In briefS-EDL: Using Sequence Likelihoods to Calibrate LLM Meaning-Level Confidence
SummaryCalibrating large language models in open-ended generation is hard because many different token sequences can express the same meaning. S-EDL is a method that draws evidence from the model's own sequence likelihoods and builds a differentiable Dirichlet prior over prompt-specific semantic classes. Across four open-ended benchmarks and three models, the authors report that S-EDL substantially reduces semantic calibration error while improving accuracy.
In briefFinding Hidden Chest X-ray Model Failures Using Frequency Analysis
SummaryThis paper studies how deep learning models for chest X-ray anomaly detection can fail under subtle data shifts. The authors convert images to the frequency domain, summarize signal energy across frequency bands, and apply unsupervised clustering. They report that model failures concentrate in specific spectral clusters, which helps identify blind spots and predict reliability without ground-truth failure labels.
In briefRobot Generalization Beyond Data Volume: A Diversity-Driven Path Through Zero-Shot Capability
SummaryThe paper argues that zero-shot generalization remains a hard problem for robots, and that adding more data from the same few collectors expands coverage of unseen situations only slowly. It proposes a shared latent space for robot actions combining multi-modal alignment, generative augmentation, and noisy-label learning. The authors suggest a community of developers building on one standard robot, Origami, as a scalable source of diverse data. They close with a testable hypothesis and an evaluation protocol.
In briefComparing Early and Late RGB-Infrared Fusion for Detecting Drones
SummaryThis paper studies how combining visible (RGB) and infrared (IR) images can improve drone detection in harsh environments and weather. It compares image-level early fusion, which merges both spectra into one image for a single detector, with decision-level late fusion, which blends the bounding boxes from two detection pipelines. Early fusion achieves better raw performance, while late fusion remains useful when sensors or frames are misaligned.
Alberto De Zan, Denis Tavaris, Gian Luca Foresti et al.
In briefGuidedSAM-Plume: An Expert-Guided Tool for Segmenting Volcanic Plumes in Images
SummaryThis paper presents GuidedSAM-Plume, an open-source interactive framework for segmenting volcanic plumes and clouds in RGB images. It builds on the zero-shot, spatial prompt capabilities of SAM2 and needs only minimal user input to outperform fully automatic dedicated methods. The authors also release a re-annotated dataset created with expert volcanologists.
Taddeo D'Adamo (LIRIS, imagine), Simon Thivet (UNIGE) et al.
In briefMemoryTax: The Bitter Lesson of Agent Memory
SummaryThis paper tests whether designed memory systems such as Mem0 and Zep still help AI coding agents once models can search their own history. It compares four memory methods and three coding agents on one LoCoMo conversation, recording costs for each run. The authors report that memory systems mainly help weaker models at high cost, and that memory quality is capped by the model that writes it. Pi with Claude Opus 5.5 gave the best cost-accuracy result.
In briefUserSWE: A Benchmark for Coding Agents in Multi-Turn Conversations with Simulated Users
SummaryThis paper introduces UserSWE, a benchmark for testing whether coding agents can fix repository-level bugs through multi-turn conversations with simulated users. Tests are run after every turn, and verdicts are grounded in execution results. Eleven agents were evaluated on 72 tasks from 42 repositories, and their rankings changed with the number of allowed interaction turns. The authors conclude that single-turn performance is not enough to measure interactive coding ability.
In briefPrediction Is Not Calibration: Checking Whether Knowledge Tracing Models Recover Item Parameters
SummaryThis paper examines whether knowledge tracing models trained only to predict the next student response can also recover item parameters such as difficulty and discrimination. The authors pair 2PL, GPCM and NRM decoders with LSTM, Transformer and DKVMN encoders and test them in simulations. Difficulty-related parameters are recovered well, while discrimination and option slopes need wider item representations. Separating the sequence representation from the item parameter key improves recovery without reducing test accuracy.
Wenrui Yuan, St\'ephanie M. van den Berg, Johannes Steinr\"ucke
In briefCOMET: A Multi-Layer Runtime Monitor for Human-Supervised AI Safety Systems
SummaryThis paper presents COMET, a runtime monitoring framework with three layers (signal integrity, physics consistency, and the AI's competence envelope) whose outputs are combined and shown to human supervisors on a traffic-light display. Tested on a compressed-air leak dataset, the bare AI system fails silently under all shift types, while COMET reduces trusted-wrong-answer rates to zero for covariate-shifted no-leak data and three of four sensor faults. Residual silent misses under semantic shift mark the limit of passive monitoring.
In briefNeuroCodec: EEG-Guided Generative Model for Clearer Target Speaker Extraction
SummaryThis paper introduces NeuroCodec, a generative framework that uses EEG brain signals to extract a target speaker's voice in multi-talker settings, aimed at hearing assistive devices. It generates speech in the latent space of a neural audio codec, using graph convolutions for EEG and attention plus Mamba modeling to fuse neural and speech features. On the Cocktail Party and KUL datasets it outperforms baselines, and a 20-participant listening test gave it a mean score of 82.24 versus 49.63 for M3ANet. The authors also report 44.1 kHz reconstruction and a causal streaming variant for real-time use.
Julian Silva, Jaliya Kodithuwakku, Navindu Sirimanna et al.
In briefReal PMIDs but Irrelevant Papers: Harness Design for Biomedical Claim-Verification Agents
SummaryThe study runs a controlled ablation on biomedical claim verification (SciFact), using only the live PubMed API as a tool, and compares four setups: closed-book answering, single-shot retrieval, a ReAct-style agent that writes its own queries, and an oracle given gold abstracts. The self-querying agent found gold papers most often and gave the best verdicts, though differences were not statistically significant at this sample size. Giving gold abstracts raised accuracy to 79-86%, indicating retrieval is the main bottleneck. A citation audit found closed-book models cite PubMed IDs that almost always exist but are almost never relevant, so existence checks alone are insufficient safeguards.
In briefRecognizing new object-attribute combinations by reusing primitive visual concepts
SummaryThis paper studies compositional zero-shot learning, which recognizes unseen combinations of seen objects and attributes. The authors show that directly using decomposed primitive representations from vision-language embeddings suffices for recognizing unseen compositions, without explicit re-composition of pairs. They analyze how Euclidean, spherical, or hyperbolic embedding geometry and dataset structure affect compositional reasoning. Experiments span several models and benchmarks, including MIT-States, VAW-CZSL, Waterbirds, and CelebA.
In briefAsymmetric Likelihood-Free Hypothesis Testing: Telling Which of Two Distributions a Sample Comes From When Reference Sample Budgets Differ
SummaryThe paper studies likelihood-free hypothesis testing where the two reference distributions may be sampled with different budgets, and characterizes which budget combinations are feasible up to logarithmic factors. It shows that unequal reference budgets can yield a polynomial reduction in the samples needed from the unknown distribution compared with existing baselines. The results are applied to non-adaptive distribution 2-clustering, with matching upper and lower bounds, and a worst-case adaptivity gap of order k^{1/4} is obtained by comparison with adaptive allocation.
WhyIntroduces asymmetric likelihood-free testing with polynomial sample savings, matching upper and lower bounds, and applications to non-adaptive clustering.
In briefMAPS: A method that adjusts search momentum in retrieval-augmented generation through parallel reasoning.
SummaryThis paper introduces "search momentum," which describes how retrieval tendencies accumulate from earlier sub-queries in a reasoning process. The authors observe that query generation tends to keep evolving along existing retrieval paths and propose MAPS, which retains retrieved knowledge and refined reasoning states and runs retrieval in parallel. The authors report that MAPS outperforms representative RAG methods across different backbone models on six question-answering benchmarks.
In briefUncertainty-aware pseudo-labeling of object detection datasets using open-vocabulary models and label propagation
SummaryThis paper presents a semi-supervised framework that assigns soft labels to unlabeled objects in object detection datasets. It uses open-vocabulary detectors to produce region proposals and object representations, then propagates probabilistic labels over an object-level similarity graph while quantifying aleatoric and epistemic uncertainty. On COCO, the method reaches 40.5% AP at about 1% of labeling cost, nearly matching the fully manually labeled model at 40.9%.
SummaryThis paper examines why large language models often recommend AI-related options when asked about study, careers, or products. The authors compare models before and after fine-tuning and vary how questions are phrased. Pretrained models already include explicit AI/ML terms in 27% of responses, rising to 59% in instruction-tuned versions. Removing the 7% of instruction examples about AI lowers the rate by 14 percentage points.
In briefEvolving with Open Worlds: Recursive Self-Improvement for Future Forecasting
SummaryThe paper introduces EvoForesight, an agent system that improves through future event prediction by interacting with a continuously changing environment and turning its reasoning experience into reusable capability. On the weekly-updated external FutureX leaderboard, its self-improved forecaster ranked first among all systems for three consecutive weeks, with rating gains of up to 195%. On the authors' synthesized evaluation, it ranked in the top three across four base LLMs and outperformed the strongest baseline.
WhyProposes a new recursive self-improvement framework for future forecasting, ranking first on the external FutureX leaderboard for three weeks.
In briefJointLight: Relighting Human Videos via Asynchronous RGB-Irradiance Joint Modeling
SummaryThis paper presents JointLight, a method for relighting human-centric videos with finer illumination control. During two-stage training, the model jointly denoises RGB frames and ground-truth irradiance frames on separate timestep schedules, and a domain matching strategy bridges the gap between irradiance cues and ground-truth maps. The authors built a multi-illumination synthetic video dataset with paired irradiance maps using a game engine and trained only on it. The model is reported to generalize to real-world scenes and to outperform prior approaches in lighting fidelity and generative quality.'
In briefKey findings for small drones assessing post-disaster damage from the CURSE'26 exercise
SummaryThis paper evaluates the CLARKE AI classifier for building and road damage assessment during the 2026 Catastrophic Unmanned Remote Sensing Exercise (CURSE'26). The exercise trained 23 small drone teams in Florida and Missouri and tested the model on unseen orthomosaics. Building damage assessment reached a Micro F1 of 0.834, while road damage assessment performed poorly. The paper also reports findings on operations, flight timing, edge computing, and ethical and legal considerations.
Thomas Manzini, Stephen Johnson, Priyankari Perali et al.
In briefUniRAP: A Unified Model for Part-Level Physical Affordance Reasoning and Actionable Perception
SummaryUniRAP is a model that links language instructions to part-level physical interaction geometry. It combines visual inputs, text, and optional prompts in a shared representation and uses a Unified Affordance Decoder to perform object detection, part affordance segmentation, and 4-DoF interaction pose estimation. The authors also introduce a curriculum-based transfer training strategy to adapt the model from general visual parsing to interaction-aware perception.
In briefALIAS: Evaluating Performance and Memorization in LLM-Written TLA+ Specifications
SummaryThis paper tests whether large language models rely on memorized answers in benchmarks, using consistent identifier renaming on TLA+ tasks. A model checker confirms whether renamed tasks keep the original behaviors. The authors find that much of the score drop after renaming comes from grading that requires exact names, so a lower score mixes memorization, grading effects, and the cost of new names.
Arslan Bisharat, Mujtaba Nazari, Khushboo Bhadauria et al.
In briefLinking basic language model skills to agent performance
SummaryThis paper studies how the foundational abilities of large language models translate into their performance as agents. The authors propose an operation-level framework that builds a foundational profile from capability benchmarks and an operation profile from interaction trajectories, and defines capability realization gaps as deviations between observed and predicted operation performance. Experiments on fifteen model services in τ²-Bench show that foundational profiles only partially predict agentic operations, while operation profiles substantially improve task-score prediction.
In briefMetamodel-guided LLM decoding that keeps domain-specific language outputs consistent
SummaryLarge language models are used to generate software artifacts in domain-specific languages, which must follow both concrete and abstract syntax defined by a metamodel. The paper presents MeMoDecode, a decoding framework that checks, each time a complete metamodel concept is generated, whether the output can still be completed into a consistent instance, and retries from the last completable point if not. Tested on four DSLs and six open-weight models from 0.6B to 70B parameters, it improves consistency by 23% on average over unconstrained decoding and 8% over grammar-constrained decoding, with modest runtime overhead.
Boqi Chen, Aren A. Babikian, Jos\'e Antonio Hern\'andez L\'opez
In briefLEAF-Mamba: A method that adapts and fuses features for analyzing whole slide pathology images
SummaryThis paper addresses the difficulty of learning from gigapixel whole slide images (WSIs) with limited annotations by proposing LEAF-Mamba, a framework for slide-level learning on frozen pretrained features. It uses a reconstruction-regularized adapter to learn compact, task-specific representations, and combines parallel Mamba and convolutional streams through an input-dependent gate. Across several feature extractors, the method reportedly achieves better average performance than strong multiple instance learning baselines on both diagnostic classification and survival prediction.
In briefAPED-CXR: Image-only chest X-ray lesion localization via adaptive privileged evidence distillation
SummaryThis paper introduces APED-CXR, which distills report-informed spatial evidence from a frozen multimodal teacher into an 11.1M-parameter student model. At test time the student predicts diseases and bounding boxes from the image alone. During training, the teacher scores regions proposed from the student's current evidence maps, and these scores serve as targets. Trained on PadChest and evaluated on PadChest-GR, box-level localization F1 rises from 0.302 to 0.338 while diagnostic AUC is maintained.
Md Shihabul Islam Shovo, Nishi Kanta Paul, Jamil Fayyad
In briefLearning to Measure and Diagnose: A Clinical-Dictionary Radiomic Bottleneck Model for Thyroid Nodule Classification
SummaryThis paper proposes a model that classifies thyroid nodules from ultrasound images using radiomic measurements tied to clinical dictionary terms, so predictions can be explained in familiar clinical language. The model combines a family-aware method for estimating these measurements, a fidelity-guided bottleneck for joint optimization, and a decomposition that traces each feature's contribution to the linear classifier. Experiments on ThyroidXL and TN3K show competitive classification performance, and case studies illustrate feature contributions for benign and malignant nodules.
In briefMonitorability Can Be Trained: Post-Training to Make Chain-of-Thought Easier to Monitor
SummaryThis paper studies whether an AI model's chain-of-thought (CoT) stays useful for monitors, since it must remain tied to what the model actually does. The authors propose an unsupervised post-training objective that maximizes mutual information between the CoT and the output. Across three models with degraded CoT, it recovers most or all of the lost monitorability, and it also improves monitorability for models whose CoT was not degraded.
In briefSAKURA: Bandwidth-Adaptive Attention Cache Delivery for Private On-Phone Image Editing
SummaryThis paper presents SAKURA, a cloud-edge system that delivers reusable attention key-value caches to phones under a bandwidth budget, so private image editing such as face swapping and virtual try-on can run on the device. It uses token-importance-aware compression to preserve image quality and selects cache representations per shard based on current network bandwidth. A pipelined runtime overlaps cache transfer and reconstruction with on-device execution. On a Snapdragon 8 Elite Gen 5 phone, the authors report 1.82x and 3.27x latency reductions over CacheGen adapted to DiT and on-device generation, and a 3.06x storage reduction over uncompressed cache.
In briefSingle-policy uncertainty estimates for robot vision-language-action models
SummaryThe paper proposes post-hoc kernel-variance heads that estimate epistemic uncertainty from the embeddings of a single vision-language-action (VLA) policy, requiring no regression targets or gradient updates and leaving the policy unchanged. In experiments with SmolVLA on LIBERO-10, these heads perform competitively with two-policy ensemble scores on task-success correlation, active-learning demonstration acquisition, and referral to a stronger supervisor policy. The backbone head reduces scoring time by up to about 50 times. The authors also show that controlled perturbations offer a complementary evaluation, identifying unfamiliar camera, instruction, or state inputs.
Johannes Keller, Albrecht M. Blo{\ss}e, Adrian Lindenmeyer et al.
In briefAccept or Repair: keeping passage-backed answers in contrastive decoding
SummaryLanguage models often answer from training knowledge even when a passage says otherwise. The paper proposes Accept-or-Repair (AoR), a training-free method that keeps a sampled answer if the passage supports it, and otherwise tries up to two contrastive-decoding replacements. Experiments report gains over sampling of 12.5 points on counterfactual passages and 2.3 on ordinary ones, outperforming eight baselines.
Jungseob Lee, Seungyoon Lee, Hyeonseok Moon et al.
In briefSpectral Uncertainty Split for Sparse Bayesian Learning in Brain Source Imaging
SummaryThis paper addresses uncertainty quantification when estimating neural sources from M/EEG measurements, an ill-posed inverse problem. The authors use the singular value structure of the prior-preconditioned forward operator to exactly split the posterior covariance into a data-informed part and an indeterminacy part that sensing geometry cannot reveal. They also derive a subtractive form that writes the prior covariance as the posterior plus a data-driven reduction term.
In briefImproving graph neural networks for Petri nets by capturing loops and invariants
SummaryThe paper proposes two ways to help message-passing graph neural networks better understand Petri nets, which are process models. PNRWPE uses return probabilities of directed place-transition walks to capture process loops, while DA-PNGNN gates same-type messages with diffusion kernels of the incidence matrix to expose invariants. On the 96 models from Process Discovery Contest 2023, macro balanced accuracy rises from 0.751 to 0.796 with PNRWPE and to 0.830 with DA-PNGNN. DA-PNGNN also improves net-to-tree retrieval top-1 accuracy on the ProcRosetta benchmark by 4.1 to 5.9 points without retuning.
In briefDo RAG techniques still work when combined? A study of how retrieval-augmented generation components interact.
SummaryThis paper examines whether individually effective techniques in retrieval-augmented generation (RAG) pipelines keep their gains when combined. The authors introduce LASAGNA, an open-source framework, and test 450 configurations across eight benchmarks and corpora of 5K to 200K documents. They find that combined gains are much smaller than predicted, that optimizing retrieval and faithfulness can work against each other, and that the best composition changes with corpus size.
WhySystematically tests RAG component composition across 450 configurations and corpora up to 200K documents, showing isolated gains shrink when combined.
In briefForward-only training of shared-kernel wavelet networks without backpropagation
SummaryThe paper proposes a CMA-ES-based global forward-only learning framework for shared-kernel wavelet neural networks, avoiding backpropagation. It adds structure-aware parameter decomposition and channel-specific credit assignment, splitting the search into three subspaces. On BSDS500, test MSE drops from 4.1783 to 4.0130 and PSNR rises from 42.621 dB to 42.787 dB compared with Adam/BP. Training requires only 6 training-set coverages, versus 1500 epochs for BP.
In briefMoEZip: Routing-aware KV cache compression for sparse Mixture-of-Experts language models
SummaryThis paper addresses the memory bottleneck of long-context inference in sparse Mixture-of-Experts language models, where the KV cache grows with sequence length. It notes that evicting KV entries can change which experts are selected during answer generation, which may hurt output quality, and proposes MoEZip, which uses a Fisher-information-based sensitivity metric for routing combined with attention weights to score context. Experiments report that under aggressive cache compression, MoEZip improves on the state-of-the-art baseline by 24 and 20 points on average on LongBench and RULER, respectively.
In briefFaster generative image compression via refined quantized latents
SummaryThis paper proposes GR-GIC, an efficient generative image compression method aimed at strong perceptual quality at low bitrates with fast inference. It introduces Guided Refinement to improve reconstruction and reorganized projected residual vector quantization (P-RVQ) to increase parallelism and reduce computation. On Kodak, GR-GIC reports BD-rate savings over OneDC across LPIPS, DISTS, PSNR, and FID, and encodes and decodes in a few milliseconds on an NVIDIA A100.
In briefRethinking post-hoc debiasing for frozen large pretrained models
SummaryLarge pretrained models such as CLIP are often used as frozen feature extractors, but their embeddings can carry spurious correlations from training data. The paper finds that existing post-hoc debiasing methods require the bias to be identified and described in text, and on the UrbanCars benchmark they underperform a vanilla classifier even when true bias labels are given. The authors propose identifying cross-class training pairs that lie closest together in the frozen feature space, which usually share a spurious attribute, and learning a single linear map to debias the features. The method performs competitively without bias annotations or large language models.
Divyaansh Mertia, Chirag Bhutra, Himanshu Shekhar et al.
In briefScaling Multilingual Pretraining for Pixel Language Models
SummaryThis paper trains pixel language models (MPIXEL), which render text as images, while scaling languages, scripts, and model capacity. Evaluated on five tasks across 83 languages, scaling mainly improves the added languages and scripts, and larger capacity mitigates the curse of multilinguality. The authors also find that cross-lingual alignment weakens as more languages are added.
In briefTraceForge: Distilling with Your Own Reasoning Style
SummaryTraceForge generates reasoning traces itself, from a question and a teacher's answer, in the style of the student's own model family, instead of recovering hidden teacher reasoning. Training uses a supervised stage followed by an adversarial stage in which a critic judges trace authenticity. On coding benchmarks with the same 5,000 questions and teacher answers, the final checkpoint improves HumanEval+ by 3.7 points and MBPP+ by 7.1 points.
In briefExtending PCA to Wasserstein space (Lifted PCA) for distributional data
SummaryPrincipal Component Analysis is a basic dimension-reduction tool, but no standard extension to distribution-valued data existed within optimal transport. This paper introduces Lifted PCA, which extends Euclidean PCA to Wasserstein space via convex extension of functionals on probability measures, yielding an optimal measure supported on several low-dimensional subspaces. The authors derive dual formulations, optimality conditions, and statistical consistency results, and report numerical experiments on distributional datasets.
In briefToward recursive self-improvement in physical AI: robots that safely refine how they learn
SummaryThis paper positions recursive self-improvement (RSI), where a system improves both its behavior and the process that generates that behavior, within robot learning. It traces how physical experience can strengthen these improvement mechanisms and when resulting gains can support further improvement of the mechanisms themselves. Existing studies, mostly in manipulation, show behavioral improvement, benefits to later learning, and procedure improvement, but the reviewed physical studies have not yet isolated gains attributable to recursion. The authors synthesize these findings into evaluation principles and research directions.
In briefDriveMLP: a lightweight driving planner where training strategy matters more than model size
SummaryThis paper studies motion planning for autonomous driving and proposes Expert-Guided RL (EGRL), which uses reinforcement learning as an optimizer with a reward based on distance to logged expert trajectories. The authors find that the training procedure accounts for much of the performance gain rather than model size. They introduce DriveMLP, a lightweight MLP policy trained with EGRL, which scores 86.51 on the nuPlan Val14 Reactive benchmark, 3.2 points above the prior best imitation-learning planner Flow Planner, with over 15x faster inference.
Alessandro Canevaro, Julian Schmidt, Hang Yu et al.
In briefMulti-player bandits where players can't see each other: stationarity, traps, and certified global rates
SummaryThe paper studies a multi-player bandit model in which players pick arms simultaneously, observe only a shared reward, and cannot see other players' actions or even know whether they exist. It shows that the players' independent gradient-bandit updates together form an unbiased gradient estimator of a hidden product-policy objective, yielding a finite-time first-order stationarity guarantee. The authors also show stationarity cannot in general imply global optimality, and that global optimality can be recovered under structural assumptions on the reward table via finite-table certificates, which give bounds on the policy gap and pseudo-regret.
In briefUni-Mem: Learning a Long-Horizon Agent Memory Manager via Unified Online Policy Optimization
SummaryThis paper proposes Uni-Mem, a unified memory manager that uses one shared policy to learn both memory construction and retrieval for long-term AI agents. To handle confounded group comparisons and sparse feedback during joint training, it introduces Stage-Decoupled GRPO and Event-Wise Credit Assignment (ECA). Experiments on LoCoMo and LongMemEval show that Uni-Mem outperforms the strongest prior RL-based memory method.
In briefConflicting Records: Typed Adversarial Contracts for Auditable LLM Output Quality Control
SummaryThe paper addresses LLM-generated records that can simultaneously satisfy and violate quality rules, which classical cleaning and single LLM judges cannot resolve. It proposes BiasGuard-D, in which two adversarial roles from different model families exchange typed claim documents, a three-model quorum decides by majority vote, and a referential-plus-semantic invariant validates the verdict. The authors report that removing the iterative loop sharply reduces invariant satisfaction, and that the approach transfers from stereotype to hallucination remediation on HaluEval QA, raising right-answer recovery by 49 percentage points. All claims, votes, and instructions are stored as typed records to support auditing.
In briefReasoning Like a Cytopathologist: Multiscale Evidential Learning for Cervical Cytology
SummaryCervical cytology includes ambiguous categories such as ASC-US that are hard to classify definitively. The paper proposes a multiscale evidential framework that encodes nuclear morphology, cytoplasm, and neighbourhood cell structure separately, then fuses them using Dempster-Shafer belief masses. Cases with conflicting or insufficient evidence stay explicitly uncommitted rather than forced into a label. On the RIVA and CRIC Cervix datasets, incorrect predictions carried 1.94 to 4.13 times higher uncertainty than correct ones.
Ajeet Kumar Singh, Nehal Mishra, Anurag Tripathi et al.
In briefSPARE: A training-free method that allocates expert pruning across MoE layers using routing entropy
SummaryThe paper proposes SPARE, a training-free method that decides how many experts to remove from each Mixture-of-Experts layer based on routing entropy computed on calibration data. Layers with more diffuse routing are pruned more aggressively, and the method works with any expert selector. Under the same selector and budget, SPARE improves average accuracy over uniform allocation on Mixtral, Qwen3, and DeepSeek models, especially at high pruning ratios. At 50% expert pruning, Qwen3-30B-A3B's eight-task average rises from 57.2% to 63.8%.
In briefHow model size and quantization affect social bias and abstention in question answering
SummaryThe study tests Qwen3, Gemma-3, and SmolLM2 models of different sizes and quantization levels on the BBQ benchmark for question answering. Larger models achieved higher accuracy, but bias showed no simple scaling trend and quantization effects varied across configurations. Notably, 20 of 26 configurations with less bias also showed more abstention, suggesting lower observed bias can partly reflect greater abstention, so fairness evaluation should consider these factors together.
In briefProbabilistic machine learning for forecasting atmospheric seeing over short periods
SummaryThis paper benchmarks machine learning models for forecasting atmospheric seeing up to two hours ahead, with emphasis on probabilistic forecasts and uncertainty quantification. Models were trained and evaluated on seeing measurements from the Maunakea summit, comparing Gaussian Processes and a normalizing-flow model (FloTS) against RNN and LSTM baselines. Deep learning approaches outperformed standard recurrent networks, with LSTM and FloTS achieving the best predictive accuracy, and FloTS giving the most informative probabilistic forecasts.
Mary-Joe Medlej, Rahul Srinivasan, Simon Prunet et al.
In briefPer-person demographic steering to reduce bias in multi-person text-to-image generation
SummaryText-to-image models reproduce societal stereotypes, and existing fixes apply one demographic correction to the whole scene, which makes people within an image more alike. This paper proposes Spatially Gated Steering (SGS), which extracts a mask for each person from the Diffusion Transformer's attention maps at a critical denoising step and applies a distinct demographic vector within each mask. The method needs no fine-tuning, inversion, or layout specification. On two diffusion transformers, FLUX.1-Dev and SD3.5-Large, it raises intra-image racial and gender diversity on skewed occupational prompts.
In briefFeedback That Backfires: Do LLM Game Players Repeat Failed Moves?
SummaryThis paper examines whether LLM agents in games like Sokoban, Connect Four, and Othello are more likely to repeat a failed move after the failure is recorded in their transcript. Across eight small instruction-tuned models, the feedback record increased the probability of the failed move, with the effect stronger for illegal moves than for losing ones. The authors also report that a 'do not repeat' instruction had no effect and that a decoder ban stopped exact repeats but not illegal play.
WhyShows failure feedback raises repetition of failed moves in games, with mechanism analysis, training-stage effects, and multi-model, multi-game ablations supporting it.
In briefMS2CAST: a video anomaly detection framework using multi-scale cross-attention and efficient spatiotemporal networks
SummaryThis paper introduces MS2CAST, an end-to-end semi-supervised predictive framework for detecting anomalies in surveillance videos. It fuses features from a vision transformer and a CNN with a multi-scale cross-attention encoder, and models spatiotemporal dynamics in latent space without optical flow. Evaluated on UCSD Ped2, CUHK Avenue, and ShanghaiTech, it reports frame-level ROC AUC competitive with top methods while keeping peak GPU memory under 1 GB.
In briefReading the Room: Scale-Ambiguous Semantic Alignment of Sparse Visual SLAM Maps to Partially Built Floor Plans
SummaryThis paper addresses aligning a visual SLAM map, which lacks true scale, with a 2D floor plan on partially built construction sites. It proposes Reading the Room (RR), a global-local registration method that uses a single robust, differentiable, class-balanced semantic-geometric energy function throughout, with two global initialization schemes. Evaluated on 33 sequences across three active construction sites, RR reduces mean absolute trajectory error by 4.5 times over the strongest baseline without constraints, and by 7.6 times when an anchor is provided.
Bjorn Stragier, Emilia Szyma\'nska, J. Micah Prendergast et al.
In briefFrom Association to Decision Value: Testing Paired Instruction Dependence for VLA Corpus Selection
SummaryThe paper studies how to rank robot demonstration corpora, proposing paired instruction dependence, the rate at which a demonstrator changes its first decision when only the goal changes. Tested on 136 independently seeded quadruped-navigation corpora, the audit is reliable and tied to downstream paired correctness. However, its incremental decision value is limited, and it does not establish a selection gain in this setting.
Nurmukhambet Serikbayev, Abdul Basit, Muhammad Shafique
In briefUsing LLMs as survey respondents, compared with simple semantic baselines
SummaryThis paper examines whether large language models can predict answers to unanswered survey items from partial human responses, as a way to augment survey data. The authors tested three LLMs on two surveys about attitudes toward automated or light-electric vehicles and compared them with two embedding-similarity baselines. LLMs usually outperformed the baselines, but their accuracy dropped when fewer known responses were available, matching the baseline in one survey.
In briefJoint semantic-channel coding and modulation (SemQAM) for generative image transmission
SummaryThe paper proposes SemQAM, which jointly designs source coding, channel coding, and modulation by mapping vector-quantized image latents directly to QAM symbols. The method updates the codeword matching each detected symbol so channel impairments affect codebook learning, and uses the conditional expectation of received codewords to limit error propagation in autoregressive coding. Experiments report that SemQAM-C outperforms OneDC in LPIPS and DISTS under the same LDPC and QAM settings, and that SemQAM reduces channel uses by 18.57% at matched LPIPS compared with a separated counterpart.
In briefOpenEarth-OVS44: A Multimodal Satellite Image Dataset for Labeling Land Cover Without Fixed Categories
SummaryThis paper introduces OpenEarth-OVS44, a multimodal remote sensing benchmark with 290,587 expert-verified segments from 2,832 RGB images and 1,341 registered RGB-SAR pairs at sub-meter resolution. It covers 40 cities in France, Japan, and the United States, using a 44-class taxonomy grouped into 9 base categories, and provides three evaluation tracks. Experiments show that training on the dataset improves open-vocabulary segmentation over the strongest training-free baseline, by up to 7.1 mIoU on unseen categories and 27.6 mIoU on seen categories.
In briefRecursive Self-Improving Code for Task-Oriented Semantic Communication
SummaryThe paper introduces ComRSI, a framework that uses large-language-model-assisted code edits to recursively evolve the program of a semantic communication system, covering the encoder, channel adapters, decoder, loss, and training policy. A protected evaluator screens candidates under prescribed budgets and resource constraints. On CIFAR-100 over an AWGN channel at 32 complex symbols per image, the evolved design reaches 74.3% test accuracy, versus 67.2% for the initial design at 256 symbols. Across repeated searches, ComRSI's mean final score exceeded that of a memoryless non-recursive agent, while its difference from a history-equipped agent was not statistically conclusive at this sample size.
In briefTransfer Learning with a Transfer Efficiency Score for Cross-Area Visible Light Indoor Localization
SummaryThis paper addresses the performance drop of visible light communication (VLC) indoor localization across different deployment areas. It proposes a transfer learning framework, and experiments with real measurements from four areas of a Bosch manufacturing facility show that fine-tuning a pre-trained deep neural network with 40% of target-area data recovers 97% of full-data performance. The authors also introduce a Transfer Efficiency Score to select effective source models without exhaustive evaluation.
In briefFormalizing Cooperation and Competition: An AI-Assisted Addendum to a 2018 Planning Paper
SummaryThis paper offers an AI-assisted mathematical addendum to a 2018 thesis on strategic trip planning that balances competition and cooperation. It reconstructs relationships among strategic choice, cooperative stability, participant approval and enforcement, and verifies two-person results under fixed outsiders and explicit efficiency conditions. A finite negotiation procedure implemented in Lean is proved to terminate with participant approval, non-worsening joint cost and local optimality. Reproducible computations and counterinstances are provided, and the authors note that the mathematical ingredients are established results, with the contribution being a traceable reconstruction.
In briefHow space weather shapes where decameter-scale ionospheric irregularities appear
SummaryThis paper presents a data-driven method for nowcasting and forecasting the occurrence of high-latitude decameter ionospheric irregularities, using SuperDARN and SECIRA radar data together with space weather parameters. The model uses a neural network inspired by graph neural networks and was trained and tested on data from 11 radars covering 2010 to 2025. Nowcast performance is the highest, and acceptable forecast quality holds for lead times of up to 7 to 9 hours.
In briefReliability-weighted probabilistic fusion for multimodal learning with missing inputs
SummaryThe paper proposes a reliability-aware Gaussian Product-of-Experts method that uses input-dependent weights to adjust each modality's latent precision, omitting missing modalities. On two clinical multimodal tasks, the framework is competitive with complete inputs and shows its clearest mean-AUROC gain when modalities are absent, especially on HAIM-MIMIC-MM. The authors note that learning dependable quality signals remains an open challenge.
Patitapaban Palo, Abhisek Tiwari, Mattia Carletti et al.
In briefFogTTA: Online test-time adaptation to make transformer object detectors robust in fog
SummaryObject detectors for autonomous driving often perform worse in bad weather because their training data differs from real conditions. This paper proposes FogTTA, an online test-time adaptation framework that improves RF-DETR, a Transformer-based detector, in foggy scenes without source data or retraining. It uses a teacher-student design with Varifocal loss to reduce pseudo-label noise. Experiments reportedly show better accuracy and stability than prior baselines while keeping real-time performance.
In briefTernScope: Model-hardware co-design for efficient ternary endoscopic image segmentation
SummaryThis paper presents TernScope, a model-hardware co-design approach for medical image segmentation on edge devices, focusing on endoscopic images. The network uses ternary weights and 8-bit activations, trained with progressive quantization-aware training, to enable compact storage and multiplier-free computation. According to the paper, the model has 0.2483 M parameters and 0.0949 GMACs while retaining competitive accuracy, and at 45 nm it reduces compute-array area by 88.46% compared with the original Gemmini accelerator.
In briefInterpretable AI for early liver cancer diagnosis on multiphase CT
SummaryThis paper presents dLI-RADS, an interpretable AI method that estimates the probability of hepatocellular carcinoma from manually segmented liver lesions on multiphase CT. It combines neural network predictions with handcrafted measurements of LI-RADS major features using logistic regression. In a multicenter retrospective study, lesion-level accuracy was 80% for dLI-RADS versus 74% for LI-RADS on the internal set (p = 0.04), showing non-inferiority.
Emma Sarfati, Alexandre B\^one, Pietro Gori et al.
In briefWatch Your Step: Anticipated Perception for Humanoid Locomotion
SummaryThe paper proposes a lightweight foothold prediction model that forecasts future footholds from proprioception and velocity commands alone. It samples the robot's elevation map only in small patches around these predicted footholds, removing redundant input before the locomotion policy sees it. The prediction model is trained alongside the policy and periodically retrained on new rollouts. The authors report reaching a given terrain difficulty in about a third of the training iterations needed with the full map, and zero-shot transfer to a Unitree G1 humanoid.
Matthias Heyrman, Chenhao Li, Ioannis Dadiotis et al.
In briefDiff-IrisPAD: Unsupervised Iris Attack Detection Using Diffusion Models
SummaryIris recognition systems can be fooled by presentation attacks such as textured contact lenses, printed replays, and synthetic irises. This paper proposes Diff-IrisPAD, an unsupervised method that models only genuine iris distributions without attack samples, detecting attacks through reconstruction inconsistency. On a large multi-sensor dataset with over 125K images and 8 attack types, it achieves an overall ACER of 30.52 and generalizes well to unseen attacks.
Geetanjali Sharma, Harsh Yadav, Tanmay Deshpande et al.
In briefNo Frames, Just Frequencies: Bandwidth-Efficient On-Device Adaptation for Smart Glasses
SummaryThis paper presents a framework in which smart glasses compute compact frequency descriptors on-device and send them to a host device, instead of transmitting camera frames. The host uses a model replica to synthesize distribution-shift data and adapt the model. Tested on ImageNet-C corruption mixtures, Visual Wake Words, and MS-COCO across three vision architectures, the descriptor-driven adaptation recovers 93–95% of the supervised oracle gain on the architectures, degrades gracefully at lower descriptor resolution, and reduces bandwidth by 213× to 784×.
In briefOverview of CHIP 2026 Shared Task 1: Recognizing Patient-Specific Phenotype Concepts in Full-Text Biomedical Papers
SummaryThis paper reports on CHIP 2026 Shared Task 1, which promotes methods for identifying patient-specific phenotype concepts in full-text biomedical articles. The task has two subtasks: full-text phenotype concept recognition and patient-specific phenotype association, evaluated on 200 annotated articles from the PatientPheX dataset. A total of 455 teams registered, and the best system scored 0.7389 on the final test set, 8.37 percentage points above the baseline.
In briefImproved training of strong stochastic flow maps for few-step sampling of random differential equations
SummaryThis paper studies stochastic flow maps, models that learn to jump directly between times along the solution of stochastic differential equations to enable few-step sampling. The authors identify the Brownian path, which is infinite-dimensional and nowhere differentiable, as the main training obstacle, and propose a shifted Legendre polynomial expansion with a Wong-Zakai style objective to enable derivative-based losses. They also give a recipe for fine-tuning large pre-trained flow models into strong stochastic flow maps, reporting about a threefold FD-DINOv2 improvement over SSFM on ImageNet at four NFE.
Zander W. Blasingame, Antonio Franca, Raul Minan et al.
In briefLocal and global recursive ridge regression for online residual correction of frozen multivariate forecasters under label delay
SummaryThe paper proposes DRiC (DCT Ridge Corrector), which corrects a frozen multi-horizon forecaster's residuals online by ridge-regressing leading DCT coefficients of the next residual block, with recursively updated statistics. The local variant DRiC-L keeps one regression per channel, while the global variant DRiC-G pools all channels. Across seven datasets, three backbones and 126 conditions, DRiC-L8 reduces MSE by 14.55% on average, above the most accurate comparator ELF at 12.98%, though ELF is more accurate on Traffic.
In briefChecking whether diffusion language models stay faithful to tables when generating text
SummaryThis paper compares diffusion language models (MDLM, LLaDA) with autoregressive models (GPT-2, GPT-Neo) on the ToTTo table-to-text task, scoring both similarity to reference text and grounding in the source table. The two measures diverge: adding demonstrations raised LLaDA's BLEU but lowered its NLI entailment against the table. The authors conclude that reference-based metrics alone are insufficient to assess faithfulness.
Zaynab Merimi, Noussayba El Marrakchi, Arthur Satouf
In briefNear-minimax rates for estimating return distributions in quantile distributional reinforcement learning
SummaryThe paper studies quantile distributional policy evaluation, approximating the return distribution by a quantile Bellman fixed point. Assuming a generative model, it builds an estimator from an empirical MDP and proves non-asymptotic error bounds under Wasserstein distances. It also shows the estimator attains the exact local asymptotic minimax risk at each fixed quantile resolution.
WhyDerives non-asymptotic Wasserstein error bounds for quantile distributional policy evaluation and proves the estimator attains the exact local asymptotic minimax risk.
In briefEnergy-Efficient Transfer Learning for Indoor Localization Using Visible Light
SummaryThis paper proposes a transfer learning approach for indoor localization based on visible light communication, aimed at handling lighting changes and obstacles while reducing computation and energy use. It also introduces a model efficiency metric combining accuracy, energy efficiency, data efficiency, and transfer gain. Evaluated on a real dataset from a BOSCH factory, the method reports a 47% accuracy improvement over a model without transfer learning and up to 67.5% model efficiency under high noise, versus 10.4% for conventional models.
In briefCineEvolve: A self-evolving AI agent framework for generating cinematic multi-shot videos
SummaryThis paper presents CineEvolve, an agentic framework for generating cinematic multi-shot videos that combines a Director and a Critic within one vision-language policy. The two roles are trained together through alternating reinforcement learning, with a new method, Evolutionary GRPO, added to stabilize training. The authors also contribute CineTraj, a dataset of 40K cinematic trajectories covering creation, critique, and revision. The paper reports that CineEvolve outperforms existing approaches.
In briefFragment-level graph encoding for explainable molecular property prediction
SummaryGraph neural networks are widely used for molecular property prediction, but atom-level explanations are hard for medicinal chemists to act on. This paper changes the input representation rather than the model, encoding molecules as fragment-level graphs via BRICS decomposition. Across four architectures, four tasks, and two explainers, fragment-level encoding matches atom-level test accuracy on graphs about five times smaller. Integrated Gradients attributions recover known pharmacophores above chance on two enzyme targets, and fragment-level training yields a stable attribution without arbitrary aggregation choices.
Rafael Lopes Almeida (1), Ina P\"ohner (1) ((1) School of Pharmacy, University of Eastern Finland et al.
In briefImproving Spatial Control in Image Generation by Tuning SNR Distribution
SummaryThis paper addresses the difficulty of fine-grained spatial control in image synthesis and finds that the signal-to-noise ratio (SNR) distribution during training is critical for spatial grounding in flow-matching models. Using a synthetic benchmark of geometric primitives and a color-aware layout metric, the authors report that training on a centered SNR distribution yields 25.9% layout adherence, while focusing on high-noise stages raises it to 80.8%. They also report gains on COCO data achieved through distribution tuning alone, without architectural changes.
In briefFaster Bayesian Optimization Using Linear Models
SummaryBayesian optimization is usually applied only when the objective function is expensive, with algorithm speed mattering little. This paper compares Bayesian optimization with faster methods such as CMA-ES using a speed-performance Pareto frontier, and notes that some regions of that frontier are currently empty. The authors propose fast Bayesian optimization methods that exploit linear structure in existing algorithms, for continuous and combinatorial inputs from about 10 to over 6000 dimensions. According to the paper, they perform similarly to state-of-the-art methods while being 10² to 10³ times faster.
In briefReflexivity Threshold: Phase Transition in Multi-Agent Performative Prediction
SummaryThis paper studies repeated retraining in multi-agent performative prediction, where deployed models change the data used for future training and this feedback flows through a network of interacting learners. The authors introduce a reflexivity matrix Γ and use whether its spectral radius crosses one as a threshold separating stable learning from worst-case learning obstruction. Below the threshold, retraining converges geometrically to a unique performatively stable equilibrium; above it, noisy instances are constructed where local-information learners incur linearly growing cumulative excess loss.
WhyGives a spectral-radius threshold with matching upper and lower bounds, generalizing the single-agent scalar result to networked performative prediction.
In briefPrivate online prediction when the best expert has small loss
SummaryThis paper studies online prediction from experts under differential privacy constraints. Prior private algorithms achieve near-optimal regret Õ(√T), which is suboptimal when the best expert's cumulative loss is much smaller than T. The authors give the first differentially private algorithm with regret Õ(√L*) in this low-loss regime, assuming the input sequence is oblivious. They also show that Ω̃(√T) regret is unavoidable for adaptively generated inputs, even when L*=0.
WhyFirst private algorithm achieving √L* regret, substantially improving √T bounds in low-loss regimes and matching non-private lower bounds up to log factors.
In briefLabel-free method for generalizing graph models under distribution shift
SummaryThis paper addresses out-of-distribution generalization for graph neural networks, which often fails when data distributions shift. It proposes SCARCE, a label-free representation-learning framework that uses only an environment partition of the training graphs, obtained from metadata or from clustering a self-supervised embedding. The method warm-starts a variational graph autoencoder and then applies adversarial confusion so the discriminator cannot tell environments apart. The authors report state-of-the-art OOD generalization among label-free methods on synthetic and DrugOOD datasets.
In briefRethinking Underwater Image Enhancement: Text-Guided Degradation Curriculum Learning
SummaryExisting paired clean-degraded underwater image datasets are limited in scale or hard to obtain, and do not fully cover real-world degradations. This paper proposes a degradation-aware training framework that uses a text-guided diffusion model to synthesize semantically consistent degraded samples. It further applies a progressive curriculum across attributes and difficulty levels, together with a transmission-guided one-step diffusion model. The authors report superior results on multiple underwater benchmarks, with code and models promised upon publication.
In briefCertified neural network approximation of robust nonlinear MPC
SummaryModel predictive control performs well but needs heavy online computation, which makes it hard to run on low-level hardware. The paper proposes a framework that generates neural network controllers meant to inherit the safety and stability guarantees of a robust MPC policy. It treats approximation error as a bounded input disturbance handled in the MPC design, giving design conditions that preserve recursive feasibility and stability before deployment. Simulations show the learned controller reproduces the original closed-loop behavior while reducing online computation.
Hendrik Alsmeier, Maik Pfefferkorn, Rolf Findeisen
In briefSWIFT-Cell: A Fast Cell Segmentation Model Distilled from Cellpose-SAM
SummaryThis paper presents SWIFT-Cell, a 0.63M-parameter U-Net distilled from Cellpose-SAM masks for segmenting cells in H&E and IHC tissue images. It runs faster than QuPath's watershed detector on CPU and GPU. Its accuracy stays close to Cellpose-SAM across several human-annotated datasets.
Devansh Lalwani, Swapnil Bhat, Muralikrishna H H et al.
In briefSOAR: a recommendation model that uses social ties and user opinions
SummaryThis paper proposes SOAR, a model for point-of-interest recommendation that combines a user's friendship network with textual reviews to judge whether a location suits a particular user, with emphasis on review sentiment. Experiments on real-world datasets show that including reviews in the recommendation process improves precision.
In briefWRDO: Weighted Refusal Direction Optimization in Aligned Language Models
SummaryAligned large language models refuse harmful requests, but this behavior is fragile and can be disrupted by simple activation or weight interventions. The paper proposes Weighted Refusal Direction Optimization (WRDO), which builds a composite refusal direction from per-prompt directions using learned aggregation weights. It compares WRDO with single-direction methods on three open-weight models, measuring effects on general capability, and also tests using the same method to restore refusal in an abliterated checkpoint.
Grigoriy Evseev, Vladimir Solodkin, Dmitry Bylinkin et al.
In briefGeneralized CycleGRPO for Fine-Grained Multimodal Understanding and Grounding (paper)
SummaryThis paper presents Generalized CycleGRPO, a unified reinforcement learning framework that couples describing a region with localizing it from language. The model writes a description of a given region, then localizes that description to reconstruct the region, using reconstruction overlap as a verifiable reward. The method covers image, video, and spatio-temporal regions and introduces the DOVE benchmark for temporal distinctiveness. The authors report gains over baselines on GroundingSuite, GRES, Charades, and DOVE.
WhyUnifies region captioning and localization via cycle reconstruction rewards, extends to video time, and shows clear gains on GroundingSuite, GRES, and Charades benchmarks.
In briefHyperspherical simplex representations for retrieval that stays compatible across model updates
SummaryThis paper proposes hyperspherical simplex representations, derived from softmax outputs and logits, to enable backward-compatible retrieval between independently trained models without re-indexing the gallery. The method uses the simplex structure from softmax cross-entropy training as a shared geometric reference, and a deterministic linear transformation restores alignment of class prototypes when new classes are added. The authors report competitive results on standard compatibility benchmarks without extra compatibility-specific losses, learned mappings, or architectural changes.
Simone Ricci, Niccol\`o Biondi, Federico Pernici et al.
In briefPACE: Periodic, channel-aligned four-bit encoding of recurrent states
SummaryLinear-attention language models store a fixed-size recurrent state for each sequence decoded in parallel, so lower-precision storage lets more sequences share an accelerator. PACE is a post-hoc four-bit state quantizer that rounds the state only every few generated tokens, buffers the intermediate tokens, shares quantization scales within each column, and keeps the most error-sensitive heads at eight bits. On four hybrid models from 4B to 35B parameters, PACE with mixed heads is competitive with eight-bit DAMP on the three larger models at about two thirds of its memory, and it preserves the decoding speedup of the buffer.
In briefRandomly pruning DPO's computation graph to speed up training and cut peak memory
SummaryThe paper proposes a unified framework that stochastically prunes the computation graph of direct preference optimization (DPO) by sampling scoring positions, response prefixes, and preference branches to skip parts of each training step. It uses inverse-probability weighting to obtain a conditionally unbiased single-branch gradient estimator. Branch pruning gives a 1.426x step-time speedup over a concatenated DPO baseline, while position and sequence pruning mainly reduce peak memory. Mean accuracy is slightly lower.
In briefMelange: A Synthetic Visual Benchmark for Measuring Social Bias in Multimodal AI Models
SummaryThis paper introduces Melange, a synthetic image benchmark for testing social bias in multimodal large language models across seven categories. It extends the text-only BBQ dataset into visual scenarios and tests 23 open-source models under text-only, image-only, and image-with-context settings. Switching from text to images lowers fairness by 0.291 on average, and the gap varies by model family more than by size. Models recognize the depicted attributes accurately but still rarely abstain, and when they give biased answers, they usually pick the stereotypical option.
In briefBlock diffusion models for longitudinal health records: speed and accuracy trade-offs
SummaryThis paper compares block-based masked diffusion (BD) with autoregressive (AR) models for zero-shot prediction from longitudinal electronic health records, using MIMIC-IV. Small-block BD configurations retain much of AR's predictive performance while running faster, with a representative B=8 setting about 1.43 times faster in evaluation. Frozen linear probes show near-identical representation quality, though generative rollout remains below AR in most settings.
Ashvin Gupta, Joshua Placidi, Brendan Delaney et al.
In briefPhase transitions in max-pooling learning on extremal data
SummaryThe paper studies an extremal regression problem, a multiple instance learning setting where max pooling causes vanishing gradients. Modern pipelines approximate the max with a softmax whose sharpness is learned. In a large-bag, high-dimension limit, the authors derive a low-dimensional gradient flow showing two phase transitions: a freezing transition linked to the random energy model and a retrieval transition linked to modern Hopfield networks. Numerical simulations confirm the phase diagram and the predicted asymptotic gradient.
In briefCausal discovery via residuals using pretrained in-context quantile regression
SummaryThis paper proposes Loscar, which extends the RESIT framework from additive-noise models to location-scale noise models. It estimates conditional mean and scale through quantile regression, allowing pretrained tabular foundation models to serve as plug-in estimators. On bivariate and multivariate causal discovery benchmarks, it substantially outperforms a feasible generalized least squares (FGLS) variant, especially for multivariate causes.
In briefDo JEPA-style latent prediction objectives work for patient EHR sequences?
SummaryThis paper tests whether Joint-Embedding Predictive Architectures (JEPAs) can be adapted to learn representations of longitudinal electronic health records, compared with masked language modelling (MLM). Using MIMIC-IV, the authors find that all three JEPA variants train stably without representational collapse but consistently underperform MLM. They report that target size, rather than target heterogeneity, increases prediction error, yet shrinking targets does not restore downstream performance, suggesting stable latent prediction alone is insufficient.
In briefRCSD: Reinforcement and Continual Learning for Faster Domain-Specific Speculative Decoding
SummaryThis paper proposes RCSD, a training framework that speeds up large language model inference by improving the small draft model used in speculative decoding. It builds implicit preference pairs from accepted and rejected sequences produced during inference, and uses a masked loss so that tokens matching the target model are not penalized. Across multiple benchmarks, RCSD raises average acceptance length by 35% over HASS and runs end-to-end inference 1.35x faster on practical workloads without hurting generation quality.
In briefGeometric knowledge spaces, attention pairing and architectures: a paper deriving attention mechanisms from geometry and exploring new Transformer designs.
SummaryThe paper proposes a geometric approach to deriving attention mechanisms and exploring potential new Transformer architectures. It formalizes the notions of knowledge worlds and attention pairings, and shows how this framework naturally leads to standard, causal, linear and tensor attention, as well as recurrent neural networks. It finally introduces several new geometric spaces that may inspire novel neural network architectures.
In briefpplx-embed-v2-late Technical Report: Multilingual, multimodal retrieval models sharing one embedding space
SummaryThis report introduces pplx-embed-v2-late, a pair of 0.6B and 9B multilingual and multimodal late-interaction retrieval models. Both reach state-of-the-art results within their size classes on text, image, and visual document retrieval benchmarks, and the 0.6B model is competitive on visual document retrieval with models several times its size. Both were trained by representation distillation from an 18B teacher and share an embedding space. This allows documents to be indexed with the 9B model and queried with the low-latency 0.6B model.
Markus Krimmel, Louis Milliken, Ekaterina Glazkova et al.
In briefBeyond the Vote: When LLMs Add Value to Retrieval-Based Tabular Classification
SummaryThis paper compares four general-purpose large language models with distance-weighted nearest-neighbor voting (Distance-KNN) using identical retrieved neighbors across 19 tabular classification datasets. Distance-KNN achieves the highest mean accuracy at every candidate pool size. LLM gains are concentrated where neighbors show low label consensus and depend on the average feature-space distance to the test query.
