In briefLADDER: Diagnosing Domain-Level Self-Assessment Under Noisy Feedback
SummaryThis paper introduces LADDER, a framework for studying how language models assess their own abilities when feedback is unreliable. The models can accurately report performance information they are given, but they struggle to form trustworthy self-assessments from interaction experience. Seemingly accurate self-assessments can hide differences in capability and do not necessarily predict future task performance, so trustworthy self-assessment should be checked against observed task results and their predictive value.
In briefiPrefBench: A User-Centric Benchmark for Personalized Mobile Agents
SummaryThis paper introduces iPrefBench, a benchmark for testing whether personal mobile-use agents can infer missing user information from activity history and ask for details only the user can provide. It uses an executable iOS environment with a simulated user, built on 3,051 past activity events and 130 underspecified tasks across 26 apps. Evaluating seven frontier agents, the authors find high action completion but low full task success, with particular difficulty requesting necessary information from the user.
In briefFGD2HF: Frequency-Guided Dual-Domain Hypergraph Fusion for Semantic Segmentation
SummaryThis paper proposes FGD2HF, a dual-domain network that uses frequency features to guide spatial feature fusion and enhancement for semantic segmentation. It combines a hybrid attention encoder using neighborhood and hypergraph attention, a wavelet-based module for extracting frequency cues, and frequency-guided fusion modules. The authors report that the FGD2HF-B model reaches 48.7% mIoU on ADE20K and 83.2% mIoU on Cityscapes, remaining competitive with state-of-the-art methods at lower computational cost.
In briefSKALOS: Telling Similar LLM Agent Skills Apart Using Capability Graphs
SummaryThis paper addresses selecting the right skill for LLM agents when skill libraries contain many semantically similar candidates. The authors propose SKALOS, a graph-based reranker that compares skills through explicit capability evidence, using semiring-based witness learning and ranking-feedback graph calibration. On SkillRet, it improves NDCG@10 by 2.68 percentage points over a matched text-only reranker. On ALFWorld, witness-guided skill selection raises task success rate from 22% to 29%.
In briefHaMo: Combining external experience cards and model training for long-running continual learning
SummaryThe paper addresses long-horizon agents, such as Web agents, that repeatedly meet reusable multi-step procedures. It proposes HaMo, which quickly stores distilled strategy-level experience cards for retrieval while slowly training the model to interpret and execute those strategies. Experiments reportedly show HaMo outperforms harness-only and model-only adaptation under within-domain and cross-domain shifts, improving retention and adaptation.
In briefCEPO: A General Framework for Better Prompt Optimization in Multi-Agent Systems
SummaryThis paper introduces CEPO, a framework for improving prompt optimization in multi-agent systems. It separates system-level coordination, handled by a shared coordinator prompt, from agent-level task execution, handled by agent-specific executor prompts, and updates each according to its role. Experiments across 11 benchmarks in four domains show that applying CEPO to four existing prompt optimizers consistently improves their performance.
Jongha Kim, Seohyun Lee, Jonathan Minho Chung et al.
In briefPrefetch-Steered Routing: Faster Decoding for Offloaded Mixture-of-Experts Models
SummaryThis paper addresses Mixture-of-Experts models too large for GPU memory, where decoding is slowed by copying expert weights from host memory. It proposes prefetch-steered routing (PSR), which routes tokens using a deterministic logical cache filled with experts predicted by the next layer's router, so physical transfers affect only latency. On OLMoE-1B-7B, PSR decodes 2.2 to 2.3 times faster than exact routing with LRU and 25 to 30 percent faster than Cache-Prior, with small benchmark score differences.
In briefVLA-HWBench: Where Inference Time Goes for Vision-Language-Action Models on Edge Hardware
SummaryThis paper benchmarks 18 open vision-language-action (VLA) models on two edge devices, NVIDIA Jetson AGX Thor and AMD Ryzen AI Max+ PRO 395, using multiple inference backends and a seven-stage latency breakdown. It finds that the action-generation strategy matters more than model size, and that kernel fusion with 4-bit weights speeds up π0.5 and GR00T N1.7 by 2.6 to 2.9 times. After optimization, memory bandwidth limits only the stages that process few tokens on the Jetson, while software limits the rest.
WhySystematic benchmarking of 18 open VLA models on two edge platforms with seven-stage latency breakdown and optimization comparisons.
In briefSearchHorizonBench: A Benchmark for Judging the Value of Further Search in AI Agents
SummaryThis paper introduces SearchHorizonBench, a benchmark for learning and evaluating how valuable an additional search is for AI agents, based on execution experience. It contains 45,632 executions across 3,168 questions. On 1,632 held-out questions, historical calibration reduces Brier error by 16.2%.
In briefIntervene, Calibrate, Discover: Learning Causal Structure from Pixels
SummaryThis paper studies how to learn high-level causal variables and their graph from low-level image observations. It pairs image readouts with verified values from fixed round-robin interventions and uses a frozen, invertible per-dimension calibration map to align visual responses with intervention settings. The method is evaluated on graph discovery, intervention consequences, and counterfactual images across six environments. Controlled domino experiments show that interventions resolve orientation ambiguity that equal additional observations leave unresolved.
In briefSparse Field Reconstruction via Direct Measurement-to-Field Generation in Function Space
SummaryThis paper introduces DMF, a functional rectified flow framework that generates full spatial physical fields directly from sparse measurements. It learts the mapping end to end and evaluates fields at arbitrary query coordinates without grid rasterization. On benchmarks with unobserved boundaries, missing high-resolution training fields, and unmeasured quantities, DMF reduces reconstruction error and quantifies uncertainty through stochastic ensemble generation.
In briefSafe in Benchmarks, Unsafe in Practice: Lessons from 1,331 Child-Related AI Incidents
SummaryThe paper collects and codes 1,331 publicly reported AI incidents involving children, building a child-centered harm taxonomy and annotating the harm types and failure modes in each case. It finds that harms concentrate in representational, dignity-related, and psychological categories and most often co-occur with deployment-context, governance, and misuse failures rather than model defects. The authors conclude that model-level evaluation alone is insufficient and recommend measures such as age assurance and consent verification for depiction.
WhyBuilds a child-centered harm taxonomy and codes 1331 public incidents, showing harms co-occur with sociotechnical failures rather than model defects.
In briefStable continuous-time consistency distillation: an empirical study and multistep extension
SummaryThis paper independently reproduces continuous-time consistency models built on the TrigFlow framework. The students match reported FID at one and two steps on CIFAR-10 but fall short on ImageNet-64. The authors also propose MS-sCD, a segment-conditioned multistep extension, and report FID at M = 2, 4, and 8 against a discrete-time baseline.
Vinay Saji Mathew, Soundar R. Kumara, Gretta D Kellogg et al.
In briefInformation Rates of Summaries in Simulation-Based Inference
SummarySimulation-based inference often compresses observations into a fixed-dimensional summary, and the posterior depends only on that summary. The paper models the summary as a lossy channel and measures its information budget with mutual information. It shows that under a common-support condition, a stochastic summary achieves strictly lower distortion than deterministic ones at finite budgets, and tests a binary bottleneck in neural posterior estimators across four simulator benchmarks.
Vansh Bansal, Pradip Sharma Poudel, James G. Scott
In briefMovable-element multi-functional RIS hybrid RSMA-NOMA optimized with constrained reinforcement learning
SummaryThis paper proposes MM-RIS, a movable-element multi-functional reconfigurable intelligent surface that combines element mobility with reflection, transmission, amplification, and energy harvesting. It also designs a hybrid RSMA and NOMA transmission scheme to improve spectral efficiency. The authors maximize energy efficiency by jointly optimizing beamforming, power and rate allocation, RIS configuration, and element positions, using a constrained hybrid reinforcement learning method called CRAFTS. Simulations report better energy efficiency and constraint satisfaction than benchmarks in the literature.
In briefGangqinpu: A multi-genre piano dataset linking audio, MIDI, and sheet music
SummaryThis paper introduces Gangqinpu, a piano dataset with 18,786 score-performance pairs and 1,007.1 hours of audio, including automatic note-level alignments and beat/downbeat annotations. An automated pipeline transcribes audio, aligns MIDI to scores, labels genres, and filters out non-expressive performances. The authors evaluate the dataset on MIDI-to-score transcription, expressive performance rendering, and beat tracking. Results show cross-dataset gains for beat tracking and mainly in-domain benefits for the other two tasks.
In briefIsoAct: a post-hoc method that removes unwanted factors from representations while respecting their geometric shape
SummaryThis paper proposes IsoAct, a post-hoc framework for removing nuisance factors from representations that lie on curved manifolds. It finds nuisance-related directions with Fisher discriminant analysis in tangent coordinates and corrects them with actions that keep representations on the manifold. Experiments on hyperspherical, hyperbolic, and SE(3) representations show reduced nuisance recoverability with competitive utility.
In briefRSKD: Keeping Driving VLM Meaning Consistent via Distillation
SummaryThis paper presents RSKD, a distillation method that transfers knowledge from large multimodal models to compact autonomous driving models. It allocates supervision based on each answer phrase's task importance and its semantic consistency with the teacher's prediction, using teacher signals where they agree and stronger ground-truth supervision where they do not. The method is evaluated on DrivingVQA, LingoQA, and MapLMv2. It outperforms the compared student-based baselines on DrivingVQA and LingoQA and remains competitive on MapLMv2.
In briefMachine learning for the Sun's interior and atmosphere, framed as a unified state-space model
SummaryThis paper treats the Sun as a strongly coupled, multiscale dynamical system described by a global state vector of thermodynamic, kinetic, and magnetic properties. Because solar observations give only partial, projected views of this hidden state, the authors argue that machine learning's main role is as inference, emulation, propagation, and control operators that reconstruct and couple the state across physical layers. The chapter suggests data-driven methods can fill observational gaps, reveal hidden multiscale couplings, and speed up simulations of nonlinear processes.
In briefGood Rankings, Poor Choices? Rethinking Text-Only Selection of Vision-Language Models
SummaryThis paper studies selecting the vision-language model with the best zero-shot accuracy on a target task using only text. The authors argue that ranking metrics do not directly measure the performance of the chosen model, so they adopt simple regret and an outlier-robust normalized regret. They propose TC-ESR, a training-free method, which achieves the lowest mean regret across 22 target tasks and 43 models. The paper also identifies two sources of target performance leakage in the prior evaluation protocol and provides corrections.
In briefKnowing When Evidence Is Enough: Requirement-Directed Search with Horn-Closure Certificates
SummaryThis paper proposes Certify, a method for deciding whether a search-augmented model has enough evidence after retrieval. It decomposes a question into a requirement hypergraph, searches per requirement, accepts links only after entity checks, and uses Horn closure to produce a certificate of minimal sufficient passages or an abstention naming open requirements. Across three corpora it decides more accurately than search agents, graph retrieval and sufficiency judges, and its certificate-based answers are 80% correct.
WhyStructures evidence sufficiency via requirement hypergraphs and Horn closure, validated across three corpora and 21M passages, with notable gains over baselines.
In briefTopic-Level Fairness in Retrieval-Augmented Generation
SummaryThis paper argues that group fairness in retrieval-augmented generation systems, when measured only by aggregate metrics, can hide disparities within individual topics. The authors define topic-level fairness and adapt the reranker separately for each topic using a Bradley-Terry objective trained on preference pairs built from the generator's own output. On RoMQA and TREC Fair Ranking, the method reduces the mean absolute per-topic disparity by 51% and 40% respectively while retaining accuracy, outperforming baselines.
In briefBeyond Deduplication: Exact Retrieval Limits for Attention Under Correlated Distractors
SummaryThe paper studies how redundant context affects attention, separating the effect of repeated items, which add normalization mass, from score variation within correlated groups, which can create extreme individual scores. Using a planted retrieval model with Gaussian distractors arranged in large known clusters, the authors derive exact retrieval thresholds. They show that subtracting log cluster size from pre-softmax scores removes multiplicity but keeps within-cluster extremes, so the required signal stays at least √(2α₁); averaging scores before exponentiation achieves √(2ρα₁). A controlled experiment with frozen sentence embeddings shows averaging helps for singleton targets but can discard evidence when relevant groups contain multiple matches.
In briefPro-Active Inference: choosing which outcomes to label next to estimate a population mean efficiently
SummaryThe paper proposes Pro-Active Inference (PAI), a without-replacement method for estimating a population mean when only a small subset of outcomes can be labeled. It uses covariates and machine learning predictions to proactively select the next observation to label from the remaining pool, updates predictions online as labels arrive, and builds valid confidence intervals. The authors establish finite-sample unbiasedness and asymptotic validity, and compare PAI against several baselines across eight datasets.
In briefAT-AT: Automatically Improving Tools from Agent Execution Logs
SummaryThis paper introduces AT-AT, a method that improves shared tools offline by analyzing past agent execution traces while keeping the agents themselves unchanged. A lightweight coding agent reads trajectories and evaluation results, then edits tool descriptions and, when allowed, implementations. On τ²-bench-verified, it achieved the highest task completion rates in the Airline and Telecom domains, and it also showed gains on TerminalBench-2, TBLite, and three production MCP servers.
Maya Ravichandran, Pragya Tripathi, Dharamendra Kumar et al.
In briefExamining how straight learned visual representations are over time
SummaryHumans encode the changing natural visual world with a "straightened" representation, where frame-to-frame changes follow a more linear path at deeper processing levels. This paper studies how network architecture, different robustness types, biologically inspired filtering, and representational straightness relate when networks process time-varying input. It finds that adversarial training yields straighter CNN and transformer representations, but this benefit is task-dependent and does not help segmentation or frame prediction. Straighter representations also make class predictions more temporally stable.
Anne Harrington, Vasha DuTell, Ayush Tewari et al.
In briefAutomatically Generating Learning Curricula from LLM Solutions to Acquire Hierarchical Task Knowledge
SummaryThis paper studies how curriculum guidance can help acquire hierarchical task knowledge within an HTN learning framework. It compares online learning during problem solving, human-designed curricula, and an automated method that builds curricula from verified LLM-generated solution traces. Across several planning domains, the automatically generated curricula recover much of the benefit of expert-designed ones while substantially reducing the token cost of online knowledge acquisition.
In briefBenchmark testing whether household robots adjust motion to obstacle safety types
SummaryThe paper introduces OSMA-Bench, a benchmark using scenarios that differ only in obstacle type, and a Semantic Sensitivity Index (SSI) that measures whether clearance, speed, and smoothness follow obstacle caution ordering. Evaluated VLM-based and vision-language-action policies did not consistently adapt motion to obstacle safety, even with explicit safety prompts. The authors report that policies distinguish obstacles correctly but fail to reflect this in motion, and that adding an explicit parameter-inference step improves VLM-based policies.
In briefAgent Memory as Recommendation: Recommender-System Design for LLM Agent Memory
SummaryThe paper formalizes selecting stored memories for LLM agents as budgeted slate recommendation, treating the request as a user and memories as items. It proposes AgentMemRec, which adapts candidate generation, learning to rank, and slate construction from recommender systems and selects memories without any LLM calls. On LongMemEval's multi-session questions it outperforms several memory and context-selection methods, and on LoCoMo it outperforms Mem0 and Zep.
WhyFormalizes agent memory as budgeted slate recommendation and beats several baselines on multi-session questions without LLM calls, with solid empirical support.
In briefUsing hierarchical task network planning to manage goals as they change
SummaryThis paper examines whether Hierarchical Task Network (HTN) planning can uniformly implement the Goal Lifecycle, a process for goal reasoning in dynamic environments. The authors present a new goal reasoning agent architecture that uses HTN planning in formulating, expanding, and re-expanding goals, and an enhanced HTN action representation to handle discrepancies between expected and actual outcomes. Empirical results show that simulated autonomous vehicles can adaptively conduct missions in dynamic environments with this architecture.
In briefPaper arguing video world models need predictable latent spaces
SummaryThis paper studies how the latent space of video world models trained on robot data should be designed. The authors find that temporal objectives for representation learning, combined with a past-conditioned decoder, improve generative video world models at the same training budget. The reported gains include up to 5x cheaper inference, over 7x faster training, and lower pixel and perceptual errors out of distribution and over long horizons.
In briefStreamVibe: Proactive Video Agents That Track Changing User Intent
SummaryThe paper introduces StreamVibe-Bench, a benchmark with 4.1K evaluation and 85K training samples covering explicit, implicit, and dynamic user intents. It presents StreamVibe-VL, a streaming model that infers implicit intents from continuous observations and adapts its responses as intents change, along with StreamVibe-Harness, which keeps intents as persistent states in hierarchical memory. The authors report a 10.7% improvement in response timing and answer quality over baselines for StreamVibe-VL, and a 12.3% gain in proactive interaction with the Harness component.
In briefAdaTutor: A paper on adaptive tutoring that picks the best teaching strategy at each step.
SummaryThe paper introduces Comparative Pedagogical Supervision (CPS), a training framework that learns to choose the best teaching strategy and feedback for a student. CPS was evaluated on math and science reasoning datasets. After five tutoring turns, it outperformed the strongest fixed-strategy baselines. The learned tutor transfers across datasets, and its strategy selection can also inform a separate API-based tutor.
In briefThinking-in-Bets: a training-free way to stop LLM agents from caving to user pushback
SummaryThe paper introduces Thinking-in-Bets, a training-free method in which an agent privately records its main claims, supporting checks, and confidence before reading user feedback. When challenged, it uses this record to keep verified work intact. Across five Gemini models and eight benchmarks, recording claims before criticism nearly eliminated false reversals and reduced unnecessary file edits, while also improving answer accuracy and confidence estimates in non-interactive tasks.
In briefActing While Listening: Early Tool Calls in Spoken Dialogue
SummaryThis paper studies when a spoken agent has heard enough to call a tool, and proposes a synthetic spoken-dialogue pipeline that labels the earliest point at which a request and its arguments become actionable. The authors build 26,694 dialogues (786 hours across 48 domains) and train an omni-modal language model that makes causal, chunk-level decisions. It reports tool-state macro-F1 of 0.48 and 0.47 on held-out and external data, and estimated leads of 1.1 to 2.0 seconds over turn-end calling for correct calls.
In briefDream2Assist: Intent-Conditioned World Models for Shared Vehicle Control
SummaryThis paper presents Dream2Assist, a model-based reinforcement learning framework for shared human-machine vehicle control. It extends a recurrent state-space world model with intent inference so the assistive policy can anticipate and support discrete human goals without overriding human agency. Validated through human-subject studies in high-speed racing (N=30) and urban driving (N=10) plus a preliminary on-vehicle deployment, intent conditioning improved torque alignment and reduced collision rate.
Jonathan DeCastro, Thomas Balch, Emily Sumner et al.
In briefRobotAgentBench: A benchmark for open-world robot autonomy
SummaryThis paper introduces RobotAgentBench (RAB), a benchmark for evaluating robot agents in realistic open-world household environments. Using a human-in-the-loop real-to-sim pipeline, it builds 50 household layouts with 90 atomic tasks, 18 short-horizon tasks, 8 long-horizon tasks, and 4 open-ended scenarios. The benchmark assesses task discovery, context reasoning, manipulation, failure recovery, and robustness, and compares standalone VLA/WAM models with agent architectures. The experiments reveal substantial gaps between manipulation skill and overall autonomy.
In briefPredicting hematoma expansion by brain region from non-contrast CT in spontaneous intracerebral hemorrhage
SummaryThis study built and validated an interpretable, region-based model using non-contrast CT images to predict hematoma expansion in spontaneous intracerebral hemorrhage. Derived from the ERICH study and validated on VISTA data, it uses transport-based morphometry and linear classifiers across lobar, basal ganglia, and thalamic regions. The combined model achieved an AUROC of 0.787 in validation and outperformed clinician-based and machine-learning alternatives.
Natasha Ironside, Hongyu Kan, Kristofor Pas et al.
In briefSlowLab: Testing AI Agents on a Simulated Year of Experiments They Cannot Undo
SummaryThis paper introduces SlowLab, a benchmark in which an AI agent runs a one-year greenhouse experiment campaign across four simulated compartments with hidden site conditions. The agent chooses planting times, climate policies, crop stops and measurements, then recommends one policy that is scored in unseen weather. In a preregistered study on 48 unseen sites, three low-cost language-model agents performed near a standard grower setting and below a Bayesian-optimisation baseline, and in-season measurements did not change their planting decisions.
WhyProposes a year-long, irreversible-experiment agent benchmark validated in a preregistered study on 48 unseen sites; novelty is notable, impact mainly within agent evaluation.
In briefInferPriv: Testing LLM Assistants for Indirect Privacy Leaks
SummaryThis paper introduces InferPriv, a benchmark that checks whether a curious recipient can infer a user's sensitive attributes from an LLM assistant's reply to a routine task. Across five frontier models, inferential leakage substantially exceeds direct leakage. Even with a privacy-aware system prompt, several models show near-zero direct leakage while still leaking inferentially at notable rates.
In briefCross-Loop Query Handoff: Later Loops Redirecting Two-Hop Retrieval in Ouro
SummaryThis paper studies how later loops in the pretrained looped language models Ouro-1.4B and Ouro-2.6B repair two-hop retrieval, by transplanting attention queries between loops at the same layer, head, and final prompt position. On synthetic questions, injecting loop-3 queries into loop 1 rescued most questions wrong at loop 1 but right at loop 3. The results identify a transferable set of late-loop queries supporting bridge-conditioned retrieval, but do not locate where the bridge is computed or establish a single-vector handoff.
In briefIntrinsically motivated AI agents that match fruit fly exploration and future-action neural coding
SummaryResearchers combined whole-brain recordings from fruit flies with a neuromechanical digital twin to test whether optimizing intrinsic goals produces embodied agents matching fly behavior and neural representations. Trained without external reward using 20 intrinsic motivation algorithms, the hierarchical competency method DIAYN achieved the strongest joint match to flies in behavioral and neural prospective coding. Its latent activity also captured the sum-difference rule for interhemispheric kinematic encoding.
WhyCombines intrinsic-motivation RL with whole-brain fly recordings and a digital twin, showing DIAYN reproduces prospective neural turn coding; novel and well-supported, but impact mainly within computa
In briefAgentProofBench: A Benchmark for Reusable Web Agent Verifiers
SummaryAgentProofBench is a benchmark for evaluating reusable verifiers for web agents, which must apply a fixed acceptance rule even as requirements change and support later execution. It contains 97 tasks across 79 families and four web services. Experiments with DeepSeek Flash and Pro show that complete-panel and decision-replay tests retained 82 of 96 source programs for Flash and 79 of 96 for Pro across 48 families.
In briefCodeMeepo: A Multi-Agent Workflow for Migrating and Optimizing Whole Code Repositories
SummaryThis paper presents CodeMeepo, a multi-agent pipeline that migrates entire large code repositories to a new programming language while preserving their behavior. Each migration milestone is checked against a fixed external test oracle that the agents cannot change, and a parity review adds any missing functionality back into the plan. The pipeline also includes a benchmark-guided performance optimization phase. On nine subjects across three language pairs, the authors report the best pass rate on every subject and 18% more passing tests than a state-of-the-art multi-agent baseline using the same model.
In briefRIDE-Splat: 3D reconstruction that removes windshield reflections from real driving video
SummaryRIDE-Splat is a method for reconstructing 3D scenes from posed driving video while removing windshield reflections, without needing clean images, reflection masks, or pretrained networks. It works in two passes: first it learns the 3D geometry and a signed reflection image, then it fits a clean 3DGS model using the reflection-removed training images. The authors also release the RIDE-Splat Dataset, the first 3D reconstruction benchmark for reflection removal in driving video.
SummaryEntity matching decides whether two records describe the same real-world entity. This paper examines decision models, which return probabilities over predefined answer options, as an alternative to generative LLMs and fine-tuned pre-trained language models. Six decision models were evaluated zero-shot on eight benchmarks against GPT-6 Luna and Ditto. The four strongest decision models achieved mean F1 scores close to GPT-6 Luna, and the best model outperformed Ditto on six of the eight benchmarks.
In briefOnline fair division of goods using frequency predictions
SummaryThis paper studies online allocation of indivisible goods among agents when the algorithm knows each agent's multiset of values in advance but not the arrival order or how values match across agents. It shows that a positive approximation of envy-freeness up to one good (EF1) is achievable for three or four agents, but impossible for five or more. The paper also presents results on EFX, polynomial-time EEFX allocations, and noisy predictions.
WhyGives a complete feasibility boundary for online fair division with frequency predictions, with tight impossibility proofs for five or more agents and rigorous positive results for fewer agents.
In briefOne mean log-mel vector per phoneme: a simple, fast way to inject text into speech recognition training
SummaryThis paper replaces a neural text-to-speech model with a lookup table that gives each phoneme a mean log-mel vector, used to inject unpaired text into speech recognition training. Training is 1.8 times faster than with the neural TTS, with comparable word error rates. With an external language model, the system reports state-of-the-art results on LibriSpeech using only LibriSpeech paired data, text, and lexicon, according to the authors.
WhyReplaces neural TTS with a per-phoneme mean log-mel lookup for text injection, achieving 1.8x faster training at comparable WER; thorough experiments and ablations on LibriSpeech and 25,000-hour Loqua
In briefP2P-Former: Closing the Gap Between Point-Level and Proposal-Level Supervision for Temporal Action Localization
SummaryThis paper addresses the performance gap between point-supervised and fully supervised temporal action localization, where each action instance needs only one annotated frame. It proposes P2P-Former, a Transformer-based framework with a point-aware proposal generation strategy, a point-guided semantic enhancement module, and a point-to-proposal constraint loss. The authors report that the method outperforms several existing state-of-the-art approaches on THUMOS14 and ActivityNet-1.3.
In briefEarly detection of memorization in diffusion models via sequential monitoring and trajectory-level calibration
SummaryThis paper studies how to detect when diffusion models reproduce their training examples during the denoising process. The authors compare two ways of combining split conformal calibration with sequential monitoring: per-step Bonferroni calibration and direct trajectory-level calibration of the maximum memorization score. Experiments on an unconditional CIFAR-10 EDM show that sequential monitoring beats fixed-time detection, and trajectory-level calibration achieves comparable power while using the false-alarm budget less conservatively and detecting earlier at higher noise levels.
In briefCSI-TextBench: A Dataset and Benchmark Linking WiFi Sensing Signals to Language
SummaryThis paper introduces CSI-TextBench, a large-scale dataset and benchmark that aligns WiFi Channel State Information (CSI) with natural language. It combines three sensing benchmarks into 285,790 samples across seven tasks, paired with 2,291 hierarchical text descriptions. The authors compare five alignment methods and report that contrastive CSI-text alignment matches supervised classifiers without class labels and handles open-vocabulary prompts. A unified model reaches 86.9% top-1 accuracy on 140 composite scenes, outperforming task-specific classifier ensembles.
WhyAligns WiFi CSI with language via a large dataset and compares five alignment methods, with substantial experimental support for a subfield.
In briefA test set for measuring AI propensities, with rubrics, synthetic datasets, and annotations
SummaryThe paper releases a battery of 28 AI propensities, with expert-written rubrics, synthetic datasets, and annotated demand intervals. The propensities are drawn from psychometric, psychological, and behavioral research so results can be read alongside familiar human-behavior constructs. Each synthetic dataset is designed to be sensitive to only one propensity, enabling quick prototyping of propensity profiles.
In briefHow Much Graph Is Too Much? Graph Construction and Exposure in Spatio-Temporal Forecasting
SummaryThis paper asks whether graph neighbors improve forecasts beyond each node's own history, and how the choice of graph and the way graph information enters the model affect accuracy. Experiments on U.S. state unemployment, traffic speed, and electricity demand compare a GRU using only own history and randomized graphs against graph-based models. Late integration with a fusion gate yields lower mean error on all three tasks, and a residual-correlation graph has the lowest mean error among the tested graphs.
In briefDo Text Prompts Matter? Scale and Geometry in Vision-Language Anomaly Detection
SummaryThis paper asks whether text prompts in vision-language anomaly detectors contribute semantic meaning or only scale and geometry. For a single anchor pair, the authors prove that replacement anchors with the same difference-vector length yield identical outcomes, so effective scale alone matters. Experiments across seven medical datasets and other settings show that scale-matched random or orthogonal anchors perform like correct prompts, and rotating category axes can improve industrial benchmark results.
WhyProves single anchor pairs act only through difference-vector scale, overturning semantic attribution, with broad multi-dataset experimental validation.
Arda Baris Basaran, Mesut Ceylan, Elgar Fleisch et al.
In briefAudio-R1: reinforcement learning that uses self-supervised audio tasks to improve audio reasoning in AI models
SummaryThis paper introduces Audio-R1, a reinforcement learning post-training framework for large audio-language models. It uses answers automatically produced when building self-supervised audio tasks as verifiable rewards. Experiments across several models and audio benchmarks report consistent gains, including improvements on downstream tasks different from the training tasks.
In briefFlexiMIND: Training-free deformable medical image registration combining pretrained vision features and local structure
SummaryThis paper presents FlexiMIND, a training-free, label-free two-stage framework for deformable medical image registration. Frozen FlexiCT-2D encoder features establish contextual correspondence, and MIND-SSC refines the deformation within ConvexAdam. The authors compare five feature-coupling strategies across CT-CT, MR-MR, and MR-CT datasets. Only four of fifteen comparisons outperformed the reference method, with the largest Dice gain being 0.0043.
In briefAn AI workflow that segments liver lesions and writes structured radiology reports
SummaryThis paper presents an evidence-grounded agentic AI framework that links liver and lesion segmentation with deterministic quantification and structured reporting. nnU-Net models segment the liver, Couinaud segments, and lesions, and computational tools extract volumes, dimensions, counts, and locations, which a large language model organizes into a predefined schema with numerical consistency checks. On 51 LiTS cases, liver Dice was 0.947, lesion sensitivity 75.6%, and positive predictive value 68.9%. Exploratory screening of 25,494 MERLIN reports found 10,208 (40.0%) with liver-related abnormalities.
Fatma Ahmed Mansour, Rahatara Ferdousi, Manar Fathy Elshenawy
In briefFine-tuning a prompt-based image editor to remove furniture from rooms while keeping the room's geometry consistent
SummaryThis paper addresses erasing furniture from a single room photo to produce an empty version, for uses such as diminished reality and virtual staging. The authors propose a training recipe that finetunes a LoRA adaptor for FLUX.2 on 2,500 furnished and empty room pairs. They report that the adapted model significantly improves geometric consistency while retaining high texture fidelity.
Prakhar Kulshreshtha, Brian Pugh, Zachary Nussdorfer et al.
In briefDoes the outfit change the price? Context effects in AI valuation of a cheap necklace
SummaryResearchers photographed the same $2.43 gold-tone necklace in four settings and asked six frontier language models to estimate its price. Among the four image-capable models, the formal-attire photo was valued 1.9 to 3.3 times higher than the flat-lay photo, all significant (p < 0.01). A text-only version of the task showed similar ratios, indicating the effect does not depend on pixels. Flat-lay estimates were already 7.8 to 18.7 times the receipt price.
SummaryThe paper proposes Sparse SOS Layers, differentiable components that use semidefinite programming to certify polynomial inequalities, with backward passes that reuse the forward solve's sparse structure. Compared with existing methods, it reports up to 2.7 times faster total warm-call time and about 100 times lower memory use (1.7 vs. 160 GiB) on relatively large dense-parameterization test problems. Experiments span pendulum control, robotics projection, and pose graph synchronization, and a design study reports zero certified violations on unseen instances with about 7 times fewer solves than finite differences.
In briefHARP-LoRA: Header-Hierarchy-Guided Low-Rank Adaptation for Scientific Tables
SummaryThis paper introduces HARP-LoRA, a low-rank adaptation method that conditions its projection on the row and column header hierarchies of table tokens, aimed at understanding scientific tables. The pretrained model and random projection tensors stay frozen, and only a small number of parameters are trained. The authors evaluate it on 950 annotated pharmacokinetic tables and on WikiTableQuestions, reporting gains over LoRA on both extraction and reasoning.
Hossein Sholehrasa, Mohammadhossein Amini, Doina Caragea et al.
In briefLarge Language Models: A Plain-Language Guide for Engineers
SummaryThis paper gives engineers a plain-language overview of large language models (LLMs), covering the Transformer, pre-training, fine-tuning with human feedback, retrieval-augmented generation (RAG), chain-of-thought prompting, and tool-using agents. It also describes cost-reducing methods such as FlashAttention and QLoRA. A checklist helps readers judge evaluation claims, flagging data contamination, LLM-as-a-judge grading, long context, and hallucination.
In briefHobbyLM: A Hobby-Budget Sparse Mixture-of-Experts Model Family
SummaryThe paper presents HobbyLM, a family of small sparse Mixture-of-Experts language models trained from scratch on rented GPU time costing a few hundred to a few thousand dollars. A 500M-parameter model is post-trained into chat, diffusion language, agent, and multimodal variants that run on a laptop CPU. The authors also report HobbyLM-1B, a 1.04B-parameter model, and HobbyLM-Image, a 333M-parameter image generation diffusion transformer.
Harish Santhanalakshmi Ganesan, Hariharan M. S., Prabhurajhan R
In briefRethinking Verification in Agent Harnesses: Independent Review as a Practical Test-Time Scaling Method
SummaryThis paper studies how verification should be organized in agent harnesses, comparing 17 designs across four lifecycle layers on seven frontier coding models. Independent review by a critic offers a more favorable accuracy-cost tradeoff than other options, improving average accuracy on Terminal-Bench and SWE-bench Verified. The authors also introduce VisualBench for evaluating visual and interactive artifacts.
In briefCumulative Marginal Matching: a training objective that aligns one-step generators with the full diffusion trajectory
SummaryDiffusion and flow-matching models need hundreds of denoising steps, which makes sampling slow. This paper proposes Cumulative Marginal Matching (CMM), which directly optimizes the Wasserstein-2 discrepancy between base and few-step marginal distributions across the whole generation sequence, using a tractable surrogate built from intermediate marginal samples. The authors show that the existing Shortcut-CSL objective is a first-order approximation of CMM. Experiments on standard benchmarks report better generation quality than Shortcut-CSL and other one- and few-step baselines under identical training budgets.
In briefGenerating Synthetic DVL and Inertial Signals with Diffusion Models for Underwater Navigation
SummaryAutonomous underwater vehicles navigate without satellite positioning and rely on Doppler velocity log and inertial sensors, but training deep learning methods for them usually requires costly sea trials. This paper proposes a conditional diffusion framework that generates synthetic DVL and inertial signals to replace real recordings in training. Across thirteen real trajectories, the generated signals matched the requested conditions with Pearson correlations of 0.72 to 0.997 and transferred to unseen trajectories. A velocity estimator trained only on generated signals achieved error comparable to training on the full real dataset.
In briefRefinedGAW: A refined German academic web dataset for language model training
SummaryThis paper presents RefinedGAW, a processed and annotated German text dataset built from 26 semi-annual crawls of the German Academic Web. It contains 122 million documents and 314 billion tokens, with academic research content making up 84% of the German tokens. The authors also release a queryable metadata index and a tool for retrieving and processing source documents.
Maximilian Idahl, Oleh Astappiev, Benedikt Droste et al.
In briefAgentic AI framework for spatial decisions, tested in a supply-chain deployment
SummaryThe paper presents ANASSA, a reference architecture that combines structured reasoning, multi-agent workflow orchestration, authoritative spatial grounding, and feedback-driven execution. It reports an operating supply-chain deployment covering multi-hazard risk assessment on an ocean-and-inland freight lane, over a three-month proof of concept. Reported results are a 25% improvement in data quality, a 50% reduction in spatial workflow runtime, and 60% lower decision latency versus the prior manual workflow. The authors note these are uncontrolled observations and specify instrumentation and ablation protocols for future mechanism-level evaluation.
In briefError Diffusion Replaces Latent Weights for Training Binary Language Models
SummaryBinary-weight language models usually keep a full-precision latent weight and Adam state for each parameter, making training costly and on-device adaptation impractical. This paper replaces the latent weight with a 2-8-bit stochastic-rounding residual buffer, cutting per-weight training state from 97 bits to 3-9 bits. The proposed error-diffusion (ED) optimizer approaches latent-weight reference accuracy when the buffer's quantization grid aligns with the flip threshold.
WhyTransfers error diffusion from image halftoning to weight updates, replacing latent weights with few-bit buffers, validated across several language model scales.
In briefFlockWorld: Simulating Large Crowds of Agents to Train Multi-Agent Video Models
SummaryThis paper introduces FlockWorld, an artificial-life environment based on Reynolds's Boids algorithm for studying video world models with large populations of agents. It renders synchronized first-person views for camera agents and lets their number be set independently of the surrounding crowd. The authors also present FlockDiT, a flow-matching latent diffusion transformer that jointly generates synchronized egocentric videos for multiple agents, and report that joint generation helps agent detection at matched densities and diffusion forcing improves rollout fidelity.
In briefRecursive CLIP: Reusing Looped Transformer Blocks for Vision-Language Distillation
SummaryThis paper examines whether CLIP-style vision-language models can reuse a small set of transformer blocks repeatedly instead of using independent blocks at each depth. The authors build rCLIP, with a recursive vision encoder, and r²CLIP, which also makes the text encoder recursive, and train them with direct contrastive learning and distillation from a ViT-B/16 teacher pretrained on DFN-2B. The models are evaluated on zero-shot classification, retrieval, few-shot learning, visual grounding, and concept unlearning. The results show CLIP tolerates substantial block reuse, with effects that depend on the encoder and task, and that recurrence reduces the number of distinct parameters.
In briefGRASP: Group-adaptive RL-SFT training for medical VQA without reasoning labels
SummaryThis paper addresses medical visual question answering, where training data has only short answers and no reasoning annotations. It proposes GRASP, which combines reinforcement learning and supervised fine-tuning in every update using only short answers, with their balance adapting to per-question performance. The authors report that tuning about 1% of parameters with LoRA yields better results than compared objectives, and they release the MedTerse dataset of 12.9K short-answer questions.
In briefHow Much Facts a Transformer Can Memorize: An Experimental Study of Memory Capacity and Model Size
SummaryThis paper examines how efficiently transformer models can store factual knowledge. Using Bayesian optimization for hyperparameter search, the authors measure the smallest model that perfectly memorizes synthetic datasets of random, independent fact triples. They find that memorization capacity grows exponentially with the transformer dimension, so the minimum dimension needed grows logarithmically with the number of facts. Increasing depth at fixed width yields sublinear growth in the memorizable dataset size.
WhySystematic Bayesian-optimized experiments show memorization capacity grows exponentially with transformer dimension, revising prior linear findings with solid empirical evidence and notable subfield i
Amirhossein Basareh, Rishabh Sanjay, Daniel de Mello et al.
In briefHIFI: An efficient framework for statistical inference with foundation models
SummaryThe paper proposes HIFI, a two-stage fine-tuning framework for statistical inference on parameters such as conditional means. Stage 1 adapts a foundation model once on a larger sample to control estimation bias, while Stage 2 applies resampling to a smaller independent sample, fitting multiple lightweight PEFT estimators to the residual to quantify variability. Across vision and text experiments, HIFI yields accurate point estimates and well-calibrated confidence intervals, reducing training compute by up to 91.9% relative to direct resampling.
In briefMOFjson: A Structured Format Helping Language Models Read Metal-Organic Framework Topology
SummaryExisting text formats for metal-organic frameworks (MOFs) either list raw atomic coordinates or use non-spatial identifiers, which limits language models. This paper proposes MOFjson, which splits structural data into a raw atomic layer with explicit bond shift vectors and a derived layer showing building units and network topology. The authors also introduce a crystal-derived benchmark across four question categories, where MOFjson outperforms three baselines in every category. Ablating the shift labels sharply reduces accuracy, which the authors take as evidence that models read the periodic graph itself.
In briefVTC-AUDIT: Checking Visual-Token Compression Reliability Under Paired Image Shifts
SummaryThe paper argues that visual-token compression is usually judged only by clean-image accuracy and efficiency, which can hide failures that appear under image shifts where the uncompressed model remains correct. The authors propose VTC-AUDIT, which estimates this added risk on image pairs where both the full and compressed models are correct on the clean image. Across three open vision-language models, four tasks, and eight shifts, reselecting retained positions on the shifted image raised failure rates in all six model-compressor settings. Support-set instability proved a weak ranking signal, while the compressed model's clean-input uncertainty was substantially more informative.
In briefTesting whether several latent-position random graphs are jointly independent
SummaryThis paper proposes the graph Hilbert-Schmidt independence criterion (GHSIC), a measure of joint independence for M≥2 latent position random graphs. The estimator relies on spectral decomposition of adjacency matrices and is shown to converge at a rate of Õ(n^(-1/3)), without requiring kernel or hyperparameter choices. Experiments on synthetic data and on brain functional connectivity subnetworks in natural language perception demonstrate its effectiveness.
In briefSTEP: Scoring Prediction Correctness via Per-Step Soft Probabilities
SummaryThis paper proposes STEP, a training-free method for estimating whether a large language model's prediction is correct without ground truth. It scores each reasoning step with lightweight forward passes to obtain soft probability distributions over candidate answers, then derives trajectory features capturing per-step confidence and its progression. Experiments on various benchmarks show consistent gains in AUROC over output-based baselines and hard-indicator trajectory methods.
In briefResource allocation for multi-view semantic communication in edge AI
SummaryThis paper proposes an end-to-end framework for semantics-aware resource allocation in multi-user Edge AI. It extends the Opportunistic Information Bottleneck to a multi-view Gaussian version (MGIB) that fuses complementary sources without variational approximation or retraining. Compression level, transmission mode and PRB assignment are jointly formulated as a MILP and solved with a Lagrangian Relaxation algorithm. Simulations show near-optimal utility and roughly double the PRB efficiency of average allocation.
In briefAcute Stroke Classification with Asymmetric Standardization and CNN-ViT Fusion: Multicenter External Validation
SummaryThis study presents a hierarchical diagnostic system for distinguishing ischemic from hemorrhagic stroke, combining CNN and ViT backbones with cross-attention fusion and an orthogonal-focal loss. The system reached a 95.48% F1-score internally, but performance fell to the 58-62% range in multicenter external validation across three datasets. The authors report low sensitivity for ischemic stroke and a tendency toward hemorrhage overdiagnosis.
In briefHow humans learn unfamiliar worlds: a think-aloud dataset on ARC-AGI-3
SummaryThis paper presents a dataset recording how humans discover and use mechanisms in unfamiliar environments, using the ARC-AGI-3 benchmark. It includes 144 human trajectories from 64 participants with think-aloud audio, plus 250 trajectories from LLM agents. The authors compare cognitive patterns between the two groups, finding qualitative representations more common in successful human trajectories (50.0% versus 14.3%).
In briefConviction Without Evidence: Verdict Stability in Medical Vision-Language Models
SummaryThis paper examines whether medical vision-language models truly rely on diagnostic images when answering questions. The authors replace real images with blank placeholders and apply two levels of adversarial pressure to 60 Italian State Examination items, comparing how often answers are overturned. They find that ungrounded verdicts are not more fragile, and that common epistemic signals do not distinguish grounded from confabulated diagnoses.
Federico Felizzi, Francesco Andrea Causio, Bianca Destro Castaniti et al.
In briefTRIAGE: Speeding Up Robot Vision-Language-Action Models by Reusing and Pruning Visual Tokens
SummaryThis paper addresses slow inference of vision-language-action (VLA) models on onboard robot hardware, where skipped visual tokens can be either reused from cache or pruned. It finds that reuse is preferable only when a token's drift stays below a calibrated threshold. The proposed TRIAGE method splits a fixed recomputation budget across recompute, reuse, and prune decisions. On LIBERO and SimplerEnv it stays close to full recomputation in success rate and cuts prefill latency on a Jetson AGX Orin by up to 48%.
In briefAmbiguous Fits: Rashomon Effect and Challenges in Scaling Law Estimation
SummaryThis paper examines the fitting stage of neural scaling law estimation, which is often handled with standard local optimization. The authors note that such heuristic searches cannot guarantee a nearly optimal fit and may hide different parameter choices that fit the data almost equally well. They show that standard fitting exhibits a Rashomon effect, where near-optimal parameter regions produce extrapolations that diverge substantially at scale. The paper proposes a deterministic branch-and-bound algorithm that guarantees near-optimal fitting loss and maps the resulting range of scaling curves.
In briefThe Price of Scale in Proactive Memory Recall: A Paper on False Alarms in Memory Retrieval
SummaryThis paper studies when a long-lived assistant should proactively surface stored memories without being asked, framing each decision as a multiple-testing problem. It shows that at a fixed false-alarm rate, the hit rate drops sharply as the memory store grows, using streaming replays of LongMemEval and LoCoMo. The authors also propose a label-free cue-swap conformal gate with a finite-sample false-alarm guarantee under exchangeability.
WhyFormalizes proactive recall as multiple testing with a proven power-law decay constant, validated on LongMemEval and LoCoMo across several embedding models with a conformal gate.
In briefRecovering traceability links between requirements that include images in industrial settings
SummaryThis paper extends a retrieval-augmented generation approach for recovering traceability links between requirements, including requirements that contain images. It retrieves text and image candidates separately, classifies pairs with a vision-language model, and filters predicted links by verbalized confidence. On industrial test sets it reaches an F1 score of 0.310, versus 0.114 for the replicated Hey et al. baseline. Recovery is still limited when image queries fail to retrieve related text requirements.
Haoming Huang, Shinpei Hayashi, Kazuya Kohata et al.
In briefBest Policy Tracking in Reinforcement Learning: Examining How to Select the Best Policy When Performance Fluctuates During Training
SummaryThis paper finds that reinforcement learning policies often fluctuate in performance during training, so common reporting practices, such as using the final policy or the best periodically evaluated one, may not identify the best policy visited. The authors introduce post-hoc diagnostic metrics and compare selection strategies on standard continuous-control benchmarks with an on-policy actor-critic method. They report that widely used practices can deviate considerably from the best policy and recommend a computationally constrained, validation-based selection strategy.
Judith Echevarrieta, Etor Arza, Aritz P\'erez et al.
In briefCaliMem: Search-Efficient Data Agent via Calibrated Memory
SummaryThe paper argues that data curation must adapt to the downstream model and task, since fixed pipelines can help one model and hurt another. It proposes CaliMem, an agent that predicts the outcome of each curation policy before running it, then compares the prediction with the result to retain, revise, or qualify its claims. Across biological and LLM post-training tasks, CaliMem outperforms fixed curation baselines and published methods. The code is open-sourced.
In briefApproximate Unlearning for Prototype-Based Classifiers: Removing Sample Effects from Trained GLVQ Models
SummaryThe paper proposes a robust approximate unlearning method for prototype-based classifiers, specifically Generalized Learning Vector Quantization (GLVQ), to remove the effects of specified samples such as deletion requests or noisy data. It examines how forget-set size and data characteristics affect the unlearned model and adapts the technique to be robust to these factors. On six tabular benchmark datasets, the method shows near-zero deviation from a retrained model and is about 6x to 10^5x more efficient. Experiments cover only binary classification problems.
In briefCertifying robust decisions from a single projected measurement
SummaryThe paper studies finite Gaussian projection experiments with a fixed mean-plus-standard-deviation target, showing that a single scalar measurement can yield a robust decision without identifying the latent covariance or every risk value. It gives a confidence-set certificate that either returns a decision or an unresolved outcome, and proves soundness, necessary abstention, and a fixed-checkpoint guarantee. Synthetic experiments report sample budgets 21 to 52 times explicit information lower bounds, but these are not matching bounds, and the guarantees require Gaussian replication.
In briefHow Much Reasoning Effort Should an LLM Agent Use? An Economic Study
SummaryThis paper studies the reasoning effort setting in large language models, which lets users spend more inference compute for possibly better results. The authors measure the effect of different effort levels on proprietary and open-weight models across four benchmarks, including single-call reasoning and long-horizon agentic tasks. They find that higher effort often fails on tasks that lower effort solves, can cost up to 42 times more on easy tasks, and changes how agents solve problems. They propose starting at the lowest effort and escalating only after a verified failure, which matches or beats fixed top effort at lower cost for most proprietary models.
In briefEfficient Proof Generation for Lean Statements with Multiple Candidate Answers
SummaryThis paper studies how to efficiently generate verifiable proofs in Lean 4 when an upstream solver provides several candidate answers to a math problem. The authors propose a joint framework that selects or proposes answers and generates a proof for at least one of them in a single process. On MATH-500, AMC 2023, and AIME 2024/2025/2026, it matches the proof success rate of per-candidate generation at about one sixth of the token cost. It also flags at least 6.1% of formalized statements as unfaithful without extra judges.
In briefModels Check Sufficiency, Not Validity: Usable Falsehoods Evade Gap Checks
SummaryThe paper finds that whether a model flags missing information depends substantially on whether usable material is present in the question, even when source support is held fixed. Making an unsupported premise usable cut flagging by 8 to 30 points and raised answer commitment 1.5 to 2.3 times on GSM8K. When acting as judges, models passed solutions with a required step deleted 78 to 99 percent of the time.
WhyShows models skip verification when an unsupported premise becomes usable, with many paired contrasts across four models and manipulations.
In briefIntrinsic gradient barriers in pushforward generative models
SummaryThis paper studies why pushforward deep generative models struggle to map a simple source distribution, such as a standard Gaussian, onto complex targets. The authors prove lower bounds on the gradients and Jacobians of transport maps from log-concave sources to heavy-tailed or multimodal targets, attributing the difficulty to the transport problem itself. They also propose simple bijective reparametrization classes that ease these geometric obstacles when the goal is density approximation, and support the theory with numerical experiments.
Edric Tam, Qingyang Liu, Barbara Engelhardt et al.
In briefDiscovering Lie symmetries and first integrals from data using latent ODEs
SummaryThis paper introduces LatentLieODE, a neural ODE method that learns a low-dimensional latent representation from time-series observations such as video. It jointly fits linear infinitesimal Lie generators and a latent ODE, then trains a network to produce the system's first integral. Across six benchmark systems, the discovered invariants match ground-truth conserved quantities with correlation magnitude at least 0.978, and on the sphere the method recovers a non-abelian so(3) symmetry.
In briefFrom Outranking Flow Scores to Linear Programs: Group Decision Making in PROMETHEE II
SummaryThis paper studies inverse and consensus analysis for PROMETHEE II ranking when the performance table is fixed. The authors rewrite net flow scores as a linear expression based on weighted preference increments, showing how weights and preference functions interact. They derive linear optimization models for three settings, and an example shows that letting preference functions vary can substantially change conclusions from weight-only analysis.
In briefEMMA: a method that treats multi-agent system setups as structured memory that is updated in controlled, bounded steps.
SummaryThe paper presents EMMA, which treats the configuration of a multi-agent system, including controller routing, specialist agents with their prompts, and per-step procedural skills, as episodic memory. A reflector assigns each diagnosed failure to the scope responsible for correcting it and proposes one edit within that scope, while backbone-independent guards keep accepted edits scoped, incremental, and bounded. Across KramaBench, AppWorld, and BBEH with five backbone models and four baselines, EMMA achieves the best overall performance on every benchmark, up to 32.5% above the strongest baseline. It also uses 17-38% fewer training executions than the most efficient baseline.
In briefA study of how sample size affects kernel metrics' ability to detect generative model changes
SummaryThis paper examines whether kernel distances such as KID and KAD can detect changes in generated samples at a given sample size. Using permutation two-sample tests on real embeddings of music, audio and images, the authors find that default settings at 780 samples per side rarely detect rotations, though they detect mean shifts in every run. They recommend that reported kernel distances state their sample budget and the changes they detect at that budget.
In briefPhysCreator: Generating 3D Assets with Physical Properties from Images
SummaryPhysCreator is a framework that directly generates simulation-ready 3D representations from images, with physical properties that vary across space. Instead of treating physics as a post-hoc prediction or relying on costly per-instance optimization, it jointly models geometry, appearance, and physical properties in a voxel-aligned latent space in a feed-forward manner. Vision-language distillation transfers material-aware semantic priors into voxel-level supervision, avoiding the need for dense physical ground-truth labels. The authors report better visual quality, physical plausibility, and generalization than prior approaches.
In briefExtrapolating learned chaotic dynamics across bifurcations via generalized synchronization
SummaryThis paper studies how trained models can be extrapolated to predict dynamics across bifurcations, reconstructing unseen chaotic behavior. Building on generalized synchronization in reservoir computing, the authors derive a formula that combines pretrained readouts. Using classical Lagrange interpolation theory, they provide rigorous bounds on readout errors and conjugacy residuals. Numerical examples show accurate prediction of chaotic dynamics across bifurcations.
