In briefSupermarket product detection: deep learning with corrected images to handle angled shots of densely packed shelves.
SummaryThe paper addresses the difficulty of recognizing grocery products on densely packed supermarket shelves photographed at varying angles. The authors supplement object detection models with Hough transform and homography estimation to rectify angled images, and they introduce a new dataset for testing different angle and object-density scenarios. Experiments across several detection models show that rectifying angled images improves grocery detection accuracy. The results also show that the benefit is limited by the capture angle and the number of objects per image.
In briefDialogue Is a Two-Person Problem: Two-Way Evaluation of Full-Duplex Dialogue Models
SummaryFull-duplex spoken dialogue models can listen and speak simultaneously, but existing evaluations use either pre-recorded audio or an automated examiner that only runs fixed tests, so they assess only one side. This paper proposes DyaFDB, which has two models converse directly under assigned roles with cooperative or conflicting goals, and an external judge scores both offline. Across 140 scenarios and 7,560 conversations, the authors find that each model's behavior continually reshapes its partner.
In briefBeyond Waypoint Regression: Query-Based Cost Learning for End-to-End Driving Plans
SummaryThis paper proposes an end-to-end driving planner that estimates bounded costs for dynamically reachable ego trajectory queries instead of regressing waypoints. It uses compact joint scene tokens to capture multimodal agent futures and converts the learned cost into feasible plans via contingency-aware aggregation and cost-guided MPPI mixing. The authors report improvements over cost-estimation planners such as ST-P3 and NMP on nuScenes, lower collision rates than most regression baselines, and fewer collisions than SparseDrive and Alpamayo on real-world logs without fine-tuning.
Ahmed Abouelazm, Rupert Polley, Qingyuan Zhang et al.
In briefUnder-responding in-context models for what-if forecasts; fixed by synthetic fine-tuning
SummaryThe paper tests whether covariate-aware time-series foundation models can give training-free what-if predictions for forced engineering systems. TimesFM-2.5 and TabPFN-TS respond to input changes only instantaneously, while Chronos-2 learns dynamics in context but attenuates their magnitude. A 26-minute fine-tune on synthetic forced systems restores response strength, yet classical system identification remains clearly better on three of four measured plants.
In briefEnergy-Aware Path Following: Comparing Reinforcement Learning and NMPC for Electric Cars
SummaryThis paper compares four controllers for keeping electric vehicles on a reference path while managing speed smoothness, using a kinematic model and an energy model that includes regenerative braking. The controllers are NMPC, PPO reinforcement learning, a gain-scheduled state-feedback baseline, and a Stanley geometric baseline. The PPO policy, trained on straight and S-curve tracks, transfers without modification to unseen tracks and, with brief fine-tuning, to a dynamic single-track vehicle model.
In briefFoundation Models for Data-Efficient Cardiac MRI Reconstruction
SummaryThis paper examines whether pretrained vision foundation models can serve as useful priors for accelerated cardiac MRI reconstruction. The authors insert frozen or lightly adapted encoders (CLIP, BiomedCLIP, DINOv2) into a transformer-based reconstruction architecture and test them on the CMRxRecon2023 and CMRxRecon2024 benchmarks. Pretrained representations consistently beat a transformer trained from scratch across acceleration factors, and they also perform better with limited data and cross-dataset transfer, with DINOv2 achieving the strongest overall results.
Anam Hashmi, Mayug Maniparambil, Julia Dietlmeier et al.
In briefReusing autoregressive post-training weights to improve diffusion language models
SummaryThis paper studies whether diffusion language models (dLLMs) can reuse weight updates from post-training done on their autoregressive (AR) ancestors. The authors find that adding an AR post-training update directly to a diffusion base model remains effective, bringing performance close to direct diffusion post-training. The two kinds of updates are nearly orthogonal in weight space but produce more aligned representation changes in the diffusion model, and composing them can further improve results. Based on this, they propose A2D, a training-free framework.
In briefTesting Acoustic and Content-Based Speaker Verification Attacks on Multilingual Voice Anonymization
SummaryThis paper studies attacker speaker verification systems against voice anonymization in multilingual settings, which prior work has mostly explored only in English. The authors evaluate acoustic-oriented and content-oriented attackers on multilingual anonymized speech and build a multilingual voice-converted dataset to improve cross-lingual generalization. They find that attack effectiveness depends on how well linguistic information is preserved, with acoustic-oriented attackers generally performing better. The paper calls for attacker modeling and evaluation that consider both privacy and utility.
In briefChartBmkAgent: Multi-Agent Generation of Chart QA Benchmarks from Sparse Error Specifications
SummaryChartBmkAgent builds complete chart question-answering samples from sparse error-taxonomy specifications, turning an identified model capability gap into targeted diagnostic test data. A central harness coordinates specialized agents and requires each stage to show its output aligns with the original error category, recording the basis for each acceptance decision. On 300 samples, multimodal model accuracies ranged from 32.7% to 84.3%, and targeted follow-ups scored 50.0% versus 82.2% on matched controls. Evaluators judged 86.4% of samples as testing their specified error category.
In briefDSV-Mem: A benchmark testing whether AI agents remember evolving visual information in professional tasks
SummaryThe paper introduces DSV-Mem, a benchmark that evaluates how well multimodal LLM agents remember dense, stateful visual information in professional workflows. It includes expert-reviewed scenarios and 1,000 questions across five categories, including current state, past state, change history, and conflict/refusal. Across 27 model and memory-method configurations, the strongest baseline scores below 45%. The authors find that state evolution, especially the number of governing updates, is the main difficulty, and that state-aware designs help more than extra reasoning effort.
In briefBeyond Corrected Memory: Execution Consistency in Multi-Agent Systems
SummaryThis paper argues that correct shared-memory records do not prove that agents' actions meet task requirements. It defines execution consistency through duties on state use, information handoffs, and final-state agreement, with explicit evidence conditions for judging fulfillment. Controlled experiments report that removing key evidence leaves 82.4% of opposite-label pairs indistinguishable, and restoring it separates 97.9% of merged pairs.
In briefNatural-Language Questions for Knowledge Graphs: QRAKEN Graph Distillation and Self-Healing
SummaryThis paper presents QRAKEN, a training-free, ontology-agnostic neurosymbolic pipeline that converts natural-language questions into SPARQL queries over RDF knowledge graphs. An offline step distills the graph into TTQL, a compact summary of populated multi-hop patterns and frequencies plus a class-property co-occurrence matrix, which guides an LLM and feeds iterative refinement checks. On the CK25 benchmark, it reports strict F1 of 0.643 and 0.652 with two GPT models, about 30% and 32% above the strongest recomputed participant. Ablations identify TTQL patterns as the main source of gains.
In briefSAGE: Semantic Anchor-Guided Evolution for Synthesizing Medical QA Training Data
SummaryThis paper introduces SAGE, a framework that lets small, locally deployed models generate medical training data by using public taxonomies such as MeSH as semantic anchors. It iteratively combines concept-based and relation-based synthesis, starting from minimal seeds, without needing large document collections or external APIs. Across several medical question-answering benchmarks, models fine-tuned on SAGE-synthesized data outperform those trained on self-derived or document-based data.
In briefWhen Plans Change Answers: Formalizing Cost-Accuracy Trade-offs in Semantic Queries
SummaryThis paper addresses semantic query engines, where predicates are evaluated by machine-learned models and the chosen query plan affects both cost and results. It gives a formal definition of cost-accuracy optimization, using the calibrated confidence of decision models to estimate expected errors and their propagation to output quality without labeled data. The work proposes an oracle semantics for relational algebra with semantic operators, with theoretical results on plan invariance and pushdown soundness. Evaluation is limited to simulations on a synthetic workload, with real-engine evaluation left for future work.
In briefTesting whether clinical visual concepts help AI diagnose skin conditions across several datasets.
SummaryThe paper introduces SkinLex, a harmonized dataset of 48 clinical morphological attributes drawn from four public dermatology datasets, totaling 20,411 records. Supervised nine-class classification of skin conditions shows that restricting inputs to specific visual groups, such as shapes or colors alone, lowers diagnostic accuracy. Bootstrapped backward elimination indicates that the 48 visual concepts contain some redundancy for this diagnostic task.
Linda Wermelinger, Simone Lionetti, Fabian Gr\"oger et al.
In briefDirectSpeech2LLM: A simple end-to-end method to reduce prompt overfitting in speech LLMs
SummarySpeech LLMs trained only on speech recognition instructions often fail to follow new instructions, such as speech translation, and mainly act as recognizers. The paper proposes DirectSpeech2LLM, which computes a distance-based CTC loss over the frozen LLM embedding matrix and uses greedy CTC labels to build speech embeddings that keep instruction-following on unseen tasks. Trained on 960 hours of LibriSpeech, it outperforms a cascaded system on recognition and generalizes zero-shot to speech translation and emotion recognition.
Hemant Yadav, Sunayana Sitaram, Roger Zimmermann et al.
In briefPOLAR: A Guardrail That Checks Whether Tool-Calling AI Agents' Actions Can Be Reversed
SummaryThe paper introduces POLAR, a guardrail framework for small tool-calling LLM agents that judges action reversibility using a two-layer structured ontology. Each action receives a graded reversibility score based on a derived candidate inverse sequence, and calls below a threshold are pruned before execution. Evaluated on τ²-bench across six agent models, POLAR raised mean task reward on airline for four agents, but improved only eight of eighteen model-domain cells overall, with retail and stronger agents often regressing. The authors argue it offers an auditable structural check and characterize its task-utility trade-offs, noting that reward is not a direct measure of prevented harm.
In briefProximalFM: Amortized Proximal Causal Inference Under Hidden Confounding
SummaryStandard causal methods assume no unmeasured confounding and can fail when relevant confounders are hidden. Proximal causal inference uses proxy variables to identify effects under hidden confounding, but nonparametric estimation is data-hungry, sensitive to hyperparameters, and hard to optimize. The paper proposes ProximalFM, a prior-data fitted network trained on synthetic data from structural causal models, which targets the Bayesian posterior of the conditional average treatment effect (CATE) in a single transformer forward pass. The authors report consistently strong CATE estimation across diverse proximal settings, with the largest gains when confounding is strong and proxies are weak.
Christophe Muller, Ayub Kharel, Alex Luedtke et al.
In briefSelf-Retrospection Distillation: Using After-the-Fact Experience to Build Advance Foresight
SummaryThis paper proposes Self-Retrospection Distillation (SRD), a method that uses information from completed trajectories to train the same policy to make predictions before acting. SRD serves only as a training target, so foresight does not need to be generated explicitly at inference time. Evaluated on 10 tool-integrated reasoning and long-horizon agentic tasks, SRD complements RLVR and self-distillation baselines, with its benefit most pronounced when reward contrast is scarce. In the 2B setting, adding SRD raised success from 0.0% to 60.6%.
WhyDistilling hindsight trajectories into pre-action foresight is novel; validated on 10 tasks with large gains in sparse-reward settings.
Haoxiang Zhang, Qinglin Chen, Hiroaki Hayashi et al.
In briefSpeedrunBench: Testing LLM Agents with Video Game Speedrunning
SummaryThis paper introduces SpeedrunBench, a benchmark that uses video game speedrunning, the practice of completing games as fast as possible, to evaluate how well frontier LLM agents form strategies. It spans 9 games and requires agents to repeatedly improve their strategy, reflect on performance, and reason over long action sequences. Experiments show agents approach human world records in simple platformer games but still trail human performance on longer, more complex games under practical budgets.
Yoshinari Fujinuma, Keisuke Kamahori, Ryuto Koike et al.
In briefOptimization Encoders: Rethinking second-order meta-learning for neural fields
SummaryThis paper treats the inference of latent representations in neural fields as an optimization encoder, unifying second-order differentiation, latent parameterization, and task supervision so that encoding and decoding can be trained end to end. It proposes MetaLF, a transformer-based equivariant model that uses self-attention to contextualize a latent pointcloud. The authors report improved reconstruction fidelity for images and 3D shapes within three to five gradient updates, plus semantic prediction across images, shapes, and volumes.
Rudolf L.M. van Herten, Soufiane Ben Haddou, Rachit Saluja et al.
In briefTwo half-sized experts beat one: faster, high-quality image generation via phase-wise distillation
SummaryThis paper proposes Phase-wise Velocity Distillation (PVD), which splits image generation into a coarse phase and a fine phase, each handled by a half-sized expert model, so total computation matches one full backbone pass. It reports an FID of 1.48 on ImageNet 256×256 class-conditional generation. On text-to-image tasks, distilled models reach results competitive with their multi-step teachers while reducing active parameters and peak VRAM compared with those teachers.
WhySplits generation into coarse and fine phases with half-sized experts, a substantive improvement over one-step distillation, validated on ImageNet and several T2I backbones with competitive results.
In briefDetecting a Shift Is Not Enough: Exact Limits for Repairing Representations
SummaryThe paper studies how to remove a mean shift between two data sources by learning one linear map under a fixed distortion budget. It derives the exact finite-sample minimax risk over all such maps and shows that projecting out the mean calibration difference attains this limit. It also identifies a gap between detecting a shift and removing it, and checks the formula's predictions on paired clinical and wearable sleep EEG data.
WhyCasts linear representation repair as a decision problem, derives an exact finite-sample minimax risk, and validates it on paired sleep EEG data.
Anuar Aimoldin, Yankai Chen, Ayana Mussabayeva et al.
In briefPhysTacGen: Generating Optical Tactile Images from Vision with Physics Awareness
SummaryThis paper presents PhysTacGen, a framework that generates optical tactile images from visual inputs to reduce the cost of collecting paired visual-tactile data. It combines a vision-language model refined with reinforcement learning (GTPO) for material descriptions, DINOv2-based pair selection with monocular relative-depth estimation, and an SDXL ControlNet image generator. Experiments on curated SSVTP data show better structural similarity than baselines, a blinded user study preferred GTPO-generated descriptions, and generated tactile inputs improved a force-coefficient prediction proxy task.
In briefComparing Two Approaches to Who-Spoke-What Speech Recognition in Multiple Languages
SummaryThis paper describes the HINTT system submitted to the 2nd MLC-SLM Challenge, which asks which speaker said what and when in multilingual conversations. The authors compare a cascaded pipeline of speaker diarization and speech-LLM-based recognition with a unified speech LLM that outputs speaker labels, timestamps and transcripts directly. Their submission uses the cascaded approach, which they report remains more reliable under Task 1 conditions. They present unified models as a promising direction for future work.
In briefLanguage Carries the Expert's Impression: LLM Judges Transfer Counseling-Quality Assessment Better Than In-Domain Training
SummaryThis paper studies automatic assessment of counseling conversation quality by transferring expert overall-impression predictions across three German simulated counseling corpora (195 expert-rated sessions). Training on other domains outperformed in-domain training, reaching a Spearman ρ of about 0.54 versus at most 0.48. The strongest features were session-level construct scores produced by small open-weight LLMs reading transcripts, and ensembling judges from several model families improved language-only performance.
In briefA new geometry-aware method that makes low-rank adaptation fine-tuning behave more like full fine-tuning
SummaryThe paper studies low-rank adaptation (LoRA), a parameter-efficient method for fine-tuning pre-trained neural networks, whose parameterization contains redundant equivalent directions. The authors propose a new Riemannian metric tailored to LoRA that is used to precondition each gradient step. They show theoretically that this preconditioning makes weight updates closer to those of full fine-tuning, and experiments on language and vision fine-tuning tasks show its effectiveness and efficiency.
Shoichiro Takeda, Shin'ya Yamaguchi, Satoshi Suzuki et al.
In briefDAEDALUS: Building Reusable Agent Memory from Self-Generated Practice Tasks
SummaryLLM agents often lack operational knowledge of new environments and repeat the same mistakes without memory of past attempts. DAEDALUS pairs an explorer agent that generates challenging but solvable tasks with a solver agent that attempts them, deriving heuristics from failures that are accepted only after repeated success. These heuristics are consolidated into a memory bank for test-time use. The paper reports improved success rates over a no-memory baseline on AppWorld, τ²-bench, and AutomationBench without requiring existing training tasks or oracle verifiers.
WhyProposes bootstrapping reusable agent memory without existing tasks or oracle verifiers, validated across three benchmarks with multiple metrics and ablations.
In briefSepsisLens: structure-preserving sequence model for decomposable early sepsis warning (arXiv paper)
SummaryThis arXiv (cs.LG) paper introduces SepsisLens, a model for early sepsis warning from irregularly measured ICU data. It keeps variable-indexed temporal states until risk composition, then combines variable-level and organ-level components through a structured risk head. The model was evaluated on three public ICU cohorts and one private-hospital cohort, reporting strong discrimination on all four and lower alert burden at matched event recall on MIMIC-IV.
In briefDo language models know which script to write in?
SummaryThis paper examines whether language models can recognize and use different writing scripts, which affects whether users can read their outputs. Using two experiments on multi-script languages, the authors test whether small and large models match the script of the input and follow instructions to write in a specified script. All tested models reached over 98% Latin script fidelity, and the larger models scored higher, including on non-standard script combinations.
David Kletz, Sandra Mitrovi\'c, Ljiljana Dolami\'c et al.
In briefSame Feedback, Different Answer: Measuring Repeat-Run Instability in AI Customer Feedback Analysis
SummaryThis paper proposes a repeat-run evaluation framework that measures how much the themes and counts change when the same customer feedback is analyzed multiple times. The authors compare eight frontier models and three execution designs, finding that a taxonomy-grounded agent reduces theme churn by 86–88% relative to raw generation and hierarchical decomposition, with zero volume disagreement for matched themes. The framework is also intended for repeated synthesis of other unstructured corpora, such as financial reports and legal documents.
In briefLearning in dreams, winning in real games: a continuous Dyna loop for a ten-hero MOBA
SummaryThe paper builds a structured multi-agent world model of a complete ten-hero MOBA and trains a policy only inside it using 1,400-tick imagined episodes, then tests it in the real game. A continuous asynchronous Dyna loop lets the policy win 70.2% of real games as radiant, versus 0% with dream training alone. The authors also find that model exploitation is invisible from inside the dream, and that the policy inherits the world model's fidelity profile, modeling macro play well but not crowd control or cast timing.
WhyLearns a full ten-hero MOBA world model and validates policies on real games, showing Dyna repairs model exploitation with strong real-game evidence.
In briefLabel Problems in Hallucination Detection Benchmarks: An Empirical Study
SummaryThis paper examines whether automated labelers used in hallucination detection benchmarks are reliable. The authors compare lexical similarity metrics, an NLI baseline, and seven LLM judges under different prompts against 900 human-labeled question-answer pairs. They find substantial disagreement among automated strategies and with human labels, and show that label outcomes depend strongly on how the target criterion is specified. The authors argue label source should be explicit, validated, and matched to the benchmark's goal.
Jorma Valjakka, Juhani Kivim\"aki, Juha Myll\"ari et al.
In briefSpectra: An exact method for adjusting simulation-based inference models to new priors at test time
SummarySimulation-based inference (SBI) performs Bayesian inference for scientific models whose likelihoods are hard to evaluate. This paper introduces Spectra, a test-time adaptation method for diffusion-based SBI that uses an exact score-transport identity to obtain the adapted score in closed form from a frozen model for structured prior changes. It requires no additional simulation or training. Across six SBI benchmarks, the authors report accurate adaptation under strong prior shifts at low online sampling cost.
In briefLearning consistent molecular force fields directly from quantum reference data
SummaryClassical force fields are widely used for large-scale molecular simulations, but their parameters are usually assigned by atom or bond type, which limits adaptability across configurations. This paper introduces grappa-fullFF, which learns both bonded and nonbonded parameters consistently and simultaneously from ab initio reference data. It uses electrostatic potential supervision and an architecture that supports charge equilibration, and reportedly achieves state-of-the-art accuracy on geometry optimization benchmarks without externally assigned nonbonded parameters.
In briefStructured but Silent: Checking Whether LLMs Internally Encode What Capabilities a Query Needs
SummaryThis paper asks whether large language models already encode, in their hidden states before generating text, the external capabilities a user query requires. The authors introduce TACIT, which splits capability requirements into Source, Transformation, and World Effect, yielding eight capability classes, and train linear probes on 1,600 balanced queries across four open-weight model families. The probes decode these fine-grained capability structures with high accuracy, but the same models are much less reliable when asked to classify the queries explicitly in natural language. The authors call this gap "structured but silent."
In briefFOSLS-deRhaNN: native de Rham neural classes for H(div) and H(curl) and their use in first-order least-squares neural PDE solving
SummaryThe paper constructs neural approximation classes native to the H(div) and H(curl) spaces, in two and three dimensions for H(curl) and in any dimension for H(div), so every network realization lies in the correct space. The classes are built from scalar or componentwise networks mapped by fixed de Rham complex operators, without meshes. These classes underpin the FOSLS-deRhaNN method, whose loss is a least-squares functional in the natural spaces of the weak formulation, applied to elliptic equations, curl-curl problems, transport, and conservation laws.
In briefFrom writing advice to manuscript governance: a framework for revising academic papers without changing claims
SummaryThe paper treats academic revision as an admissibility problem under ordered constraints, separating task invariants, contextual constraints and mutable presentation choices. It introduces a claim-preservation contract linking evidence and judgement to an authorised claim boundary, and sets out four governance propositions on preservation, precedence, context and sufficiency. The Academic Writing Filter implements the model by screening candidate revisions and is explicitly not a universal definition or classifier of good academic writing.
In briefCan drug effects on cells be predicted without running imaging experiments?
SummaryThe study tests whether molecular encoders pretrained contrastively on paired molecule and cell-morphology data, such as CLOOME and CellCLIP, can replace Cell Painting assays for predicting phenotypic activity. After controlling for pretraining-data leakage and the correlation between activity and cytotoxicity, six representations were compared on two Cell Painting screens. The pretrained encoders showed no clear advantage over plain physicochemical descriptors, and toxicity was generally easier to predict than phenotypic activity. The authors argue that leakage-aware, confound-controlled evaluation should be standard before such encoders are trusted in phenotypic drug discovery.
In briefTICDA: A Method for Attributing In-Context Tabular Data Examples
SummaryTabular foundation models make predictions by conditioning on labeled examples in their context, without updating parameters, but how each example affects a prediction is poorly understood. The paper introduces TICDA, which measures each demonstration's influence in a single forward pass using linear surrogates trained on the model's latent embeddings. The authors report that TICDA offers the best trade-off against competing methods across four tasks: detecting labeling errors, curating context, transferring attribution scores across models, and supporting active learning.
Yacine Benihaddadene, Milan Bhan, Eliot Dugelay et al.
In briefRethinking Model Merging: Implicit Regularization from Task Arithmetic
SummaryModel merging combines the weights of separately trained task-specific models to build a multi-task model cheaply. Most existing methods search for linear combination coefficients of task weight updates using an extra dataset, which implicitly restricts candidates to the subspace those updates span. The authors find that removing this restriction and directly optimizing merged weights significantly improves common merging methods across architectures, domains, and even with one example per class.
Sin-Han Yang, Shih-Cheng Huang, Chieh-Yen Lin et al.
In briefVisionWeave: Teaching multimodal LLMs to allocate visual detail elastically
SummaryMultimodal large language models encode images into dense, fixed-size patch tokens, which is costly, while visual detail is unevenly distributed. VisionWeave combines a gated spatial pooler that builds coarse and fine representations in a shared coordinate space with a granularity router that learns content-adaptive allocation. On Qwen3.8-27B, it saves 43.0% of tokens on average while retaining 98.9% of native performance across eight benchmarks, and reports a 2.3x throughput gain on the SGLang serving engine.
WhyProposes gated spatial pooling and granularity routing for elastic visual tokens, validated via large-scale training and SGLang serving benchmarks.
In briefZero-shot tabular learning using LLM-guided hypergraph reasoning
SummaryThis paper proposes TABHGR+, a zero-shot framework for tabular data that uses LLM-generated rules to assign weak labels and group semantically related rows. These groupings form a hypergraph, on which label propagation refines the noisy weak labels. Hyperedges are also given label-adaptive weights to improve the LLM-generated structure. The method needs only a constant number of LLM calls regardless of dataset size, reducing inference cost and achieving higher accuracy than zero-shot and strong few-shot baselines.
In briefDeciding Which Retrieved Memory Images Get Full Pixels
SummaryMultimodal assistants answer questions using retrieved long-term memories that include images, which can be costly to pass as full pixels. The paper introduces PixelTriage, a small plug-in model placed after retrieval that reads dialogue, a short note and a thumbnail of each memory to predict how much its pixels would help. With a 7B answering model, it uses 11–23% of visual tokens on M3Exam, DMV and MemEye without significant accuracy loss, and answers 2.9 times faster than opening all images on DMV.
In briefM3SunAgent: An Agent for Monocular Metric Depth Estimation and 3D Visual Grounding
SummaryThis paper proposes M3SunAgent, a unified agent that uses a large language model to plan tasks and coordinate tools for monocular 3D spatial understanding. For instance-level metric depth estimation, it reports the best performance among compared models, with 52.61% of predicted instances having depth error below 0.25 (δ < 0.25). The authors also build the M3SI benchmark with 2,910 samples. For monocular 3D visual grounding, it reaches a 3D mIoU of 41.73%, exceeding the MonoVLM model by 3.62%.
In briefDo Higher-Order Hypergraph Models Win for Higher-Order Reasons?
SummaryHigher-order hypergraph models often beat lower-order baselines, and this advantage is usually credited to higher-order information. The paper introduces a controlled framework that perturbs higher-order information while keeping pairwise information, and across 25 benchmarks most models keep much of their advantage afterward. Simple additions to lower-order baselines, such as richer pairwise weighting, more propagation steps, and normalization, shrink the remaining gaps, suggesting part of the advantage has lower-order explanations.
WhyProposes a controlled perturbation framework across 25 benchmarks, showing most hypergraph gains are explainable by lower-order information.
In briefLearning from Revision Consequences: Hindsight Meta-Experience Distillation for Self-Improving Agents
SummaryThis paper introduces HMED (Hindsight Meta-Experience Distillation), a method that helps self-improving AI agents learn from how they revise the processes used to create skills. It revisits the completed event where a revision originated and re-runs both the existing and revised meta-skills from the same restored starting state, so the effects of the change can be compared fairly. Each comparison is stored as a reusable meta-experience record. The authors report that across three interactive agent benchmarks and both open-source and closed-source models, HMED consistently improves skill discovery over strong baselines.
In briefLearning Rankings from Human Feedback in Log-Concave Random Utility Models
SummaryThis paper studies how to recover the ranking of a fixed set of items with unknown utilities from human feedback. Feedback is modeled with a random utility model using log-concave noise, under either full-ranking or winner-only feedback. The authors prove worst-case sample-complexity lower bounds for both feedback types and give algorithms that match them up to logarithmic factors, without knowing the noise distribution. The results show that the winner-only setting is intrinsically harder.
Diego Alovisetti, Marco Mussi, Alberto Maria Metelli
In briefCan Agents Work for Everyone? Reliability of Mobile GUI Agents Across Users in Personalized Interfaces
SummaryThis paper examines whether mobile GUI agents perform reliably across different users when interfaces are shaped by each user's history and preferences. The authors introduce PAIR, a pipeline for building user-conditioned app states, and RePAIR, a reinforcement learning method that learns from cross-user differences in subgoal outcomes. Tests on six agents show large variation in task success across users and lower subgoal achievement in personalized interfaces. RePAIR improves performance over its supervised fine-tuning baseline on unseen users.
In briefEmbodiedSmith: Generating Robot Training Data in Simulation via Recursive Self-Improvement
SummaryThis paper introduces EmbodiedSmith, a framework that generates embodied training data for robots in simulation. Scene and task generation iteratively refine each other through an agentic loop, which also supports mobile manipulators, humanoids, dexterous hands, and deformable objects and fluids. Experiments report good data quality, diversity, and generation efficiency. Downstream policy experiments show that greater data diversity improves generalization.
In briefDecepEval: A Benchmark for Measuring Deception in LLM Agents
SummaryThis paper introduces DecepEval, a benchmark of 1,532 instances across 3 task families and 28 professional scenarios for testing whether LLM agents deceive. It proposes the LLM Deception Diamond framework, based on classical fraud theories, identifying pressure, incentive, opportunity, and conflict as conditions that may induce deception. Evaluations of nine frontier LLMs found that inducements increased deception across models and task families, including models with low baseline rates.
In briefHow VAE-based Audio Decoders Encode Musical Features Across Layers
SummaryThis paper examines how the internal layers of RAVE, a VAE-based neural audio synthesis model, encode musical features such as pitch and BPM. The authors analyze three RAVE models trained on different musical domains and compare them with the general-purpose EnCodec decoder. Synthetic stimuli are encoded strongly, natural audio more weakly but more clearly under nonlinear probes, and joint encoding peaks in middle layers. The findings aim to advance interpretability of neural audio models and inform control strategies.
In briefReGraph: A computational model of generalization in the visual 'what' and 'where' streams
SummaryThis paper introduces ReGraph, a recurrent dual-stream graph model with biologically inspired inductive biases that simulates the dorsal ('where') and ventral ('what') visual pathways. Trained on the Something-Something V2 action benchmark, the model showed context-invariant codes and grid-like spatial bases emerging specifically along the extended dorsal stream. The authors argue that generalization may take shape gradually as sensory information is parsed into hierarchical visual streams rather than emerging abruptly in one dedicated region.
In briefDensiTok: Helping feed-forward 3D Gaussian Splatting reconstruct scenes from fewer input images
SummaryFeed-forward 3D Gaussian Splatting reconstructs a scene in one pass, but its quality drops when fewer input images are given. The paper presents DensiTok, a plug-in module that densifies the internal geometry tokens of a frozen pretrained model. It compresses the tokens into a latent space and completes the latents for unobserved viewpoints in a single flow-matching step conditioned on camera geometry. Across three backbones and two benchmarks, it consistently improves sparse-view reconstruction.
In briefRefining Language-Based Trajectory Prediction With Numerical Expert Rewards
SummaryThis paper introduces MoRE, a framework that transfers coordinate-level knowledge from five frozen numerical predictors into a pretrained language-based trajectory predictor via reinforcement learning. Expert predictions are turned into rewards, combined through an uncertainty-weighted consensus that penalizes disagreement, and anchored by a ground-truth reward. On ETH-UCY, ADE drops from 0.22 to 0.20 m and FDE from 0.32 to 0.29 m, with ADE reductions of 17.9% on SDD and 12.7% on NBA relative to the base policy.
In briefConfidence Reasoning Graphs: a method that estimates how likely an LLM agent succeeded at its task
SummaryThe paper introduces Confidence Reasoning Graphs (CRGs), an inference-time method that estimates the probability an LLM agent completed its task from a single execution trajectory, without internal model access or training data. CRGs break the success claim into sub-claims grounded in trajectory evidence, estimate confidence for each, and aggregate them into an overall score. Across three benchmarks, three backbone models and three agent frameworks, CRGs outperform verbalized, sampling-based and white-box baselines in calibration and risk-aware decisions, and their claims and evidence can be audited at decision time.
Brendan King, Farima Fatahi Bayat, Jean-Flavien Bussotti et al.
In briefAdapting vision-language-action models to unknown visual disruptions during execution
SummaryThis paper introduces SALT, a method that lets a robot's vision-language-action policy adapt at test time when visual disruptions occur mid-task, using the unexecuted leftover trajectory from the previous action chunk as self-supervision, without disruption labels or expert actions. On LIBERO-10, average success across five persistent visual corruptions rose from 43.9% to 53.2% with SmolVLA and from 58.7% to 66.0% with GR00T N1.7. On a real robot, task progress averaged over disruptions improved from 0.49 to 0.61.
In briefUltraDiff: Differentiable Ray Tracing for Ultrasound Shape Optimization
SummaryThis paper introduces UltraDiff, a framework that treats ultrasound image formation as transient rendering, modeling echoes binned by travel time as a path-space integral. It derives Monte Carlo estimates of the forward model and its gradients, enabling shape optimization by matching simulated echoes to measured ones. The method recovers vertebral surfaces from simulated B-mode sweeps and a real robotic scan of a spine phantom, starting from a sphere-initialized signed distance function. It works unsupervised directly on B-mode images without pre-segmentation and achieves competitive geometric accuracy.
Felix Duelmer, Magdalena Wysocki, Nassir Navab et al.
In briefHybrid Latent Attention for Looped Language Models: a cheaper attention cache for models that reuse layers
SummaryLooped language models reuse the same layers several times per token, which multiplies the key-value cache and slows decoding. The paper proposes Hybrid Latent Attention, which keeps exact keys and values for the last W tokens and stores older tokens as compact latents read directly by each loop. Applied to 1.4B and 2.6B Ouro looped models with frozen pretrained weights, it shrinks the cache 10.7x per token and improves decoding throughput by up to 7.4x at 16K tokens. It retains over 97% of original accuracy on most benchmarks.
In briefEnergy-efficient spiking neural networks for beam prediction using sensing data
SummaryThis paper studies beam prediction for millimeter-wave systems, using position and vision data to reduce beam-training overhead. It builds spiking neural networks (SNNs) and compares them with artificial neural networks (ANNs) under position-only, vision-only, and multimodal inputs. Compact ANNs perform better with position-only input, while SNNs achieve higher accuracy than ANNs at similar energy costs for vision and multimodal inputs, using up to an order of magnitude less energy. The authors note that SNN accuracy is limited by binary communication between layers rather than model size.
Nhan Thanh Nguyen, Phuong Nam Tran, Mengyuan Ma et al.
In briefCCDF: A Benchmark Dataset for Spotting Fake Surveillance Videos
SummaryThis paper introduces CCDF, a dataset for testing whether deepfake detectors can identify fabricated surveillance footage. It contains 1,840 videos (460 real and 1,380 generated) across 16 crime and accident categories, with generated clips made by three commercial systems. Ten recent detection methods were evaluated, and they did not reliably separate generated videos from real ones.
WhyTargets surveillance footage with three leading commercial generators, filling a gap, but the dataset is small at 1840 videos and evaluates existing detectors.
Baptiste Chopin, Thomas Swearingen, Arun Ross et al.
In briefGeneralized Matheron Variational Implicit Processes: Posterior Inference for Implicit Priors
SummaryThis paper introduces GMVIP, a pathwise variational family for posterior inference under implicit-process priors such as Bayesian neural networks and stochastic simulators, whose function-space densities are usually unavailable. Posterior samples are built by drawing a function from the prior and applying a correction anchored at inducing inputs. Experiments on regression, classification, and forecasting reportedly show it is broadly competitive with existing methods.
Luis A. Ortega, Andr\'es R. Masegosa, Thomas D. Nielsen
In briefFour-quadrant evaluation of Redpine Science, a literature access tool for AI models
SummaryThis report evaluates Redpine Science, a tool giving models and agents access to peer-reviewed literature via MCP and an API. It assesses both the relevance of retrieved results and answer quality with and without Redpine Science, using public benchmarks and an expert-validated question set. On ScholarQABench SciFact, an agent with Redpine Science answered 94.4% of claims correctly versus 87.6% without retrieval, and on the expert-validated set it stated 80.1% of required claims versus 70.2% for web-search-only.
In briefNonword probes show LLMs lack humans' sublexical sensitivity to pseudowords
SummaryThe study tested five large language models on two Italian two-alternative forced-choice pseudoword experiments and compared them with a human behavioural baseline. Models aligned more reliably with humans when real-word options gave a lexical familiarity cue, but fell well below the character-n-gram model fastText in the pseudoword-only condition. The sublexical cosine-similarity cue that drove human–fastText agreement did not consistently carry over to human–LLM alignment.
In briefSIGMA: A Method for Models to Improve Their Own Safety Alignment from a Spec
SummaryThis paper proposes SIGMA, a pipeline that lets a language model improve its own safety alignment using only a "Model Spec" describing desired behavior. The model first acts as a task designer to generate alignment dilemma scenarios, then trains through supervised fine-tuning and rubric-based reinforcement learning with itself as the reward model. Trained only on single-turn chat data, SIGMA reportedly reduces harmfulness in multi-turn agentic environments and outperforms Deliberative Alignment and Constitutional AI baselines while retaining general capability.
Jingyu Zhang, Shruti Palaskar, Daniel Khashabi et al.
In briefDynamic Alignment and Calibration for Multimodal Learning (ACML)
SummaryThis paper addresses two limitations in multimodal learning: static cross-modal alignment applies uniform constraints to all samples, which can cause unreasonable over-alignment, and confidence-based fusion does not fully account for differences in feature magnitude and confidence across modalities. The authors propose ACML, which includes a dynamic cross-modal triplet alignment module and a difference-aware attention calibration strategy. The paper reports that ACML consistently outperforms recent state-of-the-art methods in performance and robustness on multiple multimodal benchmark datasets.
In briefTF-PRVR: Training-Free Partially Relevant Video Retrieval
SummaryThis paper addresses partially relevant video retrieval, which finds untrimmed videos containing moments matching a text query. It proposes TF-PRVR, the first training-free framework for this task, using frozen vision-language features to build adaptive multi-scale segments and a graph that propagates query relevance. A moment-aware scoring step aggregates temporally aligned relevance across scales. The authors report consistent performance across datasets with diverse visual and temporal characteristics.
In briefOpenWAM: An Open Framework for Flexible Robot Models That Predict Video and Act
SummaryThis paper introduces OpenWAM, an open framework for world-action models that combine future video prediction with robot control. It builds on the Wan2.2-5B video model, pretrains it causally on over 10,000 hours of robot video, and adds an action expert through a shared Mixture-of-Transformers design supporting several generation orders. On LIBERO-Long, causal adaptation raised VTA success from 68.4% to 97.8%, and the framework is also used to study how counterfactual data improves inverse and forward dynamics models.
In briefPhysics-based sphere packing for morphing volumetric meshes used in design
SummaryThis paper presents JamTet, a sphere-packing method for volumetric tetrahedral meshing and morphing in differentiable simulation and design. It combines a GPU mesher, morphing that keeps interior node identities, and a JAX-based differentiable simulator. In soft-robot design experiments, interior-node gradients improved swimming fitness by 0.73 to 1.07 over a surface-only variant.
In briefCan artifacts be modeled explicitly? Disentangling them with pairwise edit relations for image manipulation localization
SummaryThis paper reframes image manipulation localization as a latent-variable problem, arguing that current models fail because they model editing artifacts only implicitly. It proposes a two-stage method, Pairwise Artifacts Learning followed by Standard Localization, which uses edit relations to disentangle artifacts and predict manipulation masks. The authors also release EditGroup-45K, a dataset organized into edit groups, and report consistent gains across several architectures.
In briefMultimodal Knowledge Distillation for Classifying Gastric Adenocarcinoma from Whole-Slide Images
SummaryThis paper addresses histopathological subtype classification of gastric adenocarcinoma from whole-slide images. It proposes a multimodal knowledge distillation framework in which a teacher model fuses image and pathology report text using Low-Rank Multimodal Fusion, and a student model learns from it to perform image-only inference. On the PatchGastric benchmark, the method achieves at least 3.35% higher mean accuracy than state-of-the-art approaches, without transformer-based fusion, multi-task learning, or large language models.
In briefA paper on customizing video motion while avoiding copying the reference video's appearance.
SummaryMotion customization in video generation often suffers from content leakage, where the reference video's appearance carries into the output. The authors propose Control-based Motion Customization (CMC), a training framework using Stochastic Optimal Control to steer generation toward the target motion without collapsing onto the reference. Experiments report reduced content leakage, competitive motion fidelity, preserved base-model diversity, and a 2.5 times training speedup.
In briefRevisiting temporal regularization for smoother robot control in reinforcement learning
SummaryDeep reinforcement learning policies often produce jittery actions that are hard to deploy on physical robots. The paper proves that a temporal penalty also bounds action differences between states sharing a next state, giving spatial smoothness. It proposes CATS, which combines this temporal penalty with linear ramp-up. Experiments in simulation and the real world report substantially reduced action oscillation without degraded task performance and with little computational overhead.
In briefContinuous Memory Machines: A Recurrent Network with Separate Short- and Long-Term Matrix Memories
SummaryRecurrent neural networks usually compress information into one vector state, so short-term computation and long-term retention share the same representation. This paper introduces the Continuous Memory Machine (CMM), which uses matrix-valued short-term and long-term memories with distinct roles, jointly updated by a Transformer. Across algorithmic, in-context learning, and recurrent reasoning tasks, the CMM outperforms a range of baselines and generalizes better than prior memory-augmented networks, while keeping interpretable attention patterns.
In briefIsotropic but Undecodable: A Sequential Content-Sufficiency Gap in Text Representations
SummaryThe paper studies whether text representations retain the ordered target information present in their input, and proposes CANOPE, a nonautoregressive framework using ordered latent canvases, canonical-token supervision, and geometric regularization. Experiments on text sequences and LJSpeech show that a representation with similar pooled rank can differ sharply in positional recall and downstream speech recognition error. The authors conclude that geometric regularity alone does not guarantee recoverable sequential content in the settings studied.
In briefApexQuant: Data-free quantization that re-randomizes residual errors to compress models
SummaryApexQuant is a calibration-free method that repeatedly quantizes the leftover error of an existing quantizer, using a fresh random rotation on each pass. The authors show that this rotation returns each residual to a uniform distribution on the hypersphere, which predicts how quickly the error shrinks. The number of passes a layer needs can therefore be chosen before reading any weights. The method was tested on four open-weight LLMs and on Earth-observation and medical data, reaching near full-precision results at four bits and the best two-bit results measured, without any data.
In briefAdapting vision-language models to long-tailed unlabeled data without supervision
SummaryThis paper introduces Unsupervised Long-Tailed Adaptation (ULTA), a setting where vision-language models are adapted using unlabeled data with a long-tailed distribution. The authors find that head-class performance drops sharply in this setting, unlike supervised long-tailed learning. They propose MARS, which combines Boundary-Preserving Alignment and Margin-aware Self-Refinement. Experiments on nine benchmark datasets show MARS improves average accuracy by 4.71 percentage points over state-of-the-art methods.
In briefARIA: Audio-based modeling of melody-tone relations for Cantonese lyric writing
SummaryCantonese lyric writing requires lexical tones to match melodic pitch closely. ARIA is a two-stage framework that generates Cantonese lyrics directly from singing recordings with character-level timestamps, rather than symbolic melodies. Its TRATE module predicts 0243 tone sequences from the singing audio, and DRA-TCLG generates lyrics conditioned on these predicted tones with retrieval-enhanced lexical guidance. The authors also built a large aligned audio-Jyutping-0243 dataset, and experiments report strong performance in tone prediction and tone-consistent lyric generation.
In briefIEEE 802.11bx: WLAN intelligent networking standard for AI-ready Wi-Fi 9
SummaryThis tutorial surveys recent progress toward Wi-Fi 9 within IEEE 802.11 standardization, covering the drivers and technologies behind an AI-ready design. It examines AI along three dimensions: as a protocol, as a platform, and as traffic, discussing candidate features and open challenges for each. A case study on AI traffic differentiation explores extending the current EDCA to support new AI traffic flows.
Francesc Wilhelmi, Katarzyna Kosek-Szott, Szymon Szott et al.
In briefVariance-averse n-step offline RL for sparse long-horizon tasks (VAN-Flow)
SummaryThis paper proposes VAN-Flow, a framework for generative offline reinforcement learning that favors reliable actions. It combines a categorical distributional critic, a variance-averse expectation operator that reweights return distribution atoms, and a flow-matching actor guided by rejection sampling. The authors report that across more than 40 tasks from D4RL and OGBench, VAN-Flow consistently outperforms strong baselines, with the largest gains in long-horizon and high-variance settings.
In briefFC-SWE: Reinforcement learning that uses failed patches to train coding agents
SummaryThis paper presents FC-SWE, a reinforcement learning framework for training software engineering agents on long, multi-step repository tasks. After a patch fails verification, the repository is reset and the failed patch plus verifier feedback is used as context for a recovery attempt. On all 500 SWE-bench Verified tasks, FC-SWE with Qwen3.5-4B reaches 41.7% Resolved@1, compared with 38.9% for GRPO.
In briefTextual Environmental Context and Spatial Graphs for LLM-Based Regional SST Forecasting
SummaryThis paper studies how to present heterogeneous information to a large language model for regional multi-step sea surface temperature (SST) forecasting without serializing the full SST grid as text. Historical SST and anomaly sequences, date-aligned environmental records, and static ocean knowledge form the textual context, while regional spatial state is supplied through graph-derived continuous prefixes produced by graph neural networks. On South China Sea forecasting, the complete configuration achieves the best MAE and R² among compared methods over ten forecast steps. A rule-based module also links predicted trends and environmental-factor directions to knowledge entries to provide source-linked, post-hoc explanations.
In briefRethinking LLM Faithfulness: Answering Versus Abstaining Across Paired Contexts
SummaryThis paper introduces PFaithBench, a benchmark that tests whether language models answer the same question correctly when context is sufficient and abstain when it is not. Evaluating 39 models from seven families, the authors find most models are biased toward answering, with most faithfulness errors coming from over-answering. They also show that training results depend heavily on how answering and abstaining data are composed.
WhyProposes pairwise faithfulness evaluation contrasting sufficient and insufficient contexts, validated systematically across thirty-nine models.
In briefTeaching Multimodal Models to Decline Impossible Image Edits
SummaryThe paper studies whether unified multimodal models recognize when a requested image edit is impossible and refuse it. The authors introduce Draw-or-Decline, a benchmark of 1,050 feasible-infeasible request pairs, and find that the strongest of eight evaluated models refuses only 0.4% of infeasible requests. They propose VisTA, a training method that pairs feasible and infeasible examples, and the resulting VisTA-BAGEL refuses 93.0% of infeasible requests while completing 74.3% of feasible edits.
In briefShanLiangRen: A Nutrition Agent for Personalized Daily Meal Planning
SummaryThe paper formulates personalized, fully quantified multiobjective dietary planning and presents ShanLiangRen, a nutrition agent for it. The system converts diet specifications, nutrient data, user attributes, and natural language requests into a constrained planning instance, then uses exact retrieval-augmented generation to narrow candidate ingredients and recipes. An LLM iteratively refines plans guided by Pareto principles, with deterministic nutrition computation and constraint verification. The system has been released as a WeChat Program.
In briefLabel-efficient deep learning for ECG waveform boundary detection, benchmarked against common delineation tools
SummaryThis paper studies ECG delineation, which identifies waveform boundaries in electrocardiogram signals. It compares self-supervised pretraining with supervised and semi-supervised fine-tuning across one internal and four external datasets. The selected deep learning model is then benchmarked against open-source tools (NeuroKit2, Prominence, ECGdeli) and the commercial tool CalECG, and it ranks best on every metric and dataset.
In briefLearned Adaptive Multiresolution Diffusion Imaging: RL Policy for Refinement in Image Methods
SummaryThe paper introduces Learned AMDI, which keeps the fixed-tree propagator and hierarchy constraints of AMDI but replaces the post-propagation selector with a shared local policy trained by proximal policy optimization. Across nine held-out cases, it reduces mean terminal reference discrepancy from 0.17496 to 0.13657 while raising occupancy from 0.13737 to 0.26660. The authors report that most of the gain on this static benchmark comes from the initial allocation rather than from sequential decisions.
Christian Tantardini, Stig Rune Jensen, Roberto Di Remigio Eik{\aa}s et al.
In briefComparing three BERT pretraining methods on a tiny Chinese model
SummaryThis paper compares three ways of pretraining a very small Chinese BERT model: standard MLM, Whole Word Masking, and MacBERT-style replacement. All models share the same architecture, corpus of about 1.29M Chinese Wikipedia sentences, and hyperparameters. At this tiny scale, MLM performs best overall, while WWM achieves lower perplexity and higher MLM hit rate, and MacBERT performs poorly due to a limited synonym dictionary.
In briefRevar3r: gauge-aware perturbation uncertainty for feed-forward 3D reconstruction
SummaryThe paper shows that training-free perturbation uncertainty can mistake an unobserved rotation in a frozen 3D model's output frame for error, making correct distant points look uncertain. The authors derive an error-independent variance term that grows with scene extent, and propose ReVar3R, which aligns repeated predictions to a common similarity frame before computing per-point variance, without retraining. Across VGGT, π3, and MASt3R on six datasets, it lowers AUSE below built-in confidence in 15 of 18 conditions. Compared with a trained evidential head, it trades off: the head calibrates magnitude better, while ReVar3R transfers across backbones without adaptation.
Sammam Mahdi, Fariha Binta Salim, Rakin Bin Rabbani et al.
In briefSelf-Referenced Social Preferences: Learning Cooperation Without Seeing Others' Rewards
SummaryThis paper studies cooperation in multi-agent reinforcement learning, where existing methods usually require agents to observe their peers' rewards. The authors propose self-referenced social preferences, in which each agent uses its own learned reward model to assess other agents' observed outcomes and feeds these assessments into social preferences. Tested on three sequential social dilemmas, agents learned cooperative behavior without observing others' rewards, and sometimes achieved more equitable divisions of returns than agents with access to true rewards.
Mohamed Ayman Mohamed, Harshil Kotamreddy, Marcos Menon Jose
In briefHow Could AI Eliminate Humanity? A Failure-Mode Analysis of Civilizational Risk
SummaryThis paper proposes a failure-mode framework for analyzing how advanced AI could contribute to human extinction, irreversible civilizational collapse, or permanent loss of human control. It argues that catastrophic AI risk need not involve consciousness, hostility, or intent to harm, and could arise through interacting pathways such as autonomous misalignment, harmful human use, organizational failure, and competitive deployment. The severity of these pathways depends on capability, autonomy, external access, persistence, institutional safeguards, and recovery capacity. The analysis is deliberately non-operational, identifying causal conditions, measurable intermediate quantities, and defensive research questions.
In briefOn-policy distillation with negative-policy rollouts: adding weaker-model samples as negative signals
SummaryOn-policy distillation (OPD) trains a student model on its own rollouts, using token-level supervision from a stronger teacher. When the teacher and student have limited distributional overlap, positive guidance alone may give weak learning signals. The paper proposes NP-OPD, which adds rollouts from a lower-capability negative policy as a negative reference while keeping teacher supervision. Experiments report improvements across model scales, generation modes, reasoning domains, and OPD variants, and analyses show that tokens the negative policy prefers over the teacher are suppressed.
In briefCueRator: Agentic Search for Rules That Adapt Frozen Multimodal Encoders
SummaryThis paper proposes CueRator, a framework where language model agents search for a decision rule that turns similarities from frozen audio-visual encoders into event predictions, without retraining the encoders. A lightweight policy sets the rule's parameters per video to balance adaptability and generalization. On OV-AVEBench, it raises total average from 57.8 to 60.2 and unseen-category accuracy from 55.8 to 59.9, shrinking the seen-unseen gap from 7.1 to 1.2.
In briefReFold: a training-free, reversible context folding method for long-horizon LLM agents
SummaryLong-horizon LLM agents re-send their full append-only interaction history at every step, so context and cost grow until the context window is exceeded. The paper proposes ReFold, a training-free rendering layer that keeps the full history but compresses only the rendered context, replacing already-displayed content with stubs and folding self-reported finished turns into one-line notes. Cached prefixes are rewritten in chunks, and every removal can be restored from history. The authors report up to 2.5x lower token consumption across five benchmarks and two frontier LLMs without degrading task success.
In briefA self-learning AI agent that improves X-ray diffraction analysis from experience
SummaryThis paper introduces Gan Jiang, a self-learning agent for powder X-ray diffraction built on the authors' XMatcher, XQueryer, XDecomposer and WPEM tools. It turns analytical experience into executable skills by diagnosing failures, revising skill instructions and code, and validating changes before reuse, without retraining the language model. The authors report higher refinement scores than the original expert-designed skills across FullProf, GSAS-II and PyWPEM, and leading single- and multiphase identification results on DeltaXRDbench.
WhyProposes a self-learning X-ray diffraction agent that validates revised skills, outperforming expert skills and baselines across several benchmarks and datasets.
In briefWorkflowOps: Learning Agent Collaboration Priors for Multi-Agent Workflow Orchestration
SummaryThe paper presents WorkflowOps, a framework for orchestrating multi-agent workflows that learns agent collaboration patterns from past workflows. It builds task graphs guided by a transition matrix of pairwise agent collaboration frequencies and creates new specialized agents when capability gaps are detected. Experiments on mixed code, math, and question-answering suites report higher end-to-end pass rates than recent workflow-construction baselines, with the largest gains on structured, decomposable tasks.
In briefTram-FL: cutting communication and computation in decentralized federated learning by passing one model between nodes
SummaryConventional decentralized federated learning gives each client its own model copy, which wastes computation and communication, especially with non-IID data. This paper proposes Tram-FL, which trains a single model sequentially as it circulates among nodes, with methods for choosing the route and the number of updates per node. It adds quantized momentum so that fewer circulations are needed while each transmission stays small. Experiments report that the algorithm converges to a global model with lower communication and computation costs, even with non-IID data.
Kota Maejima, Takayuki Nishio, Asato Yamazaki et al.
In briefA neural framework that learns driver-specific hidden route costs to reproduce individual vehicle routes
SummaryThis paper treats reproducing a driver's route as a shortest-path problem using latent link costs specific to each driver. A neural model estimates these personalized costs from individual traits, trip attributes, and network traffic conditions, then a shortest-path optimization layer produces the route. The model is trained end to end with decision-focused learning and an approximate gradient method, with a regularizer tying costs to observed travel-time scales. Results show better route reproduction than baseline route choice models, and the learned costs offer plausible explanations for varied route choices.
In briefConverting ternary language models to bf16 at export can revert most small learning-rate weight changes
SummaryThe paper audits export pipelines for ternary language models such as BitNet b1.58, Falcon-E and BitCPM, which cast high-precision latent weights to bf16 before producing ternary codes. The authors find that this cast changes many codes, sharply lowering GSM8K strict accuracy for some fine-tuned models. Two compatibility remedies are proposed and are reported to stay within a 4-point non-inferiority margin across the three models.
WhyShows bf16 casting in three labs' export pipelines flips ternary codes and sharply drops GSM8K accuracy, with two remedies tested against controlled comparisons.
In briefRA-MoWE: Clustering Queries with Workflow-Affinity Embeddings to Generate Agentic Workflows
SummaryThis paper proposes RA-MoWE, a framework that clusters queries using workflow-affinity embeddings, which record how well reference workflows solve each query. It then builds reusable expert workflows for each cluster through execution feedback, and an encoder lets new queries pick an expert without running the reference workflows. On a 300-query test set across mathematics, science, and programming, it improves average task score by 4.04 points over selecting among reference workflows while using 27.7% fewer LLM calls at inference.
