flow-state vs. linear-learning
courtesy of UC Berkeley and FAIR
consider these approaches for training LLM agents on multi-turn tool use:
1. PAE (propose-agent-evaluate) -> linear-learning
2. SCA (self-challenging agents) -> flow-state
The crucial differences between PAE and SCA are: (1) the task proposer in PAE generates tasks directly from API documentation and the initial observation instead of serving as an agent that actively interacts with the environment to gather information before creating the task; (2) PAE only generates instructions while SCA tasks contain instructions, verification functions, example solutions, and failure cases; (3) PAE prompts the same model to serve as the judge instead of relying on verification functions as in SCA.
As with most important findings, this one seems obvious in hindsight.
SCA beats PAE by 13.6%, 11.6%, 4.8 %, and 2.1%.
Of course, flow-state beats linear-learning. I mean, obviously!
For more on flow-state check out Vervaeke's Awakening from the Meaning Crisis and the original book by Mihaly Csikszentmihalyi