Prime Intellect open-source self-evolving intelligent environment: enabling AI to "fight each other" to generate over 8,000 testing tools

AIMPACT News, May 19 (UTC+8): According to Beating Monitoring, Prime Intellect has announced the open-source intelligent agent training environment, general-agent—a fully synthetic environment that can self-evolve. The core of this release is to frame task generation as a two-player game: a synthesizer and a solver alternate in adversarial play. So far, a large state database has been automatically built, containing 4,504 tasks and more than 8,000 unique tools. Starting from simple seed tasks, this framework splits tasks into five difficulty tiers, t0 to t4, using 9 strategies including conditional constraints, noise instructions, and cross-entity coupling. The synthesizer designs tasks with a database, interaction tools, and verification functions, while the solver attempts to complete (pass) them. Only tasks whose success rates fall within a specific difficulty range are retained, with the hardest level serving as the seeds for the next round of evolution. Official testing shows that fine-tuning a 30B-parameter model with more than 4,400 trajectories synthesized solely in this environment increased tool-call accuracy from 18.9% to 52.3% on the BFCL benchmark. This mechanism frees the model from reliance on manually labeled static datasets. Through direct game-like competition between models, the system can continuously and automatically generate training data with difficulty that is controllable and semantics that are validated. (Source: BlockBeats)
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
Add a comment
Add a comment
No comments
  • Pinned