ACE is a self‑correcting agentic canvas editor that operates on a hierarchical scene‑graph rather than flat document formats, enabling reliable multi‑slide presentation automation. It pairs a presentation‑specialized action space of 98 tools with CARE, a content‑aware router that reduces input tokens by about 89%, and a ground‑truth‑free instruction‑following judge that feeds natural‑language critiques back into the agent for self‑correction. In benchmarks, ACE outperforms a comparable agentic HTML pipeline on instruction following (4.23 vs. 3.81), runs 1.75× faster, costs 44% less, and is preferred by 58.7% of blind raters, with 81% favoring the self‑corrected output.
By JooYoung Jang, Taegyeong Lee, Jihyeon Park, Nojun Kwak
Paritok-4B is a 4‑billion‑parameter LoRA compressor designed for coding agents, which extracts and retains key spans of code rather than paraphrasing them. It is intent‑conditioned, selecting lines that are most relevant to the agent’s current task, and achieves high fidelity with 96% of identifiers, paths, and numbers preserved. Trained on 67,074 real OpenHands trajectories and fine‑tuned on Qwen3‑4B, it compresses agent context to about 25.7% of its original size while keeping 86.5% of the uncompressed solve quality across 300 SWE‑bench Lite instances.
By Jiayu Shi, Luzhuo Chen
MetaRAG introduces a belief-action aligned policy optimization framework for agentic retrieval-augmented generation (RAG). It incorporates Verify-first Action Generation and Internal Belief Probing to assess whether the current evidence is sufficient before taking an action, and uses a consistency reward gated by answer correctness to guide training. Experiments on seven public QA benchmarks demonstrate that MetaRAG improves the accuracy-efficiency trade-off over existing RL-based agentic RAG baselines, with benefits that transfer across research settings, optimizers, and model backbones.
By Qiuyi Qi, Tian Liang, Jiamu Wang, Jinjian Zhang, Wei Zhou, Pengcheng Zhu, Linjian Mo, Ming Kong, Jie Liu, Qiang Zhu
STRIVE is a new framework for longitudinal radiology report generation that separates clinical reasoning into distinct Diagnosis, Attribute, and Temporal Change agents, each producing explicit evidence. The Temporal Change agent is refined with a Progression-Aware GRPO reward that differentiates direction-preserving errors from reversals. Verification occurs twice: a Consistency Gate aligns agent outputs before report generation, and a Validation Agent ensures the final report is supported by the aggregated evidence. On the Longitudinal-MIMIC dataset, STRIVE achieves superior clinical efficacy and more than doubles Longitudinal Change Concordance compared to the strongest baseline.
By Junyeong Maeng, Eunsong Kang, Heung-Il Suk
SA-Bench is a diagnostic benchmark that evaluates how well large language model (LLM) agents reproduce scientific papers by measuring semantic alignment between generated code and the papers’ specifications. It covers 30 recent machine‑learning papers, decomposing each into 1,491 Semantic Alignment Units (SAUs) that assess numerical, methodological, protocol, and ordering fidelity. Across 12 generator configurations, the best model (Claude+PaperCoder) scores only 0.301 on average, highlighting that current LLMs often implement requirements incorrectly or with stubs.
By Xue Hu, Zewei Pan, Zeli Su, Zhou Liu, Wentao Zhang
ReproAgent is a four‑stage pipeline—Prepare, Plan, Generate, Repair—that uses a persistent implementation contract to guide scientific AI agents in converting research papers into executable code repositories. The system employs two channels: an implementation‑requirement channel that translates paper snippets into code obligations, and a reference‑evidence channel that pulls content and structure from related repositories. Evaluated on PaperBench Code‑Dev, ReproAgent achieves the highest mean score among same‑backbone scaffolds for both Claude‑Sonnet‑4.5 and Gemini‑3‑Flash, with ablation studies confirming the contribution of both channels.
By Xue Hu, Zewei Pan, Zhongyuan Wang, Zhou Liu, Zeli Su, Wentao Zhang
The paper investigates the cost–quality trade-offs involved when coding agents switch between low‑cost, low‑capability (LC) and high‑cost, high‑capability (HC) language models during long‑running tasks. By experimenting with different handoff directions, timings, and interfaces—full‑trajectory transfer, compaction, and trajectory removal—the authors find that full‑trajectory escalation recovers less than half of the LC‑to‑HC quality gap while adding significant cost, a penalty they call the handoff tax. Conversely, downshifting from HC to LC yields a more favorable cost‑quality balance, and the optimal interface depends on the direction of the handoff.
whyItMatters":"The study quantifies how model handoffs impact both performance and expense, offering guidance for designing more efficient coding agents that balance cost and quality."
By Roy Ganz, Mor Shpigel Nacson, Adi Kalyanpur, Ron Litman
The paper introduces Adaptive Influence Graphs (AIGs), a two‑stage framework that first converts a failed trace into a structured graph and then navigates it to pinpoint the critical error in multi‑agent large language model systems. Experiments across multiple models demonstrate that richer trace representations and adaptive graph construction improve failure attribution, with AIGs achieving state‑of‑the‑art results on the Who&When benchmark. The study shows that both the diagnosing model and the way traces are represented and explored are crucial for accurate failure attribution.
By Yarden Bakish, Amir Dudai, Roy Ganz, Oren Nuriel, Elad Ben Avraham, Mor Shpigel Nacson, Ron Litman
The paper presents a method for designing state‑aware transmission protocols that guide how information and resources are shared among agents in a collective discovery task. Using LLM‑guided evolutionary search, the authors evolve protocols that outperform existing baselines by up to 37%, and show that the advantage stems from conditioning on content and agent states rather than just network topology. The evolved protocols also generalize across different domains and agent populations, indicating that such protocols can be discovered in silico and may inform AI‑assisted coordination systems for human collective intelligence.
By C\'edric Colas, J\'er\'emy Perez, Eleni Nisioti, Akhilesh Mocherla, Pierre-Yves Oudeyer, Cl\'ement Moulin-Frier, Maxime Derex
The paper introduces SMITH, a reinforcement learning framework that jointly trains large language models to create and use tools within a single policy. By alternating between build and use tasks and employing separate reward signals for schema, code, and outcome failures, SMITH enables a 4B Qwen3 model to achieve state‑of‑the‑art accuracy on procedural reasoning benchmarks, outperforming larger untrained models and improving performance on downstream tasks when its tools are applied.
By Zhi Rui Tam, Chieh-Yen Lin, Yun-Nung Chen, Shao-Hua Sun, Hung-yi Lee
Meta$^n$ is a recursive self‑improvement framework for large language models that keeps a fixed meta‑operation Ω and repeatedly applies it to its own outputs, creating deeper layers that reason from higher perspectives. By avoiding changes to the meta‑operation, the system remains stable while the input grows, allowing depth to emerge through convergence and evolutionary search. Experiments on two backbone models show Meta$^n$ surpasses prior self‑improving agents across eight benchmark families, notably achieving positive scores on the ARC‑AGI‑2 benchmark designed to resist skill memorization.
By Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa, Dongyeop Kang
The paper introduces AtlasNav, a persistent multi‑view corpus‑navigation framework that organizes a corpus into a Corpus Atlas, enabling large‑language‑model agents to navigate efficiently under finite interaction budgets. AtlasNav reduces online inference cost by 30.21% and achieves 92.05% strict accuracy on BrowseComp‑Plus, while earlier and more rapidly realizing required evidence compared to dynamic‑workspace methods. The approach also transfers well to other corpora such as PhantomWiki and heterogeneous enterprise knowledge bases, demonstrating that effective agentic search relies on both accessible evidence and a reusable corpus representation.
By Hongyu Guo, Zhiyu Zheng, Zhao Cao
StepGuard introduces a step-level guard model that audits and checks tool actions before execution, addressing security risks in LLM-based agents. It is trained using StepGen, an automatic engine that generates safe and unsafe trajectories, and employs Balance-GRPO to dynamically balance learning between safe and unsafe actions. Experiments show StepGuard achieves high accuracy comparable to GPT-5.4 and significantly reduces attack success rates while minimally impacting utility.
By Zhijie Zheng, Yu Li, Chen Qian, Yuqian Fu, Yanwei Fu, Lu Sheng, Jing Shao, Dongrui Liu
StarHarness is a framework that evolves environment‑specific agent harnesses while keeping model weights fixed. It builds a compact evolution pool by stratifying tasks based on baseline failure behavior, separating proposer‑visible search tasks from hidden selection tasks, and reserving held‑out tasks for generalization evaluation. In enterprise benchmarks such as ITBench SRE, EnterpriseOps‑Gym ITSM, and AutomationBench Finance, harness evolution boosts full‑benchmark performance by 20‑35 percentage points after 4‑12 accepted changes per environment, with gains that persist on unseen tasks and transfer across GPT and Qwen model families.
By Esakkivel Esakkiraja, Denis Akhiyarov, Vikas Yadav, Sai Rajeswar, Patrice Bechard, Sridhar Nemala, Sagar Davasam
The paper introduces Recuris, a recursive Experiential‑Working Memory architecture that lets long‑horizon agents track task progress and select skills based on current needs rather than full history. By coupling working memory with experiential memory, execution becomes structured evidence that localizes failures to specific memory components, enabling a bounded recursive memory‑evolution loop. Across four benchmarks and ten models, Recuris improves task success in 35 of 37 model‑benchmark pairs, raising state‑of‑the‑art performance on tau‑bench and SkillFlow and reducing common long‑horizon failures by up to 80%.
By Zhaochen Yu, Yingcheng Wu, Zhenfei Yin, Kaiyuan Chen, Zhe Zhao, Mengdi Wang, Shuicheng Yan, Ling Yang
REFINE is a tool-agnostic, evidence-aware multi-agent approach that generates Java file-level refactoring candidates by combining static analysis, smell-informed planning, LLM-based transformation, and automated re-analysis. In experiments on 450 Java files from 15 open-source systems, REFINE reduced detected code smells by 68–73% across three LLM configurations, achieving higher median reductions with smaller edits compared to a direct-prompt baseline. However, the tool’s outputs still pose risks such as assert/fail-call changes and public-method removal, requiring compilation, testing, dependency analysis, and human review before deployment.
By Muhammad Waseem, Aakash Ahmad, Pekka Abrahamsson
The study investigates why small language model agents tend to repeat a tool call that just failed. By recording the failed call and its error message in the transcript, the authors measure a negative corrective gain—agents are more likely to repeat the failed action, with a drop of about 1.03 nats per token. The problem is traced to the harness design rather than the model’s understanding of errors, and the authors show that replacing the verbatim call with a runtime-generated description of the failure can reduce this backfiring effect by 76%.
By Esmail Gumaan
The paper introduces a new resilience evaluation framework for Embodied Agents Systems (EAS), defining metrics for Rebound, Stability, and Graceful Extensibility. It demonstrates that traditional outcome-centric metrics obscure dynamic behaviors, while the proposed metrics reveal process-level distinctions such as recovery costs and instability across 400 household tasks with 10 EAS. Optimizing with these metrics reduces recovery costs and improves stability and extensibility, highlighting a trade‑off among resilience characteristics that should guide deployment-specific system design.
By Yapeng Liu, Yuanzhao Zhai, Xudong Gong, Dawei Feng, Bo Ding, Lin Wang, Huaimin Wang
The paper presents a systematic security analysis of Google’s Agent Payments Protocol (AP2) version 2.0, focusing on its roles, transaction lifecycle, and deployment architectures. It identifies 48 threats across five attack families, scores them with the AIVSS, and demonstrates eight high‑risk threats with proof‑of‑concept attacks and mitigations. The study also introduces a deployment‑aware scanner to map threats to various checks, showing that signed mandates alone cannot guarantee user intent when pre‑authorization context is manipulated.
By Avital Aviv, Parth A. Gandh, Ron Bitton, Asaf Shabtai
Hybrid Semantic Tool Discovery for Enterprise MCP Gateway presents SCOUT, a system that addresses two major challenges in large language model (LLM) agent tool usage: a context‑engineering bottleneck and a tool discoverability barrier. SCOUT reframes tool exposure as a context‑selection problem, injecting only relevant tools into the model’s context window and providing two MCP meta‑tools—tool_search and execute_tool—to perform hybrid retrieval via BM25 and dense vector search. In production at PayPal, SCOUT cuts MCP tool‑token consumption by 99%, dramatically reducing per‑query inference cost while remaining model‑agnostic and requiring no client‑side changes.
By Olympia Saha, Amy Wang, Srinivasan Manoharan