Researchers propose a benchmark (NARCBench) and probing techniques for detecting collusion between AI agents in multi-agent systems, achieving high detection rates in various scenarios.

A benchmarking framework, AstroAgentBench, is introduced for evaluating agentic planning in space mission planning tasks. It assesses the performance of large language model (LLM) agents in domains like scheduling, observation planning, and constellation design. The framework provides a standardized evaluation methodology and highlights the importance of task-contract formulation, verifier feedback, and search adaptation in achieving high-quality plans.

Researchers propose a post-training method for multi-token prediction heads in language models, achieving similar speedup to joint pre-training with significantly fewer tokens. They also introduce a relaxation of draft token verification and an adaptive controller for dynamic MTP head engagement.

Researchers studied the relationship between LLM iterates and discovery success, developed 12 harnesses called 'Modular', and found that initialization improves LLM-driven discovery. Their results suggest that early discoveries are predictive of eventual success and propose a universally applicable intervention for initialization.

A self-evolving framework, RuleEvolve, for coding rules in AI coding agents uses an LLM-powered mutator module to generate variants and a judge module to evaluate and update the pool with the best-performing ones, outperforming manual engineering and existing prompt optimization baselines in functional correctness, code length, and generation cost.

EurekaBench is a benchmark for measuring AI agents' ability to make scientific discoveries. It tests agents' ability to conduct long-horizon experiments and discover mechanisms that explain observations. The benchmark evaluates three axes of scientific discovery: following scientific constraints, predictive accuracy, and deriving scientific insights. Current AI agents perform well in predictive accuracy but struggle to derive meaningful scientific insights.

A research paper proposes rethinking the evaluation of AI agents by considering them as configurable systems, not just models. The authors introduce a new benchmark and find that agent configuration choices, such as task information and time budget, have a significant impact on performance and behavior.

CompMat-Bench is a benchmark for evaluating AI agents on computational materials science tasks. It reproduces research steps and assesses agents on preparing inputs and analyzing outputs for expensive simulations, supporting four evaluation conditions. The benchmark demonstrates the ability of agents based on three LLMs to complete individual materials research steps with pass rates of 66.0-90.4%.

A research paper proposes a method for embodied agents to handle user corrections in text-based interactions. GAVA, a new method, uses observation-bounded evidence, legal probes, and a one-step expected-loss rule to improve accuracy and reduce interaction cost.

Praxa is an evidence-bound harness for governed AI agent execution, providing explicit representations of proposal, authority, dispatch, external effect, and serving promotion through deterministic admission, brokered execution, read-back, reconciliation, and reviewed promotion. It includes four evidence lanes, but current evidence does not establish several key aspects, including adversarial security, production safety, and specialist superiority.

This paper proposes PACE, a model-native VLM policy that learns to optimize long-horizon reasoning in AI agents by dynamically determining the execution depth.

SkillMaster is a training framework that enables LLM agents to create, refine, and select skills autonomously during task solving. It uses trajectory-informed skill review, counterfactual utility evaluation, and a dual-advantage learning approach to improve agent performance and capability.

SEPAL (Separated Expert Pairs with Answer-Level Fusion) is a new technique for improving large language model (LLM) collaboration by assigning private teams to direct reasoning, evidence grounding, and verification. This approach refines LLMs through role-specific training and majority voting, resulting in improved mean accuracy.

NarrativeSteward is an authoring environment that helps authors work with AI agents to create interactive narratives. It organizes outlines, worldbuilding, and narrative graphs to provide guidance and support for authors. The system has been tested and validated through technical tests and a within-subject study.

DAGent is a DAG-based multi-agent framework that introduces Evaluate-then-Grow incremental planning, allowing agents to adapt their plans as new evidence emerges. This approach outperforms existing Plan-then-Patch strategies in deep research tasks, achieving higher accuracy at lower computational costs.

Researchers propose URAI (Universal Robot-Agent Interface), a new approach to robot control that couples a programming agent with an execution agent. The programming agent writes reusable tools, while the execution agent selects and parameterizes them. This design retains model-level decision-making and improves performance on various robot tasks.

RankEvolve is an auto-research framework for evolving generative ranking models. It uses an Executable Operating Protocol (EOP) to declare phases, gates, branches, and loops, and a runtime enforces the compiled state machine. The framework composes complete black-box coding-agent products, including Claude Code and Codex, as execution-graph nodes that review and repair one another's work. This results in improved execution accuracy and reduced silent critical-defect rates.

Researchers introduce STITCH, a framework for generating task-specific harnesses for large language models at test time using reusable primitives. STITCH selects suitable primitives and compiles them into task-specific harnesses, improving adaptability, robustness, and task success rates.

Researchers propose Consistent Plan-Act (ConPAct), a method to improve coordination between AI agents in long-horizon tasks by detecting and resolving state contradictions. ConPAct improves performance in environments like MiniGrid with GPT-5.6-sol/terra.

Paper introduces Code to Control, a method for synthesizing parameterized reactive controllers using LLMs. It separates controller structure from parameters, allowing for real-time execution and faster action selection than planning-based methods.