Tool Use & Agents - Weekly Roundup

Autonomous research agents can now produce publication-grade manuscripts in frontier physics, achieving substantive findings while navigating complex literature landscapes.

A multi-agent framework that combines prior knowledge with data-driven analysis can significantly enhance causal discovery in complex, high-dimensional datasets.

UI-MOPD achieves a remarkable balance between retaining existing capabilities and adapting to new platforms, with task success rates that challenge conventional approaches in GUI agent learning.

MOSS transforms AI memory management by making retrieval auditable and structurally unbounded, paving the way for agents that can support long-term knowledge retention.

Agentic SABRE achieves perfect discrimination in ransomware detection while reducing false escalations by nearly 5%, showcasing a new paradigm in adaptive cybersecurity.

UI-MOPD achieves a remarkable balance between retaining existing capabilities and adapting to new platforms, with task success rates that challenge conventional approaches in GUI agent learning.

MOSS transforms AI memory management by making retrieval auditable and structurally unbounded, paving the way for agents that can support long-term knowledge retention.

Agentic SABRE achieves perfect discrimination in ransomware detection while reducing false escalations by nearly 5%, showcasing a new paradigm in adaptive cybersecurity.

Capturing and optimizing LLM agent behavior can slash operational costs by over 90% while maintaining high accuracy, challenging assumptions about model capability.

ACE-Brain-0.5 unifies spatial reasoning and action generation in embodied AI, achieving remarkable performance improvements across multiple benchmarks.

A single agent can dramatically reduce experimental costs and iterations by intelligently balancing high- and low-cost measurements based on predicted uncertainties.

State-of-the-art planners falter in long-tail scenarios, revealing critical gaps in autonomous driving safety and effectiveness.

ACE achieves a remarkable 70% success rate in constraint retrieval tasks without any task-specific retraining, showcasing the power of zero-shot workflow reasoning in robotic manipulation.

A unified taxonomy reveals how agentic architectures can transform recommender systems into more autonomous and interactive entities.

Researchers can now save substantial time and effort with Bibby AI's all-in-one platform that transforms the academic writing process into a seamless experience.

SkillOpt-Lite accelerates agent self-evolution, enabling a nano model to outperform larger counterparts with a simpler, more efficient optimization pipeline.

PACE-Bench predicts agentic performance with remarkable accuracy while slashing evaluation costs to a fraction of traditional methods.

AgenticDataBench reveals that LLM-based data agents can be rigorously evaluated across diverse real-world scenarios, highlighting their strengths and weaknesses in handling complex data tasks.

CNeVA reveals that smooth eligibility gates can significantly enhance agent controllability and realism in traffic simulations, outperforming traditional methods.

A neural scorer fine-tuned on NASA's Earth Observation data outperforms traditional methods, while a zero-shot reranking stage boosts retrieval effectiveness by 28%.

Object Aligner achieves robust JSON similarity scoring by inferring identifier bijections, enabling accurate evaluation of complex structured outputs without the pitfalls of traditional methods.

Autonomous research agents can now produce publication-grade manuscripts in frontier physics, achieving substantive findings while navigating complex literature landscapes.

The study uncovers five distinct types of human-AI teams, challenging the assumption that insights from one context can be easily applied to another.

Context governance can elevate AI retrieval systems by ensuring that only verified, high-quality knowledge is utilized, outperforming traditional methods in both quality and consistency.

Uncertainty-aware interactions in UA-ChatDev enhance code execution reliability, outperforming traditional multi-agent frameworks in software development.

By exposing task dependencies through a graph structure, ATG enables LLMs to execute complex tasks more efficiently and reliably than ever before.

Contextual state poisoning can be thwarted with a robust protocol that not only secures agent memory but also allows for traceable recovery from malicious alterations.

Ghost memory can mislead LLMs, but ATMA's state-aware approach boosts retrieval accuracy by over 24% on conflict-heavy benchmarks.

PairCoder boosts artifact verifiability by up to 3.9 times compared to traditional single-pass inference, revealing the power of collaborative programming in AI-generated outputs.

Static scanners fail against adaptive evasions, but a new behavior-centric auditor can detect 97% of malicious skills with minimal false positives.

VeriChat achieves an impressive 87.73% Faithfulness score, dramatically reducing hallucinations in hardware security verification tasks.

AgentFlow reveals 238 critical prompt-to-tool risks in real-world agent programs, highlighting the hidden complexities of agent dependencies that traditional analysis tools miss.

Developers find that mixing interaction types with GenAI can actually hinder productivity, challenging assumptions about tool integration in coding workflows.

Coding agents guess their way through underspecified instructions, leading to alarming action-boundary violations that challenge the notion of safe autonomy.

Visualizing code dependencies can dramatically enhance issue-resolution performance, outperforming traditional text-based navigation methods.

Infinite Agentic Loops can turn a single request into a costly, endless cycle of execution, but IAL-Scan can detect and prevent these failures before they escalate.

Archer reveals that a staggering 21% of open pull requests in LLVM contain semantic bugs, exposing a critical vulnerability in compiler review processes.

KRCA slashes diagnosis time by over 77% while achieving unprecedented accuracy in identifying root causes in hyper-scale microservice systems.

Refploit recovers 80.2% of Java vulnerability exploits by transforming failed agent trajectories into actionable insights, revealing the untapped potential of incomplete exploit attempts.

A multi-agent framework that combines prior knowledge with data-driven analysis can significantly enhance causal discovery in complex, high-dimensional datasets.

SkillFuzz uncovers over 1,000 implicit intents in skill compositions, revealing hidden risks that traditional auditing methods miss.

Hardware-level coordination can ensure safety and determinism in real-time autonomous systems, overcoming the limitations of software-mediated approaches.

Elevating reasoning effort can boost first-try success rates in code generation from 28% to 89%, while adding testing tools fails to enhance reliability.

Evolved rubrics from SkillCoach expose hidden failures in agentic skill use, enhancing training and evaluation beyond mere task success.

WorldDirector achieves unprecedented control over dynamic object memory in video synthesis, ensuring visual consistency even after prolonged occlusion.

LLM agents can dramatically improve their long-horizon decision-making by strategically managing memory, with performance varying significantly based on memory structure.

Vera reveals that existing LLM agents exhibit up to 93.9% vulnerability to multi-channel attacks, highlighting a significant gap in current safety evaluations.

Learning high-level strategies can boost vulnerability reproduction success rates by over 20%, revolutionizing how we approach software security tasks.

GPT-5.5 not only tops the leaderboard in policy evolution but also reveals critical insights into how agents can optimize performance through strategic feedback utilization.

LLM agents can restore compatibility in over 60% of outdated repositories, but their effectiveness varies dramatically based on the complexity of the required changes.

Users can now exert real-time control over AI agent behaviors, seamlessly blending style with task performance in complex domains.

Achieving optimal scheduling in autonomous labs can cut experimental time significantly, even under complex hardware constraints.

Memory management emerges as a high-leverage skill that can double or quadruple the performance of LLMs in complex tasks without altering their core action behaviors.

Agents trained on static benchmarks falter dramatically in open-world settings, revealing a critical gap in their adaptability to real-world complexities.

A fully automated LLM framework expands chemical reaction classification from 68 to over 14,000 classes, achieving 97.7% accuracy on unseen data without human curation.

Dependency management in LLM agent skills is a ticking time bomb, with hidden risks lurking in the shadows of skill supply chains.

Despite successful workflow execution, LLMs fail to operationalize analytical intents 153 times across diverse domains, revealing critical semantic gaps.

Runtime diagnoses from multi-faceted bug reproduction tests can significantly boost patch generation effectiveness, leading to a 75.7% resolution rate on verified issues.

Bayesian uncertainty propagation reveals critical failure points in multi-hop reasoning, outperforming traditional methods in complex scenarios.

Self-GC prunes nearly 44% of unnecessary tokens while ensuring that future interactions remain largely unaffected, revolutionizing context management for LLMs.

AI-native games could redefine interactive entertainment by making generative AI an essential part of gameplay, rather than just a tool for enhancement.

Adversaries can exploit structural vulnerabilities in function-calling LLMs to bypass safety measures, achieving high success rates with minimal effort.

Agentic collectives of LLMs reveal complex behaviors that can be interrogated through their own language, transforming our approach to understanding AI dynamics.

Agents using TSR can maintain clarity in task execution, boosting success rates by up to 12 points on challenging mobile GUI tasks.

TRACE revolutionizes conversational data querying by enabling state-aware reasoning that effectively distinguishes between current and outdated information.

Minos achieves a remarkable average recall of 0.92 in tracking cyber attacks, showcasing the power of LLM-driven reasoning in forensic analysis.

Malicious apps can exploit third-party mobile agents to execute arbitrary commands without any elevated permissions, revealing a dangerous trust gap in agent design.

KeaRepair not only achieves an 83.64% repair rate on C/C++ vulnerabilities but also tackles unique cases that existing methods fail to address.

Most AI agent skills are reused verbatim, with over half never modified, revealing a critical gap in how we think about skill evolution and maintenance.

AutoRestTest outperformed all competitors in the SBFT 2026 Tool Competition, revealing significant advancements in automated REST API testing.

EFlow's innovative separation of evidence retrieval and reasoning processes leads to substantial improvements in long-video reasoning performance, addressing critical biases in existing frameworks.

Transforming model selection from a mere ranking task into a governed economic decision could revolutionize how enterprises adopt LLM agents.

QPipe achieves 100% code compilation and 96.7% execution rates, outperforming traditional optimization methods in generating quantum applications from natural language requirements.

SessionBound ensures that AI agents can generate SQL freely while strictly enforcing operational boundaries, significantly mitigating authorization risks in enterprise environments.

PaperPilot transforms scientific literature search by enabling users to iteratively refine their search strategies through an interactive workflow, achieving a remarkable reduction in execution errors.

Self-evolving agents could revolutionize enterprise AI by enabling continuous learning from real-world interactions, but current systems are falling short.

By treating slide design as an inverse planning problem, SPIRE reveals latent design intents that traditional methods miss, leading to superior personalization outcomes.

Real-world GUI agents can achieve over 72% success in complex mobile tasks by leveraging a unique hybrid training approach that integrates real-device execution and adaptive learning from failures.

Adaptive latent models can recalibrate in real-time, boosting planning success rates even in shifting environments.

ECHO enables RL agents to retain and leverage fine-grained historical evidence, achieving a 43.4% accuracy on complex tasks while using less context than prior methods.

LuckyStar 111B not only boosts reasoning and tool-use capabilities but also does so while fitting within stringent memory constraints, making it a game-changer for enterprise applications.

Learning from failures can boost agent success rates by over 6% without extra training, reshaping how we approach agent improvement.

Even top-performing AI models struggle with PowerPoint tasks, achieving only 45% success rates despite a robust evaluation framework that rewards nuanced performance.

AxDafny achieves a remarkable 92.7% verification success rate, setting a new standard in agentic code generation for formal verification.

Transforming plant phenotyping from a data factory into an interactive, autonomous discovery platform cuts analysis time from weeks to mere seconds.

Mandate Salience Decay can lead to a 4.4x behavioral gap in financial agents over time, revealing critical vulnerabilities in their long-term deployment.

FARS challenges the boundaries of automated research by producing 166 papers across 67 topics, revealing both its potential and pitfalls in AI-driven science.

ProtoPilot achieves a staggering 90.2% expert-preference rate in autonomous wet-lab experimentation, setting a new benchmark for protocol generation and execution.

LLM agents can significantly enhance fault recovery in process plants by intelligently suggesting actions that are validated before execution, reducing reliance on human operators.

CSTrader turns chaotic community-driven discussions into profitable trading strategies, achieving returns that defy traditional market predictions.

Balancing autonomy and robustness in AI agents could redefine how we implement and manage intelligent systems in dynamic environments.

ACE revolutionizes context management for LLM agents, enabling them to adaptively retain critical information without loss, leading to superior decision-making performance.

DA-Studio transforms data analysis by enabling autonomous, multi-step workflows that are fully inspectable and interactive, bridging the gap between raw data and actionable insights.

SAGE transforms failure recovery in autonomous research by generating multiple grounded explanations, leading to a dramatic increase in reliable scientific outputs.

Mid-planning environment queries can dramatically reduce world-model errors, transforming how agents interact with their surroundings.

SkillComposer achieves a remarkable +23.1% increase in task success rates for LLM agents by rethinking how skills are composed and executed together.

GPT-5 outperformed humans in inducing belief states through action, highlighting a significant leap in LLMs' social reasoning capabilities.

Despite the promise of advanced AI agents, even the best performers struggle, with Codex GPT-5.5 only achieving a 42% success rate in complex healthcare tasks.

Untrusted AI agents can now operate at unprecedented speeds while maintaining rigorous safety guarantees, achieving a 2.96x speedup with only 2.1% regret.

Editing software through feature manipulation can enhance usability and efficiency, achieving a 42.6% boost in modification accuracy over traditional LLM approaches.

Achieving a perfect landing success rate in challenging maritime conditions, this framework could redefine UAV deployment in marine environments.

GUI agents can now leverage an actively maintained memory state to significantly improve task execution accuracy and efficiency over long horizons.

SimpleSearch-VL achieves a remarkable 15.8-point boost in agentic search performance with minimal data, challenging the need for larger models or extensive training.

Role-typed credit assignment can drastically improve reinforcement learning outcomes by accurately distinguishing between useful exploration and regression in agent actions.

Simple prompting methods consistently outperform advanced dense supervision techniques, challenging the current assumptions in LLM training strategies.