A 1T-parameter MoE model reportedly learned math reasoning with zero-RL and no human solutions
i_dg23 · x · 2026-07-27
A new paper claims a 1T-parameter MoE model learned math reasoning with zero-RL, without showing any human-written solution traces.
Setup
- Base model: Ling-2.5-1T-Base, a 1T-parameter MoE where only about 63B parameters are active per token.
- Training used a four-stage pipeline:
- First-stage RL to amplify low-probability reasoning tokens and surface latent reasoning.
- Self-distillation to compress long chains of thought and reset train/inference drift between Megatron and SGLang.
- Second-stage RL for stability and sustained improvement.
- Third-stage RL with adaptive depth, allowing about 4,000 tokens for easy problems and up to 64,000 tokens for hard ones.
- Infrastructure: 320 H200 GPUs.
Reported results
- After stage 1, the model already reached 84.2% on AIME 2026.
- The final model reportedly won pairwise CoT comparisons against MiniMax-M2.7 at 78% and GLM-5.1 at 76%.
- It used 6,368 tokens to solve tasks, versus 17,220 for GLM-5.1 and 14,115 for Kimi-K2.6.
- The distillation set was only 100k samples, compared with 800k for DeepSeek-R1.
The image also highlights infrastructure optimizations such as mixed-precision control and context-parallel optimization, plus emergent behaviors like anthropomorphism, structured format, parallel reasoning, and context anxiety.
More from Research
- NUS builds a soft force sensor that drives actuators without electronics or power — CurieuxExplorer · 2026-07-27
- Chelsea Finn says robot RL is bottlenecked by physical rollout cost, not algorithms — ycombinator · 2026-07-27
- ICML 2026 oral paper replication scores stay middling after a stricter re-scoring — profjamesevans · 2026-07-27
- Long-running agents will need immutable event logs, this thread argues — sebpaquet · 2026-07-27
- Seed IQ navigates Doom II, prompting questions about benchmarks beyond ARC-AGI — Fit_Transition8824 · 2026-07-27
- Agentic Data Science in Practice: Agents Write Code but Answer Wrong Questions — hugobowne · 2026-07-27