5B-Instruct using Group Relative Policy Optimization (GRPO), a reinforcement-learning method based on verifier rewards, and rejection-sampling fine-tuning (RFT).
Pass@k misses output diversity changes caused by model post-training
GRPO and rejection-sampling fine-tuning produced sharply different output distributions despite often similar pass@k scores.
Academic
Subham Rath · Raj Dandekar · Rajat Dandekar · Sreedath Panat
Accenture Strategy & Consulting · Vizuara AI Labs
Research Digest··3 min read
5B-Instruct checkpoint on grade-school mathematics.
Why this paper
From Accenture Strategy & Consulting and Vizuara AI Labs
In one line
pass@k measures correct-answer probability but not output diversity, so it cannot detect post-training's effects on capability breadth.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§