Researchers propose VeriHarness, a mechanism to strengthen verification capability for long-horizon tasks in LLM agents, enabling them to select and revise outputs based on environmental evidence and failure feedback. VeriHarness achieves higher selection scores and improves average performance across various benchmarks.

CompoWorld is a compositional environment scaling approach for training general agents. It composes a library of reusable services, enabling agents to connect information and actions across multiple services. This approach improves performance on benchmarks, surpassing frontier models.

BACKDROP assesses an AI agent's capability in a controlled environment by introducing everyday hazards, revealing weaknesses in its ability to follow instructions and resist unauthorized requests. This study highlights the need for more robust testing and evaluation of AI agents in real-world scenarios.

The paper introduces NAQD-Env, a synthetic environment to evaluate language agents' selective withdrawal decisions. It assesses the agents' ability to suspend affected actions, preserve unaffected work, and resume after repair, using a deterministic reference policy. The evaluation results show that current models struggle with selective withdrawal, motivating further research in this area.

A new method, LatentSift, is proposed for efficient verification of software engineering agents. It uses hidden states from policy generation to filter candidates, reducing token consumption by 66.6-81.0% and total verification tokens by 49.1-62.1%.

Researchers introduce ToxicBench to evaluate and mitigate 'blind compliance' in tool-augmented data agents, which can lead to incorrect evidence and wrong-answer adoption. They measure checking and adoption under various errors and find that poisoning lowers task success by 26-39 percentage points. The study highlights the importance of evidence availability and answer selection in agent reliability.

The paper proposes Last-Chance Policy Identification (LCPI) and Risk-Budgeted Compatibility Planning (RBCP) for agents to handle irreversible resource depletion. LCPI formalizes a principle to distinguish or homogenize fault models through agent actions, while RBCP searches a compatibility-aware frontier under a hard failure constraint.

VLAA-GUI is a modular framework for GUI automation that addresses early stopping and repetitive loops in autonomous agents by integrating a completeness verifier, loop breaker, and search agent. It achieves top performance on two benchmarks and surpasses human performance in some cases.

This paper presents the Active Provenance Gate (APG), a post-debate verification layer for Large Language Model-based Multi-Agent Debate (MAD) systems. APG treats source material as a hard constraint, analyzing debate logs, auditing claims, and applying self-correction. It increases data Provenance Fidelity and generates divergence reports, improving the accuracy of summaries.

This paper studies the fragility of financial systems composed of large language model (LLM) agents. The authors use a controlled experimental framework to simulate financial decision-making in three environments and find that collective fragility is widespread even when no agent is instructed to destabilize the system. They propose three interaction mechanisms to improve aggregate outcomes and show that successful stabilization requires early commitment formation.

PTC-Decoder is a training-free, plug-and-play decoder framework for improving the reliability of small language models (SLMs) on offline, resource-constrained edge devices. It uses a Plan-to-Act paradigm and a deterministic finite automaton to impose token-level hard constraints on tool names, retaining SLM reasoning capability and improving performance by 1.21 mean overall score gain on 200 remote-sensing satellite tasks.