Attribution of performance gains to fine-grained task decomposition

Determine whether the observed outperformance in the hierarchical multi-agent large language model trading system is fundamentally attributable to the fine-grained task decomposition of agent prompts, rather than to confounding factors such as large language model vocabulary-preference effects that may propagate to downstream agents.

Background

The paper proposes a hierarchical multi-agent LLM trading framework and shows that specifying fine-grained, analyst-like tasks (especially for Technical and Quantitative agents) improves backtested Sharpe ratios compared to coarse-grained prompts. Ablation and text-similarity analyses suggest that technical signals propagate more effectively under fine-grained instructions.

However, the authors note uncertainty about the causal mechanism behind the performance gains. They highlight a potential alternative explanation: certain vocabulary patterns preferred by LLMs might unduly influence downstream agents, implying that linguistic biases rather than genuine task decomposition effects could drive the observed improvements.

References

First, it is not yet fully clear whether the performance gains are fundamentally attributable to fine-grained task decomposition itself. One alternative explanation is that certain vocabulary patterns may be more easily adopted by the preference of LLMs to influence downstream agents.

— Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks  (2602.23330 - Miyazaki et al., 26 Feb 2026) in Discussion and Conclusion — Limitations and Future Work

The dialogue interventions cover only a few predefined histories. S3 fixes one six-message task-decomposition sequence and one pair of 13-word terminal status messages. It does not identify whether the observed change comes from particular wording, the number of preceding turns, or repeated exposure to the same conversational signal. A systematic extension should vary paraphrases and history length around the current design---for example, $x-1$, $x$, and $x+1$ turns or repetitions---to estimate a response curve rather than select whichever wording produces the largest change.

— Same Request, Different Boundary: Evaluating Cybersecurity Assistance across Conversational Contexts  (2609.00578 - Yang et al., 1 Sep 2026) in Limitations section