026 Real-World AVSE Challenge, with the lowest CER on both tracks.
In briefEvaluating AI agents on whether physics research reaches well-supported conclusions
SummaryThe paper introduces PhysClosureBench, 50 expert-verified physics research tasks for testing whether AI agents can organize investigations that lead to supported conclusions. It scores local scientific accomplishment with a 20-criterion Atomic Score and research completeness with Research Closure levels RC0 to RC4. Seven frontier LLMs were evaluated, and the best mean Atomic Score reached RC4 in only a minority of attempts, showing a gap between local accomplishment and full closure.
In briefQ($\gamma_H$): Entropy-guided mixing of planning horizons in Q-learning
SummaryStandard Q-learning uses one fixed discount factor, which sets a single planning horizon throughout training. This paper proposes Q($\gamma_H$), which maintains several heads with different discount factors and combines them at control time using an entropy-based confidence signal. As policy confidence rises, the mixing weights shift toward longer-horizon heads, without changing individual Bellman updates. On classical control and deep reinforcement learning benchmarks, the largest gains appear on delayed-credit tasks compared with strong fixed-discount baselines.”
In briefReducing Gender Bias in Audio LLMs via Adversarial Training and Distillation
SummaryAudio-enabled large language models let users speak to them directly, but non-verbal cues in speech may lead to biased responses based on inferred gender. The paper uses adversarial methods to make models unlearn gender perception without hurting performance on other benchmark tasks. It also distils the gender-agnostic model into one that perceives gender but responds similarly to the teacher unless explicitly asked. Tested on three audio LLMs, the gender-unlearning results were satisfactory and the distillation results partially satisfactory.
Adnan Al Ali, Jind\v{r}ich Libovick\'y, Anne Lauscher
In briefUsing multimodal AI models to detect fake, AI-generated face images
SummaryThis paper tests whether six multimodal large language models can identify synthetic images by spotting physical inconsistencies or digital artefacts. The models were evaluated on 400 images, including 50 real photos and 350 generated by seven image models, across single-image, pairwise, and self-detection experiments. GPT-5.2 reached 97.9% accuracy in pairwise comparison against specific generators, with a mean pairwise accuracy of 82.5%. The study also found that invisible SynthID watermarks reduced Gemini's detection accuracy by about 10% compared with a clean control group.
Olivia Salamon, Virginia N. L. Franqueira, Shujun Li
In briefBeyond Source Prompts: Visually Grounded Quality Checks for AI-Generated Videos
SummaryThis paper introduces VGCap, a framework for assessing the quality of AI-generated videos. It replaces the original source prompt with visually grounded captions: a global caption minimally revised from the prompt using multi-frame visual evidence, and segment captions describing localized temporal content. These captions support evaluation of spatial quality, temporal quality, and text-video alignment. VGCap outperforms representative baselines on the LGVQ and FETV datasets across all three dimensions.
In briefDeciCAD: Learning Action Policies for CAD Agents Within a Dedicated Harness
SummaryThis paper introduces DeciCAD, a domain-native harness for computer-aided design (CAD) agents that exposes semantic CAD actions, decision-oriented observations, and persistent state transitions. The authors extract causal decision supervision from interaction trajectories and improve the policy through supervised fine-tuning and geometry-grounded online reinforcement learning. Across four CAD benchmarks, DeciCAD achieves the best F1 and IoU while cutting model-output tokens by 45 to 88 percent relative to a general-purpose coding harness.
In briefHarmful partial interventions and benign leakage in concept-based models
SummaryThe paper studies how human interventions on intermediate concepts in concept-based models can counterintuitively worsen predictions, a failure called intervention harm. It proves that sufficiency and localisation conditions do not guarantee that concept interventions improve performance, and finds that incomplete concept supervision is associated with more harm. The authors introduce a monotonic intervention objective that penalises loss increases along partial interventions, which together with endpoint training substantially reduces such harm.
Gregory Verghese, Christopher J. Soelistyo, Mateo Espinosa Zarlenga et al.
In briefDormant Spiking Neuron: A State-Driven Way to Save Computation in Spiking Networks
SummaryThe paper introduces the Dormant Spiking Neuron (DSN), which lets a neuron enter a dormant state when its membrane potential drops below a set negative threshold, skipping redundant computation. The mechanism is trained end-to-end with backpropagation through time using surrogate gradients. The authors report computational energy reductions of 40% to over 70% across static and neuromorphic datasets, and on ImageNet they report outperforming static pruning methods while improving baseline accuracy by 1.34%.
In briefThe Delegation Dilemma: Economies of Scale, Accountability, and Polycentric Governance
SummaryThis paper studies collective-action systems in which agents delegate decisions to self-interested representatives who may not share their constituents' objectives. Using a repeated game with endogenous delegation, the authors identify a scale–accountability tradeoff and show that the threat of exit sustains non-exploitative delegation only up to an endogenous constituency size. Simulations with 30 LLM agents show polycentric structures repeatedly emerge but remain fluid, cycling through concentration, exploitation, exit, and re-formation.
In briefPairwise AUC Optimization Needs Corrective Power: A Unified View
SummaryThis paper examines why directly training with pairwise AUC surrogates, which match the test-time ranking metric, may fail to consistently correct severe positive-negative ranking errors during stochastic mini-batch updates. The authors introduce the notion of Corrective Power and use degenerate properties of U-statistics formed from pairwise gradients to explain the observed behavior. The analysis offers mechanistic explanations for why squared-hinge loss works well, why cross-entropy warm-up can help, and why parameter-efficient fine-tuning from a pretrained model can reduce that need. Experiments on 8 CV tasks and 8 NLP tasks using ResNet50, DenseNet121, DistilBERT-base, and Qwen2.5-1.5B-Instruct support these predictions.
In briefSparse federated echo state networks for time-series classification using Smooth-L1 regularization
SummaryThis paper proposes a Smooth-L1 regularized Echo State Network (SL1ESN) that makes the readout layer sparse while remaining differentiable for gradient-based training. The method is used both for centralized training and for a second-order Newton-step federated aggregation scheme. On seven time-series classification benchmarks, the readouts are 75 to 95 percent sparse, with accuracy close to the strongest dense baseline. The authors argue sparse readouts are practical for edge and IoT deployments.
Gengcheng Lyu, Manuel R\"oder, Frank-Michael Schleif
In briefRIKA: a method for reliable knowledge gathering in continually learning AI agents
SummaryThe paper finds that about 12–33% of entries in representative agent memory systems harm future task performance when reused. It proposes RIKA, which hierarchically extracts factual, conceptual, procedural, and metacognitive knowledge. Its Evidence-based Knowledge Admission uses source evidence and later experience to decide which claims enter the agent's context, without extra validation tasks. RIKA achieves the highest aggregate score on all four model backbones across six task streams.
In briefSceneWeave: Exploring Alternative Scenes to Guide Script Adaptation
SummarySceneWeave is a method for adapting prose into a script. It writes contrasting drafts of a key scene, compares the interactions they produce, and revises characterization, setup, later development, and scene order before generating the full script. Blind reader comparisons favor it over a baseline workflow and description-based controls, and content assessment reports less narrative redundancy and strong psychological treatment.
In briefLearning Preference-Based Pareto Solutions from Limited Offline Data Using In-Context Learning
SummaryThis paper introduces ICPSL, a framework that maps user preferences directly to Pareto optimal solutions when only a limited offline dataset is available. A hypernetwork is pretrained on large synthetic prior data and, for a new problem, generates a problem-specific preference-conditioned solution model in a single forward pass. Experiments on synthetic and real-world problems report high-quality Pareto solutions and substantial speedups in building the problem-specific model.
In briefRetrieved but Not Read: Why Search-Augmented LLMs Fail to Use Available Evidence
SummaryThis paper studies why search-augmented large language models sometimes give outdated answers even when relevant evidence is retrieved. The authors build temporal facts whose values changed after training cutoffs and evaluate one proprietary model and six open-weight configurations. They find models often answer from search snippets without opening source pages, which is linked to more outdated answers and unfaithful citations. Layer-wise analysis reveals a gap between evidence availability and its effective use.
In briefFIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning
SummaryThis paper addresses removing targeted knowledge from large language models trained federatedly, where forget and retain data stay on private clients. The authors propose FIRM, a single-round server-side aggregation method that uses client-local Fisher statistics to normalize merged updates and protect retain-sensitive parameters. Experiments on WMDP, TOFU, and MUSE-HP with Zephyr-7B and Llama-3-8B show utility gains at comparable forgetting on WMDP and MUSE-HP, with metric-dependent trade-offs on TOFU.
Reeyansh Shah, Maram Hasan, Subhasis Chaudhuri et al.
In briefMulti-Centre Validation of Unsupervised Domain Adaptation for Normal Versus Abnormal Retinal Development
SummaryThis study develops an unsupervised domain adaptation model that classifies normal and abnormal retinal development across different optical coherence tomography devices. It uses retrospective scans from five centres in four countries and the UK Biobank, with a ResNet50-based domain adversarial network trained on labelled source-device images and unlabelled target-device images. The best model reached 96.06% accuracy and an AUC of 0.994 on validation data. The authors report that classification across devices was possible without labelled target-device data.
In briefGANs with Monte Carlo dropout uncertainty for structural health monitoring: fibre Bragg grating data augmentation on a Hawk T1A aircraft
SummaryThis paper presents a data augmentation framework that learns class-conditional temporal and spectral characteristics of aerospace structural responses. It combines conditional generative adversarial network training with Monte Carlo dropout as a Bayesian approximation to synthesise signals and quantify uncertainty. The generated signals are compared with experimental measurements across several domains and assessed for use in supervised damage classification. The framework is validated on a full-scale BAE Systems Hawk T1A aircraft benchmark dataset including acceleration, fibre Bragg grating strain and nonlinear dynamic response.
In briefRobust covariance estimation for data with entry-wise outliers and missing values
SummaryThis paper studies how to estimate a covariance matrix robustly when data entries are contaminated and some values are missing. It extends the Student-t* model to incomplete observations using a stochastic approximation EM (SAEM) algorithm on a reduced latent space. Experiments on synthetic and real sensor data report lower Fisher-Rao error than Gaussian, Student-t, imputation-based and detect-then-estimate estimators at moderate to high outlier rates.
In briefUsing human interventions as labels to locate and steer why LLM agents fail
SummaryThe paper treats human interventions in LLM agent failures as measurements pointing to the internal computation that went wrong. It studies one failure, asking a clarifying question when the request was already unambiguous, in a controlled agent environment with rule-based labels validated by blind human judgment. Among 13 models, five qualified, and the ask-versus-act decision was causally localizable in every eligible family. The authors also report mechanistic signatures of over-asking, probe-gated steering results, and mixed outcomes when consolidating the fix into model weights.
Cesare Zavattari, Alessandro Tommasi, Giuseppe Prencipe
In briefSynthesizing contrast-enhanced MRI from non-contrast scans using feature differences
SummaryThis paper proposes a method to synthesize contrast-enhanced MRI (CE-MRI) from non-contrast MRI sequences, aiming to reduce the health risks and costs of contrast agent injection. It introduces a Mixture of Feature Difference (MoFD) module that uses dynamic weighting to select and combine features from individual sequences and their differences. On the UCSF-PDGM and UPenn-GBM datasets, the method outperforms baselines in image fidelity and perceptual quality, reconstructs tumor regions, and achieves segmentation performance comparable to real CE-MRI.
In briefClosing the Validation Gap for Markerless Spine Tracking Using Intraoperative X-ray
SummaryThe paper presents a clinical pipeline that derives per-vertebra 6-DoF pose from routine intraoperative C-arm fluoroscopy and aligns it to an RGB-D camera frame. This provides an independent subsurface reference for validating markerless tracking methods. Ex vivo tests on ovine vertebrae showed end-to-end accuracy of 2.06 ± 0.29 mm and 3.35 ± 0.93 degrees. The authors also evaluated two markerless registration strategies on clinical data from 11 vertebrae across three surgical sequences.
Connor Daly, Salvatore Russo, Jinendra Ekanayake et al.
In briefShiftQAT: Training Edge CNNs to Meet a Fixed-Function Accelerator's Numerical Rules
SummaryThe paper presents ShiftQAT, a quantization-aware training method for MAX7800x-family edge CNN accelerators that produces networks satisfying the hardware's numerical constraints, such as power-of-two, symmetric scales and one shift per layer. On ImageNet, its W8A8 accuracy drop matches less constrained QAT methods, and at W4A8 it stays within 2.9 points of FP32. On MAX78002 silicon, MobileNetV2 and ResNet-18 classify ImageNet images in 14.6 and 8.4 ms at 5.1 and 2.8 mJ per inference.
Oguzhan Buyuksolak, Osman Erman Okman, Mehmet Gorkem Ulkar
In briefAbstracts or full text? Testing input scope and speed in biomedical plain-language summaries
SummaryThis study compares generating plain-language summaries of biomedical papers from abstracts versus full texts, and few-shot prompting versus task-specific fine-tuning. It uses two public datasets and two recent models, evaluating readability, factual consistency, semantic similarity, and runtime. Abstract-based summarization performs better overall. Abstract fine-tuning runs 4.70 times faster, and abstract zero- and few-shot prompting 5.31 times faster, than full-text methods.
In briefMARS: A Momentum-Aware Rating System for Personalized Adaptive Testing
SummaryThis paper introduces MARS, an interpretable adaptive assessment system that models the question bank as a directed weighted knowledge graph. It updates learner proficiency using dynamic rating sensitivity, response quality, and momentum-aware learner velocity. Under a time-budgeted, negative-marking setting, MARS produces smoother rating updates than Fixed Elo and improves the total score by 8.09 over the strongest baseline, while reducing overtime violations from 50% to 5.6%.
