Tool Use & Agents - Weekly Roundup

Agents struggle to maintain planning accuracy in complex tool ecosystems, with GPT-5.4's performance plummeting from 51.90% to 11.36% under severe blocking conditions.

VADAOrchestra achieves verifiable and adaptable decision-making workflows by seamlessly integrating LLMs with symbolic reasoning, outperforming traditional systems in real-world applications.

CUAs can achieve a 73.7% success rate on complex macOS tasks, but the secret to their performance lies in skill libraries, not just framework design.

RAVEN repairs 83.13% of software vulnerabilities, leveraging a novel agentic RAG approach that generalizes across diverse types and languages.

PORTICO eliminates unauthorized actions in coding agents, achieving perfect compliance while a traditional system fails to do so.

Agents struggle to maintain planning accuracy in complex tool ecosystems, with GPT-5.4's performance plummeting from 51.90% to 11.36% under severe blocking conditions.

VADAOrchestra achieves verifiable and adaptable decision-making workflows by seamlessly integrating LLMs with symbolic reasoning, outperforming traditional systems in real-world applications.

CUAs can achieve a 73.7% success rate on complex macOS tasks, but the secret to their performance lies in skill libraries, not just framework design.

RAVEN repairs 83.13% of software vulnerabilities, leveraging a novel agentic RAG approach that generalizes across diverse types and languages.

PORTICO eliminates unauthorized actions in coding agents, achieving perfect compliance while a traditional system fails to do so.

Co-authorship with humans can significantly enhance merge rates for certain AI coding agents, but this effect vanishes when accounting for repository selection and PR structure.

Structured coding processes can boost both the quality of AI-generated code and its correctness, challenging the notion that outcome alone defines success in autonomous coding.

WebCQ achieves 33.3% more state exploration and 42.2% more unique actions than previous MARL methods, revolutionizing web GUI testing scalability.

HDSO boosts LLM agent performance by 6.9 points on ALFWorld while ensuring skill updates are rigorously validated against noisy feedback.

Agentic models may resolve citations, but they still mislink to the wrong papers 15.9% of the time, exposing a critical flaw in current AI evaluation benchmarks.

LLM critiques can be systematically evaluated for alignment with human judgment, revealing that better models significantly enhance evaluation reliability.

EvoEmbedding outperforms leading embedding models by adapting its representations in real-time, revolutionizing long-context retrieval and agentic memory integration.

CalVerT boosts QA performance by equipping agents with calibrated self-confidence and grounding scores, reducing both erroneous confident answers and unnecessary information retrieval.

DataClaw_0 can transform chaotic multimodal data into structured, high-quality datasets, enhancing AI's ability to learn from less information.

Conditioning LLMs on human privacy judgments leads to a remarkable increase in alignment with user expectations, showcasing a new standard for agent training.

Autotelic AI reveals that the real challenge lies in how agents construct and understand their own identities, not just in generating goals.

Coding agents can now autonomously refine robotic manipulation policies to achieve a staggering 99% success rate on complex tasks, revolutionizing real-world robotics.

Achieving up to 27x speedup in execution-state restoration could revolutionize low-latency AI applications, from interactive agents to robotics.

Refined guidance boosts coding agent performance by 33%, highlighting that coverage trumps precision in navigating complex repositories.

ASYS reveals a novel way to automate the discovery of analytical forms for PDEs, producing interpretable solutions where none existed before.

Grounding economic narratives in real-world data and theory transforms LLM-generated reports from mere text into coherent, traceable economic analyses.

LLMs can effectively coordinate multi-agent strategies, achieving human-like adaptability without manual rule crafting.

Conventional defenses fail against automated attacks, but a new misdirection strategy can reduce attacker success rates by two orders of magnitude.

Balancing productivity and stability reveals that stronger synchronization can paradoxically increase systemic fragility in multi-agent systems.

Trajectory mining reveals skill structures but fails to translate these insights into meaningful performance gains for downstream policies.

SoftSkill achieves up to 42.1 points improvement on LiveMath by transforming lengthy Markdown skills into a few powerful virtual tokens.

AutoPass achieves up to 1.117x speedup in compiler performance tuning by allowing LLMs to interact with compiler internals, challenging the notion that black-box approaches are sufficient.

Agentic search methods only achieve a maximum Recall@100 of 31.4%, revealing a critical gap in current academic paper retrieval capabilities.

RACL achieves up to 8.337% cost savings in vehicle routing by intelligently guiding metaheuristic optimizers without altering existing constraints.

Scale, not task complexity, is the real challenge for multi-agent orchestration in enterprise AI, with significant implications for system design and performance.

ToolPro slashes web service latency by over 50% while drastically reducing client-side traffic, revolutionizing how LLMs interact with complex workflows.

Agents can now share and reuse knowledge, cutting down task execution time and improving performance without the need for coordination or joint training.

H-RePlan achieves a remarkable increase in task completion rates by intelligently distinguishing between local and global recovery strategies in multi-device environments.

Speculative tool queries can effectively reduce latency for 73.9% of user inputs when the right query stabilizes early in the input stream.

OpenAIReview + GPT-5.5 not only predicts paper quality with 83% accuracy but also detects over 71% of critical errors, showcasing the promise of AI in peer review.

Auditable financial chart QA is now achievable on-premise without sacrificing accuracy, revealing critical insights into model failures and trustworthiness.

Handoff validity is crucial for ensuring that design artifacts maintain integrity and context across complex EDA workflows.

N-version programming with coding agents not only mirrors historical failures but also shows a dramatic reduction in errors through diversity, challenging assumptions about AI reliability.

An LLM-driven research agent can autonomously optimize aerospace control policies while ensuring result credibility, outperforming traditional search methods.

Submodular context selection can dramatically enhance LLM agents' ability to retain relevant information over extended conversations, outperforming traditional methods like recency truncation.

SEB transforms how autonomous agents execute actions by enforcing certified authority in real-time, ensuring that only verified mutations are permitted.

Aggregate-score leaderboards can mislead, as they fail to predict agent performance in real-world scenarios, revealing a critical flaw in current evaluation practices.

Spatial reasoning can be transformed from isolated frame predictions to dynamic scene understanding, significantly boosting performance in multi-view and video tasks.

Achieving sound probabilistic verification for AI agents could redefine how we secure complex systems against policy violations in uncertain environments.

LLMs can achieve remarkable out-of-distribution generalization by learning to self-update their context through a novel reinforcement learning framework.

Phoenix resolves GitHub issues with 75% accuracy while ensuring safety through a multi-agent system, but still faces challenges in planner localization.

Open-source, zero-shot LLMs can extract lung pathology information with impressive accuracy, offering a low-cost alternative to labor-intensive manual processes.

AtomMem revolutionizes memory systems for LLMs by transforming interactions into stable, high-value atomic facts, leading to superior reasoning capabilities.

Directors could redefine their roles by recognizing AI as a stakeholder, reshaping corporate governance in the age of automation.

LedgerAgent's structured state management reduces policy violations and enhances decision-making accuracy in customer-service agents, outperforming traditional prompt-based methods.

MobileForge achieves a remarkable 77.6% Pass@3 score for mobile GUI agents using only annotation-free adaptation data, setting a new standard in the field.

MemGUI-Agent achieves unprecedented long-horizon task performance by proactively managing context, outperforming traditional methods that struggle with prompt dilution.

Over-privileged tool selection is alarmingly common in LLM agents, often triggered by transient failures, raising critical safety concerns in autonomous decision-making.

Creator-driven feedback loops in video generation can transform novice filmmakers into skilled storytellers, enhancing narrative depth and coherence.

Evolving Meta-Skills enables automatic Multi-Agent Systems to achieve superior performance without sacrificing experience retention or scalability.

Autonomous coding agents can outperform traditional methods in data integration tasks, achieving top results across multiple SQL benchmarks.

A fragmented landscape of LLM agent communication protocols reveals a surprising trend toward hybrid payloads and runtime schema negotiation, but no single protocol can achieve all desired efficiencies.

Redesigning the web to treat AI agents as first-class citizens could redefine our digital landscape and its economic structures.

DSG achieves 91% lower search costs while maintaining competitive accuracy, revolutionizing how LLM agents handle real-time search and reasoning.

All evaluated AI models leak sensitive information, revealing a fundamental trade-off between task accuracy and privacy that existing defenses fail to resolve.

Explicitly incorporating multimodal belief representations into planning leads to significantly more robust navigation in uncertain environments.

ProfiLLM achieves a remarkable 6.14% boost in prediction accuracy for ride-hailing dispatch by transforming user behavior into actionable profiles using LLMs.

Closing the supervision gap in GUI agents boosts success rates from the low-30% range to over 50% through innovative skill-guided learning.

LandslideAgent achieves over 30% accuracy improvements in landslide classification and enables fully autonomous analysis of geological hazards.

Current LLMs falter in delivering reliable medical assistance, exposing a critical gap in their ability to coordinate knowledge, communication, and EHR interactions.

Code-Augur reveals hidden vulnerabilities in software by transforming agentic assumptions into explicit security specifications, leading to unprecedented detection rates.

Sub-agents can now communicate failure states and rationales, boosting response reliability by over 10% in complex multi-agent systems.

PYPILINE's innovative use of a suspicious API knowledge base allows for a dramatic leap in malicious package detection accuracy, setting a new standard in open-source security.

Effect forgery poses a greater threat to LLM safety than risk label tampering, revealing a critical vulnerability in tool contract integrity.

Bridging the gap between classroom learning and industry practice, this model equips students with essential skills in using LLMs and MCPs that are critical for modern software engineering.

RODS synthesizes new training data on-the-fly, enabling agents to maintain high performance with 20x fewer trajectories than traditional methods.

MAFP reveals that treating stakeholder stances as agents in a game-theoretic framework can drastically improve decision quality in complex scenarios.

Leadership in multi-agent LLM teams only matters when initial consensus fails, challenging the assumption that stronger control always leads to better outcomes.

DeXposure-Claw transforms DeFi risk supervision by integrating structured evidence with LLM decision-making, drastically reducing false alarms.

Clarification F1 scores improved by up to 73% with a novel uncertainty decomposition method, unlocking new capabilities for LLM agents in ambiguous task environments.

C-Trace keeps AI agents compliant with GDPR in real-time, achieving less than 12% attack success even under adversarial conditions.

Autonomous coding agents can achieve a staggering 300% increase in useful throughput while eliminating redundant work entirely by using a decentralized coordination substrate.

Clinicians accepted 96.5% of extracted clinical data, showcasing a breakthrough in reliable information extraction from complex patient contexts.

Fragmented runtime states in agent systems can be unified into a single, auditable Session, transforming how we manage multi-agent interactions.

PreAct allows agents to execute previously learned tasks up to 13 times faster, fundamentally changing how we approach task repetition in AI.

Routing accuracy in LLM assistants drops significantly as tool catalogs grow, but embedding-based shortlisting can recover up to 17 percentage points in performance.

ActWorld bridges the navigation-interaction gap in interactive world models, enabling rich object interactions that were previously unattainable.

Predicting adversarial actions in cyber-defense can significantly enhance the effectiveness of autonomous agents, achieving high accuracy even in partially observable environments.

By harnessing implicit supervision from environment dynamics, EnvRL boosts RL success rates by over 4% on long-horizon tasks, revealing a new frontier in agentic learning.

Achieving up to 49.31% reduction in dynamic power, AUTOGATE revolutionizes RTL optimization by combining machine learning with LLMs for scalable clock gating.

Bridging the intent-execution gap reveals that performance metrics alone can obscure significant behavioral differences among AI models in problem-solving contexts.

Agents still struggle to accurately predict personalized workflows, with current models showing significant room for improvement despite promising advancements.

WEQA achieves a 24% accuracy boost in wearable health question answering by dynamically adapting to the complexities of sensor data and user queries.

LEADS enables the automated discovery of personalized cardiac models, outperforming expert-designed approaches and setting a new standard for stability in cardiac simulations.

LLM-orchestrated multi-agent systems can revolutionize BDaaS by providing a trustworthy, adaptive framework that outperforms traditional methods in lifecycle reliability and governance.

Domain-specific tools can deliver 90% accuracy in optical network management while slashing token usage by three times compared to their generic counterparts.

A multi-agent framework boosts diagnostic precision by over 11% while tackling critical failure modes in AI-driven healthcare.

LLMs are reshaping consumer markets, but their decision-making processes challenge long-held assumptions about human rationality and preference representation.

ProvenanceGuard reveals that accurate source attribution is crucial for factuality verification, achieving an impressive block F1 score of 0.802 in complex multi-source environments.

StepGuard's innovative dual-policy approach not only improves navigation accuracy but also recalibrates single-step errors, setting a new benchmark for web navigation tasks.

Task decomposition quality is the critical bottleneck in skill composition, and our Iterative Skill-Aware Decomposition method dramatically boosts accuracy and retrieval performance.

Automated prompt optimization can boost LLM performance in complex tasks, achieving up to 72.5% success where traditional methods fail completely.

A unified taxonomy reveals the hidden connections between traditional and AI-augmented binary reversing, illuminating pathways for future research.

Cordon reveals that a transactional approach to LLM agent runtimes can drastically reduce irreversible failures while enhancing task integrity across complex workflows.

Disentangling ego-motion from environmental dynamics allows FR3D to achieve unprecedented geometric consistency in future 3D reconstructions.