In briefDeepOpen: Decision-oriented intent classification results on BANKING77 and CLINC150
SummaryThis paper presents DeepOpen, an intent-classification framework with dedicated configurations for the BANKING77 and CLINC150 datasets. On BANKING77, a single English classifier further trained with sharpness-aware minimization (SAM) reaches about 94.12% test accuracy, with a three-seed mean near 94.04%. On the 150 in-scope CLINC150 classes, probability fusion of four classifiers reaches about 98.02%. The authors state that the continuation does not establish a reliable accuracy gain over the historical checkpoint and note selection effects and execution-path differences.
Shuang Song, Muhammad Luqman Naseem, Zhongqiao Yuan
In briefSolving Flood-It Puzzles: Tailored A* Search Versus Declarative Solvers
SummaryThis paper presents an optimized A* search method for finding minimal move sequences in Flood-It puzzles and compares it with SAT, Integer Programming, and Constraint Programming solvers. Using an appropriate heuristic and equivalences between game states to limit the search space improves A*'s performance. A* was the fastest at finding a minimal solution in about 75% of the tested instances. For higher solution lengths, the SAT-based approach outperforms A*.
In briefRIFT: KV Cache Compression for Draft Models with Complementary Compressibility Training
SummaryThis paper addresses growing KV cache costs in speculative decoding for long outputs, noting that prefill-focused compression loses relevance over time, a problem it calls decode staleness. It proposes RIFT, selection-based KV cache compression that balances prefill information with recent tokens, and an adaptive refresh strategy to reduce score calculations. The authors also add a compressibility-driven drafter training phase using sparse attention masks. They report that Qwen3.5-4B drafting for Qwen3.5-122B retains over 98% of full-cache acceptance length under their stated setting.
In briefCo-Recursive Semantic Self-Extension: A Math Foundation for Open-Ended AI
SummaryThe paper argues that AI is limited by its fixed representational framework and models cognitive states as partial evaluations over a formal language, drawing on Kant and Gödel. Using Kolmogorov complexity, it shows that no architecture with a fixed state description, transition mechanism, and initial condition can generate histories containing irreducible algorithmic information. It proposes semantic self-extension as a design principle, in which intelligence transforms unresolved structures into new spaces of meaning rather than only resolving ambiguity within existing ones.
Simon Hegelich, Bardh Prenkaj, Habiba Sarhan et al.
In briefLAMB: A Modern Context Language Model for Latin
SummaryThis paper presents LAMB (Latin Modern BERT), a Latin-language encoder built on the ModernBERT architecture and pretrained in two phases on over 74 billion Latin tokens using a morphologically informed masking method. The authors also release two Latin corpora: Acervus, which includes OCR text repaired by a ByT5 model, and Corpus Trinum, the first Latin semantic triplet dataset. On a benchmark of 17 subtasks, LAMB outperforms all other evaluated encoders on 16 of them.
In briefPairwise Tests Miss Prompt-Level Bias in Reward Model Scores
SummaryThe paper argues that pairwise evaluation of reward models compares responses to the same prompt, so score components shared across a prompt's responses cancel out and go untested. The authors audit eighteen released reward models and six training conditions against human ratings, finding that model scores cluster by prompt far more than human ratings do. They propose checking between-prompt variation, sensitivity to a wrong prompt, and usefulness for selection separately.
In briefTransformers with chain-of-thought learn Heavy-Ball acceleration during training
SummaryThis paper asks whether chain-of-thought reasoning in transformers can go beyond imitating a gradient-descent teacher on in-context linear regression weight prediction. After initial teacher supervision, adding trajectory-level supervision with the true regression weights makes the model implement the Heavy-Ball method, which is faster than the original teacher. This accelerated algorithm appears without any Heavy-Ball teacher supervision, offering a theoretical example of transformers discovering new algorithms through training.
In briefI²-KAN: Interpretable Multi-Pathway Learning for Multimodal Sentiment Analysis
SummaryThis paper addresses the difficulty of telling whether a modality's predictive value comes from itself or from its interactions with other modalities in multimodal sentiment analysis. It proposes I²-KAN, which keeps intra-modal and pairwise inter-modal information in separate structural pathways before a Kolmogorov-Arnold Network fusion. Experiments on CMU-MOSI, CMU-MOSEI, and CH-SIMS v2, with ablation and interpretability analyses, show that a modality weak on its own can still be informative through pairwise interactions.
In briefMQAScore: Fine-grained evaluation of text-music alignment using a large audio language model
SummaryText-to-music alignment is usually judged by one global similarity score, which hides which described attributes are missing. MQAScore instead checks each musical concept across five dimensions (genre, instrument, mood/theme, vocal, and usage) by asking a large audio language model yes/no questions and using the probability of affirmative answers. The authors also introduce the Sensitivity-Invariance Ratio to measure how selectively an evaluator responds to meaning-changing concept substitutions. Experiments show concept-level scoring improves both question-answering and similarity-based evaluators, offering a more interpretable complement to global similarity.
In briefnnInteractive Agent: Using General-Purpose AI Agents to Annotate 3D Images
SummaryThe paper introduces nninteractive-agent, an open-source tool that lets general-purpose AI agents perform 3D image segmentation from plain-language requests without additional training. The agent identifies targets, prompts the nnInteractive model, inspects and corrects the results, and responds to feedback. On a benchmark of 14 images and 51 targets across CT, MRI, PET and micro-CT, Claude Code and Codex reached mean Dice scores of 0.81 and 0.84, and the paper compares them with the text-prompted model VoxTell.
WhyUses general-purpose AI agents to annotate 3D images without extra training, validated on 51 targets across modalities and more robust than a text-prompted baseline.
Fabian Isensee, Constantin Ulrich Harsy, Maximilian Rokuss et al.
In briefWhen Forgotten Facts Return: Relearning-Aware Unlearning for Knowledge Graph Embeddings
SummaryThis paper studies approximate unlearning for knowledge-graph embeddings, which edits entity embeddings to honor deletion requests instead of retraining. The authors find that under continued training, forgotten facts resurface more than in an exact-unlearning baseline, because forgotten and new facts share entity embeddings. They propose Kernel-Anticipated Unlearning (KAU), which uses the tangent kernel to model how continued training changes forgotten facts' scores and steers edits accordingly. KAU reduces leakage relative to the strongest knowledge-graph unlearner on FB15k-237 and CoDEx-L across three backbones, while retaining higher utility.
In briefIRON-Meme: a benchmark for understanding who political memes portray and how
SummaryThis paper proposes a framework grounded in social science, covering social identities, viewpoints (ingroup and outgroup), and narrative roles, and builds IRON-Meme, a human-annotated benchmark of 1,110 political memes. The authors find that most memes assign derogatory roles and that identities, roles, and templates are systematically linked. Among nine vision-language models tested, frontier-scale and reasoning models perform best, while accuracy drops for less frequent labels, and adding template background knowledge does not improve narrative role classification.
Gabriella Bollici, Stefano De Giorgis, Wouter Van Atteveldt et al.
In briefDual Exam: a controlled test of multimodal agents and world models
SummaryThis paper introduces Dual Exam, a framework that tests multimodal agents under fixed environment responses and world models under fixed action inputs, across physical, digital, social, and scientific domains. The agent track compares 10 configurations on 1,088 items, and the world model track covers 400 cases across eight scenarios. The authors note that the two tracks use different tasks and metrics and are not directly comparable, and a small coupled pilot recorded zero task successes. Results show uneven task performance and failures to preserve task-relevant state, while their value for predicting coupled success remains to be established.
In briefDPIF: A General Dual-Path Framework for Multi-View Medical Segmentation That Reduces Dependence on Large Datasets
SummaryMedical image segmentation usually needs large annotated datasets, and performance drops when samples are scarce. The paper proposes DPIF, a plug-and-play framework that works with any encoder-decoder network and uses complementary anatomical information across views to reduce annotation needs. It combines Sinkhorn optimal transport alignment, conditional mutual information feature decoupling, and persistent homology topological regularization. Experiments on 24 baseline models show DPIF-enhanced models beat single-view baselines at comparable data sizes, and most dual-path variants trained on a portion of the data perform close to single-view models trained on the full dataset.
In briefSCOPE: Turning Natural-Language Rules into Verifiable Neural-Symbolic Programs for AI Agents
SummaryThis paper presents SCOPE, a method that compiles a natural-language policy for LLM agents into an executable program made of neural predicates, deterministic extractors, and composing code. The program is optimized together with its training data, using adversarially generated counterexamples that expose where it deviates from the policy. On a benchmark of real-world policy tasks, SCOPE reports the highest average accuracy and more stable predictions, with each verdict traceable to the predicate outputs behind it.
In briefDiagnosing the Gap Between Explicit and Applied Theory of Mind in LLMs
SummaryThis paper argues that current Theory of Mind benchmarks report aggregate scores that blur recognizing what someone believes with using that belief to predict behavior. On SimpleToM, Llama-3.1-8B scores 51.5 percent overall, recognizing mental states at 99.4 percent but predicting behavior at only 28.9 percent. The authors compare structured prompting, in-context learning, and their proposed Belief-State Test-Time Training (BSTTT), finding that recovery varies by model, task, and intervention. They recommend reporting explicit and applied accuracy separately alongside intervention comparisons.
In briefDirecting to Diversify: Multi-directional search for automatically designing problem-solving heuristics
SummaryThis paper studies using large language models to automatically design heuristics for NP-hard combinatorial optimization problems. The authors propose PRS, a framework that transforms heuristic ideas in three directions: expansion, compression, and steering for refinement. It organizes candidate exploration with a breadth-first search and adaptive operator control. The authors report that PRS consistently finds high-quality heuristics across diverse tasks.
In briefHarm Is Behavioural, Moderation Is Lexical: How Language and Culture Shape Who Automated Safety Systems Protect or Silence
SummaryThis paper argues that automated moderation often judges harm by vocabulary and surface form, which over-flags African American English and under-protects low-resource languages. The authors organise the evidence into seven failure modes and propose an evaluation protocol that reports the gap between best- and worst-served language communities. Testing five open systems across eleven languages, they find none reaches native-speaker accuracy, and greater capability does not close the gap between the best- and worst-served languages.
WhyProposes behaviour-based harm framing and an evaluation protocol, tested across five systems and eleven languages with specific, though limited, empirical findings.
Gabriel Lars Sabadin, Dr Nicola Harding, Sarah Kingston
In briefConflict-aware adapters that block distillation while keeping teacher utility
SummaryThis paper proposes CALA, a lightweight training-time defense that stops reasoning traces from being distilled into downstream student models. A small residual adapter is attached to a frozen teacher and trained alone, perturbing outputs where imitation is most sensitive while staying in high-probability regions. Experiments on reasoning benchmarks report substantially reduced successful student imitation with near-original teacher performance and generation quality.
Quoc Phong Dao, Binh Anh Nguyen, Linh Ngo Van et al.
In briefAlgorithmic Drift in Generative Adaptive Learning: A Study on Temporal Fairness
SummaryThis paper introduces "algorithmic drift," a feedback loop in which small, demographically asymmetric observation errors in adaptive learning systems accumulate over time into systemic inequity. The authors formalize the learner-tutor interaction as a coupled dynamical system and support it with an agent-based Monte Carlo simulation. Their proposed temporal fairness stabilizer narrows the achievement gap in the simulated environment, but the findings still need validation with real learners.
Firuz Kamalov, Hana Sulieman, Ahmed El Sayed et al.
In briefResGVC: Residual Coding for High-Fidelity Generative Video Compression
SummaryThis paper proposes ResGVC, a residual coding framework for generative video compression. It first encodes original frames with a learned video codec, then splits the distortion between reconstruction and originals into residual segments transmitted at each denoising step, guiding denoising toward the source. The authors report BD-rate reductions of 63.7% in LPIPS and 83.8% in PSNR over the previous state-of-the-art method, with better visual quality.
In briefCANNBot-CAKE: An Agent System That Evolves Custom Kernels for Ascend NPUs
SummaryThe paper presents CAKE, a system that generates and evolves custom operators for Ascend NPUs using adaptive multi-DSL routing and reviewed knowledge reuse. On 53 CANN-Bench v1.1 operators, it passed 1,060 of 1,060 cases on the standard set and 4,167 of 4,240 on the hidden set, with an average speedup of 1.334 times. In evolution mode, EngramGateFusion reduced aggregate device time by 46.6% over 20 rounds.
In briefInverse modeling for designing laser pulse shapes in fusion experiments
SummaryThe paper introduces LPDS, a generative inverse modeling system that maps desired outcomes and target pellet setups directly to optimized laser pulse shapes for inertial confinement fusion. A multi-objective loss function is designed so generated pulses satisfy physical constraints and experimental requirements. The authors also present constraint-conditioning, inpainting, and gradient-based editing for fine control, and study transferring the framework to a limited set of real experimental shots.
Ricardo Luna Gutierrez, Vineet Gundecha, Rahman Ejaz et al.
In briefNeural Network Transistor Models for Fast SPICE Simulation and Cell Library Characterization
SummaryThis paper develops neural network-based transistor models that predict drain current and various charges for a 3nm nanosheet device, and applies them to standard cell library characterization. The models are benchmarked against an industry-standard BSIM-CMG implementation for accuracy and speedup, and are further tested in analog and digital circuit simulations. The authors report delay and total power errors below 0.11% and 0.76%, with library characterization taking up to 8 times less time.
In briefAutoregressive Multimodal Anomaly Detection with LLMs for Microservices
SummaryThis paper presents LLMMAD, a framework that uses large language models to detect anomalies in microservice systems from multimodal data. It combines the data into a unified sequence to model relationships between modalities, while using modality-specific reconstruction heads to preserve each modality's characteristics. To avoid the identity-shortcut problem, it trains the model to predict future states instead of copying current inputs. On three real-world datasets, it reports F1-score gains of 3.1%, 1.4%, and 0.9% over state-of-the-art methods.
In briefUpdate Cost of Document Chunking Methods When Source Texts Are Revised
SummaryThis paper studies how much re-embedding is needed when document chunking methods for retrieval-augmented generation are applied to revised texts. Using 120 adjacent Chinese and English Wikipedia revision pairs, it compares three stateless chunkers under a byte-identical reuse rule and a frozen encoder. At fixed parameters, mean evidence recall differs by under two points, while recomputation cost varies substantially with chunking settings. The authors recommend reporting update cost alongside retrieval quality.
In briefDifferentiable diffusion conditioning for optimizing physical designs offline (D2C)
SummaryThis paper proposes Differentiable Diffusion Conditioning (D2C) for offline model-based optimization of physical systems. It makes the reverse denoising chain of a frozen conditional diffusion model differentiable, so conditioning variables are optimized at test time using gradients from surrogate objectives and physics-informed penalties. The method is evaluated on laser pulse optimization for inertial confinement fusion, data-center workload scheduling, and four Design-Bench tasks. The authors report the strongest results on the two real-world tasks and the best average rank across the four Design-Bench tasks.
Ricardo Luna Gutierrez, Vineet Gundecha, Rahman Ejaz et al.
In brief"Mostly Reading": What Open-Book Scores Actually Measure in Document-Level LoRA Adapters
SummaryThis paper examines whether LoRA adapters, each trained on a single document, can serve as knowledge units to be merged and retrieved. Training one adapter per Wikipedia article on Llama-3.1-8B-Instruct, the authors find that open-book scores mostly reflect a reading behaviour shared by all adapters. A random adapter recovers 82% of a specialist's open-book advantage but only 2% on a cloze probe, a share indistinguishable from zero.
Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov et al.
In briefXALIGN: A human-centred visual tool for comparing AI explanation methods
SummaryThis paper presents XALIGN, a visual analytics interface that shows rule-based explanations and three feature attribution methods (TreeSHAP, KernelSHAP, and LIME) side by side for the same Random Forest model. It makes agreements and contradictions between these methods visible so users can compare them with domain knowledge. The authors apply it to the German Credit dataset and a meteorological dataset, and report an expert evaluation with three specialists. Participants found it useful for identifying influential features and avoiding over-trust, though linking rules to attributions was harder for a participant without an XAI background.
Bahavathy Kathirgamanathan, Gennady Andrienko, Natalia Andrienko et al.
In briefMulti-Tool Selection for Robots: Semantic Retrieval vs. Large Language Models
SummaryThis paper compares four methods for mapping natural-language commands to the smallest complete set of executable robot tools. The methods include three ChromaDB-based retrieval approaches and a cloud-hosted 120-billion-parameter language model baseline. On 164 manually annotated prompts, the best retrieval-only method, CDB-v3, scored 0.7822 macro-average versus 0.8869 for the LLM. The LLM gave higher routing quality, while CDB-v3 had much lower latency and can run locally without cloud connectivity.
Miguel Sousa, Luis Ramos Pinto, Jo\~ao P. Matos-Carvalho
In briefEvaluating Artificial Wisdom: A Trajectory-Based Framework for Context-Sensitive Decision Regulation
SummaryThe paper argues that current benchmarks measure what large language models can do but not whether their recommendations stay appropriate as decision conditions change. It proposes a trajectory-based measurement framework and prospective validation protocol for "artificial wisdom," defined as context-sensitive regulation of judgment. Simulations and an exploratory study are reported, but the authors state these do not validate the instrument, and whether the profile predicts decision outcomes beyond capability remains open.
In briefRevisiting Distribution Matching for Distilling State-Based Policies into Vision-Based Policies in Laparoscopic Robotic Manipulation
SummaryThis paper proposes a framework for distilling a state-based policy into a vision-based policy for laparoscopic robotic manipulation. A teacher policy with privileged state observations guides a student policy that uses RGB images and robot proprioception. Experiments on various surgical manipulation tasks in simulation show that general-purpose reinforcement learning struggles to train vision-based policies, while the proposed distillation yields more consistent and stable performance.
In briefMitigating Saliency Collapse: Robust Saliency-Aware Long-Text Image-Text Alignment
SummaryThis paper identifies saliency collapse, where models over-rely on visually prominent text and underuse non-salient context in long-text image-text alignment. The authors propose RoSA, a saliency-aware finetuning framework that splits image-text pairs into salient and contextual views and aligns them with auxiliary objectives. According to the paper, RoSA improves long-text retrieval and remains robust to saliency collapse across benchmarks, and its objectives can be added to existing tuning methods. Code and checkpoints are publicly available.
In briefAgri-SCHOLAR: LLM-bounded, simulation-grounded rolling advisories for in-season crop management
SummaryThis paper presents Agri-SCHOLAR, a framework that issues personalized in-season farming advisories every few days or week, using soil and crop data, observed and forecast weather, and historical weather analogs. Multiple LLM agents propose feasible parameter bounds, and an optimizer searches within them using a crop simulator (APSIM) as an evaluation oracle, against a reward covering yield, resource use, and economics. The authors analyze eleven years of maize data from two South Asian locations and report that it consistently outperforms regional standard practices.
Geetha Charan, Vedant Balasubramaniam, Manojkumar Patil et al.
In briefTruthful Bidding via Omission: Click Thinning for AI Chatbot Ads
SummaryThis paper studies ad auctions in which sponsored content is embedded in generative AI assistant answers, where clicks depend on the passage hosting the ad and earlier ads take clicks from later ones. The authors propose click thinning, which randomly omits ads so that each ad's overall click probability no longer depends on ordering. The resulting mechanism is truthful in expectation and achieves at least half the optimal welfare under free reordering. The approximation factor improves to 2-(1-δ)^(K-1) under stated conditions.
WhyIntroduces click thinning achieving truthfulness and 2-approximate welfare, with a tight approximation bound for order-preserving rules.
In briefPadding to k: schema minimums make open-weight models pad extraction lists
SummaryThe paper studies what language models do when a JSON extraction schema requires a minimum number of list entries but the document contains fewer items. Testing eight open-weight models on 558 real documents, plus six hosted models as reference, the authors find that a minimum shown in the schema but not enforced raises the share of over-long lists from 34.5% to 55.4%. Most added entries are real text from the document, so the padding is hard to detect.[]
WhyShows schema minimum counts induce list padding, testing eight open models on 558 real documents with preregistered analysis and several fixes.
In briefEEG-Captioner: Generating Language from Brain Signals Grounded in Signal Properties
SummaryThis paper presents EEG-Captioner, a framework that translates EEG brain signals into natural-language descriptions. The authors build a dataset of 160,781 EEG segments annotated with signal-level question-answer pairs about spectral, spatial, temporal, and lateral properties. They also propose a Dual-Domain EEG Encoder that models both temporal waveforms and spectral-spatial representations. Experiments show that combining signal-level QA with caption supervision gives complementary benefits.
In briefColaBench: Evaluating Memory-Guided Knowledge Acquisition in Personal Agents
SummaryThis paper introduces ColaBench, a synthetic benchmark in health and daily living with 349 tasks, 10 user profiles, and a 13,033-document knowledge base, linked by an Evidence Graph to users' historical interactions. It also proposes MemKB, which uses state summaries and memory-document links to retrieve knowledge suited to a user's history. MemKB reaches 86.25% accuracy with an average of 4.07 interaction rounds per task.
In briefInterpretive Variation and Discourse Structure in Detecting Enthymemes
SummaryThis paper studies how to detect and evaluate enthymemes, arguments that leave a premise or conclusion unstated, in social media tweets about vaccination and immigration, each labeled by five annotators. Fine-tuned models plateau at an F1 of about 0.67 regardless of data size, features or label type. The authors propose metrics comparing predictions with the full distribution of annotator votes, and a model that predicts this distribution directly, capturing disagreement better without losing F1.
In briefA reality check on methods that add text to prompts to improve AI accuracy
SummaryThe paper compares five inference-time context augmentation methods that insert content into model inputs, ranging from filler tokens that only add computation time to relevant elaborations. The methods were applied sentence by sentence across seven models of different sizes and three datasets. The authors find that accuracy gains are often not genuine improvements but undirected perturbations that fix some predictions while breaking others, especially when the inserted text only adds computational time. They argue that the reliability of accuracy gains should be assessed before these methods are adopted in practice.
In briefDiagnostic Loss: asymmetric training to make radiology reports more diagnostically accurate
SummaryThis paper argues that standard token-level training of medical vision-language models favors text agreement over key diagnoses, leading to missed or wrong diagnoses in generated radiology reports. It proposes Diagnostic Loss (DLoss), a differentiable objective that aggregates decoder probabilities separately for diagnoses supported and unsupported by the reference report, using asymmetric pooling across token positions. Trained and evaluated on 50,000 abdominal CT scans across several 3D architectures, DLoss improves diagnostic faithfulness while remaining competitive on conventional language-generation metrics.
WhyProposes asymmetric diagnostic loss on decoder probabilities, validated on 50,000 abdominal CTs across several 3D MedVLMs, improving diagnostic faithfulness.
In briefLAST-V2X: Language-Driven Adaptive Testing of Communication Failures in V2X Safety Systems
SummaryThis paper presents LAST-V2X, a framework for testing how communication failures affect V2X safety applications in connected vehicles. It uses large language models to turn natural-language test descriptions into executable traffic scenarios, combined with traffic simulation and message-level fault injection. Matched runs under degraded and ground-truth information are used to tell communication-caused failures apart from other safety violations, and an execution-guided refinement process improves test cases. Experiments in diverse CARLA environments show the framework generates reliable scenarios and finds failures more effectively.
In briefGraphCAD: Geometry-aware diffusion on edge-adjacency graphs for generating CAD models
SummaryThis paper presents GraphCAD, a generative framework for Computer-Aided Design (CAD) models that centers on a new graph representation called the Edge-Adjacency Graph (EAG). It uses a hybrid graph diffusion approach to jointly model discrete and continuous components, with an SE(3)-equivariant backbone that grounds node features in spatial geometry. The authors also build EAGBench, a large-scale engineering CAD dataset. Experiments report that GraphCAD generates models with more than 50 faces and outperforms sequence-based baselines in geometric fidelity and distributional alignment.
In briefZero-shot forecasting of soil moisture and temperature using a time-series foundation model
SummaryThis paper tests TimesFM 1, a 500M-parameter time-series foundation model, on soil moisture and temperature at a semi-arid land-restoration site, without task-specific training. It uses 15-minute data from 27 sensors and forecasts from 1 to 168 hours ahead. The model beats rule-based baselines at sub-daily lead times but falls behind beyond a day, which the authors attribute to diurnal alignment. The paper also compares it with a CatBoost model and measures inference cost on CPU hardware.
Mohammad Talaei, Mohammad Goudarzi, Maille Todd et al.
In briefJoint or Per-Option? Small Encoders for Choosing Among Fixed Options
SummaryThis paper studies small encoders that make closed-set decisions, such as choosing a tool, and compares reading all options jointly with scoring each option separately. At 32 million parameters, per-option scoring performed better on several measures but needed many more encoder passes. The authors also released six joint encoders from 17 million to 1 billion parameters, and the 68- and 150-million models outperformed a 435-million-parameter DeBERTa-v3 zero-shot classifier on 209 tasks.
In briefProgressive image compression via bound-conditioned latent inference (BLI)
SummaryThis paper proposes BLI, a progressive image compression framework that treats progressively decoded trit-planes as shrinking interval constraints on continuous latents. It jointly refines Gaussian entropy parameters and the latent estimate used for reconstruction. Experiments on Kodak, CLIC, and JPEG-AI report state-of-the-art rate-distortion among progressive codecs, with a shared model using about 5.3 times fewer parameters than CTC and lower latency.
Jaeseok Jang, Seungmin Jeon, Kwang Pyo Choi et al.
In briefEnergy and Latency First: Deployment Limits Reverse Time-Series Model Rankings
SummaryThis paper studies handover prediction in cellular networks, where models run on user devices once every 40 ms and must finish before a deadline. The authors first screen models for tail-latency feasibility, then rank the feasible ones by measured inference energy including feature computation. Gradient-boosted trees (F1=0.974) were the most accurate and among the cheapest to run, while a 1D CNN (F1=0.914), ranked first by accuracy alone, used over three times more energy per inference.
In briefCan Language Models Carry Hierarchical Grammar Across Languages?
SummaryThis paper studies whether language models trained on a new language can transfer hierarchical syntactic structure learned from a source language. The authors introduce a cross-lingual poverty-of-stimulus setting that controls the amount of hierarchical evidence in the source language. Results show only partial transfer, yielding a small but detectable hierarchical bias that depends on the language pair. Vocabulary overlap is found to be crucial for this transfer.
WhyIntroduces a cross-lingual poverty-of-stimulus design testing hierarchical transfer, with ablations showing vocabulary overlap is crucial for transfer.
Shin Yamaguchi, Masaya Taniguchi, Tatsuki Kuribayashi et al.
In briefDoubly Robust Distributionally Robust Offline Contextual Pricing
SummaryThis paper studies offline contextual pricing from logged data when the training and deployment distributions differ. It proposes a doubly robust framework for distributionally robust off-policy evaluation and learning with continuous prices. The authors establish a finite-sample regret bound for smoothness orders 1 and 2, and experiments under various distribution shifts validate the approach.
In briefGround-truth-first benchmark for reinforcement learning using converse optimality
SummaryThis paper proposes a benchmarking framework for reinforcement learning in which each task is built with a certified optimal value function and policy that are withheld from learners and used only for evaluation. This allows direct measurement of optimality gap, regret, and policy error rather than relative rankings. The framework covers continuous-time sensitivity studies and native discrete-time oracle comparisons, and extends to non-affine inputs, non-Gaussian disturbances, non-smooth value functions, and partially observed tasks.
Gr\'egoire Ouerdane, Sinan Ibrahim, Stefan Streif et al.
In briefTesting whether driving AI's stated reasons actually drive its choices
SummaryThis paper examines whether the reasoning produced by driving vision-language-action models actually determines their trajectories. The authors apply counterfactual interventions to four models across several benchmarks, changing the meaning of the models' own reasoning and photorealistically removing key objects named in it. They find that the generated reasoning and driving behavior are only weakly coupled, which limits the explainability these systems are meant to offer.
In briefOpenAttention: One Percent Is Enough With Tail Proxies
SummaryThis paper studies how sparse attention should handle the unread tail of the key-value cache. It finds that a single proxy token summarizing the whole tail gives lower output error than sampling at every read level tested behind a practical selector. The authors propose OpenAttention, combining clustering, ErrorGate and top-k selection with a count-aware mean proxy. On RULER-32K HARD with Llama-3.1-8B, the paper reports retaining 99% of dense accuracy at a 1% read, with decode attention about 10 times faster than dense attention.
In briefRandom Initialization as a Baseline for Time Series Models
SummaryThis paper examines how much pretrained weights improve time series foundation models beyond their architecture. It compares seven public models with randomly initialized versions of the same architecture on the UEA multivariate time series classification archive. Pretraining improves accuracy for all seven models, but the gains are modest. The authors recommend separating the effect of pretraining from architecture and including strong non-pretrained baselines.
In briefIQuest-Q1: Building Foundation Capabilities for Agentic Command-Line Systems
SummaryThe paper presents IQuest-Q1, a sparse Mixture-of-Experts agentic foundation model with 320B total parameters and 15B activated. It is developed in three stages: stable pre-training, fine-grained mid-training, and multi-scaffold post-training, followed by model merging to combine experts and checkpoints. The authors report capabilities in both coding and general tasks, and the model is released as open source.
In briefRandom projections for spectral algorithms: optimal learning rates in Sobolev norms
SummaryThis paper analyzes how random projections can reduce the cost of spectral learning algorithms while keeping their accuracy. Using properties of reproducing kernel Hilbert spaces, the authors derive learning rates in Sobolev norms that match minimax lower bounds up to logarithmic factors. For randomized sketches and Nyström sub-sampling, the projection dimension needed for optimality scales with the effective dimension. Experiments are reported to support the theory and show practical effectiveness.
In briefRetiJev: Fast, Source-Grounded Models for Eye-Care Decisions
SummaryThis paper presents RetiJev, a family of 0.8B to 9B parameter ophthalmic decision models that score candidate answers directly without autoregressive generation. The authors built RetiJev-Data, an ophthalmology corpus drawn from PubMed Central articles and a MeSH vocabulary, plus guideline-grading questions and an evaluation benchmark. According to the paper, RetiJev outperforms a 27B LLM and the commercial Jev 1.13 model on guideline grading and achieves the highest clinical-decision accuracy, answering a case in 158 ms on one GPU.
In briefTracing Hallucinations in LLM Reasoning for UML Class Diagram Generation
SummaryThis paper examines hallucinations that appear when large language models generate UML class diagrams from natural language requirements, analyzing their intermediate reasoning traces. It presents an empirical taxonomy of eleven sub-types across three dimensions, derived from DeepSeek-R1 and DeepSeek-V4-Pro traces. Adding prompt constraints for prompt faithfulness and reasoning quality raised the mean F1 score by 4.0 percentage points across three Qwen3-based models on 13 cases, while relationship modelling remained the lowest-scoring element.
Mojdeh Rahmanian, Foutse Khomh, Ashkan Sami et al.
In briefMetacognitive Scaffolds Create Productive Friction in AI-Assisted Knowledge Work
SummaryThis paper reports a randomized, between-subjects experiment with 297 workers producing 594 AI-assisted essays, testing dynamically generated scaffolds for goal specification and task decomposition. Scaffolds improved writing quality, and the gain persisted in short-lag transfer after removal, with improved topic recall three days later. Scaffolded participants, however, rated the AI as less helpful and more cognitively demanding despite performing better.
Viktor Kewenig, Lev Tankelevitch, Sean Rintel et al.
In briefAttention-based classification and tumor localization of mixed liver cancers
SummaryCombined hepatocellular-cholangiocarcinoma (cHCC-CCA) contains components of both hepatocellular carcinoma and cholangiocarcinoma, and is hard to diagnose because of strong intratumoral heterogeneity. The study trains a classification model on Hematoxylin-Eosin-Safran (HES)-stained whole slide images to distinguish these liver cancer subtypes. It also offers a potential way to localize different tumor types within combined liver cancer images.
Xiaowen Liang, H\'el\`ene Cazier, Aur\'elie Beaufr\`ere et al.
In briefGateFormer: Forecasting Short-Term Solar Irradiance from Sky Images Using Forecast-Space Decomposition
SummaryThe paper introduces GateFormer, a vision transformer that predicts global horizontal irradiance within an hour using sky images and measured irradiance. It uses two decoders, one forecasting the future irradiance sequence and another predicting the final-horizon change, whose outputs are adaptively weighted. On the NREL-TSI benchmark it reaches 125.01 W/m² RMSE at +30 minutes, a 9.6% improvement over a transformer baseline. It also reports the lowest final-horizon RMSE among evaluated baselines across NREL-TSI, Chilbolton, and SIRTA.
David Hamlyn, Sunny Chaudhary, Jonathon Hare et al.
In briefPAIR: Preference-Aligned Intent Retrieval for LinkedIn Premium People Recommendation
SummaryThe paper argues that profile-similarity recommendations often misrank candidates on professional networks, where relevance is complementary, such as job seekers engaging with recruiters. It proposes PAIR, a retrieval framework that models user intent preference using an exposure-conditioned multi-negative contrastive objective, and updates user representations from positionally encoded click histories at constant cost. Evaluated on LinkedIn's Premium retrieval task, a 0.6B-parameter model reaches a MAR@100 of 0.45.
In briefVision Transformers estimate sky irradiance from limited sky-image data
SummaryThis paper proposes a dual-framework method for estimating diffuse horizontal irradiance (DHI) from fisheye sky images when data is scarce. It combines calculated atmospheric parameters and time encodings, predicts synthetic global horizontal irradiance before refining DHI, and uses unsupervised physics-informed cloud segmentation overlays. Tested on Chilbolton Observatory data with models trained on a single month, the approaches achieve RMSE within a few W/m² of a state-of-the-art ViT trained on the full dataset.
In briefThe Geometry of Meaning in Neural Manifolds: Semantic Correspondence Enables Interoperability Between Independently Trained Models
SummaryThis thesis argues that independently trained neural networks can have incompatible coordinate systems yet still share an underlying intrinsic geometric structure. It proposes semantic correspondence as a fundamental primitive that anchors different models to a common semantic frame, enabling zero-shot transformation between them. Fourteen contributions cover model stitching, cross-modal translation, policy transfer, activation steering, and neural decoding across vision, language, reinforcement learning, and neuroscience. The work also reports that fMRI data from different human brains show alignable geometric structure.
In briefExplaining Activation Differences in Plain Language with Natural Language Autoencoders
SummaryThis paper studies whether natural language autoencoders (NLAs) can translate large language model activations into readable text and how well they generalize. Across four models from three families, the authors test compositional structure, reuse on finetuned models, recovery of finetuning objectives in model diffing, and specialization for diffing. Results largely support a positive answer, such as recovering the hidden secret word in most Taboo organisms. NLAs do not yet match state-of-the-art methods on emergent misalignment detection but outperform a blackbox baseline.
Andrea Blasi N\'u\~nez, Gianluca Barmina, Federico Torrielli et al.
In briefReflectPO: Reflection-guided policy optimization for long-horizon LLM agents
SummaryThis paper proposes ReflectPO, a reinforcement learning framework that helps LLM agents recover autonomously from errors in long-horizon tasks. It builds error-recovery hints from branch differences and repeated visits in trajectories, and organizes sampled trajectories into trees to assign action-level credit. On two long-horizon tool-use benchmarks, ReflectPO outperforms several baselines in most domains and achieves the best average performance.
In briefICGF-CXRMate: Image-conditioned clinical context compression and gated fusion for chest X-ray report generation
SummaryThis paper proposes ICGF-CXRMate, a method for generating chest X-ray reports that uses patient data. It conditions learnable queries on the current X-ray and compresses the patient-data sequence into a fixed number of clinical context tokens, which are selectively fused with visual features via gated cross-attention. On linked MIMIC-CXR and MIMIC-IV-ED data, Stage 2 results show improved report quality and reduced average prompt length from 373.89 to 288.37 tokens compared with CXRMate-ED.
In briefMaking Multimodal LLM Latents Useful for Cinematic Video Generation via Factor-Aware Predictive Learning
SummaryThis paper proposes Factor-Aware Predictive Learning (FAPL) for cinematic video generation, which aims to make latent representations from a multimodal large language model usable by a video diffusion transformer. It uses three training stages: grounding the generator on frozen MLLM latents, learning controlled changes to target factors while preserving others from paired videos, and learning temporal inference from sparse text-only inputs. The authors report consistently better results than baselines on sparse-to-video, factor-aware editing, and multi-shot generation tasks.
In briefLow-latency MoE inference in space data centers using wireless power transfer
SummaryThe paper proposes SpaceWPT, a framework that uses inter-satellite laser wireless power transfer to redistribute solar power among satellites hosting mixture-of-experts LLM inference. Because skewed expert activation leaves some active experts power-limited while others sit idle, the authors derive a two-sided water-filling strategy that decides whether each expert donates, retains, or receives power. Simulations with Mixtral-8x7B and Phi-3.5-MoE report up to 29.25% lower inference latency than without wireless power transfer.
In briefDConstrain: compiling context-aware rules so diffusion language models produce schema-valid output
SummaryDiffusion language models decode several tokens at once, so tokens that are each valid may conflict when combined under a schema. DConstrain compiles schemas into context-parameterized transition programs and tracks field histories at runtime, allowing arbitrary property order. On the BFCL benchmark it compiles 100% of schemas, reduces end-to-end decoding latency, and raises DiffusionGemma's accuracy from 37.27% to 81.56%.
In briefUsing Language Models as Optimization Modeling Agents (two lightweight inference harnesses, OPTICYCLE and OPTISTATE)
SummaryThe paper studies whether large language models can turn natural-language optimization problems into solver programs. The authors introduce two lightweight inference harnesses, OPTICYCLE and OPTISTATE, and evaluate them on nine benchmarks with 2,634 problems using 11 Qwen models from 0.6B to 27B parameters. Execution-guided revision improves accuracy over direct generation for all evaluated models, by up to 49.46 percentage points. Without additional training, both harnesses outperform six adapted agentic workflows and four post-trained optimization models of comparable size.
In briefWatch and Learn: Removing Text Shortcuts to Teach Visual In-Context Learning
SummaryVision-language models lag behind text models at in-context learning, and the authors find they lean on textual shortcuts instead of visual demonstrations. They build MIME, a task-agnostic training set without such shortcuts, where instructions do not state the rule linking demonstration images to answers. Fine-tuning on MIME improves visual in-context learning accuracy by 8.1 to 18.0 points across three VLMs, and they also propose DTAD, which withholds demonstration labels during training.
In briefHierarchical fine-grained image-text matching using adaptive regions and optimal transport
SummaryThis paper addresses fine-grained image-text retrieval, where a model must check not only overall consistency but also whether local details like objects, attributes, quantities, actions, and relations are visually supported. The authors propose HFCMA, which groups image patches and text tokens into adaptive visual regions and phrases, then matches them using entropy-regularized optimal transport to obtain a structured alignment. Background debiasing and related regularizers are added to reduce noise. Experiments on Flickr30K and MS-COCO are reported to show improved discrimination of subtle semantic differences.
In briefH-2DGS: A height-field Gaussian splatting method that reconstructs a sharp, deformable brain surface for neurosurgery
SummaryThe paper presents H-2DGS, a method for reconstructing the exposed cortical surface seen through a surgical microscope so it aligns with MRI anatomy and tractography pathways. It uses a three-stage pipeline: a robust 3D Gaussian Splatting reconstruction, fusion of its depth into a truncated signed distance field surface, and curved surface-oriented Gaussian primitives. Across five real neurosurgical cases, the authors report low multi-view depth inconsistency, a 7.97MB representation, and interactive rendering at 29–42 ms per frame.
Zhang Zichuan, Bart Brouwers, Rembrandt Bakker et al.
In briefTwinStore: A Synthetic Shopper Platform Using Vision-Language Models
SummaryThis paper presents TwinStore, a source-available platform that builds a digital twin of a retail store and fills it with synthetic shoppers. These shoppers decide what to buy by viewing rendered shelves with packs, price tags and promotion signs, so layouts, prices, packs or promotions can be tested before store rollout. In held-out validation on eleven public datasets, simulated receipts approached households' natural variation, price elasticities fell within field-measured ranges, and three independent decision models agreed.
In briefQuery-time calibration of CLIP features for open-vocabulary 3D segmentation
SummaryThis paper improves open-vocabulary 3D segmentation by changing how CLIP features lifted into 3D Gaussian reconstructions are compared with text queries. The method subtracts a shared scene direction from visual features, whitens text prototypes against a generic reference vocabulary, and uses soft clustering and CSLS-based comparison. The authors report consistent mean IoU gains across seven lifting methods and four datasets, without retraining or re-lifting features.
In briefSame Molecule, Different Knowledge: Cross-Identifier Binding Failures in LLMs for Chemistry
SummaryThis paper examines whether different identifiers for the same molecule, such as names and SMILES strings, give large language models equal access to their stored knowledge. The authors find they do not: a model may recognize a molecule from its SMILES yet fail to retrieve a fact that it can access through the molecule's name. They call this cross-identifier binding failure and argue it stems from weak links between identifiers and existing knowledge. They propose Binding Self-Distillation, which uses lexical identity during training to enable direct factual retrieval from SMILES.
In briefBeyond Phenotypic Similarity: Relation Refinement and Selective Population Graph Learning for Brain Disorder Diagnosis
SummaryThis paper proposes PopDRP, a framework for multimodal brain disorder diagnosis that refines inter-subject relationships in a population graph. It calibrates attribute-based affinities using training labels and further rectifies graph relations with fused functional-connectivity and non-imaging representations. A hierarchical selective state-space backbone then learns population representations at multiple resolutions. Experiments on ABIDE and ADHD-200 under stratified ten-fold cross-validation report improved accuracy and AUC over baselines, and ablations support the contributions of relation refinement and the hierarchical backbone.
In briefWhen Annotation Patterns Are Learned: Skin-Tone Dependence in Dermatology Concepts
SummaryThe paper audits the reliability of SkinCon, a dermatology concept dataset, across skin tones. Within the same disease, erythema is annotated much less often on dark skin than on light skin, while hyperpigmentation and hypopigmentation are annotated more often. A concept model trained on these annotations learns and reproduces the same tone-dependent pattern. The authors conclude that concept annotations should be audited within each subgroup before explanations built on them are trusted.
In briefSilicon Samples Match Answer Distributions but Miss Individual Respondents
SummarySilicon sampling uses large language models to answer surveys as people with given demographic attributes, creating synthetic respondents. Existing evaluations check individual answers, per-question distributions, and cross-question correlations, but do not compare the respondents themselves. The authors propose the respondent coverage gap (RCG), which measures how far each human respondent's full answer profile is from the nearest generated one. On two US social surveys, all six tested LLMs leave human respondents uncovered, performing worse than a baseline sampling answers from human marginals, and every LLM has a higher RCG than a Gaussian copula fitted to the human data.
In briefParameter learning in active inference via variational message passing
SummaryThe paper studies planning while learning in active inference, recasting Expected Free Energy minimization as Variational Free Energy minimization with epistemic priors solved by sum-product message passing. It gives each factor its own Dirichlet prior over observation and transition parameters, which preserves the information-gain terms that standard mean-field factorization removes. In experiments on two grid POMDPs, only the planner that learns the model reaches the goal on every episode, while the compared planners plateau or take the risky route.
In briefMaking Multigraph Neural Networks Robust Against Adversarial Attacks in Financial Crime Detection
SummaryThis paper studies whether multigraph neural networks used for anti-money laundering can resist attacks that manipulate transaction structures to evade detection. The authors propose a scalable framework for structural attacks and adversarial training based on node-edge incidence matrices, plus a cycle-based regularizer that measures attack plausibility. Experiments show learned attacks substantially reduce detection performance (PR-AUC), while adversarial training consistently improves robustness.
Dennis Heijmans, H. \c{C}a\u{g}r{\i} Bilgi, Rui Wang et al.
In briefContinuous-depth Deep Gaussian Processes: a Bayesian regression model treating depth as continuous evolution
SummaryDeep Gaussian processes are Bayesian regression models, but their discrete-depth form becomes harder to optimize as depth grows. The paper treats depth as a latent continuous evolution, driven by a stochastic differential equation with a depth-indexed drift family under a GP-based prior. It presents two instantiations, CDGP and FlowDGP, and reports that both give a stronger overall performance than plain DGP across the regimes studied.
In briefA unified training framework for generating higher-quality PowerPoint slides automatically
SummaryThe paper presents a training framework for automatic PowerPoint generation that combines cross-validated data synthesis, rule-based teacher model enhancement, and verifiable reward optimization. It adds a GRPO-based reinforcement learning stage to improve slide-level visual attributes beyond text-only supervision. Experiments report gains of 7.45 and 9.35 points on PresentBench for Qwen-3.6-27B and Qwen-3.6-35B-A3B, with further improvements across five aesthetic dimensions.
In briefFrom Perpetrators' Digital Footprints to Behavioral Communities: A Framework for Bengali Cyberbullying Analysis
SummaryThis paper proposes ODCBNA, a perpetrator-centred framework that combines ontology design, knowledge graph construction, behavioural modelling, community analysis and influence modelling to study cyberbullying. Using the BenCoDiC Bengali corpus, it identifies 1,063 perpetrators and represents each with a 64-dimensional feature vector. Louvain analysis of the resulting network finds six communities with modularity of about 0.495.
Md. Nesarul Hoque, Md Hanif Seddiqui, Miskatul Anwar et al.
In briefTracing how language models inherit unintended preferences from other models' data
SummaryThis paper studies subliminal learning, where a model trained on another model's generated data picks up traits unrelated to that data. Using Pythia, PolyPythia, and DataDecide OLMo models, the authors compare hard label and on-policy soft label distillation and find transfer persists across many training differences. They report that the ability to transmit a preference and the ability to acquire one develop separately during pretraining, and transfer can be strongly asymmetric when roles are reversed.
In briefInherited, Not Invented: Where the “wait” Token in Reasoning Models Comes From
SummaryThis paper examines where the word “wait” in reasoning models comes from, tracing it through open models’ training histories. Using label-free measures of frequency, position, form, and usage, the authors find that zero-RL rarely creates the marker, which mostly arrives through training data. Where the base model already uses it, some RL recipes amplify it, while others reduce it.
In briefOne Teacher, Many Views: Reducing Reference Sensitivity in Self-Distillation
SummaryThis paper studies on-policy self-distillation (OPSD), which improves a model's reasoning by training it on its own outputs using dense token-level guidance from a teacher that sees the full reference solution. The authors argue that a single reference is only one possible reasoning path, so conditioning on it can make targets depend on information a student lacks at inference, causing what they call reference-induced target instability. They propose Monte Carlo Privileged Distillation (MCPD), which treats privileged context as a distribution over partial reference views and distills the averaged teacher distribution. The paper reports that MCPD yields more reference-consistent supervision and consistently outperforms full-reference OPSD on mathematical and scientific reasoning benchmarks.
In briefHimmBO: A Preference-Driven Method for Merging Many Image-Generation Adapters
SummaryHimmBO is a two-stage method for combining many fine-tuned adapters for text-to-image models according to user preferences. Its screening stage identifies the active adapter set from pre-generated images without generating new ones, and its refinement stage optimizes merge coefficients within that smaller subspace. In simulated-user image-matching experiments, it reached higher final similarity than all compared methods and reached a given similarity with under one tenth of the image generations of prior work. Its performance stayed stable as the collection grew to 1080 adapters.
In briefMOER: Mix-Policy Optimization with Expert Prefix Guidance and Selective Rollback
SummaryThis paper proposes MOER, a reinforcement learning framework that combines expert chain-of-thought prefixes with the model's own exploration for reasoning training. The prefix length is gradually reduced to encourage self-exploration, and failed completions are replaced with full expert traces to ensure correctness. The authors report the highest average scores on mathematical and general reasoning tasks, with consistent generalization across tasks, model scales, and base models.
In briefAdding Shared Usage Examples to a Tool Index Lowers Retrieval of Unused Tools
SummaryAI agents often find tools by searching documentation, and agents sharing a catalog can pool past usage examples. This paper studies how folding those examples into the search index affects tools that no agent has used yet. On two of three catalogs, folding sharply reduced how often a needed unused tool reached the top five results. The authors propose ZoE, which scores documentation and usage separately, and it retrieves unused tools more often than a router across the catalogs and embedding models tested.
In briefVerus-Insight: Explainable Detection of Visual Artifacts in AI-Generated Videos
SummaryThis paper studies how to build artifact detectors for AI-generated videos by splitting detection into inspection planning, evidence verification, and evidence aggregation. The authors introduce Verus-Insight, trained with human-anchored supervision, and Verus-Bench, a benchmark with fine-grained human annotations of perceptible artifacts. Experiments across several benchmarks report strong performance in both artifact analysis and video-level detection, with better recovery of human-perceived artifacts.
In briefConsensus-based distributed filters for estimating a spacecraft's position from angle-only measurements in cislunar space
SummaryThe paper studies consensus-based distributed filtering for estimating a target spacecraft's state in cislunar space using only line-of-sight measurements from a network of observer spacecraft. It compares three distributed filters against a centralized Extended Kalman Filter benchmark and independent local filters. Simulations show the distributed approaches substantially improve accuracy over independent filters. Monte Carlo analyses then assess robustness to finite consensus iterations and prior misalignment among agents.
Giovanni Romagnoli, Fabrizio Pilone, Giordana Bucchioni
In briefCross-Spectral Feature Fusion for Better Multispectral Historical Document Binarization
SummaryThis paper addresses binarization of historical documents, where faded ink, bleed-through, and background degradation vary across visible and near-infrared wavelengths. It uses a multi-branch conditional GAN on four SMADI bands (500, 600, 700, and 1000 nm), fusing visible, NIR, and structural features with adaptive softmax gating and edge maps from selected band pairs. Nine configurations were compared using five-fold document-level cross-validation and DIBCO metrics. The F3 plus NIR edge-guided configuration performed best, with mean F1 of 0.7916, and its gains over the fixed Mid AVG baseline in F1, p-FM, and IoU remained statistically significant after Holm correction.
In briefBeyond Correctness: Checking Whether Formal Library Documentation Reflects Actual Function Behavior
SummaryThe paper addresses natural-language documentation for functions in formal libraries, which developers rely on because formal notation is hard to read. It argues that a documentation statement can be provably true yet still hold after a function's behavior changes, and defines this dependence on specific behavior as attribution, with a method to check it. The authors present DocAuth, which uses attribution to decide what to publish and attaches checkable formal properties and evidence to each sentence. On 150 ACL2 Community Books functions, DocAuth raised the share of published sentences confirmed as attributable from 82.9% to 94.7% compared with a truth-only pipeline.</summary>
WhyProposes attribution as a new criterion for formal-library documentation, validated on 150 ACL2 functions against truth-only and mutation-based baselines.
In briefDirect Forcing: Single-stage training for fast, few-step autoregressive video generation
SummaryThis paper introduces Direct Forcing, a single-stage framework for few-step autoregressive video generation that removes the separate ODE initialization stage. It combines Causal Continuation Alignment with Rollout Distribution Matching, which uses maximum mean discrepancy, and coordinates the two with Dynamic Gradient Balancing. Experiments with Wan2.1-1.3B report a four-step generator trained in 250 updates, reaching 84.63 VBench Total and 86.01 Quality.
In briefSynthetic-to-Real Nepali-English Code-Switched Speech Recognition Benchmark
SummaryThis paper introduces a new benchmark of 13,044 spontaneous Nepali-English podcast utterances, audited to remove machine translation leakage and acoustic loops. It proposes a two-stage method: supervised fine-tuning on about 0.95 hours of multi-speaker synthetic speech, followed by Group Relative Policy Optimization with transcription and script rewards. The authors report that synthetic fine-tuning lowers Mixed Error Rate from 23.40% to 21.52% and switch-point error rate from 35.89% to 28.77%. The subsequent reinforcement step yields only small adjustments.
In briefCAVC: A Training-Free Cache That Speeds Up Joint Audio-Video Diffusion Generation
SummaryThis paper addresses the high computational cost of diffusion-based audio-visual generation, which jointly denoises audio and video at every sampling step. The authors find that caching audio computation mid-sampling causes drift that a few final joint steps can correct, whereas caching video harms motion and temporal detail. They propose CAVC, a training-free method that reuses the audio output in the middle stage while keeping full video refinement, then restores joint steps at the end. Reported experiments show 1.10 to 1.35 times wall-clock speedup with competitive audio fidelity and audio-visual alignment.</summary>
In briefTraining neural networks that can be verified quickly, without set propagation
SummaryThe paper proposes a neural network architecture that represents mixed-monotone functions as the difference of two monotone subnetworks. Using mixed-monotone system theory, an output enclosure is obtained from just two forward passes at opposite corners of the input interval, avoiding complex set propagation. The enclosure size is a closed-form function of the weights, so it is minimized during training, and a lower bound on enclosure size is proven. Experiments show the enclosure comes within 0.5% of this bound without losing prediction accuracy, and it is computed much faster than CROWN and α-CROWN.
WhyReduces tight output enclosure to two forward passes via mixed-monotone networks, with a proven lower bound; evaluated against CROWN and zonotope baselines on regression and dynamical systems.
In briefADAPT for Science: a multi-agent system that evolves its own scientific optimization
SummaryThis paper presents ADAPT for Science, a multi-agent framework that treats scientific optimization as two linked decisions: where to search and how to search. It uses historical designs to pick promising, diverse starting points and refines the optimization procedure based on execution feedback. Tested on gravitational wave detector design with a differentiable physics simulator across eight unseen topologies, it reports a mean loss of 0.204 ± 0.101 under matched budgets, a 54.9% reduction versus the best classical optimizer and 62.2% versus the best general agentic evolution baseline.
In briefReading Bifurcations from Weight Space: JEPAWG on Parametric Dynamical Systems
SummaryThe paper builds model collections by fine-tuning a forecasting network at each point of a parameter grid for four dynamical systems, giving a test-bed with known bifurcation structure. It analyses these collections with JEPAWG, a joint-embedding hypernetwork, and reports competitive forecasting performance. The authors state that the norm of the parameter derivative of decoded weights locates Hopf points, gives a qualitative Feigenbaum constant estimate, and detects a global bifurcation.
