LLM-based agents reasoning // CoT, ToT, GoT!

Summary

Medium explainer arguing that no single prompt solves hard agent tasks and that sequences of prompts (CoT, ToT, then GoT) let you manipulate intermediate LLM outputs. Summarizes Graph of Thoughts (Besta et al., AAAI 2024): reasoning as an arbitrary graph of thoughts with merge/refine/feedback operations closer to recurrence than linear CoT or tree search. Claims GoT improves sorting accuracy 62% over ToT while cutting cost >31%, and helps set operations, keyword counting, and document merging by decomposing subtasks and combining solutions. Not original research; a blog walkthrough of the prompting literature.

Keywords

blog · cot · tot · got · llm-agents · prompting · reasoning

Topics

LLM agents, prompting, reasoning

Research notes

• Primary: Medium post (same slug also at https://evoailabs.medium.com/llm-based-agents-reasoning-cot-tot-got-c46d792f776a; sbagency is a custom Medium domain). Direct Medium fetch timed out; cataloged from the evoailabs/sbagency listing and GoT paper claims quoted there. Underlying papers: Wei et al. CoT; Yao et al. Tree of Thoughts; Besta et al. Graph of Thoughts AAAI 2024 (https://arxiv.org/abs/2308.09687, DOI 10.1609/aaai.v38i16.29720). Discord title LLM-based agents reasoning CoT, ToT, GoT (edited). Overview blog, not a new corpus, so no datasets_local row.