2d ago

OpenDeepThink Scales LLM Reasoning Via Parallel Population Evolution

——0——
Original post
OPTeortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)#420Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@TEORTAXESTEX

> verifier-free > +405 points CodeForces Elo in eight sequential LLM-call rounds (∼27 minutes wall-clock). Man, I want to test this on CritPt and harder MathArena sets. If this works with cheaper models, this is insanely powerful

9:33 AM · May 17, 2026 View on X
QUOTE POSTTeortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)#420Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@TEORTAXESTEX

This is pretty fun vibecoded an implementation of the paper, added a webui and freeform mode, plus model split (deepseek-Flash for generation, Pro for judge). Still chokes on hard problems, but would be trivial to chain with decompositions/other hacks. GPT-Pro on a budget?

7:18 AM · May 18, 2026 · 10.8K Views
REPLYTeortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)#420Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@TEORTAXESTEX

yess… I love this

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

This is pretty fun vibecoded an implementation of the paper, added a webui and freeform mode, plus model split (deepseek-Flash for generation, Pro for judge). Still chokes on hard problems, but would be trivial to chain with decompositions/other hacks. GPT-Pro on a budget?

7:18 AM · May 18, 2026 · 10.8K Views
7:47 AM · May 18, 2026 · 1.6K Views