Speculative Macro Commit (SMC) is a runtime technique for tool‑using language‑model agents that separates an authoritative actor model from a faster speculative drafter model. The drafter predicts and executes future action chains on a snapshot, storing recurring multi‑action patterns in a macro library. When the actor’s next tool call aligns with a drafted action, SMC commits the pre‑executed steps, reducing latency by up to 18.59% on certain benchmarks while maintaining accuracy.
By Zeyu Liu, Souvik Kundu, Peter A. Beerel
The paper introduces LUGL (Local Updates, Global Learning), a framework that separates data collection from model fitting, allowing non‑incremental learners such as gradient‑boosted trees (LightGBM) to be used in reinforcement learning for games. LUGL alternates between a local update phase—where agents play self‑play games and store tabular updates—and a global learning phase—where a function approximator is trained on the accumulated table before it is reset. Experiments on both perfect‑information and imperfect‑information games show that LightGBM‑based agents perform competitively or better than neural‑network baselines like DQN and DeepCFR.
By David Milec, Spyridon Samothrakis, Michael Fairbank, Dennis J. N. J. Soemers
AgentRM proposes a generalizable reward model to guide LLM-based agents during test-time search, outperforming direct policy fine-tuning. Three reward modeling strategies—explicit, implicit, and LLM-as-a-judge—are explored, and AgentRM improves base policy performance by an average of 8.8 points across nine tasks, surpassing top general agents by 4.0 points. It also shows strong weak-to-strong generalization and can boost specialized agents, with plans to release code for further research.
By Yu Xia, Jingru Fan, Weize Chen, Siyu Yan, Xin Cong, Zhong Zhang, Yaxi Lu, Yankai Lin, Zhiyuan Liu, Maosong Sun
CORAL is a framework that enables autonomous multi‑agent evolution for open‑ended discovery, replacing rigid heuristics with long‑running agents that explore, reflect, and collaborate via shared memory and asynchronous execution. It incorporates safeguards such as isolated workspaces, evaluator separation, and resource management. In experiments across mathematical, algorithmic, and systems optimization tasks, CORAL achieves 3–10 times higher improvement rates with fewer evaluations than traditional evolutionary baselines, and improves the best known score on Anthropic’s kernel engineering task from 1363 to 1103 cycles.
By Ao Qu, Han Zheng, Zijian Zhou, Yihao Yan, Yihong Tang, Shao Yong Ong, Fenglu Hong, Kaichen Zhou, Chonghe Jiang, Minwei Kong, Jiacheng Zhu, Xuan Jiang, Sirui Li, Cathy Wu, Bryan Kian Hsiang Low, Jinhua Zhao, Paul Pu Liang
The paper evaluates CASCADE, a fully local layered defense for Model Context Protocol (MCP)-based systems, by conducting a component ablation and corpus audit on a fixed 5,000-sample dataset. It demonstrates that the choice of aggregation convention heavily influences reported metrics, that detection performance varies with provenance, and that the released configuration does not fully disclose the operating point. The study also shows that a local review model invoked for a third of requests does not alter classification outcomes, highlighting the importance of reproducibility and transparency in defense evaluations.
By \.Ipek Abas{\i}kele\c{s} Turgut, Edip G\"um\"u\c{s}
The paper introduces set difference captioning for autonomous driving datasets, aiming to generate natural‑language descriptions of differences between two image subsets. It adapts a two‑stage approach to focus on object‑centric patches, allowing attribution of differences to specific objects or categories. A new benchmark, AD‑Diff Bench, is presented to evaluate these methods, especially for sparse, real‑world differences, with open‑weight models to ensure reproducibility.
By Julian Truetsch, Felix Hauser, Christoph Stiller, Frank Bieder
The paper introduces a contrastive curriculum for teaching large language models to use a zoom‑in tool without supervised fine‑tuning, leveraging an InfoNCE‑style intrinsic reward that progressively hardens negative examples. Experiments on V*, HRBench, and MME‑RealWorld demonstrate competitive performance and efficiency, with the method even surpassing baselines when replacing traditional SFT. A new synthetic dataset, Muffin&Chihuahua (M&C), is created to directly evaluate zoom‑in capability, revealing that recall best correlates with task success.
By Falko Helm, Iryna Gurevych
The paper introduces LaPla, a Vision‑Language‑Action framework that uses a latent‑aligned planning approach to convert discrete semantic reasoning into continuous, physics‑constrained driving actions. It employs a residual VQ‑VAE to encode vehicle kinematics into a structured latent space, then projects multimodal inputs—images, past actions, and text—directly into this latent space, allowing a frozen decoder to generate physically plausible trajectories without quantization errors. Experiments on nuScenes and NVIDIA AlpaSim show LaPla reduces long‑horizon L2 error by 15.52% and improves closed‑loop success rates by 33.34 percentage points while cutting inference latency.
By Ruoyu Yao, Yusen Xie, Qingzhao Liu, Pei Liu, Zewei Yang, Yipeng Zhu, Xiaolong Wang, Jun Ma
The paper proposes a deployment‑focused framework for deadline‑constrained network control, introducing the Effective Congestion (EC) metric family and Uniform Path Grouping (UPG) heuristic to better capture traffic urgency and balance load. It integrates these with a Multi‑Agent Deep Reinforcement Learning architecture (MADRL EC (p*)) that combines a distributed scheduler and a centralized RL router. A unified training objective merges live‑reward, pre‑collected‑reward, and policy‑imitation terms, leading to the Model‑Guided Annealed Reinforcement Learning (MGA‑RL) protocol built on DDPG, which generalizes offline‑to‑online learning for demonstration‑driven training.
By Vincenzo Norman Vitale, Mohammad Solki, Antonia Maria Tulino, Andreas F. Molisch, Jaime Llorca
LightEMMA is a longitudinal evaluation framework that tests the autonomous driving performance of vision‑language models (VLMs) without fine‑tuning or prompt engineering. Using this protocol, the authors evaluated 15 VLMs from five major families on the nuScenes prediction benchmark and found that larger, more capable models do not consistently outperform earlier generations. The study identifies common failure modes such as overreliance on historical actions and difficulty reconciling conflicting visual cues, underscoring the need for domain‑specific adaptation to enhance VLM safety in autonomous driving.
By Zhijie Qiao, Haowei Li, Zhong Cao, Henry X. Liu
FlexMap is a vectorized high‑definition map construction framework that works with flexible camera configurations without needing calibrated rigs or explicit 2D‑to‑BEV transformations. It replaces geometric projection with a geometry foundation model that encodes cross‑view 3D structure, and uses a spatial‑temporal enhancement module and a camera‑aware decoder to separate spatial reasoning from temporal aggregation. Experiments on nuScenes and Argoverse 2 show that FlexMap outperforms pose‑dependent baselines and remains accurate even when camera views are missing or pose estimates are inaccurate.
By Run Wang, Chaoyi Zhou, Amir Salarpour, Xi Liu, Zhi-Qi Cheng, Feng Luo, Mert D. Pes\'e, Siyu Huang
The paper reviews how Large Language Models (LLMs) are being adapted for medical reasoning, moving beyond single-step answers to systems that can systematically, transparently, and verifiably reason. It introduces a taxonomy of enhancement techniques, split into training-time methods such as supervised fine‑tuning and reinforcement learning, and test-time methods like prompt engineering and multi‑agent systems. The review examines their application across text, image, and code modalities in key clinical areas—diagnosis, education, and treatment planning—and tracks the shift in evaluation benchmarks from simple accuracy to more nuanced assessments of reasoning quality and visual interpretability.
By Zizhan Ma, Wenxuan Wang, Meidan Ding, Shiyi Zheng, Shengyuan Liu, Jie Liu, Jiaming Ji, Linlin Shen, Yixuan Yuan, Wenting Chen
EmoDistill is an offline framework that distills emotional negotiation skills from large language model interactions into smaller agents. It separates emotion selection, handled by an Implicit Q‑Learning selector, from emotion‑conditioned expression, learned by a LoRA‑adapted 7B policy via supervised fine‑tuning and judge policy optimization. Experiments across four negotiation domains show that the full EmoDistill policy outperforms vanilla and IQL‑only baselines, while removing the explicit emotion channel markedly reduces negotiation utility and reveals partial, domain‑dependent transfer to unseen counterparties.
By Yunbo Long, Haolang Zhao, Lukas Beckenbauer, Liming Xu, Alexandra Brintrup
DuplexSpeechBench-IFEval (DSB-IFEval) is a new benchmark that evaluates how full‑duplex voice agents follow implicit instructions during real‑time spoken interaction. It contains 1,038 test cases across eight assistant roles and tests five conditioning protocols, measuring floor management with an Instruction Adherence Score (IAS) and persona consistency with a Persona Adherence Score (PAS). Experiments on six speech systems reveal architecture‑dependent trade‑offs, showing that some models are more sensitive to explicit versus persona‑only instructions and that even when following conflicting directives, they struggle to override them under safety conflict.
By Puneet Mathur, Dinesh Manocha
The paper introduces a governed approach to enterprise analytics in which a language model interprets user queries and a deterministic policy selects and runs pre‑approved analytical programs that return both results and evidence. The authors demonstrate that this restriction remains expressive for a defined analytical class—including relational operations, aggregation, comparison, windows, ranking, and similarity—while ensuring reproducibility through fixed meaning, policy, data, and execution rules. In experiments with 440 runs, three 8B models generated SQL and selected tools at runtime, whereas a policy‑executed analyzer achieved a perfect 110/110 match across all test datasets, though no runtime‑planning episodes matched the full answer‑and‑evidence contract.
"whyItMatters":"The study shows that a governed, policy‑driven framework can reliably produce accurate, reproducible analytics results, highlighting a viable path for controlled AI‑driven data analysis."
By MasterControl AI Lab
The paper introduces PlanFence, a dependency-scoped action‑validation protocol for distributed large language model (LLM) agent teams. PlanFence requires plans to cite the exact public records they rely on, and executors validate only those records that could affect the pending action, replanning or blocking if validation is incomplete. In 30 controlled live workflows, a freshness‑only executor always acted on obsolete plans, whereas PlanFence completed all tasks without invalid actions, demonstrating controlled safety and system‑cost benefits.
By Evan Chen, Shiqiang Wang, Christopher G. Brinton
Dude is a Dual-Detection Multi-Agent System designed to detect discrepancies between research papers and their accompanying code. It addresses limitations of single-agent LLM approaches by aligning granularity between paper and code through negotiation and a two-stage salience-filtering mechanism, reducing false positives. Experiments on real-world datasets show that Dude improves recall and precision by up to 22.8% and boosts the F1 score by up to 18.7% over baseline methods.
By Weijie Liu, Running Zhao, Wenhao Yuan, Jinfeng Xu, Zhanfeng Xu, Xiaoxi Zhang, Edith Cheuk-Han Ngai
The paper introduces the Necessary Tool‑Evidence Path (NTEP) annotation scheme and its associated reward mechanism (NTEP‑R) to better supervise vision‑language models that use external tools. By explicitly specifying which evidence is needed and penalizing redundant tool calls, the authors train an 8B‑parameter model that shows improved accuracy and tool‑use efficiency across seven image‑grounded benchmarks. The approach demonstrates that fine‑grained supervision of tool‑evidence paths is essential for robust agentic VLM performance.
By Xingming Long, Yu Liu, Zhiwei Yang, Hanqi Feng, Shaojie Zhang, Barnabas Poczos, Chao Jiang, Zhenbo Luo, Lei Jiang, Pei Fu
SENTINEL‑RL is an agentic SOC architecture that separates topological reasoning from semantic reasoning. It uses a heterogeneous graph attention encoder to compress a large authentication subgraph into a fixed‑dimensional state, a PPO policy to map that state to constrained investigative actions, and an LLM loop that only consumes policy recommendations and produces analyst‑readable narratives. Experiments on LANL and Indiana University datasets show fast graph ingestion, reliable alerting, high PPO performance, and a median 6.3‑second containment cycle.
By Uday Vallabhaneni, Cassie L. Cagwin, David J. Wild
The paper introduces AgentAuditor, a method that improves multi-agent large language model (LLM) reasoning by structuring agent outputs into a Reasoning Tree that captures agreements and divergences, rather than relying on simple majority voting. AgentAuditor resolves conflicts by comparing evidence at key divergence points, enabling efficient localized verification. The authors also propose Anti-Consensus Preference Optimization (ACPO) to train the adjudicator with evidence-verified supervision, reducing reliance on misleading majority cues. Across four MAS frameworks and multiple reasoning benchmarks, AgentAuditor consistently outperforms majority voting, achieving up to 5% absolute accuracy gains while remaining token‑efficient.
By Wei Yang, Shixuan Li, Heng Ping, Peiyu Zhang, Paul Bogdan, Jesse Thomason