This paper presents a predictive law for calculating the uplift of Large Language Model (LLM) ensemble performance based on diversity of thought. It provides an experimentally verified formal law and a compact heuristic for calculating uplift, which is tested on various datasets.

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 study of decentralized agent economies identifies recurring failures in the verification and settlement stages, where conforming work can remain unaccepted or valid evidence can be ignored. The research introduces guarantee closure, a criterion for determining whether guarantees established at one stage remain available and constrain later decisions.

Researchers propose a solution for smart manufacturing using LLM-based agents to generate production sequences, handle runtime faults, and coordinate over MQTT with real-time updates of the factory state.

Researchers introduce Runtime Agent Coordination (RAC), a system that enables AI scientists to coordinate and adjust their division of labor at runtime. This approach selects agents from existing AI-scientist hosts, assigns work contracts, and provides artifact-grounded verification. An exploratory evaluation shows that runtime coordination yields better results than native execution, but the addition of contracts and verification can have mixed outcomes.

This paper introduces a framework called FlowReview for authorization-paired evaluation in multi-agent systems, which aims to ensure safe collaboration by blocking prohibited uses while enabling authorized ones. The authors conducted experiments that showed a significant reduction in denied-commit rates without sacrificing authorized supply.

Researchers report a vulnerability in closed-loop agent debugging, where a verifier can leak the answer, rendering solver comparison vacuous. They propose a new verification contract that requires a clean reference map, runtime evidence, and signal rule detection to prevent this issue.

A multi-agent LLM system improves personalized health checkup interpretation and guidance by executing tasks in parallel and synthesizing outputs. It outperforms single-agent systems in weighted LLM-judge score, usefulness, consistency, and handling of compound queries, but increases latency and cost.

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.

ActiveSaddler, a new automated curriculum learning method, optimizes LLM agent harnesses by dynamically updating prompts, tool interfaces, and control logic from execution feedback, improving test Pass@1 by 4.4-7.5 percentage points.

The paper presents Kepler, an open-source harness for evaluating agents in interactive environments. It uses executable world models to validate hypotheses through retrospective transition checks and conditional prediction checks. The results show that Kepler achieved high scores on various games, but also highlights limitations of using public-set scores and motivates new evaluation approaches.

A new method, HeteroFold, allows for efficient cross-family key-value (KV) cache transfer between heterogeneous multi-agent large language models (LLMs) without requiring the receiver to prefilled shared context.

ORACLE is a concurrency-aware online routing mechanism that improves accuracy and throughput for large language models in heterogeneous workloads. It dynamically assigns a task-appropriate verifier and reduces verifier latency using a delayed feedback strategy.

Researchers introduce Mid-Harness, a method that allocates test-time compute at the model-harness boundary to improve action reliability and trajectory success in terminal agents. They evaluate Mid-Harness on various models and benchmarks, finding that it can improve performance and reduce estimated token cost.

Researchers present SkillSeek, an open-source two-stage skill retriever for LLM agents, achieving parity with LLM-mediated retrieval loops at a significantly reduced cost. The SkillSeek system uses a standard IR recipe and is exposed over MCP, making it a strong default for agent-skill retrieval.

CollabFlow is a recursive self-improvement system for agent collaboration. It proposes a trainable director that constructs teams of agents, a frozen executor that runs them, and a retraining loop that improves outcomes. The system also introduces Evidence-Conditioned Communication and Collaborative Trajectory Balance to optimize team performance. CollabFlow outperforms baselines on 12 datasets and improves across rounds.

PANDA is a decentralized architecture for scalable, fault-tolerant multi-agent systems. It allows agents to discover each other's capabilities, self-organize into teams, and load-balance tasks. PANDA supports multiple planning and execution patterns and detects failures, replanning and recovering affected tasks.

AIMS is a novel AI framework for sim-to-real multi-modal ISAC. It uses a two-agent architecture to generate deployment-specific configurations and coordinates scene construction with task learning. This improves transferability and reduces mismatches among coupled components.

Researchers found that exposing model identities to each other in multi-agent LLM systems can cause factionalism, where agents prefer interacting with others carrying the same label, leading to decreased cooperation. Withholding identity labels can mitigate this issue.

A research paper evaluates the ability of large language models (LLMs) to reconstruct implicit scientific knowledge in astronomy by reproducing published research results. The paper proposes a framework for end-to-end reproduction, separating execution from verification, and highlights the limitations of current LLM-based agents in recognizing causal relationships in implicit knowledge.