[flyingdoubleG]

Description

RECGroupedAdvantage.calculate_group_advantage (trinity/algorithm/advantage_fn/rec_advantage.py) builds rewards only in the len(exps) > 1 branch, then uses it for every group:

if len(exps) == 1:
    group_reward_mean = torch.tensor(0.0)
    group_reward_std = torch.tensor(1.0)
else:
    rewards = torch.tensor([exp.reward for exp in exps], dtype=torch.float32)
    ...
is_pos = rewards >= group_reward_mean   # UnboundLocalError for a group of one

So the rec advantage crashes with UnboundLocalError: cannot access local variable 'rewards' whenever a task group has a single experience (repeat_times: 1, or a partial group).

The fix builds rewards before the branch. A single experience keeps the existing convention of the len(exps) == 1 branch (mean 0, std 1, as in GRPO), so its advantage is its reward.

Test: tests/algorithm/advantage_fn_test.py::test_rec_advantage_single_experience_group covers the default args, std_normalize=True and drop="balance". It fails before the change; the file passes after (9 tests, CPU).

Checklist

• Code has passed all tests
• Docstrings have been added/updated in Google Style
• Documentation has been updated
• Code is ready for review

🤖 Generated with Claude Code