In briefMatching training hints is not enough: a premature-flood failure in recurrent graph reasoners
SummaryThis paper studies recurrent algorithmic reasoners trained on breadth-first search, which reproduce every training step yet on deep graphs declare nodes reached before the search arrives, a failure the authors call a premature flood. They trace two routes to this failure and test remedies such as clamping the degree input and adding dense training graphs. The authors recommend auditing the range of every input and aggregate and testing on deep graphs.
In briefTwo-layer sound-word canvas model for generating Thai klon pad verse
SummaryThe paper presents a masked language model, initialised from WangchanBERTa, that generates one Thai klon pad stanza per topic by decoding over a fixed two-layer canvas with one sample per topic. On 78 held-out topics judged by a frozen LLM, 70.5% of generated stanzas pass the combined form, cleanness and alignment check versus 76.9% for human stanzas. Low sampling temperature is reported to matter for reducing lexical errors.
In briefMetric-Driven Reinforcement Learning to Reduce Object Hallucination in Image Captioning
SummaryLarge vision-language models sometimes describe objects that are not present in an image, a problem called object hallucination. This paper proposes a reinforcement learning framework that relies only on natural language processing metrics instead of costly human feedback. It introduces an Object Consistency Score, KL regularization, and a lightweight PPO variant that tunes adapters on a frozen backbone. Experiments on InstructBLIP report hallucination reductions of up to 41% on COCO and 29% on Visual Genome.
In briefHodgePhaseNet: Splitting Transformer Attention Flows into Consistent, Circulating, and Harmonic Parts
SummaryThe paper proposes HodgePhaseNet, which treats semantic transitions between tokens as an antisymmetric 1-cochain on a finite oriented simplicial complex. It gives a unique orthogonal decomposition into a globally consistent direction, local circulation, and harmonic flow around larger unfilled cycles. The work provides projectors, a sparse implementation, gauge and integrability results, a numerical example, and a redesigned attention interface. It is a mathematical and architectural proposal and reports no natural-language performance results.
In briefFleetPilot: Fine-grained steering for personalizing LLMs in real-world use
SummaryThis paper proposes FleetPilot, a steering framework that personalizes large language models by generating sample-level interference vectors from user data and injecting them into the forward pass. It combines match contrast, a fine-grained steering component hooking attention and MLP activations, and a query-conditioned aggregation module. Experiments on short-to-long text generation and web function calling report consistent gains, especially in fast-shifting and data-sparse settings, with no training cycles and light storage or latency overhead.
Kounianhua Du, Jianxing Liu, Kangning Zhang et al.
In briefSkillHorizon: A benchmark for long-horizon agent tasks built by composing skills
SummaryThis paper introduces SkillHorizon, a benchmark that builds long-horizon tasks by composing reusable skill procedures across seven business environments. Tasks are linked through input-output dependencies, shared resources, and state updates, so workload and coordination demands grow together. Evaluating five language models on 1,050 tasks, mean local-goal completion drops from 90.4% to 46.8% as workloads rise. The benchmark is intended for studying sustained progress, dependency coordination, and complete task delivery.
WhyBuilds scalable long-horizon tasks via skill composition, systematically evaluating five models on 1,050 tasks across seven business environments with programmatic verification.
In briefCross-material magnetic core loss prediction by matching operating conditions and partial optimal transport
SummaryThis paper addresses transfer learning for predicting magnetic core loss across different materials when only limited target measurements are available. It proposes T-PCOT, a Transformer-based method that matches comparable operating conditions using frequency, flux, temperature, and waveform descriptors, and aligns source and target data via capacity-constrained partial optimal transport. Across three material pairs and three target label budgets, it reports lower mean relative error and 95th-percentile error than T-MKMMD in all nine comparisons.
In briefAD-GS: Anatomy-density-guided 3D Gaussian splatting for reconstructing dental CBCT scans from few X-ray views
SummaryThis paper addresses 3D reconstruction of cone-beam CT scans from sparse projections, which can reduce radiation dose. It proposes AD-GS, which treats adding 3D Gaussians as attenuation-aware MCMC sampling so that representation focuses on high-attenuation anatomical regions. It also adds a semantic regularization using zero-shot features from a medical segmentation model to enforce multi-view consistency. Experiments on the ToothFairy3 dental CBCT dataset report better structural fidelity and boundary preservation than existing 3DGS-based methods under sparse-view settings.
In briefHiPAD: A Gait, Audio-Visual and Personality Fusion Network for Depression Detection, Studying Personality Ablation and LLM Calibration Limits
SummaryThe paper presents HiPAD, a single fusion network submitted to all six sub-tracks of the MPDD-AVG 2026 depression-detection challenge, combining gait, audio-visual data and personality descriptions. It finds that a roughly 150-token personality description carries more depression signal than audio, video and gait combined. Removing all personality-derived features leaves the model at chance, and domain-adapted LLMs lower the overall score despite improving binary kappa.
In briefHumans vs. Machines: Comparing AI Agents with Human Analysts in Reverse Engineering
SummaryThis paper compares human analysts and LLM-based AI agents on software reverse engineering, the work of understanding compiled binary programs. Nine models were tested on two static-analysis challenges, with their interactions with IDA Pro recorded. The best agents reached human-competitive understanding, but they investigated differently: they fetched information broadly and rarely revisited code, a behavior the authors call greedy fetching. Measurements across 4,000 real-world binaries show that context-window limits constrain scaling to larger programs.
In briefWhat remains when activation steering is used as memory
SummaryThis paper studies using activation steering as an agent's memory, storing the shift an experience causes in a model's internal states as a vector and adding it back when the conversation's topic matches. Pre-registered experiments on Qwen3.5-2B and 9B show the memory is selective, leaving about 98% of unrelated replies unchanged. However, stored facts are almost never used, and preferences transfer only within a narrow strength window.
In briefCatPortrait: Using Video Generation Models to Animate Portraits via In-Context Conditioning
SummaryThis paper introduces CatPortrait, a framework that makes portrait animation controllable by feeding 3D face guidance into a video generation model as in-context conditioning. It keeps the original video backbone unchanged and injects the 3D facial structure through spatial concatenation, so pretrained image-to-video models can animate portraits while retaining their learned priors. The authors state that one unified framework supports both video-driven and audio-driven animation, with competitive or superior results reported in both settings.
In briefMCPDisCo: Seesaw-Mitigated Tuning for MCP-Native Agents
SummaryThis paper presents MCPDisCo, a reinforcement learning framework for agents that use the Model Context Protocol. It trains on real, long-horizon interactions across 70 MCP servers and 527 tools in six domains. To address the reasoning–tool seesaw, where gains in one objective can hurt the other, it combines token-masked reward attribution with bidirectional gradient surgery. The authors report higher success with fewer chain-of-thought tokens and continued gains on the out-of-distribution MCP-Atlas and GAIA benchmarks, with model weights and implementation released.
In briefCounterfactual rollouts for offline contrastive goal-conditioned reinforcement learning under confounding
SummaryContrastive reinforcement learning assumes no hidden variable jointly affects recorded actions and their outcomes, an assumption that fails when demonstrations come from a demonstrator with information the learner cannot see. The paper formalizes this setting with confounded Markov decision processes and introduces counterfactual transitions that separate what the data identify from what they do not. The proposed method, Causal Contrastive Reinforcement Learning (CauCRL), learns a pessimistic counterfactual transition from offline demonstrations and trains the critic on counterfactual rollouts. It reports substantially higher success rates than standard contrastive RL on confounded goal-reaching tasks.
In briefDo Personalized LLMs Need Full Chat History? Relevance Filtering Matches It With 5% of the Text
SummaryThis paper asks whether a personalized assistant needs the whole conversation history to respond. On PersonaMem-v2, a sentence-level retrieval method that ranks history by similarity to the current request and stops at a fixed budget matched full-history accuracy using only 5% of the text, cutting judge input about 20-fold. An LLM summary of similar length performed comparably, but it requires an LLM to read the entire history, while retrieval needs only a small embedding model.
In briefDSM-free drone visual localization using latent-depth projection and adaptive structure coupling
SummaryThis paper addresses absolute positioning for drones in GNSS-denied areas without requiring Digital Surface Models. The proposed Latent-Depth Variable Projection treats ray scales as latent depths, eliminates them analytically, and recovers horizontal camera position in closed form given rotation. Structure-Adaptive Coupling selects this method or Planar-PnP based on scene structure inferred from monocular depth. Experiments report consistent gains over existing DSM-free methods and a narrower gap to DSM-based localization.
In briefText-guided medical image segmentation using residual and dense alignment methods (research paper)
SummaryThis paper presents RADAR, a framework that uses natural language descriptions to specify targets for medical image segmentation. It combines BiomedCLIP's medical semantic grounding with MedSAM2's prompt-driven mask decoding, and uses LoRA for low-cost task adaptation. The method includes guided residual depth aggregation, structured dense representation alignment, and multi-modal prompt generation. Experiments on three public datasets reportedly achieve competitive performance with only 0.74M trainable parameters.
In briefProbabilistic Multi-Source Data Selection with Learnable Categorical Priors
SummaryThe paper addresses multi-source bilevel data selection by proposing a new probabilistic relaxation that uses multiple learnable categorical priors supported only on feasible data subsets, yielding a lossless rounding scheme. It introduces a two-point gradient estimator and a reducible representation that lowers the number of lower-level retrainings needed. Building on this, the authors present the ReSGD stochastic gradient algorithm, which has an approximation guarantee under weak submodularity, and experiments are reported to validate it.
In briefCompositional average-velocity modeling for fast diffusion-based offline RL (one-step generative policies)
SummaryDiffusion policies perform well in offline reinforcement learning, but multi-step denoising adds heavy computation and requires backpropagation through time during training. The paper proposes Instant Flow Q-Learning (InsQL), which replaces marginal velocity modeling with average-velocity modeling to map noise directly to actions in a single step, avoiding iterative sampling and Jacobian computations during training. The authors report that InsQL matches or exceeds state-of-the-art diffusion-based methods on standard offline RL benchmarks and real-world robotic manipulation tasks while substantially reducing training and inference time.
Thanh Nguyen, Abu Hanif Muhammad Syarubany, Hwanhee Kim et al.
In briefA memory-engram-inspired adaptive memorization method for lifelong learning
SummaryThis paper addresses lifelong learning, where artificial systems must keep adapting as their environments change. The authors integrate the full life cycle of engram neurons from biological memory and build structured memory from gradient representations, which regulates parameter updates and guides memory management. Evaluations on four benchmarks show the approach stays stable over long task sequences while supporting flexible adaptation.
In briefVC-OPD: Visual Counterfactual On-Policy Distillation for Visually Grounded Reasoning in Vision-Language Models
SummaryThe paper proposes VC-OPD, a distillation method for improving vision-language reasoning. It runs on-policy distillation on student trajectories and compares teacher outputs under the original image and a counterfactual image with corrupted evidence, using the gap to estimate each token's visual dependence and reweight the loss. Experiments on multimodal reasoning benchmarks report improved visual reasoning performance.
In briefPredicting How Model Traits Spill Over to Other Behaviors: A paper studying how boosting one trait in a language model affects its other traits on new inputs.
SummaryThis paper studies out-of-distribution generalization in language models. The authors increase a target trait using system prompting, few-shot learning, SFT on reference answers, and GRPO, then measure how the change spills over onto other traits. Spillover is highly correlated across models under the same elicitation method, but SFT and GRPO produce different patterns. Some benign traits, such as claiming sentience, can increase undesired traits like power-seeking. The authors also find trait clusters and a general drift independent of the targeted trait, concluding that no single predictor dominates and generalization remains hard to predict.
In briefCFG-Policy: A Unified Diffusion Policy Learning Framework Using Classifier-Free Guidance
SummaryDiffusion policies can represent multimodal action distributions, but reinforcement learning optimization may reduce their behavioral diversity. The paper proposes CFG-Policy, which separates exploration and exploitation into two generative policies and combines their score or velocity fields during sampling via classifier-free guidance. In online RL, a frozen exploration branch preserves diversity while an exploitation branch learns from rewards; in offline-to-online RL, a pretrained exploitation branch stays frozen while a trainable branch steers the composed behavior. The authors report performance gains on 5 MuJoCo tasks and 4 RoboTwin manipulation tasks.
In briefBSIPipe: Bidirectional Sequence-Level Pipeline Parallelism for Training Long-Context Language Models
SummaryThis paper proposes BSIPipe, a pipeline parallelism method for training large language models on long sequences. It co-schedules two opposing V-shaped interleaved pipelines at the subsequence level to improve throughput over existing unidirectional methods. Experiments on GPT and LLaMA3 report average speedups of 1.11× to 1.13× over state-of-the-art sequence-level baselines.
In briefFidelityQ: Keeping Language Model Predictions Accurate After Ultra-Low-Bit Quantization
SummaryThis paper presents FidelityQ, a scalar quantization method that compresses language model weights to as few as 2 bits while keeping predictions close to the original model. It combines Information-Aligned Analytic Coding, which fits a four-level quantizer to transformed weight rows, with Knowledge-Intensive Calibration, which selects calibration samples with the largest teacher-student disagreement and updates only quantization levels and scales. On LLaMA-2-7B, the reported WikiText2 perplexity is 6.62 at W2A16 and improves from 8.31 to 6.96 at W2A4KV4.
In briefAn Algebraic Method for K-means Clustering with Pairwise Constraints
SummaryThis paper proposes an algebraic approach to k-means clustering that incorporates pairwise constraints requiring objects to be in the same or different clusters. Before optimization, the method simplifies the constraints using Gröbner bases, which improves cluster assignment efficiency, and it can detect whether a feasible assignment exists. Comparative experiments reported comparable or better clustering quality at lower computational cost, especially when the numbers of constraints and clusters are small.
In briefTLA-Verus: LLM agents generate and solve temporal proof obligations for Rust verification
SummaryTLA+ is a formal specification toolkit widely used to model distributed systems, but it cannot detect drift between implementations and specifications, and its tools struggle to prove arbitrary temporal properties. This paper translates TLA+ specifications into Verus, a verifier for Rust, using both a deterministic transpiler and an LLM-based agentic translator. The authors show their pipeline generates proof obligations at scale and produces machine-checkable proofs of safety and liveness properties. The work also yields an evaluation dataset used to measure current LLM capabilities in Verus theorem proving.
Kseniia Strelbytska, Bal\'azs Szil\'agyi, Anna M\'esz\'aros et al.
In briefRepresentation drift in continual learning: LoRA vs. N-gram memory
SummaryThis paper studies which part of a language model should absorb continual updates by measuring how each update method changes internal representations. Using a 48-layer mixture-of-experts model and a seven-task sequence, the authors compare updating only an N-gram memory table, shared LoRA, and per-task merged LoRA. Updating only the N-gram memory produced the smallest, non-growing drift in base representations, but stored tasks interfered with one another.
In briefSkillForgeBench: A Benchmark for Agent Reliability Based on Skill Specifications
SummarySkillForgeBench is a benchmark that uses agent skills as execution blueprints to build tasks across three stages: resource preparation, skill-driven task synthesis, and holistic evaluation. It contains 842 sandboxed tasks across five professional domains and tests eight mainstream agents with requirement-level scoring. The authors report that requiring all constraints exposes capability bottlenecks hidden by individual pass rates, and that completion-only checks overestimate real-world reliability.
In briefWorkBorne: Discovering Agent Tasks from Real-World Artifact Transitions
SummaryThis paper introduces WorkBorne, a framework that synthesizes training tasks for terminal-based agents that run over long horizons. It builds artifact corpora from web retrieval, finds meaningful input-output transitions between artifacts, and turns them into grounded task environments. The authors report 7,882 task environments and 15,485 trajectories, with experiments on Qwen3.5 models showing gains over base models and other synthesis baselines.
In briefImproving task selection in reinforcement learning by better predicting task risk
SummaryThis paper addresses two weaknesses in risk predictive models used for model predictive task sampling in reinforcement learning: sparse optimization histories and a one-step lag in risk approximation. The authors recast risk prediction as an in-context sequence modeling problem and add a one-step look-ahead mechanism with a temporal-weighted sliding window. According to the paper, the approach gives more reliable difficulty estimates and consistent performance gains in robust Meta-RL, domain randomization, and prompt curriculum RL post-training of LLMs.
In briefProbably Plausible, Not Valid: Auditing Decision Models as Synthetic Survey Respondents
SummaryThis paper tests whether language models can replace human respondents in a psychometric survey. The authors evaluate fourteen decision models and 47 generative LLMs on a Lithuanian organisational-psychology survey (263 respondents, 68 items) using 100 real respondent profiles and a Psychometric Similarity Score. The best generative model matches a Gaussian copula fitted only on other respondents, but decision models predict individual answers worse than other respondents' answer frequencies. The authors conclude that synthetic respondents should aim for sample-level fidelity rather than individual-level prediction.
In briefScalable multi-human cooperative object manipulation using commander-executor coordination
SummaryThis paper addresses multi-human cooperative object manipulation for humanoid control, focusing on role-dependent, long-horizon tasks that existing methods handle poorly. The proposed hierarchical framework uses a commander to plan tasks jointly from instructions and scene information, dividing them into role-specific subtasks and shared object-centric coordination plans. An executor then learns a physics-based low-level policy through a solo-to-cooperative curriculum. The authors report that the approach scales from two to multiple agents and achieves higher success rates than state-of-the-art methods.
In briefMinimum-volume optimal transport method for multivariate conformal prediction
SummaryThis paper addresses optimal transport-based multivariate conformal prediction, noting that a fixed Euclidean transport cost can yield suboptimal prediction-set volume. The authors propose Minimum-Volume OTCP, which learns a Mahalanobis ground cost to directly minimize prediction-region volume and establishes finite-sample marginal coverage. An individualized extension with covariate-dependent metrics is also developed, and simulations and real-data experiments report substantial volume reductions.
In briefPredicting with expert advice when losses have heavy tails: a study of optimal algorithms for this setting.
SummaryThis paper studies prediction with expert advice over T rounds and K experts, where losses have bounded (1+ε)-moments (ε in (0,1], bound θ). The authors establish a minimax pseudo-regret rate of Θ((θT)^{1/(1+ε)}(log K)^{ε/(1+ε)}). They propose adaptive FTRL and OMD algorithms that need neither ε nor θ, achieving near-optimal regret in adversarial and self-bounded environments.
WhyEstablishes a minimax pseudo-regret rate under heavy-tailed losses, removing extra polynomial dependence on K, with rigorous theoretical analysis supporting the claims.
In briefAutonomous 5G edge slicing resource allocation using multi-agent systems and LLMs
SummaryThis paper proposes a two-timescale architecture for allocating resources in 5G edge network slicing. An LLM-based Strategist converts natural-language intents into economic policy parameters, while SPADE-based slice agents run a deterministic second-price auction to resolve contention. Allocations are enforced by patching Kubernetes CPU limits and Calico bandwidth annotations. The architecture is evaluated on a real Kubernetes/Open5GS edge cluster across Smart-City and Hierarchical-slicing scenarios, each repeated 30 times.
Nathan R. Prediger, Rodrigo Moreira, Juliano Araujo Wickboldt et al.
In briefTeachGrade: Training-Free Short Answer Scoring Using Teacher-Graded Examples
SummaryThis paper presents TeachGrade, a training-free framework that scores short answers using examples previously graded by teachers. It uses two frozen memory modules, Skill and CaseBank, to capture teacher preferences and local evidence, then regrades answers without preliminary scores. On 3,829 SAS-Bench answers across 12 tasks, it reports higher macro collaborative consistency and quadratic weighted kappa scores than several compared methods.
SummaryThis paper examines low-frequency noise from AI data centers, which may affect nearby residents' health. It presents an FDTD model simulating sound propagation over the terrain near a data center. The topography strongly influences noise levels, with areas of sudden elevation beside the river being more exposed, while distant areas are largely shielded. The findings can help identify where noise mitigation measures would be most effective.
In briefPro-t-VAE: Closed-form Student's t mixtures for robust prototype discovery in images
SummaryThis paper proposes a deep generative model that uses a heavy-tailed Student's t mixture prior to discover representative image prototypes without labels. It builds the latent space via a Bayesian scale-mixture formulation, yielding a closed-form augmented ELBO, and treats precision as a latent variable so outliers are downweighted. The authors report validation on standard vision benchmarks and brain MRI analysis of neurodegenerative diseases, where the model recovers clean, anatomically coherent prototypes.
In briefDescription complexity bounds for analytic approximations of Lipschitz functions
SummaryThis paper studies how many bits are needed to write an elementary analytic expression that uniformly approximates a Lipschitz function. Such expressions are finite trees over addition, subtraction, multiplication, division and exp, and their size is the length of their shortest description. For one-variable functions the authors prove matching upper and lower bounds of order L/Δ + lg(1/Δ+2), and they extend the results to functions on the d-dimensional unit cube.
In briefAgent-guided evidence gathering for finding emotion-cause pairs in conversations
SummaryThis paper studies emotion-cause pair extraction, which identifies the causes of emotions in conversations. A student agent selectively asks experts about borderline relations and revises only the queried predictions. At a nominal 10% query rate, selective fusion reaches mean pair F1 scores of 65.14 on ECF and 49.74 on MEC4.
In briefMARL-BattleGrounds: A New JAX-Native Benchmark for Competitive Multi-Agent Reinforcement Learning
SummaryThis paper introduces MARL-BattleGrounds, a JAX-based benchmark for competitive multi-agent reinforcement learning, with a team deathmatch mode featuring two teams of five heterogeneous agents. It provides 52 maps, eight endgame generalization scenarios, and a monthly round-robin tournament with Elo rankings, plus scripted opponents and tools for human-AI play. The authors report that one RTX 5090 runs 50,000 5v5 game steps per second across 1,024 parallel games.
Ulixes Tariq Hawili, Subramanian Ramamoorthy, Yoann Altmann et al.
In briefSocial Bias in Chinese Language Models Using Social Media Expressions
SummaryThis study evaluates social bias in six Chinese-capable large language models using 73 paired probes covering gender, region, and occupation. The probes include neutral, social-media-style, and reversed-identity versions, with the task framed as social consensus estimation. All models show significant identity-bias effects, and five show significantly negative effects from stylistic packaging. The authors conclude that lower scores under sarcasm should not be read directly as bias mitigation.
In briefBeyond Outcome Feedback: Trajectory-Quality-Guided Experience-based Learning
SummaryThis paper proposes TQG-Exp, a framework that lets LLM agents learn from past executions by evaluating process quality as well as final task outcomes. It introduces a Trajectory Quality Loss that jointly considers task completion and process quality, and uses paired rollouts with and without retrieved experience to update each experience's utility. Experiments on ALFWorld, WebShop, and ScienceWorld show it outperforms existing experience-based methods.
In briefContrastive Linearized Laplace: uncertainty estimates for contrastive embeddings without retraining
SummaryContrastive learning yields deterministic representations without built-in uncertainty estimates. This paper proposes a post-hoc framework based on the Linearized Laplace approximation that propagates encoder uncertainty through downstream tasks without retraining the encoder. The authors characterize the curvature of contrastive objectives and add positive-semidefinite repairs where needed. The method is tested on metagenomic binning and image classification.
Francisco J. S\'aez-Maldonado, Luis A. Ortega, Andr\'es R. Masegosa et al.
In briefDependency Reasoning for Aspect Sentiment Quadruple Prediction (DRIFT)
SummaryThis paper presents DRIFT, a method for Aspect Sentiment Quadruple Prediction (ASQP), which identifies combinations of aspect, opinion, category, and sentiment. It linearizes dependency parse trees and injects them directly into LLM prompts as explicit relational context, alongside retrieval-augmented in-context learning and a syntax-aware adjudication mechanism. On SemEval Rest15, Rest16, and ACOS, DRIFT achieves the strongest results among compared generative ASQP methods, with 54.47% and 64.10% F1 on Rest15 and Rest16. The authors also report that exact-match metrics systematically underestimate semantically correct predictions.
Shima Esfandyari, Seywan Soleymanbaigi, Fardin Akhlaghian Tab et al.
In briefMORE: Learning motion-oriented registers for driving world models and planning
SummaryThe paper proposes MORE, a V-JEPA framework that learns motion-oriented registers during video pretraining as a shared state interface for driving world models and planning. Temporally aligned ego motion is injected into a small set of registers, which serve as the sole context for masked patch prediction, routing dense supervision through a compact bottleneck. On NAVSIM-v2, MORE outperforms the state of the art by 1.6 in EPDMS with one-tenth the inference latency of DriveLaW.
In briefScalLoc: Continuous Oblique Thermal UAV Geo-Localization Using 2D Satellite Maps
SummaryThis paper presents ScalLoc, a framework for locating UAVs with thermal cameras in GNSS-denied settings, using only 2D satellite maps instead of elevation data or 3D maps. It reconstructs local geometry from thermal sequences, refines camera poses, anchors windows to the map, and propagates results to recover frame-level trajectories. The authors also release Thermal-Oblique, a dataset of 216 day-night trajectories with 28,112 annotated frames. Reported results show the mean localization error dropping from 112.69 m to 26.97 m, a 76.07% reduction relative to the VGGT-only configuration.
WhyProposes oblique thermal UAV localization using only 2D satellite maps, with a new 216-trajectory dataset and a 76% error reduction.
In briefWeatherBridge: Forecast-aware adapter that makes operational weather data usable by pretrained models
SummaryWeather forecasting models trained on ERA5 data struggle with initial states from operational numerical weather prediction systems, which have different distributions. WeatherBridge is an adapter that maps GFS, HRES, and CMA initial states toward ERA5 reference states without changing the pretrained forecasting model. It combines ERA5 reconstruction with one-step rollout alignment through a frozen FuXi model. The adapter reduced FuXi's mean RMSE over 6 to 240-hour forecasts for 17 of 18 source-variable combinations.
In briefAdding experimental electron density to protein-ligand interaction modeling
SummaryThis paper presents CoDE, a coordinate and density encoder pre-trained on experimental electron-density maps. It uses a vision transformer on a voxel grid combining atomic occupancy, electron density, and density gradient, trained on about 100K curated complexes to reconstruct masked channels. The frozen encoder improves over a coordinate-only encoder on binding-affinity regression and pocket-conditioned ligand generation, and remains competitive with task-specific baselines.
In briefWorld2TS: A dataset linking real-world events to time series for better forecasting
SummaryThis paper introduces World2TS, a large-scale dataset and construction framework for time-series forecasting that includes real-world event context. The authors built a reusable tool combining agent-led reasoning, Web search, command-line validation, and human review to discover events and link them to relevant series. The dataset spans 31 datasets across six domains, with 4,179 series, 97.1 million observations, 1,159 events, and 56,846 event-series links. Controlled experiments are reported to show that the collected context can improve forecasting accuracy.
In briefTerrarium: Training Personal Assistants by Synthesizing Simulated App Worlds
SummaryThis paper introduces Terrarium, a framework that builds executable simulated app environments from real application information and generates training tasks with evaluation rubrics. The authors create a training suite of 275 app worlds and 13,344 tasks for outcome-based reinforcement learning on Qwen3 models at 4B, 8B, and 14B scales. Models trained with Terrarium outperform the base Qwen3 models by an average of 18.34 points on τ²-bench and 10.94 points on AppWorld.
In briefCAST: Real-time motion capture that works on any skeleton shape
SummaryThis paper addresses recovering human motion from a single video when the target skeleton differs from the person filmed. The authors propose CAST, which combines learned local joint rotations with geometric correction guided by predicted global bone directions. Experiments under the TopoCap protocol show improved motion recovery across target topologies, including unseen ones. A causal variant, CAST-B, runs at over 51 FPS on one NVIDIA A100 GPU.
In briefPEAM: An On-Device Agent for Personalized, Fast Mobile Network Configuration
SummaryThis paper presents PEAM, an on-device LLM-based agent that autonomously creates and executes personalized network-configuration rules within milliseconds on smartphones. The system fine-tunes a compact model on manually annotated real-user data, generates rules from historical records during idle periods, and retrieves and applies rules when the communication context changes. Experiments with a fine-tuned 4B model report 99.13% correct configuration decisions and 92.50% valid generated rules.
In briefSeparating Transport Capacity from Geometry in Optimal-Transport Label Transfer
SummaryThis paper studies why optimal transport label transfer between unpaired datasets can systematically under-predict rare classes when source and target class compositions differ. The authors attribute the failure to the transport marginal, where each class's capacity is proportional to its source frequency and may be insufficient even when classes are well separated. Matching the transport marginal to the target composition substantially improves transfer, raising macro-F1 on single-cell data from 0.322 to 0.623.
Ali Khodabandeh Yalabadi, Azim Dehghani Amirabad, Artem Moskalev et al.
In briefSpatial Image Generation Improved by Agentic Rewards and Curriculum RL
SummaryThis paper proposes a curriculum-guided reinforcement learning framework for spatial text-to-image generation. It introduces SpatialCritic, an agentic reward model with vision tools, and organizes training with a capability graph. On SpaRW-Eval, SpatialCritic reaches 95.1% verification accuracy with an 8B backbone, outperforming a 320B model, and raises baseline scores for SD3.5-Medium and BAGEL.
In briefPrivacy-friendly thermal sensing for indoor human perception, from dataset to real-world deployment
SummaryThis paper introduces Indoor RGB-T, a large paired RGB and thermal dataset covering 20 indoor rooms for human perception research. It establishes benchmarks for thermal occupant-state detection, occupancy estimation in unseen rooms, and translation between RGB and thermal images. The thermal perception model was also deployed on a commercial sleep-monitoring system, with field tests involving 100 participants and about 1,000 devices.
WhyBuilds a large paired RGB-thermal indoor dataset and validates it in real deployment with 100 participants; novelty is a useful extension, evidence strong.
In briefPhysics-informed random neural networks on compact manifolds for random-operator models
SummaryThe paper introduces stochastic physics-informed neural networks (SPINNs) for stochastic boundary-value problems with random operators. A shared Gaussian random source couples the problem's random operator with a random neural architecture built on a compact, boundaryless, multiply connected manifold. Training uses only evaluations of the boundary-value problem's solution mapping, without PDE-residual or boundary-penalty losses. An example on a periodic cylindrical surface illustrates the approach.
In briefAudited Machine-Generated Operators in a Physics-Discovery System
SummaryThe paper proposes a pre-registered, two-tier accounting method to tell whether a self-improving system reuses designer-provided mechanisms or authors new ones itself. On real NIST hydrogen Balmer measurements, the system's compositional search produced the operator sq(inv(n)) = 1/n², the Rydberg-level form, which no standard library transform could express. A second event recovered Kepler's velocity law from JPL ephemeris data. The authors state their contribution is measuring and governing machine-authored mechanisms rather than the search algorithm, and note that both events show capability to author operators, not yet autonomous genesis.
In briefLearning branching and node selection in branch-and-bound from a sequence of search events
SummaryThis paper proposes Seq-BnB, a framework that learns branching and node selection for mixed-integer linear programs by modeling the search process as a sequence of typed events. A self-attention encoder summarizes the global search history, and a bipartite graph neural network encodes the current LP relaxation. On four MILP benchmark families, Seq-BnB processes fewer BnB nodes than heuristic and learning-based baselines, with strong solution times.
In briefReproducing a Method for Making CLIP's Image Predictions More Interpretable
SummaryThis paper is a reproducibility study of an ICLR 2025 paper that proposes unsupervised adversarial fine-tuning with norm regularisation to improve the visual interpretability of CLIP's image encoder. The authors attempt to reproduce the original claims about saliency map quality, concept alignment, and out-of-distribution transferability. They also add a saliency-guided extension using an Energy Pointing Game loss and evaluate how it affects localisation and zero-shot accuracy across different loss weights.
Anton Nuzhdin, Andrei Gruzitski, Bal\'azs Egyed et al.
In briefEvoCowork: A Benchmark for Self-Improving Agents on Real Office Tasks
SummaryThis paper introduces EvoCowork, a 240-task benchmark that tests whether AI agents can improve their own working procedures on office work such as presentations, documents, spreadsheets, and cross-file workflows, with model weights kept fixed. The protocol separates learning from feedback, revision selection, and final testing. The authors find that gains on single-file-type tasks do not necessarily transfer to coordinated cross-office workflows.
In briefCross-lingual speech synthesis for Indian languages: a benchmark and inference-time duration estimation
SummaryThis paper addresses how alignment-free text-to-speech models suffer duration mismatch and quality loss across languages, since character durations differ. The authors build a cross-lingual benchmark covering English and nine Indian languages. They propose a language-aware inference-time duration estimation method using per-character duration statistics, requiring no retraining or architecture changes. Experiments on bilingual and multilingual A-DMA models report better naturalness, intelligibility, and speaker similarity with reduced duration mismatch.
In briefAI-Native Product Management: A Vision for AI as a Persistent Product Participant
SummaryThis paper examines how product management should be re-envisioned as AI-native R&D accelerates software delivery. The authors synthesize research on software product management, product learning, and human-AI collaboration. They propose a vision where AI persistently participates across sensing, exploration, decision-making, realization, and learning, supported by six principles and an operating model.
In briefA reference architecture for integrating agentic AI into enterprise software
SummaryThis paper proposes a technology-neutral reference architecture for integrating agentic AI into enterprise systems. It groups architectural capabilities into ten responsibilities and defines boundary invariants, typed interactions between responsibilities, and variation points. The authors instantiate it as an ESG portfolio-analysis and pre-trade compliance agent over an existing ESG system, which realizes seven of the ten responsibilities, and map an independent agentic architecture onto the same framework. The authors state these results are initial evidence of the architecture's organization and applicability, not of its completeness.
In briefExplore This! Beat and Downbeat Tracking from a Learned Tatum Grid
SummaryThis paper proposes a new music beat-tracking system that combines deep learning with differentiable music-theory modules. It first predicts a tatum grid, then groups tatum events into beats and downbeats, explicitly modelling multiple tempo and time-signature options in a bar-pointer-like way. The method improves continuity metrics over existing DBN-free systems and is more flexible than the standard DBN, but on average still trails the best DBN-based methods. The authors publish their code and invite further exploration of the design choices.
Robert Kihlborg, Andr\'e Holzapfel, Jan Schl\"uter
In briefEdge AI for Factory Predictive Maintenance: A Real-World AWS Greengrass Test
SummaryThis paper proposes a hybrid cloud-edge architecture built on AWS IoT Core and AWS IoT Greengrass v2, running machine learning inference locally at the edge and synchronizing selectively with the cloud for manufacturing predictive maintenance. The authors compared a cloud-only baseline with the hybrid design on real AWS infrastructure. The hybrid configuration reduced cloud-bound bandwidth by 86.8% and cut decision latency from 23.1 ms to 0.09 ms, while keeping an anomaly-class F1 score of 0.84 and 91% recall without cloud involvement. A quantitative experiment on cloud disconnection is left as planned future work.
Prince Kwabena Appiah Boadu, Jeffery Baafi Kyei, Enoch Kofi Koranteng et al.
In briefMachine learning prediction of dengue in Bangladesh using clinical and blood test data
SummaryThis paper compares Logistic Regression, Random Forest and XGBoost for predicting dengue from clinical and hematological data from two Bangladeshi hospital datasets, using SMOTE to address class imbalance. In the 2024 data, extended models outperformed benchmark models, with Random Forest ROC-AUC rising from 0.640 to 0.661 and RDW-CV the most consistently influential added predictor. Benchmark performance on the 2025 data was higher (ROC-AUC 0.904), which the authors note is not explained by class imbalance and may reflect data quality. External validation is still needed.
In briefDepth-by-depth 3D seismic velocity model building using 2D diffusion models
SummaryThis paper presents a framework that builds 3D seismic velocity models using 2D generative diffusion models, treating the depth dimension as channels of a 2D U-Net. Overlapping depth patches are generated sequentially from shallow to deep, with deeper predictions conditioned on shallower results and any available constraints such as well logs and structural attributes. Experiments on Overthrust, SEAM Arid, SEG/EAGE, and out-of-distribution Marmousi models show reduced parameter count, training cost, and inference cost compared with an equivalent 3D benchmark, while maintaining high accuracy.
Shijun Cheng, Randy Harsuko, Mohammad H. Taufik et al.
In briefGroups don't move as one: a clustering view of group influence
SummaryThis paper studies how removing or reweighting training data affects a group, a question answered with influence functions instead of retraining. The standard first-order estimate sums individual contributions, which can hide differences among group members and cancelling effects. The authors model the aggregate as a Bregman centroid, quantify omitted variation with Bregman information, extend this to multiple clusters, and propose an estimator that jointly optimizes cluster weights. Experiments report higher R² than standard group influence when predicting target changes.
In briefASIWorld: A persistent survival sandbox for testing long-running autonomous agents
SummaryThe paper introduces ASIWorld, a persistent survival sandbox where characters have limited observations and agents can run asynchronous programs. Nine agent configurations were tested on interdependent production and exploration tasks, with eight completing production and six completing both. Configurations with similar task results differed in preparation, combat, recovery behavior, and survival costs.
In briefYOUDrive: Driving Style as a Composable Direction for End-to-End Driving
SummaryThis paper proposes YOUDrive, a vision-language-action planner that gives continuous, composable control over driving style in end-to-end autonomous driving. Each teacher's behavior is distilled into a task vector that defines a style direction, which can be scaled and combined at inference without retraining the backbone. On NAVSIM navtest, a single coefficient monotonically controls assertiveness across six teacher-derived personas while PDMS stays above 0.90, reaching 0.954.
In briefEAIR: Interaction Risks in Enterprise LLM Multi-Agent Systems
SummaryThis paper examines how LLM agents that split enterprise work by role, each seeing only a local view, can take individually reasonable actions that together break organization-wide rules. For example, several service agents may each issue credit within their own permissions yet exceed a shared compensation cap. The authors introduce EAIR, a framework that formalizes these interaction risks, provides executable micro-worlds for testing them, and offers a pipeline to diagnose and repair workflows using execution traces. Large-scale experiments show recurring interaction failures across the organizational properties studied.
In briefCESNET-TimeSeries26: A multi-year network traffic dataset for testing long-term forecasting and anomaly detection
SummaryThis paper introduces CESNET-TimeSeries26, a dataset of 995 days of continuous network traffic time series from the CESNET3 network, plus a single subnet spanning about 4.5 years. It covers 278,466 IP addresses, 285 institutions, and 566 institutional subnets at three aggregation intervals. The authors find that annual seasonality is the dominant periodic component. They also show that a forecasting model can use this annual pattern only when its training window covers at least one complete year.
In briefGREAT-ESDD: A graph-based framework that uses neural routing and learned customer preferences to set dynamic prices for home versus out-of-home same-day delivery.
SummaryThe paper addresses online same-day delivery, where operators can offer home delivery or out-of-home (OOH) pickup options. The authors train a Graph-Edge-Attention-Network-based neural solver for capacitated vehicle routing that quickly reoptimizes routes and performs competitively with traditional baselines. Dynamic prices are set from estimated routing cost increases, and Thompson sampling with a Laplace-approximated posterior learns customer preferences online. In simulations, the framework increases profit by about 11% over a traditional routing solver and outperforms a static pricing strategy after 5 simulated days.
Attila Lischka, Ze Zhou, Xuegang (Jeff) Ban et al.
In briefCounterfactual regret bounds for reflective language agents, verified in Lean
SummaryThis paper studies reflective language agents that reuse feedback and memory to adapt across tasks without updating model weights. It derives sufficient conditions under which local generation and deployment guarantees imply sublinear counterfactual regret for the complete agent under resource constraints. All thirteen numbered theoretical results are formalized in Lean, including conditional proof chains.
In briefTraining-Free BioNEL-RAG: Linking Biomedical Terms to Standard Concepts Without Training
SummaryThis paper addresses biomedical named entity linking, which maps terms in biomedical texts to standardized entities in a knowledge base. The authors propose BioNEL-RAG, a training-free framework in which a retrieval-augmented generation module retrieves candidate entities and a large language model selects the best match based on the mention's context. According to the experiments, it performs comparably to existing methods without training on subdomain datasets such as genes, diseases, or chemical compounds, which those methods require.
In briefFormalizing Forgetting: A Theory of Forgetting in Online Continual Learning
SummaryThis paper studies catastrophic forgetting in online continual learning, where a model learns a sequence of tasks in a single pass over the data. The authors analyze single-pass Shuffle SGD under strongly convex and general convex objectives and derive explicit upper bounds on expected final forgetting. They examine how per-task data amount, task similarity, and task ordering affect these bounds, which can differ across objective classes. Experiments with neural networks on Split-CIFAR-100 and PMNIST show that the trends suggested by the theory also appear in practice.
In briefLYNX: Analytical Program Synthesis via Grounding and Closed-Loop Refinement
SummaryThis paper presents LYNX, a multi-agent framework that treats LLM-driven data analytics as a schema- and knowledge-constrained problem rather than a single-pass generation task. Three role-specialized agents work in a closed loop of planning, generation, verification, and refinement, supported by Semantic-Structural Distillation and Stage-Aware Allocation and delta-guided iterative revision. Evaluated on BIRD-SQL Mini-Dev and Spider 2.0-Lite, it reports 10.66% to 59.54% execution accuracy and a 70.53% token reduction versus the prior most efficient method.
In briefTraining a Compact, Strong, and Realistic Text-to-Speech Model
SummaryThis paper presents Audio8-TTS, a compact 0.6B-parameter zero-shot voice cloning model that reaches state-of-the-art intelligibility and speaker similarity, with especially strong perceived naturalness. The authors train a realism critic on human recordings to align training with human perception rather than signal fidelity. Training proceeds in three stages: multilingual supervised fine-tuning, GRPO with WER and speaker similarity rewards, and GRPO guided by the realism critic.
In briefProgAgent: Learning Tool Orchestration via Programmatic Training
SummaryThis paper presents ProgAgent, a training approach where language models call tools through executable code and store intermediate results in runtime variables instead of the model context. It includes a pipeline for synthesizing verifiable long-horizon tasks, a lightweight runtime, and a two-stage training recipe. Trained on only 215 tasks, ProgAgent achieves higher pass rates than comparable baselines across four benchmarks, reaching 30.7% on LOCA-bench at a 64k-token state where the baselines score zero.
In briefHow a camera-first AI agent should be designed: splitting memory and reasoning into two deadlines
SummaryThe paper argues that the main obstacle for future camera-first agents is time rather than recognition, because many visual targets appear and disappear within a short window. The authors formalize this as Ephemeral-Target Interaction and a Perceptual Deadline Contract that separates a hard memory deadline from a soft reasoning deadline. They propose a two-deadline architecture with an always-on fast path, event-driven active memory, and a deferred deliberation layer. The design is instantiated on MOT17 track windows and calibrated against measured frontier-model latency.
In briefPrior-fitted and frozen Transformer scorers for multimodal contrastive learning
SummaryThis paper studies the scoring functions that measure similarity between representations from different modalities in multimodal contrastive learning. The authors use a Transformer as the scorer and find that even a randomly initialized, frozen model can outperform strong handcrafted baselines. They further propose the Prior-Fitted Scorer (PFS), pre-trained on synthetic multimodal data generated from structural causal model priors and then transferred to real data. Across four real-world datasets, frozen Transformer scorers are competitive with contrastive baselines, and prior-fitting further improves results.
In briefWho Gets Respect? A Counterfactual Benchmark for Honorific Bias in Bengali Translation
SummaryThis paper examines how large language models assign respectful versus ordinary forms when translating English into Bengali, where honorific choice is grammatically required but absent in English. The authors built a benchmark of 2,016 sentences across 72 occupations with counterfactual gender pairs and computed a 0–100 Honorific Bias Score for nine models. Occupation explains far more variation in register choice than gender, and only two models show significant gender gaps.
In briefCUSE: A Fine-Grained Method for Making Multimodal AI Models Forget Specific Information
SummaryMultimodal large language models can reproduce sensitive, copyrighted, or unsafe information from their training data. This paper proposes CUSE, which builds counterfactual inputs by removing target-related evidence to find which tokens need stronger forgetting. It then distills token-level forgetting and retention targets on responses sampled from the current model. On MLLMU-Bench, CUSE cuts forget-set ROUGE-L by 33.6% and scores 1.2% higher on retain-set ROUGE-L than the strongest baseline.
In briefCrossMean: merging models by pairing merged and expert activations
SummaryThis paper studies model merging, which combines several fine-tuned expert networks into one without retraining. The authors note that existing closed-form methods compute layer statistics from a single expert, while merged layers receive activations from preceding merged layers. They propose CrossMean, which uses a bilinear statistic pairing merged and expert activations. Experiments on an eight-task benchmark for vision backbones and on BERT encoders show gains over expert-measured statistics, and over re-measuring statistics in the merged model when all linear layers are merged.
In briefTeaching a Machine to Translate Nheengatu: Open Models and a Reproducible Corpus for an Amazonian Language
SummaryThis paper addresses poor machine translation support for Nheengatu, an Amazonian language with roughly 6,000 to 10,000 speakers that is co-official in São Gabriel da Cachoeira, Brazil. The authors assemble a 14.5k-pair Portuguese–Nheengatu parallel corpus anchored by the 2023 official translation of the Brazilian Federal Constitution, and build the first evaluation set from transcribed spontaneous speech. They release what they describe as the first open-weight bidirectional Portuguese–Nheengatu MT models. Portuguese-to-Nheengatu systems reach 31.7–32.8 chrF++ on constitutional text and 45.2–47.5 on speech.
Fabr\'icio Carraro, Rodolfo Joel Zevallos, Rodrigo Gon\c{c}alves de Souza et al.
In briefBetter interfaces for collecting human opinions during iterative optimization
SummaryThis paper studies how to design interfaces that collect a human decision-maker's feedback in human-in-the-loop optimization. The authors compared six representative opinion measurement interfaces in a study with 360 participants, using text summarization and image color enhancement tasks. They recommend a listwise approach, which addresses limits of both individual ratings and preferential choices and conveys more information, and they discuss trade-offs among interface designs.
In briefAmharic-English Cross-Language Question Answering Using Machine Translation and Semantic Indexing
SummaryThis paper presents a bidirectional question answering system that lets users ask questions in Amharic or English and retrieve answers from documents in the other language. It combines statistical machine translation for queries, maximum-entropy named entity recognition for Amharic, semantic indexing, and passage retrieval. Reported precision and recall were 72% and 79% for Amharic-to-English retrieval, and 65% and 70% for English-to-Amharic retrieval.
In briefGateSight: Recurrent memory predicts attention cache needs for query-agnostic KV eviction in hybrid LLMs
SummaryThe paper studies hybrid language models that combine fixed-size recurrent memory (Gated DeltaNet layers) with full-attention layers whose key-value cache grows with context length. It shows that signals emitted by the recurrent layers during prefill can predict how useful each attention KV entry will be later. The authors propose GateSight, which uses lightweight per-head multilayer perceptrons on these signals to decide which entries to keep, scoring the cache once after prefill. Reported experiments on SCBench and RULER-4096 show it keeps accuracy with a tenth of the attention KV cache, ranking highest among five query-agnostic methods on most model-benchmark pairs.
WhyPredicts future KV entry utility from GDN gate signals during prefill, a novel query-agnostic eviction approach validated against baselines on two models.
In briefComputing random Fourier features in near-linear time
SummaryRandom Fourier features approximate shift-invariant kernels such as the Gaussian kernel, but generating m features normally takes O(md) time. Earlier structured methods sped this up but with weaker concentration guarantees. This paper presents a feature map computable in near-linear time O((d+m)log d) that achieves near-subgaussian concentration. It is obtained through a tighter analysis of the Fastfood feature map.
In briefFair-comparison evaluation of credit assignment in modular financial RL
SummaryThis paper proposes an evaluation framework for credit assignment in modular financial reinforcement learning, combining equal-information reconstruction controls, fixed-utility comparisons, and prespecified downstream guardrails. The framework is tested in an attribution-guided modular trading architecture across three controlled studies on update allocation, supervision-target representation, and portfolio risk control. Results show that residual-based allocation is matched by an ordinary loss-budget control, and some supervision methods improve component prediction but fail to preserve decoded outcomes.
In briefMore Domains, No Gains? FLO: Fast-Slow Reciprocal Learning for Multi-Domain Graph Pre-training
SummaryThe paper finds that existing graph foundation models struggle to benefit from more domains during multi-domain pre-training, citing reliance on a single message-passing paradigm and early knowledge acquisition disturbed by cross-domain interference. The authors propose FLO, which adds a propagation-free attribute branch that predicts relations to refine the model, and pairs a fast per-domain process with a slow cross-domain integration process. On 12 datasets, FLO achieves the best average performance in cross-domain and cross-task transfer.
In briefBridgeWAM: Linking Video and Action Experts Through Latent Bridge Queries
SummaryThis paper introduces Bridge-of-Experts (BoE), implemented as BridgeWAM, which uses 32 Latent Bridge Queries to combine visual, language, and proprioceptive information into a shared, limited latent space. The action expert selectively retrieves task-relevant information via cross-attention, and the two experts have decoupled depths in a 5.06B-parameter model. It reports 98.25% success on LIBERO and 72.24% pooled success on LIBERO-Plus. Compared with Fast-WAM, it reduces single-sample inference FLOPs by 28.72% and reaches 77.1 ms real-robot latency.
In briefSAGE: Making 2-bit vector-quantized LLMs more practical
SummaryThis paper proposes SAGE, a weight-only vector quantization method that compresses large language models to 2 bits to reduce deployment costs. It introduces Energy-Stratified Grouping to protect sensitive weight channels and Adaptive Bit Split to balance codebook capacity and radius precision under a fixed bit budget. The authors report state-of-the-art results in the 2-bit setting, retaining 95.3% of the BF16 average score on Qwen3-30B-A3B and up to a 2.55x non-streaming speedup over FP16 with a custom CUDA kernel.
WhyAdds energy-stratified grouping and adaptive bit split to 2-bit vector quantization, validated across diverse benchmarks, real models, and a custom CUDA kernel speedup.
In briefJapanese handwriting generation trained with human preference feedback via Direct Preference Optimization
SummaryThis paper proposes a framework for generating Japanese handwriting that uses Direct Preference Optimization (DPO) to incorporate human preference feedback, aligning the model with human perception and style. Handwriting is non-deterministic, so the same character varies in shape, and collecting enough handwritten data is costly, whereas humans can easily compare two generated samples. Experiments on four Japanese handwriting styles show that DPO training improves overall structural accuracy. However, for highly variable styles such as brush pen writing, the model fails to capture style-specific features.
Wakana Kuwata, Ryota Mibayashi, Masanori Tani et al.
In briefsGLP: Simulating Grassroots Platforms with Stochastic Agents
SummaryGrassroots platforms are apps run by their participants, specified in a volition-guarded Grassroots Logic Program (vGLP). This paper introduces stochastic GLP (sGLP), a rate extension of GLP that replaces people and their user interfaces with stochastic agents so platforms can be simulated at scale before deployment. The authors demonstrate sGLP on a grassroots social graph and grassroots currencies, reporting how different mixes of behavioural profiles shape the resulting social graph and circulating currencies. They also prove that every run of a vGLP program with sGLP agents in place of people is a valid vGLP run.
In briefPTGPO: Prefix Tree Group Policy Optimization, a critic-free reinforcement learning method for training language agents on long tasks
SummaryThe paper addresses the difficulty language agents face on long-horizon interactive tasks, where sparse terminal rewards give weak credit signals for intermediate actions. It proposes PTGPO, a critic-free method that builds prefix trees over trajectory suffixes within anchor-state groups and assigns each edge the difference between its child and parent empirical mean returns. Experiments on ALFWorld, WebShop and Sokoban report that PTGPO consistently improves over RL baselines.
In briefTool-augmented agent that plans strategies to make LLM-based optimization traceable
SummaryThe paper addresses a limitation of Optimization by PROmpting (OPRO), where LLM-driven search is hard to track and cannot reuse information from earlier iterations. It proposes the Tool-Augmented Strategic Agent (TASA), which plans a strategy for each step from the optimization trajectory and uses LLM-built problem-specific tools to update solutions. The authors establish local convergence under persistent exploration and evaluate the framework under a shared evaluation budget, reporting solution quality and traceability of strategy-guided search.
In briefAlign-VAE: aligning attention and geometry in VecSet VAEs for 3D shape reconstruction
SummaryThis paper studies set-based variational autoencoders (VecSet VAEs), which represent 3D shapes as compact sets of latent tokens, and checks whether those tokens truly align with surface geometry. The authors identify four alignment problems and propose Align-VAE, which adds shared decoder tokens, distance-biased cross-attention, normal-free encoder queries, and a screened Poisson-inspired regularizer. On the ABO dataset, Align-VAE reduces Chamfer distance by 26.6% compared with the standard VecSet VAE.
