Probe repo: GPT-4, with pre-GRPO cutoff, predicts DeepSeekMath's GRPO from anonymized text
CatAstro_Piyush · x · 2026-10-20
Developer ruhzi ran an intriguing probe: GPT-4's training data ends September 2021, before DeepSeekMath introduced Group Relative Policy Optimization (GRPO) in February 2024. Using the paper's method description with the name removed, the author asked GPT-4 two questions:
- Predict the result: given a 7B math model's post-instruction-tuning baseline scores, predict its scores on eight benchmarks after this RL stage;
- Give its take: how does the algorithm compare with actor-critic methods like PPO with GAE?
Both prompts and verbatim replies are published in the GitHub repo gpt4-grpo-probe. The experiment probes whether older models showed research foresight into later RL directions.
More from Fun
- Prototyping with Gemini 4 Argon: WebGL toys, auto-edited reels, procedural C++ music — alexanderchen · 2026-10-20
- Agentic coding triggers the same neurons as a 14-hour Civilization addiction — aniketapanjwani · 2026-10-20
- "0.0 days since last GitHub outage" sign photo goes viral as GitHub fails again — kscottz · 2026-10-20
- Meme: by late 2026 even a glider game ships a true wind solver — MoonL88537 · 2026-10-20
- Google Flow rejects harmless prompts but greenlights a video full of profanity — Frustratedtwice · 2026-10-20
- Chinese humanoid robot shows off pole dancing skills in viral demo — LexiLove · 2026-10-20