In briefDynaFlow: A method for learning free-form multi-agent orchestration
SummaryThis paper introduces DynaFlow, in which a main agent generates task-specific executable workflows that coordinate sub-agents through sequential or parallel execution, information exchange, and verification. Only the main agent is trained, while sub-agents remain fixed. Across six benchmarks covering mathematical reasoning, scientific question answering, and multi-hop retrieval, DynaFlow reaches 88.08% overall accuracy in Forced mode, 10.90 percentage points above the strongest evaluated baseline with the same Qwen3.5-27B backbone.
In briefIntNet: A Communication-Driven Multi-Agent Learning Framework for Safer Cooperative Self-Driving
SummaryThis paper introduces IntNet, a multi-agent reinforcement learning framework that shares vehicle intent and schedules communication adaptively to improve safety in cooperative autonomous driving. It combines an observation prediction module, a Graph Attention Network, and a communication scheduler. Tested in complex urban settings against other communicative methods, it reports better coordination and fewer collisions while reducing communication effort.
In briefNeural modeling of electromagnetic response and design performance for free-form RFIC passives
SummaryThis paper proposes an end-to-end neural framework that predicts the electromagnetic response and downstream design performance of free-form passive matching networks in RFICs. It combines a frequency-conditioned neural simulator with a circuit-conditioned neural solver, and uses a dependence regularizer so the two response components capture complementary information. Trained on a large dataset of layouts built from expert contours and random paths, the 21M-parameter model reaches R² above 0.98 for almost all targets and evaluates orders of magnitude faster than traditional EM simulation.
In briefTask-relevance-driven prompt expansion for federated continual learning
SummaryThis paper proposes TRPE, a model for federated continual learning that does not require prior knowledge of new tasks' arrival times or category counts. It uses an out-of-distribution detector on local clients and a radius-adaptive DBSCAN on the proxy server to identify new tasks and their categories. Task relevance is measured by the affinity between gradients, and prompts from highly related old tasks are concatenated with new-task prompts for training, with a soft orthogonality constraint balancing learning and forgetting. Experiments across diverse datasets and backbones are reported to confirm its efficacy.
In briefFrom Explaining to Enabling: Actionable Transparency in Algorithmic Insurance Pricing
SummaryThis paper examines how to explain to policyholders the premiums set by machine learning models. The authors define three requirements for explanations (validity, actionability, and reliability), formalize them as axioms, and propose a method that uses a Conformal Prediction constraint to screen out recommendations the model is uncertain about. Applying four Counterfactual Explanation techniques to 57,832 Florida flood-insurance claims and 98,058 synthetic telematics policies, the authors find that constrained recommendations for high-risk policyholders are more reliable and linked to larger predicted risk reductions.
In briefToward Neuroadaptive Communication Interfaces: Personalized Speech and Language Support
SummaryThis paper proposes a framework that adapts how language is presented to a user's abilities and ongoing interaction, including sentence segmentation, pictograms, and optional spoken output. Behavioral feedback is the primary basis for adaptation, while EEG is treated as an additional, uncertainty-aware signal rather than a direct measure of comprehension. The authors outline a staged evaluation comparing fixed, behavior-adaptive, and neuroadaptive interfaces. The central question is whether neural information improves comprehension and communication efficiency beyond a strong behavioral baseline.
In briefEvolveMAS: A Unified Self-Evolution Framework for Decentralized Multi-Agent Systems
SummaryThis paper proposes EvolveMAS, a framework that lets decentralized multi-agent systems of language model agents learn from their interactions. Each main agent is paired with a MemoryAgent that manages structured memory, and adaptation happens both within episodes and across episodes through distilled knowledge. Across five benchmarks and six configurations covering cooperative, mixed-motive, and adversarial settings, it reports the highest normalized average score of 77.2 among compared methods. Ablations indicate that memory governance and knowledge reuse each contribute to performance, without updating model parameters.
In briefSpeculative Decoding for Mixture-of-Experts Models: How Expert Data and Kernel Speed Set the Best Draft Length
SummaryThe paper studies when speculative decoding is faster than normal decoding for mixture-of-experts (MoE) language models. It models each step's time as the bytes requested by each kernel divided by its effective bandwidth, plus fixed time, and derives the conditions for speedup and the best draft length. Measurements on llama.cpp Metal, MLX-LM and llama.cpp CUDA show that the deduplicating engine is faster only at 5 to 8 tokens per pass, or never. The predictions match the measured savings within their error bounds in most tested configurations.
In briefStochastic mirror descent for Markov-chain expectations with unbounded noise
SummaryThis paper studies stochastic mirror descent for minimizing a convex objective defined as an expectation under a Markov chain's stationary distribution, using noisy subgradients observed along a single sample path. The authors decompose Markovian noise via Poisson's equation and control it under regularity and moment conditions that allow unbounded subgradients and Poisson solutions. For the weighted-average iterate, they establish an O(T^{-1/2}) bound on the expected suboptimality gap with an appropriate step size, without requiring aperiodicity or marginal convergence.
In briefWhether Detector Boxes Can Safely Limit Small-Object Segmentation
SummaryThis study examines whether predicted bounding boxes are reliable for restricting where small-object segmentation is performed without causing false negatives outside those boxes. It uses adenovirus particles as small objects, with YOLO26x for detection, YOLO26x-Sem and Mask R-CNN for segmentation. Preliminary experiments show that restricting segmentation to detector-predicted boxes can exclude matched segmentation predictions outside them, potentially introducing false negatives.
In briefScaleRLT: exact log-probabilities and faster deferred rollouts for RL on looped Transformers
SummaryThe paper addresses reinforcement learning for looped language models such as Ouro, where a fast deferred rollout schedule produces log-probabilities that differ from a dense learner pass in floating point. The authors propose a hybrid log-probability whose value comes from a no-gradient replay of the rollout's own schedule and whose gradient comes from the dense learner. They report that the replay matches the rollout bit for bit across their tested cells and that deferral and CUDA graphs raise rollout throughput and shorten update time. They also report accuracy gains on held-out GSM8K over 100 updates.
In briefCTXAgent: Language Model Agents That Manage Their Own Context
SummaryThis paper studies letting the model itself manage its context, deciding when to compact history, what to keep, and what to write as a hand-off, instead of compacting at a fixed token threshold. The authors compare seven frontier models with and without these context tools, and compare their harness with three common harnesses on two models. Most strong models manage context well and reduce cost, and with Claude Opus 5.5 the self-managed harness matches Claude Code's task success at 34% and 49% lower cost. The agent, benchmark and all results are released.
WhySelf-managed context is a fresh angle; seven frontier models and multiple benchmarks with harness comparisons give solid evidence, with moderate practical impact on agent systems.
Matthias Busch, Marius Tacke, Roland C. Aydin et al.
In briefMigrating ML Projects from GPUs to TPUs with Coding Agents: Successes, Failures, and Costs
SummaryThe paper uses Claude Code to migrate ten PyTorch machine learning projects from NVIDIA GPUs to Google Cloud TPUs, then independently checks whether each one runs, preserves numerical behavior, and meets performance goals. Of 70 execution attempts, 53 failed (76%), and 26 of those failures raised no error. In one four-chip migration, the agent reported success even though a change in gradient clipping lowered accuracy from 86.53% to 84.89%.
Hui Sun, Anderson Uch\^oa, Wesley K. G. Assun\c{c}\~ao
In briefMulFCoder: A framework-aware multi-agent method for generating front-end code across React, Vue, and Angular from UI screenshots.
SummaryThe paper addresses generating runnable front-end code from UI screenshots for multiple frameworks, where prior MLLM-based work mostly targeted HTML/CSS. It proposes MulFCoder, a multi-agent system with Grounder, Planner, Writer, and Judger agents that encode framework-specific constraints and check updates without expensive builds. Experiments reportedly show higher compilation success rates and fewer framework-specific errors, with especially strong gains on Angular.
In briefAMBIT: generating and editing spatial audio (First-Order Ambisonics) from executable scene plans
SummaryThis paper presents AMBIT, a unified framework for generating and editing First-Order Ambisonics spatial audio. It uses executable ScenePlans that specify sources, their activity, and 3D trajectories, within a shared Transfusion model that combines autoregressive planning with rectified-flow rendering. The authors also add an FOA-aware continuous VAE and joint on-policy self-distillation training. Evaluations on speech, music, and sound report improvements over compared baselines in generation quality, editing fidelity, and preservation of unedited sources.
In briefGenerating Instruction-Specific Adapters to Improve Audio Model Generalization
SummaryThis paper extends Text-to-LoRA to large audio language models, using a Transformer hypernetwork to generate instruction-conditioned LoRA adapters instead of one fixed shared adapter. It proposes T2L-ICAG, a token-level method called in-context adapter generation. The approach was tested on two audio language models with a controlled suite of five task classes, and it outperformed static LoRA and the global embedding-based T2L baseline, with larger gains on structurally out-of-domain tasks.
In briefAdaptive Multi-Step Decoding for Time-Series Foundation Models
SummaryThis paper addresses the limitation of fixed decoding strategies in time-series forecasting, where one-shot prediction and autoregressive rollout each have drawbacks. It proposes a learned selector that chooses a rollout schedule for each forecast using hidden states of a frozen pretrained foundation model, with dynamic programming keeping candidate evaluation tractable. On the fev-bench benchmark with Chronos-2 and Moirai-1.1-R-small backbones, the authors report improved point-forecast accuracy (MASE) over one-shot decoding, and claim gains only in point accuracy.
In briefNo Cones, No Duals: A Fast, Scalable Solver for Non-Rigid Structure from Motion
SummaryThis paper addresses non-rigid structure from motion, which recovers camera motion and time-varying 3D shape from tracked 2D image points. It replaces costly semidefinite programming cone and ADMM solvers with a differentiable Euclidean correction, row-polar projection, and an unrolled spectral-proximal solver. Across sparse, dense, noisy, and long-sequence benchmarks, the method reports 7x to 1000x runtime speedups on dense sequences with 28,887 tracked points, with competitive reconstruction accuracy.
In briefA pipeline that structures arXiv superconductivity papers and answers questions over them with retrieval
SummaryThis paper presents an end-to-end pipeline that turns arXiv superconductivity publications into a searchable corpus with structured metadata and retrieval-augmented question answering. It extracts materials and superconducting transition temperatures with domain-aware methods and combines dense embeddings with structured fields in hybrid retrieval. On a manually verified gold set, domain-aware pair accuracy is 45.5% versus 9.1% for a regex baseline. On a 100-question benchmark, adding extracted metadata did not improve hybrid retrieval recall (39% versus 41%), and BM25 achieved the highest Recall@5 at 47%.
Ivan Iorsh, Aleksandr Kapitonov, Ivan Filianin et al.
In briefControlled block proximal linearized method for nonconvex nonsmooth optimization
SummaryThis paper studies multiblock nonconvex and nonsmooth optimization without assuming bounded iterates or coercivity of the original objective. The authors propose the Controlled Block Proximal Linearized (CBPL) method, which adds a control potential to block updates while keeping the original objective, variables, and constraints. Under explicit control conditions and the KL property, the method is shown to yield bounded iterates, finite length, whole-sequence convergence of the original variables, and convergence of objective values. The approach is applied to graph-regularized nonnegative matrix factorization, CP decomposition, and matrix tri-factorization with l0 sparsity, with experiments demonstrating convergence.
In briefFrom human intent to in-orbit action: a self-learning LLM agent for 6G satellite control
SummaryThis paper presents ALMOND, a framework for autonomous control of 6G low Earth orbit satellite networks that translates natural-language demands into reinforcement learning reward functions. It also mines implicit constraints from failed policy trajectories and uses an LLM to guide policy architecture search, while satellites run only distilled small policies without an onboard LLM. In a 350-satellite digital twin, polar handover success rose from 55% to over 95%, convergence was about three times faster, and reward engineering time fell from 48 to 4.5 hours.
In briefParetoRSI-Gym: Can AI Agents Improve Language Models Across Multiple Domains?
SummaryThe paper introduces ParetoRSI-Gym, an evaluation where AI agents search for and curate training data, train a language model, and iterate using per-domain results across five domains: math, code, finance, medicine, and science. Tested over three agent systems, three base models, and five rounds, the strongest system (Claude Code with Opus-5) raised its proxy-best checkpoint's generalization scores by 9.2% relative to the starting model. Single-domain relative gains reached up to 63%.
In briefOutcome-aware agents: does causal structure help forecast action consequences?
SummaryThis paper studies when causal structure helps a learned forward model predict what an agent's actions will cause, in synthetic structural causal models where true consequences are known. When the model's factors are the observed variables, the proposed AC-LSCM recovers the true graph exactly on all 10-variable graphs and has lower three-step prediction error than every baseline. With learned latent factors, however, the same model performs worse than the baselines. The authors conclude that factor choice matters more than the explicit do-operator or counterfactual data, and that every planner remains far from the oracle.please note
In briefAndroidFlux: Testing How Phone-Use AI Agents Recover from Mistakes
SummaryThe paper introduces AndroidFlux, a benchmark built on AndroidWorld that tests whether mobile-use multimodal agents can recover from errors. It restores recorded trajectories to checkpoints around the first recoverable error so different agents continue from the same device state. The authors also release a reward-modeling benchmark and an offline reinforcement learning dataset. They find that recovery ability only partially correlates with clean-start task success, and explicit reasoning helps only models trained with GUI-specific reinforcement learning.
In briefWhether or Where? Tail Calibration Reveals Reward Hacking in Finite-Budget Best-of-N Selection
SummaryThis paper studies a failure risk in Best-of-N selection, where picking the top proxy-reward candidate from N samples can hurt true quality as N grows. The authors show that BoN depends only on score ranks and can decline only when the tail of the calibration curve turns downward. Experiments on real candidate pools find tail failures are rare and occur in code generation. A held-out tail calibration statistic predicts where a budget cap helps.
WhyProves exact BoN value is a Beta-weighted integral of conditional true reward, with a tail-driven taxonomy; strong theory, with moderately sized empirical validation across several domains.
In briefComBeautify: Automatically writing standardized Git commit messages using code change summaries
SummaryThis paper presents ComBeautify, a method that automatically generates commit messages following the Conventional Commits Specification (CCS). It combines context-preserving summaries of Git diffs with large language models to improve commit classification, learning category definitions from historical commits rather than manual rules. The authors evaluated it on 13 languages, 6,500 open-source repositories, and over 20 million commits, finding rising CCS compliance over time. When trained on project-specific data, ComBeautify reportedly outperformed baselines by up to 3x.
Rishi Raj, Aashray Tandon, Nikolay Matyunin et al.
In briefSEEK: A Red-Teaming Agent That Extracts Protected Skills From Science Agents
SummaryScience agents often package domain expertise as reusable skills, which can leak through normal interactions. The authors built a 40-skill benchmark across eight platforms and found that all of them leak at least one protected skill, with four reaching 100% attack success on a single direct request. They propose SEEK, a black-box, self-evolving attack agent that uses only user-visible responses, and it achieves 100% attack success on 12 skills from two platforms. SEEK also extracted one of six skills under a confidentiality policy that blocked the benchmark protocols, suggesting defenses must be tested against adaptive multi-turn attackers.[]
In brief6G Reasoning Agent Societies: Composition, Scaling, and Consistency-Time Laws
SummaryThis paper studies how Large Language Model agents in future 6G networks, which exchange interpreted conclusions, can propagate hallucinations even over error-free transport. The authors model the agent society as a cellular sheaf and establish three laws: a composition law, a scaling law, and a consistency-time law. These are validated on simulated topologies and deployed LLM societies, providing worst-case latency ceilings for near-real-time operation.
Hatim Chergui, Carolina Fern\'{a}ndez-Mart\'{i}nez, Mustapha Benjillali et al.
In briefAn efficient non-interpolation virtual trellis method for analyzing gradient patterns in mammograms
SummaryThis paper proposes a non-interpolation virtual trellis-based (NVHT) operator for extracting features from mammogram images. The method applies a filter kernel directly to the square image without converting it to a hexagonal trellis. The authors state that it reduces computational and storage costs while improving feature extraction efficiency and accuracy for disease classification.
Shahid Rasheed, Muhammad Rashid, Seifedine Kadry et al.
In briefBeyond the Opening: How the First Token Keeps Shaping Reasoning Length
SummaryThis paper studies how the first reasoning token in reasoning models continues to affect the length of later reasoning. Replacing only the first token and resuming generation still substantially changes the remaining reasoning length. The authors find that this sensitivity is concentrated at a few generation positions, and blocking attention from those positions to the first token reduces the length gap.
