Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Systems

Background and Context

In the architecture of multi-agent reasoning systems, decision-making mechanisms have traditionally relied on consensus, confidence scores, or automated ratings to determine which messages should influence the final output. This filtering logic operates on a fundamental assumption: that a message likely to be correct is inherently worth retaining. However, empirical reality suggests a more complex dynamic where an incorrect final answer may still contain valuable decomposition steps, critical constraints, or underlying scientific principles. To test this distinction, researchers have introduced the Diverse Hypothesis Deliberation (DHD) protocol, a controlled measurement framework designed to redefine the standards for evaluating message value. The core contribution of DHD lies in shifting the focus from binary answer correctness to the potential impact of a message on subsequent reasoning processes. By caching five independently generated messages and replaying them through the same downstream solver, known as the integrator, the protocol precisely measures the "trajectory value" of each message, determining whether its presence aids or hinders the reasoning chain.

This research addresses a significant gap in existing evaluation frameworks, which have largely ignored processual value in favor of outcome-based metrics. The DHD protocol provides a new perspective on information flow in multi-agent collaboration, demonstrating that the utility of a message cannot be reduced to the correctness of its final conclusion. By isolating the causal impact of individual messages on the overall reasoning trajectory, the study offers a rigorous method for understanding how agents interact and how information propagates through a system. This approach moves beyond simple aggregation of answers, focusing instead on the structural and logical contributions that individual agents make to the collective problem-solving process.

Deep Analysis

The technical methodology of DHD is designed with a strict controlled experimental strategy to ensure causal clarity. The system first generates five independent messages for a given problem. During the replay phase, the protocol executes the reasoning process under two distinct conditions for each message: one where the message is available to the integrator and one where it is hidden. By comparing the final results under these two states, researchers can quantify the specific impact of a single message on the overall reasoning chain. This design effectively eliminates confounding variables, allowing for a high degree of causal interpretation regarding the trajectory value of each message. The experiments utilized two publicly available model families, gpt-oss-120b and gemma-4-31B-it, ensuring that the results are reproducible and broadly applicable across different architectural paradigms.

Intervention experiments further explored the specific mechanisms by which message content influences reasoning outcomes. The findings indicate that retaining the full message yields the best performance, while retaining only the reasoning component preserves more success probability than retaining only the final answer. This suggests that the context and logical structure within a message are crucial for maintaining reasoning coherence, although the exact source of the full message's advantage remains an open research question. This fine-grained intervention analysis provides concrete technical grounds for optimizing information filtering strategies in multi-agent systems, highlighting that the structural integrity of the argument is as important as the final conclusion.

Industry Impact

The experimental results were derived from five distinct math and science benchmarks, covering a wide range of reasoning tasks. A key finding is that "wrong but beneficial" messages appeared in every benchmark-model combination, indicating that this phenomenon is not accidental or specific to a particular domain. Among incorrect answer messages that altered the final correctness, more than 40% of the changes were beneficial in each model. To verify the reliability of these results, researchers conducted controlled repetition experiments, finding that the number of repeatable message effects was unlikely to be caused by replay variance alone. Statistical tests yielded a p-value of 0.0002, confirming the significance of the observed effects. Furthermore, within the same problem, repeated trajectory value evidence enabled better retention or deletion choices than relying solely on answer correctness.

These data strongly support the core argument that while answer correctness is informative, it does not determine trajectory value. The DHD protocol generates reusable labels that allow systems to learn when to listen to specific agent outputs, thereby optimizing overall reasoning performance. For the open-source community and industrial applications, this challenges the prevalent "correctness bias" in current multi-agent systems, prompting developers to rethink information filtering algorithms. In industrial settings, systems can leverage DHD-generated labels to train smarter integrators capable of identifying and utilizing messages that appear incorrect but contain critical insights. This not only enhances the robustness of reasoning but also uncovers potential logical paths that human experts might overlook.

Outlook

For future research, DHD provides a standardized measurement framework that enables the comparison of trajectory value across different models and tasks. This standardization is crucial for the development of more sophisticated multi-agent architectures that can dynamically adjust their information intake based on real-time value assessments. The unresolved mystery regarding the advantage of full messages also offers a clear direction for future inquiry, such as exploring which specific features of a message, like logical structure or constraint formulation, contribute most significantly to trajectory value. Understanding these granular contributions will allow for the design of more efficient communication protocols between agents.

Overall, this work drives the evolution of multi-agent systems from simple answer aggregation toward more complex optimization of reasoning processes. By proving that answer correctness is not the sole determinant of trajectory value, the study lays the theoretical foundation for building more intelligent and efficient collaborative systems. As multi-agent systems become more prevalent in complex problem-solving domains, the ability to accurately assess the value of intermediate reasoning steps will be a critical differentiator. The DHD protocol represents a significant step in this direction, offering a practical tool for developers and researchers to enhance the performance and reliability of their systems.