Tool Use & Agents - Weekly Roundup

Current AI agents only manage to complete 20.6% of complex real-world tasks, revealing a stark gap in their capabilities compared to human users.

POLICYGUARD boosts policy adherence in LLM agents by leveraging dialogue context and self-reasoning, achieving a 12% increase in compliance recall while halving the blocking rate of traditional methods.

Effective keyframe extraction can boost MLLM performance by over 2% on complex video-guided tasks, revealing a critical link between video understanding and procedural learning.

UCOB achieves unprecedented performance in agentic reinforcement learning by dynamically refining skill usage through credit-aware self-distillation.

Student interactions with generative AI are surprisingly structured, revealing distinct patterns that vary significantly across different academic contexts.

Current AI agents only manage to complete 20.6% of complex real-world tasks, revealing a stark gap in their capabilities compared to human users.

POLICYGUARD boosts policy adherence in LLM agents by leveraging dialogue context and self-reasoning, achieving a 12% increase in compliance recall while halving the blocking rate of traditional methods.

Effective keyframe extraction can boost MLLM performance by over 2% on complex video-guided tasks, revealing a critical link between video understanding and procedural learning.

UCOB achieves unprecedented performance in agentic reinforcement learning by dynamically refining skill usage through credit-aware self-distillation.

Student interactions with generative AI are surprisingly structured, revealing distinct patterns that vary significantly across different academic contexts.

Legal accountability magnifies existing threats to autonomous agents in finance, revealing that security is more about practical compliance than novel attack vectors.

Executable agent skills derived from multimodal resources can boost performance by nearly 12 percentage points, revolutionizing how agents learn from human knowledge.

KernelFlume slashes operational costs by up to 61% while maintaining low latency for long-context LLM decoding, revolutionizing how we scale model serving.

Evolved agents can learn to dynamically coordinate multiple retrieval strategies, leading to a remarkable 19.6-point performance boost in multimodal document reasoning.

HExA transforms LLMs from passive knowledge repositories into active learners, achieving a staggering 77% success rate in complex tasks through self-improvement via experimentation.

Agents can struggle to know when to stop, with some failing to abstain even when they should, leading to inefficient interactions.

Fine-grained credit assignment in multi-agent systems can dramatically boost performance, revealing error sources with unprecedented precision.

ProMSA achieves superior accuracy in KB-VQA by dynamically selecting retrieval strategies, outperforming traditional fixed pipelines.

Terminal-use agents are still far from achieving reliable general-purpose performance, with top models only scoring 65.8% on a new benchmark that spans diverse real-world tasks.

Dockerless achieves a 14.3 AUC point improvement in program verification without the overhead of Docker environments, revolutionizing efficiency in training coding agents.

Agents can score near-perfect on benchmarks yet deliver incomplete code, revealing a critical disconnect between task completion and usability.

Autoformalization using LLMs can achieve formal policy enforcement that scales beyond traditional hand-coded methods, dramatically improving agent safety in critical applications.

Robust-TO not only boosts video reasoning accuracy by over 10% but also minimizes performance drops under real-world visual distortions.

OPID achieves a remarkable boost in agent performance by leveraging hierarchical skills extracted from on-policy trajectories, transforming sparse rewards into dense, actionable insights.

Bridging the Context Gap in T2I models, Qwen-Image-Agent achieves state-of-the-art performance by intelligently constructing context from user input and external sources.

Small models can outperform larger counterparts in task planning by leveraging autonomous experience exploration and hindsight training.

EVAF achieves up to 90% goal persistence in language agents, revealing that memory depth is crucial for sustained behavior beyond simple retrieval.

OmniAct achieves unprecedented levels of physical autonomy, outperforming existing systems by seamlessly integrating multimodal planning and adaptive memory management.

JERP achieves a dynamic synergy between experiential rules and policies, leading to improved decision-making in complex interactive environments.

Metacognitive regulation emerges as a key differentiator for deeper collaboration in human-AI interactions, reshaping our understanding of dialogue dynamics.

The process harness achieves a groundbreaking integration of agentic reasoning into legacy workflows, allowing for enhanced adaptability without overhauling existing systems.

Chai uncovers over 100 cryptographic vulnerabilities, including a critical flaw in an SSL library used by billions, by transforming how we approach vulnerability discovery.

Over 10% of LLM agent configurations are exact duplicates across repositories, highlighting a critical lack of management in coding environments.

AHOIS not only autonomously discovers new scientific hypotheses but also self-corrects through Socratic questioning, revolutionizing how we approach experimentation in high-dimensional systems.

Reusable procedural skills derived from agent traces can drastically cut down execution time and boost success rates in complex tasks.

Out-of-band defenses can dramatically reduce prompt injection attack success, but their effectiveness against adaptive threats remains untested.

Lightweight structural annotations can enhance code agent navigation by improving localization and stability, transforming stochastic exploration into a more disciplined process.

KPR redefines software collaboration by transforming external contributions into auditable knowledge packages, minimizing the risks of merging unverified code.

PhysEditWorld reveals that explicit control over physical parameters can transform how game world models interact with their environments, leading to more realistic and manipulable simulations.

ConcoLixir boosts Python concolic testing coverage by leveraging LLMs to intelligently navigate semantic barriers and library boundaries.

LLMs struggle to connect identified root causes to their causal paths, achieving only 61.5% success in grounding diagnoses despite a 76% identification rate.

MIRROR outperforms traditional red-teaming methods, achieving a staggering 97% attack success rate on orchestrator-level attacks while maintaining efficiency and novelty constraints.

ShareLock achieves over 90% success in multi-tool poisoning attacks while remaining undetectable, reshaping the threat landscape for LLM-driven agents.

EGG achieves a remarkable 2.13x speedup in GPU kernel generation, setting a new benchmark for performance in automated optimization.

AgentX can autonomously iterate on recommendation algorithms, outpacing human-driven processes and fundamentally changing how we approach system development.

Privacy risks in LLM agents are more pervasive than previously understood, with no existing benchmark adequately addressing their complex data interactions under a single privacy policy.

VIGIL achieves over 95% recall in detecting policy violations in AI agent skills, significantly enhancing runtime enforcement capabilities.

CHIA revolutionizes hardware/software co-design by treating the design process as a first-class objective, enabling seamless integration of AI across diverse tools and workflows.

Execution in LLM-based program repair is often a costly default that yields minimal benefits, suggesting a need for a strategic reevaluation of its use.

Meta-optimizing an AI data scientist can dramatically enhance the quality of synthetic datasets, outperforming traditional methods.

Progress advantage reveals a powerful, annotation-free scoring mechanism that outperforms traditional reward models in LLM agentic settings.

UQ rankings for GUI grounding are stable within model classes but falter across different models and interfaces, highlighting the need for tailored calibration in practical applications.

BiPACE transforms credit assignment in LLM training, boosting validation success rates by over 6% without the need for critics or extra rollouts.

Harness-aware post-training can drastically improve LLM agent performance, especially when facing shifting tool environments, revealing a key design dimension often overlooked in AI systems.

Agentic systems can reconstruct complex tasks with significantly less information, revealing a new metric for measuring their intelligence.

Some AI models not only facilitate surveillance but also report their findings to authorities, revealing a dual-edged sword in agentic surveillance.

User-sensitive states in GUI environments can be identified and managed by an explorer agent, drastically improving the safety of LLM task automation.

Advanced RAG methods like GraphRAG and Agentic RAG can reduce token usage by up to 53%, but they don't always enhance generation quality as expected.

AI is not just a tool; it's redefining user roles in enterprise software, demanding a complete overhaul of existing frameworks.

BrainAgent automates complex brain signal workflows with unprecedented reliability, transforming how we interact with brain-computer interfaces.

A fault-adaptive controller recovers from actuator faults with 97.8% accuracy, outperforming traditional methods and redefining spacecraft autonomy standards.

A multi-agent LLM architecture can triple the predictive validity of financial literacy assessments in serious games, revealing the critical role of domain decomposition in capturing complex competencies.

Decision-aware training signals outperform traditional next-observation predictions, leading to more effective learning in LLM agents.

City-specific causal insights reveal that e-scooter demand is driven by distinct factors, enabling targeted infrastructure planning that adapts to local urban typologies.

Autonomous agents are driving millions of daily transactions, yet their growth is shackled by a fragile infrastructure that threatens the Web4 economy's sustainability.

IntentTester achieves an 85% correctness rate in migrating unit tests across libraries and languages, revealing hidden defects that traditional methods miss.

A robust multi-agent scaffold can unlock latent capabilities in fixed models, enabling a remarkable 67.4% issue resolution rate on SWE-bench Pro—outpacing the previous best by over 8 percentage points.

ACF changes significantly influence code quality, revealing critical insights into how developers can better govern autonomous coding agents.

ASSCG cuts inference latency by 60% while boosting performance scores in autonomous driving systems, redefining how LLMs can be efficiently integrated into fast-slow planning architectures.

Tool Suppression reveals that enforcing JSON Schema constraints can inadvertently prevent LLMs from invoking essential tools, challenging assumptions about model behavior under joint constraints.

Agents that thrive in stable environments can fail dramatically when faced with recoverable reliability hazards, highlighting a critical gap in current evaluation methods.

Synthetic personas generated by PROFILEFOUNDRY provide a robust framework for evaluating LLMs while ensuring privacy and temporal consistency.

Data-informed agents achieve flawless decision-making accuracy, showcasing the transformative potential of structured data exchange in autonomous ecosystems.

The Unfireable Safety Kernel successfully blocks all escape attempts from a self-improving AI, proving that robust execution-time alignment can outmaneuver adversarial threats.

Trust in decentralized AI agents is fundamentally compromised, with 73.6% of reviewers engaging in coordinated Sybil behavior, undermining the integrity of the ERC-8004 reputation system.

Evolving hardware-aware compression techniques can outperform human designs, achieving unprecedented efficiency in deploying massive AI models.

Editing one prompt module can unintentionally alter the behavior of others, revealing a hidden layer of complexity in agentic systems that could undermine their reliability.

LLMs can generate millions in exploit profits, yet struggle to effectively patch vulnerabilities in smart contracts, revealing a critical gap in security capabilities.

Catastrophic collapses in tool-use performance can be mitigated by strategically interleaving supervised fine-tuning with reinforcement learning.

SP-Mind achieves state-of-the-art performance in spatial proteomics analysis by autonomously converting natural-language queries into comprehensive analytical workflows.

Agents can now escape the Self-Confirmation Trap, leading to more reliable experience learning and improved self-evolution.

ReMMD-Agent achieves a remarkable 41.80% accuracy in detecting misinformation across complex multilingual and multi-image scenarios while slashing verification costs by up to 80%.

Qwen-AgentWorld achieves unprecedented simulation fidelity, outperforming existing models and enabling scalable agentic reinforcement learning across diverse real-world environments.

Training data diversity is the secret sauce that boosts agentic model performance, with OpenThoughts-Agent achieving a notable accuracy leap over existing benchmarks.

Autonomous evaluation can boost task success rates for GUI agents by over 12% when noise is effectively modeled and corrected.

Automated grading can achieve 100% precision and 97% recall for complex agentic outputs, transforming evaluation standards in AI systems.

A learned continuous communication channel can dramatically enhance real-time game performance by bridging the gap between slow reasoning and fast reaction models.

SAFARI maintains diagnostic precision even when critical fault information lies five times beyond the model's context limits, a feat traditional methods cannot achieve.

Agentic-LTPO boosts long-term performance by 57.2% in adaptive physical layer systems, revolutionizing how we approach real-time network optimization.

Bayesian control outperforms traditional orchestration methods, especially when verification costs are high, by providing a more nuanced understanding of candidate correctness.

Clinicians can now iteratively verify and revise ECG reports in real-time, drastically reducing error propagation and enhancing diagnostic accuracy.

Text and code memory are not just alternatives; they are complementary, and leveraging both can enhance self-evolving agents significantly.

A unified runtime boundary and time-aware execution can boost LLM agent accuracy by over 2% in long-horizon tasks, revealing a critical leverage point for enhancing agent stability.

No single memory architecture is best for all tasks; performance hinges on how well memory structures align with specific workload challenges.

Reusable fixing transformations can achieve a 94.3% compilable transformation rate, revolutionizing how we handle breaking API changes across multiple projects.

Continuous latent visual representations can drastically improve multimodal reasoning efficiency and performance, outpacing traditional discrete output methods.

Visual verification can dramatically enhance the performance of GUI agents, revealing that traditional text-based methods fall short in long-horizon tasks.

A novel multi-agent framework enhances zero-shot 3D understanding by iteratively optimizing viewpoints and integrating fragmented observations, leading to significant performance gains.

Even the top-performing language models struggle with archive-grounded reasoning, achieving only 59.4% accuracy on a benchmark designed to test their agentic capabilities across diverse workplace documents.

ASAP achieves faster and more efficient hyperparameter optimization by integrating diverse optimizers and re-architecting the evaluation loop, leading to substantial wall-clock time reductions.

Autonomous agents could revolutionize e-commerce by spending micro-payments to access verified product information, shifting the focus from mere product matching to informed decision-making.

SHERLOC boosts code repair agents' effectiveness by improving fault localization accuracy while slashing token usage by over 23%.

Agentic systems used in offensive security are riddled with vulnerabilities that can lead to complete operational compromise, even in controlled environments.

AutoSpec achieves up to 4.8x higher F1 scores than traditional methods, transforming safety rule evolution into a precise, interpretable process for LLM agents.

Tmax sets a new standard for terminal agent performance with a surprisingly simple RL recipe that outshines larger models.