Synthesizing RL Envs from Open Trajectories: LFM2.5-24B Self-Scores Task Difficulty

yb2698 · x · 2026-11-06

The author shares how they synthesize new RL training environments for agentic RL, using public agentic trajectories from the Hugging Face Hub as seed data, to increase the difficulty and diversity of previous environments that focused only on data analysis tasks.

Key practices:

Difficulty labeling: each task is scored by LFM2.5-24B itself (under terminus-2), with solve rate kept as a difficulty label — hard below 0.3, medium up to 0.6, easy above.
Task mix weighting: GRPO only learns from groups where some rollouts pass and others fail; tasks the model always solves give zero signal, so the mix is weighted toward medium and hard.
Test quality: each task's hidden tests are mutation-tested, catching test suites that miss broken code.

Original post →

Synthesizing RL Envs from Open Trajectories: LFM2.5-24B Self-Scores Task Difficulty

More from coding & agent