ExComm: An Exploration-Phase Communication Protocol for Error-Resilient Agent Reasoning

ExComm is an exploration-phase communication protocol for error-resilient agent reasoning. Its core is to effectively block error propagation and significantly improve the accuracy of long-range reasoning tasks by detecting and resolving cross-agent factual conflicts during the exploration phase. This article will introduce it from aspects such as background, mechanism, experiments, contributions, and applications.

Problem Background: Error Propagation in Agent Reasoning

In long-range agent reasoning tasks, error propagation is a fatal problem—factual errors or invalid inferences in intermediate steps remain in the belief state, contaminating subsequent reasoning and forming a "snowball" effect.

Existing test-time expansion methods have limited control: relying on agents to detect errors on their own, selecting from defective trajectories, or correcting after errors have shaped the path, resulting in poor post-hoc remediation effects.

Core Idea of ExComm

ExComm is based on the observation that most intermediate errors produce detectable cross-agent factual conflicts in parallel reasoning. Its core mechanisms include:

Periodic Belief Auditing

Regularly cross-audit the belief states of each agent to detect conflicting views on the same fact.

Toolized Verification Cycle

Conflicts are checked through a tool chain: calling external knowledge bases, executing code verification, retrieving authoritative data sources, etc.

Soft Belief Update

Verification feedback is integrated into beliefs in an "append" manner, preserving reasoning history and avoiding information loss.

Trajectory Diversity Protection

When detecting that agent paths are converging, guide some to switch to orthogonal strategies to maintain exploration breadth.

Experimental Verification and Results

Test Benchmarks

• AIME 2024 (real questions from the American Invitational Mathematics Examination)
• AIME 2025 (latest competition questions)
• GAIA (General AI Assistant Evaluation Benchmark)

Model Configuration

• Gemini-2.5-Flash-Lite
• Qwen3.5-4B

Core Results

• Gemini model: average improvement of 5.7% compared to the strongest baseline
• Qwen model: average improvement of 5.0% compared to the strongest baseline (statistically significant)

In-depth Analysis

• Error recovery success rate increased by nearly 40%
• Advantages expand as the number of agents/reasoning steps increases
• Maintains higher trajectory diversity
• Optimal cost-performance ratio

Technical Contributions and Methodological Insights

1. Cross-agent Factual Conflict Detection: For the first time, cross-agent consistency checks are introduced into the test-time expansion framework, similar to cross-validation in human teams.
2. Tool-enhanced Verification Paradigm: Introduce external tools for objective verification to improve the reliability of error detection.
3. Balance Between Soft Update and Autonomy: Update beliefs in an append manner, respecting agent autonomy and conforming to distributed characteristics.

Application Prospects

ExComm can be applied to:

• Scientific research assistance (literature review, hypothesis generation)
• Code generation and debugging
• Complex decision support (finance, medical care)
• Educational tutoring systems

The research team has open-sourced the ExComm implementation and provided integration interfaces for mainstream agent frameworks to facilitate the deployment of the technology.