Bewaard in:
| Hoofdauteur: | |
|---|---|
| Formaat: | Recurso digital |
| Taal: | |
| Gepubliceerd in: |
Zenodo
2026
|
| Onderwerpen: | |
| Online toegang: | https://doi.org/10.5281/zenodo.18869913 |
| Tags: |
Voeg label toe
Geen labels, Wees de eerste die dit record labelt!
|
Inhoudsopgave:
- <p>Current Large Language Model (LLM) agents primarily rely on the ReAct (Reason + Act) paradigm for task execution. However, these agents often suffer from "episodic amnesia," where successful reasoning chains are discarded after task completion, leading to redundant computation and high token latency in recurring scenarios. We propose <strong>ReAct+DT</strong>, an advanced framework that introduces <strong>Deposition (D)</strong> and <strong>Tree-structured Retrieval (T)</strong>. By distilling successful reasoning paths into permanent Python scripts and Markdown manuals, and organizing them into a self-growing hierarchical tree, the agent transitions from unstable zero-shot reasoning to stable, programmatic skill execution. Experimental results demonstrate that ReAct+DT significantly reduces inference costs and improves long-term task success rates.</p>