send step level advantages to prime rl

[eligotts]

Description

Added opt-in support for verifier-provided step-level advantages:

• Introduced use_verifiers_advantages on the base Environment and auto-populate it into state so environments can explicitly signal this mode to prime-rl
• Added completion_advantages on each trajectory step, computed as per-completion-token values by expanding the step scalar advantage across that step’s completion_ids (one value per token).

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes

---

Note

Medium Risk
Changes rollout/state/output schema and modifies grouped scoring post-processing, which can affect downstream training/serialization expectations despite being additive and covered by tests.

Overview
Adds opt-in plumbing to propagate use_verifiers_advantages from Environment into each rollout State (and makes it available via states_to_outputs) so downstream consumers can detect/verbalize this mode.

Enhances grouped scoring (Rubric.score_group) to populate each trajectory step with a new completion_advantages field, expanding the step’s scalar advantage into a per-completion-token list (and filling missing step reward/advantage from the state as a fallback). Includes tests covering default fallback behavior vs preserving an existing step advantage.

Also hardens gepa.py when reading result.best_candidate by handling non-dict candidates.

^{Written by Cursor Bugbot for commit 0b96ae2. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

RubricGroup produces incorrect completion_advantages after aggregation

Medium Severity

When RubricGroup.score_group is used, each sub-rubric's score_group calls _populate_step_completion_advantages with that sub-rubric's partial advantage. Since step advantages are only set when None, the first sub-rubric's partial advantage "sticks" on the steps and completion_advantages, while RubricGroup never recomputes them from the final aggregated reward. The result is completion_advantages reflecting only one sub-rubric's contribution rather than the total.

 

[eligotts]

this was already the implementation before, the rubric group only does aggregated trajectory level rewards, if someone wants to do step level advantages they will probably have a single reward function in a rubric that does this. probably a cleaner way to get around this but not sure if this is worth fixing right now