Limits of Difficulty Scaling: Hard Samples Yield Diminishing Returns in GRPO-Tuned SLMs

Develop a benchmarking toolkit that maps the plateau points of SLMs during GRPO/RL tuning. This tool should provide automated 'difficulty-scaling' analysis to help developers stop training once they hit diminishing returns.

Suggested repo: grpo-limit

"Know exactly when your SLM stops learning: Automated difficulty-plateau detection."

Estimated effort: 20h