In briefUniversal Test-Time Training: One Shared Fast Memory Across All Layers
SummaryThis paper introduces Universal Test-Time Training (uTTT), where all network layers read and write a single shared memory while keeping layer-specific backbone parameters. The shared memory recurs across both time and depth. Two variants, uTTT-MoE and uTTT-Dense, are presented. In language modeling, uTTT-MoE outperforms its layer-private counterpart on RULER accuracy at equal state and active compute, and in novel view synthesis the shared design improves object PSNR.
In briefFLEX-WAM: A Flexible World-Action Model for Long-Horizon Imagination and Planning
SummaryFLEX-WAM is a unified world-action model that predicts future states conditioned on actions, generates actions, and supports planning in imagination. It supports variable-length contexts and infinite autoregressive rollouts through a block-causal, KV-cacheable design, and it balances training gradients and regulates sampling using Forward-Dynamics elasticity to improve action responsiveness. The paper reports better multi-step prediction quality and latency, solving PushT and all five OGBench Puzzle-4x4 tasks in imagination via MCTS, and real-time use on a bimanual OpenArm robot.
R. Khorrambakht, Joseph Amigo, F\'elix Lebel et al.
In briefReflections on creative limits with an AI doodling partner
SummaryThis paper examines how negotiating creative boundaries with a co-creative AI system can prompt personal reflection on one's creative practice. The authors study Froggi-Draw, a single-initiative co-doodling system where users control when and how much the AI "collaborator" contributes. In a one-week pilot with eight novice and experienced artists, they report how participants handled creative risk and uncertainty and how their use of and views of the system changed over time. The authors argue that a supportive co-creative AI should offer users control so they can find and negotiate the boundaries of the power dynamic.
Samia Menon, Samyukta Jayaram, Chetan Goenka et al.
In briefCodeForge-MA: Multi-agent, execution-verified training for multilingual code generation
SummaryThis paper presents CodeForge-MA, a framework for improving code generation in large language models. Four agents iteratively refine instruction-code pairs, verify them with tests, and filter duplicates and benchmark leakage, while training combines masked supervised fine-tuning with a test-driven reinforcement objective. For the larger model, sparse expert routing over language-conditioned LoRA adapters is used, keeping the base model unchanged at inference. The authors report that the joint data, objective, and adapter design yields robust gains across programming languages.
In briefPopulation Scaling or Data Dilution? Local Topology Evolution in Decentralized Learning
SummaryThis paper examines how increasing the number of clients in decentralized learning affects accuracy and consensus, comparing Ring, Static Random graph, and a locally adaptive topology (LFHE) on CIFAR-10 with 10 to 200 clients. Holding each client's local dataset size fixed substantially reduces the apparent population penalty seen when a fixed total dataset is split. Remaining degradation depends strongly on communication structure: Ring enters a high-disagreement regime, while Static Random and LFHE stay close to consensus. Raising LFHE's degree threshold further improves accuracy and consensus but at a substantially higher model-transmission cost.
In briefHierarchical RL with stable temporal abstraction for language agents
SummaryThe paper observes that an explicit hierarchy alone does not ensure stable temporal abstraction, as subgoals may be replaced almost every turn or kept after becoming inappropriate. It proposes STAC, a constrained boundary-policy optimization method that treats premature replanning and stale persistence as constraint costs applied only to sampled boundary decisions. Across two backbones and two benchmarks, ALFWorld and WebShop, STAC improves success over a strong hierarchical baseline.
Shayan Mohajer Hamidi, Yize Cheng, Yuanda Xu et al.
In briefReasoning Hallucinations in Large Models: Visibility, Causal Evidence and Release Control
SummaryThis survey reviews 312 papers and first-party reports on hallucinations in text-based reasoning by large reasoning models. It examines which evidence can be observed, which study designs establish causation, and which corrective actions the evidence supports. The authors propose the UIPCA framework covering unsupported premises, invalid inferences, dependent reuse, answer-trace consistency, and action-policy failures. Across 58 reviewed sources, no matched comparison showed that an intervention improves reasoning without reducing factual reliability.
In briefShaping the Wind: Nested Potentials for Physically Consistent Urban Wind Prediction
SummaryThis paper presents Sculpt, a nested potential framework that builds mass conservation and wall impermeability directly into a neural surrogate for predicting transient urban winds. It generates divergence-free velocity updates via the discrete curl of a volume vector potential on a three-dimensional staggered grid, with a shared scalar potential enforcing boundary constraints. Multi-resolution potentials are used to better capture large-scale flow structures. The authors also introduce UrbanWindFlow, an LES dataset spanning urban morphologies and inflow conditions, for evaluating accuracy and kinematic admissibility together.
In briefWriting as a self-organized critical process
SummaryThe paper explains autocorrelation decay power laws found across texts using self-organized criticality. The authors analyze the recently released KLiCKe keystroke dataset and find that final texts' autocorrelations follow a power law with finite-size scaling. Text revisions also produce restoring dynamics, dependent on revision size, that move toward this manifold, suggesting human writing dynamically regulates semantic correlations toward a critical state.
In briefDo multimodal AI models search images like humans? A psychophysical comparison.
SummaryThis study compares visual search performance of 1,250 humans and multimodal large language models using identical 2D and 3D stimuli at varying set sizes. Both groups performed efficiently on feature searches and slowed on conjunction searches as set size grew. Error rates correlated strongly (ρ = 0.82), but humans spent extra time on target-absent trials while models showed extreme present/absent response biases. The authors conclude the models replicate high-level human performance patterns, but their underlying computations differ significantly.
Renchi Zhang, Joost C. F. de Winter, Dimitra Dodou et al.
In briefDebate-based AI safety can be undermined by human cognitive biases
SummaryThis paper examines whether AI safety via debate, where two agents argue opposing positions for a human judge, remains reliable against manipulation. The authors built 68 LLM-generated detective-mystery dialogues and applied four rhetorical interventions: anchoring, fallacy oversight, pro-jargon, and verbosity. In a study with 369 participants, these interventions significantly shifted judgments toward the manipulated side, showing that human adjudication is vulnerable to such strategies.
Gefei Liu, Sonya Rashkovan, Sophia Lloyd George et al.
In briefMeasuring and reducing mismatch of the same language model across GPU vendors
SummaryThe same language model can produce different outputs on GPUs from different vendors, even with identical weights and inputs. The paper analyzes two dense and two mixture-of-experts models using five metric families and traces part of the mismatch to accumulation order inside vendors' matrix instructions. Upcasting to FP32 cuts the dense model's logit error by 43% at three times the runtime, while keeping only the MLPs in BF16 keeps 94% of that gain at about 1.3 times the runtime.
WhySystematically measures cross-vendor GPU mismatch with multiple metrics; MLP-only BF16 retains 94% of gains, validated across models and distillation experiments.
Erland Hilman Fuadi, Chong Tian, Xiaosong Ma et al.
In briefTaylor representations for model-free reinforcement learning in networked MDPs
SummaryThis paper studies Networked Markov Decision Processes, where transition dynamics are unknown and the state-action space grows quickly with the number of agents. Under smooth expected local rewards with controlled derivatives, the authors show that Taylor coefficients decay with graph distance, allowing distant-agent terms to be discarded and yielding scalable local representations. They propose a model-free actor-critic method with finite-sample guarantees for a linear LSTD critic, plus a neural TD variant. Across three control benchmarks, the method matches or outperforms spectral baselines and scales to large graphs.
Salah Chikhi, Abdelhaq Chaoui, Asuman Ozdaglar et al.
In briefCompressing biomedical translation models: combining distillation and quantization for faster French-to-English translation
SummaryThis paper studies combining knowledge distillation and quantization to compress neural machine translation models for French-to-English biomedical text. The authors compare several fine-tuning strategies to adapt compressed student models to specialized terminology and limited parallel data. They report that a collaboratively distilled and quantized student model is 69% smaller, has 98.21% faster inference, and 98.46% lower CO2 emissions than the baseline without sacrificing translation quality.
In briefArtemis: Geometry-grounded multi-agent driving world model with shared 3D state
SummaryThis paper presents Artemis, a multi-agent driving video world model that builds an explicit 3D world map from multi-agent observations to keep a unified 3D state across agents. It distinguishes controllable agents from uncontrolled background dynamics and updates the 3D map progressively using keyframes from video rollouts. The authors also introduce MA-CARLA, a dataset sampled from the CARLA simulator, and report better visual fidelity and cross-view consistency than prior methods.
In briefLLM-based multi-agent collaboration for placing containers under constraints while balancing power and affinity
SummaryThis paper proposes a framework in which four LLM-driven agents collaborate in a closed loop to decide where containers run in a data center. The agents balance power consumption against affinity preferences while respecting multi-resource capacity and anti-affinity constraints, with candidate actions filtered for feasibility. On the Google Cluster Trace with 100 applications and 25 machines, it uses 8.36% less power and achieves 39.30% higher affinity than a power-greedy baseline.
Suvarthi Sarkar, Abhishek Hazra, Akshat Mittal et al.
In briefVERA: Verdict-Conditioned Reliability for Adaptive LLM Judges
SummaryThis paper addresses how LLM judges can estimate the reliability of their judgments when adapting to newly verified feedback without forgetting prior behavior. It proposes VERA, which estimates reliability from hidden activations by distinguishing correct from incorrect judgments within each predicted-verdict group, and uses it to guide a periodic adaptation framework. After adaptation on Chatbot Arena, 8B and 14B judges outperform the strongest baseline on four held-out public benchmarks, with relative gains of up to 23.01%, and improve focal-class recall by up to 16.1% on a proprietary temporal auditing task.
In briefGroup-Specific Policy Learning for Fairer Preference-Based LLM Alignment
SummaryThe paper studies preference-based alignment methods such as RLHF and NLHF, which learn from pairwise comparisons but may not maximize social welfare. It proposes GLHF, an algorithm that learns one group-conditioned policy from one comparison per user, matching the optimal population distortion bound across a prespecified set of possibly overlapping groups. Experiments with human coffee ratings and synthetic LLM ratings show lower distortion in every evaluated group and reduced worst-group distortion compared with NLHF and other baselines.
WhyExtends distortion guarantees to overlapping groups via group-conditioned policies, with logarithmic sample growth; supported by coffee-rating and synthetic LLM experiments.
In briefHierarchical Time-aware Bootstrapping for Learning Subgoal Values Off-Policy
SummaryOff-policy hierarchical reinforcement learning must estimate high-level decision values while the low-level policy keeps changing. The paper proposes Hierarchical Time-aware Bootstrapping (HTB), which evaluates specified subgoals under the current low-level policy, keeps accumulated task rewards, and uses remaining execution time to separate subgoal continuation from new high-level decisions. On AntFall at 10M environment steps, HTB reached a 32.8% success rate versus 9.6% for matched local HIRO over five paired seeds.
In briefSublevel Flood bifiltration: a scalable way to compute two-parameter persistent homology
SummaryThis paper introduces a method for efficiently computing 2-parameter persistent homology on large point sets, called the sublevel Flood bifiltration. It extends the single-parameter Flood filtration and serves as a scalable approximation of the sublevel offset bifiltration. The authors report theoretical stability properties and describe how to compute it efficiently. They demonstrate its performance on classification tasks using low-dimensional synthetic datasets and real-world time series datasets.
In briefFault vulnerability study of ReRAM-based in-memory CNN accelerators
SummaryThis paper studies how faults affect convolutional neural networks running on ReRAM-based processing-in-memory accelerators. The authors built a fault injection framework that maps CNN parameters to ReRAM crossbars and injects stuck-at-high and stuck-at-low faults. They report that vulnerability depends on layer type and depth, parameter location, and fault type and value, and that stuck-at-low faults reduce classification accuracy more than stuck-at-high faults.
In briefTeleTune: Learning Agent Skills from Offline Usage Logs
SummaryThis paper presents TeleTune, a framework that builds a textual skill library for computer-use agents from offline user telemetry, even though logs lack recorded goals, cannot be replayed, and may mix several tasks. It proposes library edits from action-prediction errors and keeps only those that improve held-out action-prediction accuracy. On WorkArena and Online-Mind2Web, TeleTune reports average success rates of 77.1% and 80.6%, exceeding the strongest baseline on each benchmark.
Justin Chih-Yao Chen, Elias Stengel-Eskin, Yan Chen et al.
In briefIdentifiable world models from frozen diffusion features: a lightweight alignment layer makes latent coordinates match true state variables
SummaryThe paper asks whether a frozen pretrained diffusion model can be given latent coordinates that recover underlying state variables without retraining its generative backbone. It proposes Contrastive Diffusion Alignment (ConDA), which learns only a lightweight alignment map on top of frozen diffusion latents, and shows under standard TCL/GCL assumptions that the aligned representation is identifiable up to permutation and componentwise invertible transformations. Experiments on physical and robotic video systems show near-perfect blockwise state recovery and competitive lagged graph recovery for TCL and GCL variants, with exact recovery in a simulated falling-body system.
In briefPharmAgent: Constraint-Aware Molecular Optimization Using Frozen Language Models
SummaryPharmAgent is a method for optimizing molecules so they improve target activity while meeting developability constraints within limited evaluation budgets. It uses a Lagrangian controller that accumulates constraint violations from accepted states and a structure-indexed replay of relevant past transitions to guide new candidate proposals. Across five tasks, the paper reports higher target-score AUC than MOLLEO and higher property-adjusted AUC than ExLLM in constraint-aware search. The authors state that all five molecular cases reached feasible final states.
In briefLearning in Continuous Games from Pairwise Preference Feedback
SummaryThis paper studies how players learn in continuous games when they only observe which of two actions is preferred, without payoff values or preference strengths. The authors show that standard regret and coarse correlated equilibria cannot be identified from such ordinal feedback, and they introduce an ordinal directional regret benchmark with corresponding equilibrium notions. Using a single pairwise comparison per player and round, their algorithm achieves sublinear finite-resolution ordinal regret and, in ordinal potential games, almost-sure last-iterate convergence to the Nash set.
In briefHow Unmentioned Checklist Findings Affect Reinforcement Learning for Chest X-ray Report Checking
SummaryThe study trains a vision-language model with reinforcement learning to fill a 12-finding checklist from chest radiographs, and a separate model judges report sentences against it. On held-out patients, a rule-based check and an independent medical checker showed discrimination gains of 12.6% and 11.8%, but only the rule-based check met the prespecified false-alarm criterion. Switching to a training format that fixes finding order and marks unmentioned findings as absent changed the two checkers' results in opposite directions. Labels were derived from reports rather than adjudicated by radiologists.
In briefEvoMem-VLA: State-Evolution Memory for Robots' Long-Task Manipulation
SummaryMost vision-language-action models rely only on current observations, so task-relevant information is lost once it leaves view, which hurts long-horizon, memory-dependent tasks. This paper proposes EvoMem-VLA, which builds state-evolution memory by encoding and retaining observed changes between past states, plus a routing scheme that uses a subtask route for multi-stage tasks. The authors report success rates of 80.7% on RMBench, 82.0% on RoboMME, and 83.8% across four real-world tasks, improving over the previous state of the art in all three settings.
In briefRender to Reason: Predicting Unseen Views Helps Vision-Language Models Understand Space
SummaryPrior work added geometry features from pretrained 3D models to vision-language models, but simply fusing them and training on standard spatial QA gave only small gains on multi-hop tasks. The authors attribute this to weak training signals that do not require geometry, and propose novel-view semantic rendering, an auxiliary task where the model predicts the semantic layout of an unobserved viewpoint. The method gives consistent gains over the geometry-augmented baseline on three benchmarks, and the full model surpasses prior open-source methods on VSI-Bench and ReVSI.
In briefPrism: A Dynamic Sparse Attention Method for Training 2K Joint Video-Audio Generation Models
SummaryThis paper presents Prism, a dynamic sparse attention framework for natively training joint video-audio generation models at 2K resolution. It divides token sequences into spatiotemporal macro-zones and assigns each zone a tailored block shape based on video feature variance and audio-to-video cross-attention norms. According to the authors, Prism trains 2.5 times faster than full attention while achieving better generation quality.
In briefA simple, training-free benchmark for judging vision encoders in multimodal LLMs
SummaryThis paper examines how to evaluate vision encoders by predicting their performance inside multimodal large language models. The authors run large-scale experiments, identify limitations in prior cross-modal evaluation methods, and introduce improvements. They propose RAVEL, a training-free method based on cross-modal nearest-neighbor retrieval, which achieves state-of-the-art results and outperforms earlier approaches by a substantial margin.
In briefCleanMDM: A Clean Motion Diffusion Model for Multimodal Motion Cleanup
SummaryMotion capture data often has missing segments, jitter, drift, and contact errors, so it needs cleanup before use. CleanMDM treats cleanup as masked conditional generation that accepts noisy 3D motion, sparse 2D or 3D keyframes, and text in any combination. It adds a Latent Motion Quality Discriminator and Mesh-Aware Contact Projection to reduce skating, jitter, and interpenetration. The authors report that it outperforms prior baselines and that text and 2D keyframes improve controllability.
In briefBeliefGraph-JEPA: A Structured Latent World Model for Forecasting Time Series Under Actions
SummaryThe paper introduces BeliefGraph-JEPA, a structured latent world model that represents the influence of future drivers as typed latent-effect states. These states are rolled forward over the forecast horizon and routed through a graph to target-specific nodes. On four multi-target clinical, agricultural, environmental, and industrial systems, the framework reportedly outperforms several pretrained and supervised known-future-covariate baselines.
In briefNo Concept Escapes the Audit: Auditing-Aware Unlearning for Verifiable Concept Erasure in Diffusion Models
SummaryThis paper argues that common concept-erasure methods for text-to-image diffusion models act mainly at the text interface and are judged only by outputs, so erased concepts can remain encoded inside the network, as latent-space auditing reveals, even for methods robust to adversarial prompts. The authors propose AVCE, which audits each concept's embedding neighborhood, consolidates vulnerable directions into an anchor, edits attention projections in closed form there, and uses orthogonal gradient projection to handle multiple concepts. Experiments on SD v1.5, SDXL, and Flux 1.0 report 5.07x lower attack success rates and 3.84x higher auditing scores than the strongest baseline, with competitive generation quality.
In briefTask Vector Descent: learning from batches that arrive in clusters, without losing old knowledge
SummaryThis paper studies continual learning for language models when training data arrives in temporally clustered batches from different distributions, which can cause forgetting of earlier learned data. The authors compare fully applying the parameter update (the task vector) with scaling it by a coefficient λ before applying it to the model. Across continual pretraining, pretraining from random initialization, supervised post-training, and reinforcement post-training, intermediate λ values often outperform full application, especially after longer same-distribution sequences.
Anton Baumann, Jonas H\"ubotter, Zeynep Akata et al.
In briefCASE: A Cost-Aware Stopping Framework for Efficient Long-Video Agents
SummaryLong-video agents gather question-relevant evidence but often do not clearly decide when they have seen enough to answer. CASE is a plug-in termination framework that uses a lightweight Ridge regressor to learn the gap between answering now and stopping later, accounting for correctness and reasoning cost, and outputs STOP or CONTINUE decisions. On Video-MME, it changes end-to-end accuracy by +0.67 points on average while reducing model-token use by 53.63%. The frozen policies also transfer zero-shot to LongVideoBench and MLVU.
In briefMMPostTrainBench: a benchmark for autonomous multimodal model post-training
SummaryThis paper introduces MMPostTrainBench, a benchmark of eight tasks spanning image, audio, video, and joint audio-video understanding plus image-grounded software repair, testing LLM agents that improve models within fixed budgets. Across tasks, 52.1% of model-task means fall below the base model, and submissions show regressions on non-target abilities. Agents also fail to reliably submit their best evaluated candidate. The authors propose MMResearch, a framework using hierarchical memory and development evaluation, which raises submitted-model accuracy for two tested agent setups.
WhyProposes a multimodal autonomous post-training benchmark and hierarchical-memory framework, systematically evaluating eight tasks and revealing regressions and selection gaps.
In briefFaster Graph Generation via Direct Prediction and Flow Matching (DiGFM)
SummaryThis paper introduces DiGFM, a graph transformer that generates graph-structured data using continuous flow matching and directly predicts clean graphs instead of noise. It transforms a prior noise distribution into the clean graph distribution over multiple steps. According to the paper, it produces high-quality samples with 2.5% to 15.6% of the steps used by diffusion models, yielding 5.3x to 257x faster inference. It reportedly matches or outperforms prior state-of-the-art models on general graph and molecular benchmarks.
In briefThe Hidden States Cookbook: Study on Noise-Robust Intent Classification
SummaryThis paper examines how to keep intent classification accurate in conversational database interfaces when user queries contain greetings, politeness, or off-topic remarks. The authors ran 360 controlled experiments with Llama-3.2-1B-Instruct on the BANKING77 and CLINC150 datasets, comparing several pooling strategies. Under noisy conditions, attention pooling performed best, while mean pooling degraded performance by up to about 5 F1 points.
In briefOffline voice AI system with hardware and quantization benchmarks for African languages
SummaryThis paper presents an offline, voice-first AI system designed for African language communities with unreliable internet access. The authors benchmark three instruction-tuned models of 2-5B parameters across four quantization formats on an NVIDIA Jetson Orin NX and a Raspberry Pi 5. They evaluate deployment metrics and multilingual topic classification, along with speech recognition on Amharic and Oromo. The main finding is that Q4_K_M quantization offers the best size-to-quality trade-off on both hardware tiers.
In briefGNN-CB: A Graph Neural Network Competition Benchmark for Humans and LLMs
SummaryThis paper introduces GNN-CB, the first competition-based benchmark for evaluating humans and large language models on graph neural network coding tasks. It includes 18 competitions covering node, edge, and graph-level prediction, scored through an automated pipeline with hidden test sets. Under the evaluated protocol, LLMs rarely matched the top human performance and were less stable across competitions, though a few competitions were won by LLMs. The benchmark and framework are publicly available.
In briefSibyl: A Small-Large Model Collaboration Framework for Long-Horizon Tasks
SummaryThe paper proposes Sibyl, a method that trains small language model agents to selectively consult larger cloud models at key steps and internalize their guidance for later decisions. Training follows three stages, jointly optimizing task success and the number of cloud calls. On ALFWorld and WebShop, a 0.6B-parameter model outperforms baselines by 95.2% and 80.4% in success rate, averaging only 0.8 and 3.9 cloud calls per trajectory.
In briefHarness-Search: Multi-agent coordination for long-horizon search
SummaryThis paper introduces Harness-Search, a multi-agent framework for long-horizon search that splits proposing retrieval actions, updating persistent state, and deciding when to stop among three permission-bounded agents. Across seven long-horizon search benchmarks, it improves retrieval and answer generation over the strongest harness-based baseline under the same policy backbone. Trajectory analyses indicate it keeps accumulating useful evidence and expanding coverage with less redundant retrieval as search history grows.
In briefVHDL-REPOBENCH: A benchmark testing how well AI models generate VHDL hardware designs across whole repositories
SummaryThis paper introduces VHDL-REPOBENCH, a large-scale benchmark built from about 100 open-source VHDL repositories containing roughly 2.5k files and 500 testbenches. It evaluates large language models on syntax, semantic correctness, cross-file dependency resolution, and functional verification. Tested models achieved moderate line- and block-level accuracy, but struggled with multi-file reasoning, hierarchical design understanding, and generating modules from specifications.
Prashanth Vijayaraghavan, Akul Malhotra, Ashutosh Jadhav et al.
In briefDoes Explainability Survive Data Drift? A study on whether model explanations stay faithful when input data changes
SummaryThe paper examines whether explanations of machine learning models remain faithful to the models that replace them after data drift. Using the IEEE-CIS Transaction Fraud Detection dataset, the authors find significant covariate shift but no significant concept drift. Prior explanations stay relevant to a retrained model but are consistently less faithful than newly generated ones, so the authors argue explanations need their own monitoring and should be tied to the model that produced them.
In briefDistributed Subliminal Learning: Sharing Knowledge via Random-Input Model Outputs Instead of Weight Updates
SummaryThe paper proposes Distributed Subliminal Learning (DSL), where participants adapt a common model locally, probe it with task-unrelated inputs, and send only the resulting outputs. A coordinator pools these outputs and distills them into a shared model. The authors report that, compared with LoRA averaging in LLM composition and FedAvg in federated classification, DSL achieves higher preference retention and substantially lower upload or uplink communication.
Dario Fenoglio, Gabriele Dominici, Martin Gjoreski et al.
In briefReconstructing incomplete physical fields by globally correcting local estimates
SummaryThis paper studies how to reconstruct physical fields when every training sample is incomplete. The authors propose an estimator that first learns local field estimates from incomplete patches with a shared coordinate-conditioned predictor, then reconciles them into a consensus field. A full-domain model learns a residual correction from the original observations to revise this local structure. On three real ocean datasets with authentic observation gaps, it achieves the lowest MSE and highest PSNR on withheld values, reducing MSE by 28.9%–34.5% against the strongest external baseline.
In briefUnbiased On-Policy Distillation for Block Diffusion Language Models
SummaryOn-policy distillation is a post-training method for language models, and recent work has applied it to block diffusion language models, mostly with small block sizes. This paper identifies two biases that cause unstable training when student models use larger blocks: mismatched block boundaries that misalign context, and an optimization tendency that drives overconfidence collapse. It proposes Un-OPD, which filters context-misaligned decoding steps, recalibrates confidence at high-support positions, and reuses rollouts to cut generation overhead. Experiments on math reasoning and code generation benchmarks report more stable training, better performance, and roughly half the wall-clock training time.
In briefWiSPER: WiFi-based 3D Pose Estimation for Multiple People
SummaryWiSPER is a two-stage framework that estimates 3D poses of multiple people from WiFi channel state information (CSI). It combines pose-supervised masked pretraining (PAMEL) with a Transformer-based conditional residual flow that refines pose candidates (ReFT). On the PiW3D dataset it reaches a mean per-joint position error of 63.72 mm, 40.0% lower than WiFi-JEPA. Inference requires only CSI input.
Gabriel Lee Jun Rong, Shanhong Liu, Pai Chet Ng et al.
In briefRobust Ensemble Guidance (REG) for Scientific Inverse Problems: Weighting and Clipping to Improve Diffusion-Based Reconstruction
SummaryThe paper proposes Robust Ensemble Guidance (REG), which improves ensemble guidance for scientific inverse problems by weighting observations using predictive spread and adaptively clipping standardized residuals. This lets pretrained diffusion priors work with black-box forward models without differentiating through the simulator. The method reuses existing particles and predictions, and the authors report better reconstructions than the underlying ensemble solver on Navier-Stokes inversion, black-hole imaging, and acoustic full-waveform inversion. Reported gains include 6.2-8.2 dB PSNR improvement for black-hole imaging and a 26.4% error reduction for Navier-Stokes in a matched-budget comparison.
In briefEnGRICH: Improving Generative Reward Models Using Human Critiques
SummaryGenerative reward models produce written critiques along with preference judgments, but existing training only checks whether the final preference is correct, so unreliable critiques can be reinforced. EnGRICH trains a generative reward model alongside a training-time MetaCritic learned from a small set of human critiques. MetaCritic builds response-specific rubrics to assess the evidence coverage and correctness of generated critiques and generalizes these criteria to outcome-only preference data. The trained reward model runs independently at inference, and the authors report consistent gains over competitive baselines across seven reward-model benchmarks.
In briefEffect of Missingness-Pattern Mismatch on Time-Series Classifier Selection: A Controlled Study
SummaryThis paper studies whether a mismatch between the missing-data pattern in validation and at deployment affects classifier selection for time-series classification. Using 64 univariate UCR datasets in a controlled 2×2 design with random point and circular block missingness, the authors compare three candidate classifiers. Mismatched validation reduced test balanced accuracy by 1.14 percentage points on average, with the effect growing at higher missingness rates.
In briefAIProver: Certificate-driven evolving agent for auto-formalizing math research into Lean
SummaryThis paper presents AIProver, an agentic framework that automatically translates natural-language math theorems and proofs into formal Lean code and synthesizes proofs. It post-trains a 119B open-weight language model and evolves the agent's tool-calling harness using HarnessEvolve, guided by verifier certificates for type correctness, proof completeness, and semantic correctness. The authors also introduce LoCoBench, a benchmark of 58.9k instances drawn from several formal libraries. According to the paper, pass@4 semantic correctness rose from 15.7% to 36.7% over its base model, and the system is reported to be 24% cheaper than Numina-Lean-Agent.
WhyCombines certificate-driven harness evolution with post-training, beating many baselines on research-level formalization with fairly thorough comparisons, a meaningful but subfield-focused impact.
Prithwish Jana, Viet Bach Hoang, Logan Luna et al.
In briefIrọyinSpeech Text Corpus: 24,905 Curated Yorùbá Sentences for Speech Technology
SummaryThis paper releases the text component of IrọyinSpeech, a Yorùbá read-speech corpus. It contains 24,905 unique, hand-verified, tone-marked sentences curated in 2022 as recording prompts, with about 11,000 adapted from openly licensed news. The authors also document and correct Unicode normalisation failures affecting over 60% of lines. The corpus supports diacritic restoration, grapheme-to-phoneme conversion and TTS front-end work.
Kola Tubosun, Aanuoluwapo Aremu, Tolulope Ogunremi et al.
In briefOptimal control using learned critics when dynamics are incompletely modeled
SummaryThe paper proposes a learning-based model predictive control framework that augments an analytical dynamics model with a residual network to learn missing state-dependent effects. A learned action-value critic is added to the local iLQR optimization to incorporate long-horizon structure. The authors also develop a GPU-accelerated batched iLQR solver to handle many trajectory optimization problems in parallel. Experiments on biased and incompletely modeled control tasks report improved closed-loop control performance while keeping the model-based structure.
In briefFACET: A Method That Generates Realistic Protein Localization Microscopy Images More Efficiently
SummaryThis paper introduces FACET, a probabilistic generative framework that synthesizes fluorescence microscopy images from a protein's amino-acid sequence and the cell's morphological context, enabling prediction of localization for unimaged proteins. It models the sequence and context asymmetrically and uses a pretrained diffusion predictor for efficient stochastic transport. On held-out proteins, the authors report improved spatial overlap and lower FID than a backbone-matched baseline, with 75% fewer network evaluations.
In briefRethinking Tabular Foundation Models on Data Streams
SummaryThis paper systematically studies tabular foundation models (TFMs) on continuously arriving, evolving data streams, adapting through in-context learning rather than parameter updates. It examines memory management, computational cost, concept drift, and delayed labels. The authors find TFMs achieve the highest predictive performance, that keeping only the most recent examples works as well as existing memory techniques, and that they recover faster after drift. However, re-encoding a nearly unchanged context at each prediction makes serving costly, so they point to architectural efficiency as the path forward.
Nilesh Verma, Daniel Nowak-Assis, Afonso Louren\c{c}o et al.
In briefBetter Task Inference for Behavioral Foundation Models Beyond Least Squares
SummaryBehavioral Foundation Models solve downstream tasks without test-time policy learning by inferring a task vector from the reward function. The standard ordinary least squares method only minimizes reward reconstruction error and ignores reward ordering, which can make the zero-shot policy's successor measure diverge from the optimal one. The authors propose BLS, a test-time inference method that balances reward reconstruction error with successor-measure mismatch, and provide a suboptimality bound. BLS outperforms existing task inference baselines on locomotion, manipulation, and humanoid control benchmarks with negligible extra computation.
Kuan-Hsun Tu, Chien-Sheng Chiang, Hsin-Wei Chen et al.
In briefLearning conditional source distributions via flow reversal for temporal flow matching
SummaryThis paper introduces CNP-Flow, a flow matching framework for temporal generation. A conditional noise predictor produces an isotropic Gaussian source for each temporal condition, supervised by samples obtained by reversing a pretrained flow matching model. Across video prediction, video interpolation, and 7-DoF Franka robot motion planning, the method improves generation quality and matches baselines with fewer function evaluations.
In briefMemStrata: A Local Conversational Memory System Evaluated on Two Long-Term Memory Benchmarks
SummaryThis paper presents MemStrata, a conversational memory system that uses a local Qwen 3.8 27B model as its reader. It reports 475/500 (95.0%) on LongMemEval-S and 1,400/1,540 (90.91%) on LoCoMo under source-aware grading, where a judge checks references against the full source, and compares these with reference-only grading and several baselines. The authors state that superiority of source-aware grading over human adjudication is not established, and that development exposure and automated judging preclude independent replication or a leaderboard claim.
In briefIRSTD-Agent: Zoom-guided agent for detecting tiny infrared targets
SummaryThis paper proposes IRSTD-Agent, a framework that lets a multimodal large language model find infrared small targets through dynamic visual search. The model uses five visual tools, PROPOSAL, ZOOM, DETECT, DROP and REFINE, to inspect original-resolution images and localize targets step by step. It is trained with Zoom-guided Interaction Learning, which uses annotation-derived trajectories to supervise tool selection and arguments. Experiments on WideIRSTD-Full and IRSTD-1k show it outperforms the evaluated vision-language models.
In briefWILLIE: A Unified Benchmark for Wound Classification, Segmentation, and Localization
SummaryThis paper presents WILLIE, a unified framework and benchmark that evaluates wound classification, segmentation, and localization under one common protocol. The authors merge three public wound datasets and compare unified models at three scales against 10 single-task baselines. The best model produces all three outputs in one forward pass, reaching 91.88% classification accuracy, 91.41% Dice, and 96.23% AP@0.5. The authors also report that localization derived from segmentation outperforms dedicated detection baselines.
Gopi Trinadh Maddikunta, Shannan Hamlin, Hsin-Mei Chen et al.
In briefStudy of Semi-Markov Suboptimality in Hierarchical Reinforcement Learning
SummaryThis paper examines how committing to temporally extended subtask execution in hierarchical reinforcement learning causes suboptimality, separating execution losses from policy losses. It introduces a Unified Value Function and a four-stage Generalized Hierarchical Bellman Equation, with improvement results under bounded rewards and uniform termination. Option-based and goal-conditioned experiments show that changing execution and changing the learning target yield complementary gains, making execution design an explicit part of hierarchical policy optimization.
In briefGreen-Routed Neural Operators: Equations Decide Where the Network Reads
SummaryThe paper notes that many PDEs use transport fields to pick read coordinates, while neural operators usually treat them only as input values. It proposes the Green-Routed Neural Operator (GRNO), where the governing equation determines where latent features are sampled. Across five 2D and 3D PDE systems under 40-step autoregressive evaluation, GRNO achieves the lowest mean final relative L2 error on four and stays competitive on Keller-Segel.
In briefAgentDiscover: Autonomous Discovery with Minimal Search Scaffolding
SummaryThis paper introduces AgentDiscover, a framework in which a coding agent plans its own search for scientific discovery and records ideas and candidates in a database used as long-term memory. The authors report that it is more cost-efficient than existing frameworks and outperforms them on kernel engineering, biology, algorithm design, and mathematics tasks. They also state its programs would have placed first in seven past AtCoder heuristic contests.
WhyLetting a coding agent plan search with a database memory is a fresh idea, validated against baselines across kernel, biology, algorithm, and math tasks.
Mahdi Farahbakhsh, Ilan Sela, Fatemeh Doudi et al.
In briefDiffusion Transformers as Provably Optimal In-Context Generators
SummaryThis paper theoretically analyzes how a Diffusion Transformer pretrained on diverse tasks learns and generates predictive distributions for new queries from inference-time demonstrations. It shows that the appropriate generation target from finite demonstrations is a predictive distribution that preserves the remaining task uncertainty, not a single estimated task output. With sufficient pretraining and diffusion sampling steps, the model is proven to achieve the minimax optimal rate over a Hölder class of test-time tasks.
In briefCompact set-valued deep ensembles for multi-class classification
SummaryThis paper addresses training and storage costs in deep ensemble learning, and the robustness of set-valued predictions. It proposes compact ensembles, such as Bayesian neural networks and CNNs with Monte-Carlo dropout, to produce probabilistic predictions that define a representative distribution for each query. This distribution is used to derive Bayes-optimal predictions for various utilities, including a family of set-utilities intended to make predictions more robust than single labels. Empirical results are presented to illustrate the framework's potential advantages and disadvantages.
Kim-Dung Tran, Dang-Man Nguyen, Vu-Linh Nguyen et al.
In briefVAMPS: Learning Visual and Motor Robot Policies via Sampling-Based Planning
SummaryThe paper introduces VAMPS, a framework that uses Model Predictive Path Integral (MPPI) control to train robot policies without human demonstrations. In simulation, the policy and MPPI refine each other iteratively, and the learned locomotion policy is transferred to a Unitree Go2. On a Flexiv Rizon 10S, visuomotor policies are trained from autonomously collected real-robot data for pick-and-place and force-aware whiteboard erasing.
Mohamed Yassine Kabouri, Pietro Noah Crestaz, Quang-Nam Nguyen et al.
In briefCT-Miner: Fast, coarse-grained time-series pattern mining using Cartesian trees
SummaryThis paper presents CT-Miner, an algorithm that mines frequent time-series patterns with equivalent Cartesian tree structure, using a Cartesian suffix tree to share structural information. Cartesian tree equivalence keeps hierarchical ordering while ignoring exact values. The authors report reducing exhaustive CT-pattern occurrence collection from O(n^4) to O(n^2) time and prove correctness and complexity bounds. Across several time-series datasets, a small set of mined patterns preserves meaningful clustering structure.
In briefHow weight averaging during LLM training affects post-training quantization
SummaryLarge language models are usually pretrained in high precision and then deployed with low-precision post-training quantization. Prior work found weight averaging during pretraining can improve quantization performance over learning-rate decay, but the reason was unclear. This paper frames weight averaging as a trade-off between retaining training progress and robustness to perturbation, and derives a family of averaging kernels that reaches the Pareto frontier of the two goals. A theoretical framework shows coarser quantization is more sensitive to perturbations, and experiments support these predictions and the proposed strategy.
In briefQuantitative universality of approximate message passing under rank-one quadratic sensing, explained for general readers
SummaryThis paper studies how approximate message passing (AMP) algorithms behave when sensing matrices are built from rank-one quadratic terms. Over a fixed number of iterations, the authors show that these matrices and a covariance-matched isotropic Gaussian ensemble give normalized overlaps and smooth spectral statistics that differ by O(d^{-1/2}). This yields universality for normalized mean-squared error, spectral moments, and empirical spectral distributions. For prescribed smooth spectral denoisers, the result proves the AMP universality conjecture of Maillard et al. (arXiv:2408.03733).
WhyProves quantitative universality of AMP for rank-one quadratic sensing, confirming a conjecture; substantial theory advance with narrow application scope.
In briefShape analysis of the corpus callosum in Kendall space, examining aging and Alzheimer's disease.
SummaryThe paper proposes a Riemannian shape-space framework to analyze age-related changes in the midsagittal corpus callosum, using the OASIS-1 cohort. Contours are represented by 128 landmarks, and geodesic regression localizes age-related deformation, with the anterior regions (Genu and Anterior Body) accounting for most of the change. Geodesic regression outperforms a Euclidean linear benchmark in cognitively normal participants, while age explains less than 0.5% of shape variance in Alzheimer's disease, suggesting the disease disrupts the healthy aging trajectory.
In briefBeyond Refusal Patterns: Robust, Generalizable LLM Safety via Safe-Role Internalization
SummaryThe paper introduces SSRFT (Supervised Safe-Role Fine-Tuning), which frames safety alignment as internalizing a predefined safe role rather than learning explicit refusal patterns. It builds a Safe-Role Question-Answer dataset from psychometric questions, limited jailbreak prompts, and a role description, with synthesized and validated role-consistent responses. Experiments on multiple Base and Instruct models report greater robustness to prefilling attacks, better generalization to unseen jailbreak domains, and less over-refusal on benign queries than standard SFT, while preserving general capabilities.
In briefShadow Feature Refinement Network: Knowledge-distillation-based progressive feature refinement for shadow removal
SummaryThis paper proposes SFR-Net, a shadow removal network that combines supervised learning, a feature refinement loss, and knowledge distillation. A post-processing algorithm is added to restore natural color consistency in the shadow-free outputs. On the ISTD+ and SRD public datasets, it reports whole-image RMSE of 3.4627 and 4.3781 and SSIM of 0.9382 and 0.9341, respectively.
In briefCoDance: Teaching humanoid robots to dance reactively and compliantly with a human partner from video
SummaryCoDance is a framework that learns reactive and compliant human-humanoid interaction from a single video of two dancers. It converts motion demonstrations into force-aware training data by adapting the robot reference under structured forces at both hands. In simulation, the policies adjust footsteps to partner changes and reproduce about 80% of the wrist displacement in the augmented demonstrations. On a physical humanoid, it sustains two-hand dancing with a human partner, including repeated forward and backward transitions.
In briefWhen Longer Reasoning Helps: Forecasting Math Reasoning via Discovery and Execution
SummaryThis paper introduces a Discovery–Execution (DE) framework that uses short-budget attempts to forecast how mathematical reasoning success scales with more test-time compute. It is evaluated on four models using 35 fresh Olympiad problems and non-geometry problems from IMO-ProofBench Advanced. The authors report that measuring conditional execution improves held-out forecasts for one model, and a regularized variant yields lower regret for continue-or-restart decisions.
In briefRobust low-rank adaptation of LLMs on analog in-memory hardware
SummaryThis paper studies parameter-efficient fine-tuning of large language models on analog in-memory computing hardware. Pretrained weights stay fixed on analog arrays while only LoRA weights are trained to handle downstream tasks and hardware non-idealities. The authors propose input reshaping to reduce input-induced matrix-vector multiplication errors and update accumulation to preserve small updates before programming finite-state devices. Experiments on Llama-3.2-1B-Instruct and Llama-3-8B with Muon and AdamW show improved adaptation under noisy computation and finite-resolution programming.
In briefTransferable Adversarial Robustness for Speech Models via Layer-Wise Stabilization
SummaryThe paper proposes a way to make frozen speech foundation models robust without generating adversarial examples or updating the backbone for each downstream task. It stabilizes representations across hidden layers while keeping clean representations, then fixes the selected layer mixture and enlarges only the classifier margin. Tested on Wav2Vec2, HuBERT and WavLM Large across four tasks under adaptive 30 dB attacks, hierarchical robustification improves robust accuracy by 46.4 points and margin refinement adds 4.0 points.
Aref Mousavi, Shahab Sherafat, Kiarash Kiani Feriz et al.
In briefRubricArmor: Adversarial Evolution to Improve LLM-Generated Evaluation Rubrics
SummaryRubric-based reinforcement learning rewards language model responses using query-specific evaluation criteria, but rubrics generated directly by LLMs can be exploited through reward hacking. This paper proposes RubricArmor, an adversarial framework that alternates an attack step, which builds responses satisfying the rubric while failing the task, with a repair step that revises the rubric. The authors report that RubricArmor outperforms competitive rubric generation baselines and leads to more effective downstream rubric-based RL.
In briefMotion-based continuous sign language retrieval using gloss-guided alignment of 3D body movements
SummaryThis paper addresses aligning sign language with text by using 3D SMPL-X motion as the main representation instead of appearance-heavy RGB video. It uses gloss temporal spans as weak supervision to split continuous signing into segments and match them to words. The authors report state-of-the-art bidirectional retrieval on CSL-Daily and competitive results on PHOENIX-2014T.
In briefAnalytical study of compute-communication trade-offs among parallelism strategies for LLM inference
SummaryThis paper presents an analytical framework for distributed LLM inference under tensor, pipeline, and hybrid parallelism. It splits end-to-end latency into computation, inter-GPU communication, and pipeline bubble overhead, and models how each depends on hardware, model, and workload. Experiments on multi-GPU platforms validate the framework and show how prefill and decoding phases favor different strategies. The authors suggest it can guide parallelism selection and capacity planning for LLM serving systems.
In briefASCENT: Online test-time training of long-horizon AI agents by distilling verified experience into their weights
SummaryThis paper studies how a large language model agent can keep learning while deployed, updating its own weights from the trajectories it executes on a stream of related tasks. The proposed method, ASCENT, has a frozen copy of the model receive the verified trajectory as privileged information and distills its next-token predictions into persistent LoRA fast weights used on later tasks. The authors report that across ALFWorld, WebShop, and AppWorld, task success and interaction efficiency improve as experience accumulates, and the gains transfer to held-out scenes.
WhyProposes self-distilling verified trajectories into weights during deployment, with comparisons across three benchmarks; impact is mainly within agent training.
In briefMESH-Harness: Self-Improving Agent Harnesses Using Bandit-Guided Module Recombination
SummaryThe paper introduces MESH-Harness, a method for improving agent harnesses, the code that manages context, state, and tool calls for a language model, while keeping model weights fixed. It splits harnesses into role-specific modules and uses LinUCB-based bandit search with coordinate ascent to select and recombine module combinations for evaluation. Under matched evaluation budgets, it outperforms Meta-Harness on text, retrieval-augmented math, code generation, and interactive science tasks. The authors also report lower aggregate test-time cost than Meta-Harness.
In briefRevisiting ground-truth synthesis from high-speed video: validity conditions and audited smartphone clips
SummaryMotion-deblurring datasets are often built by averaging consecutive high-frame-rate frames and labelling the result with the central frame. The paper shows this label is unbiased only when the window is odd and frame durations are equal; an even window shifts every label by a fixed fraction of the blur length. Auditing 51 smartphone clips nominally recorded at 240 fps, the authors find 19 actually captured near 176 fps, recorded only in the container's timing tables. They provide a tool that checks both conditions without decoding and plan to release the clips and timing tables.
In briefFlexCast: Weather Forecasting Adaptable to Any Set of Atmospheric Variables
SummaryFlexCast is a weather forecasting model that uses one set of parameters to predict any chosen subset of 69 ERA5 atmospheric fields. It encodes each field's metadata, fuses the available fields into a fixed-width representation, and decodes forecasts for the requested fields. The model learns a six-hour increment applied recursively for longer lead times. Tests on the 2020 ERA5 set show that compatible cross-field context lowers forecast errors, while mismatched context raises them.
In briefReadable Role Signals Versus Effective Intervention in Indirect Prompt Injection
SummaryThis paper studies how LLM agents follow commands embedded in external data, a vulnerability called indirect prompt injection. The authors use counterfactual role probes, component-wise activation patching, and interventions on AgentDojo trajectories. They find that the role signal is readable, but intervention directions that change the next action transfer poorly across channels, and their effect depends on depth, position, and context.
In briefBossouChimpanzee: A Long-Term Video Dataset of Wild Chimpanzees
SummaryThis dataset is a video archive of wild chimpanzees recorded at an outdoor field laboratory in Bossou, Guinea, from 1988 to 2018, totaling over 1,200 hours of continuous footage. The paper describes the history of the experimental setup and archive, and releases an initial roughly 74-hour subset, BossouChimpanzee70h, covering 23 identified individuals for individual and action recognition. It is intended for ethology research and as a benchmark for machine learning on wild audiovisual data.
Daniel Schofield, Susana Carvalho, Vladimir Iashin et al.
In briefErased, Rerouted, or Rescaled? How Post-Training Changes What a Language Model Still Knows
SummaryThe paper asks what happens to information a pretrained language model already encodes when post-training no longer rewards using it. It separates three fates: erasure, rerouting away from decisions while still represented, and rescaling to lower variance. Using a reward-null kernel in controlled settings, the authors find that post-training mostly reroutes or rescales such information and keeps it decodable, with erasure appearing only under prolonged weight decay.
In briefCalibrated Evidence Inference With a Small Open Model
SummaryThe paper fine-tunes Qwen3-4B-Base to classify whether a randomized trial shows a significant increase, significant decrease, or no significant difference for a stated intervention and outcome. Training used only articles under CC BY, CC0, or public domain licenses, and the model is released under Apache 2.0. It reports an expected calibration error of 0.017 and a macro-F1 of 0.925, with a 2.6 percent error rate among its most confident 80 percent of answers.
In briefA smoothed gradient method for nonconvex federated bilevel optimization
SummaryThis paper proposes a stochastic doubly smoothed gradient method for nonconvex federated stochastic bilevel optimization, using first-order methods to avoid costly second-order Hessian and Jacobian computations. The method decouples upper- and lower-level learning rates and does not require a strongly convex lower-level loss. The authors prove a convergence rate of O(κ^{15/2}/ε^5) and communication complexity of O(κ^4/ε^3), with notably better dependence on the condition number κ than existing methods. Extensive experiments validate the algorithm's effectiveness.
In briefWhen to Rethink: Multi-Perspective Self-Verification for Vision-Language Models
SummaryVision-language models can give plausible but wrong answers, and this paper studies how verifier strength and prompt design affect self-verification, finding results highly sensitive to prompt choice. It proposes MOTIVE, which judges each candidate answer from several complementary perspectives and learns a reliability score aligned with correctness. The score decides whether to accept an answer or trigger history-guided rethinking. Experiments across multiple benchmarks and model backbones report that MOTIVE outperforms strong self-verification and self-correction baselines..
In briefMobile-4DGS: real-time static and dynamic Gaussian splatting on phones
SummaryThis paper presents Mobile-4DGS, a lightweight framework for real-time rendering of static and dynamic 3D Gaussian Splatting on mobile devices. It compresses appearance information into first-order spherical harmonics and prunes redundant Gaussians while maintaining multi-view consistency. For dynamic scenes, it uses a compact explicit 4D representation and reuses depth orders during playback to reduce computation. Experiments reportedly show substantially lower storage and rendering overhead with competitive visual quality.
In briefDAGO: Bandit-guided fusion of AI workflows to optimize agent designs under limited evaluation budgets
SummaryThis paper introduces DAGO, a framework for automated optimization of agentic workflows under a limited evaluation budget. It treats combinations of parent workflows as bandit arms, uses a LinUCB policy to choose which parents to fuse, and has an LLM generate child workflows whose validation scores serve as rewards. A shared directed acyclic graph tracks discovered workflows and their multi-parent lineage. On six benchmarks, DAGO reports the highest macro-average score among baselines, improving AFlow from 80.3 to 81.7 under matched budgets while cutting search expenditure by 11.2%.
In briefWhen agent memory goes stale: keeping tool observations consistent with changing files
SummaryAgents often reason over tool outputs such as file contents, but those sources can be edited later while the model keeps the old text in its context. The paper proposes Concord, a framework that links each observation to its source, detects changes, and updates, annotates, or suppresses stale context before reuse. On the authors' ConcordBench, where file contents become stale after edits, Concord gave answers consistent with the restored workspace state across three frontier models, using 46.4% fewer tokens than the strongest non-oracle baseline.
