Ring-2.5-1T-Zero scales RL reasoning to 1T parameters
kastnerkyle · x · 2026-07-21
A research thread highlights Ring-2.5-1T-Zero, describing a multi-stage RL pipeline that scales reasoning training to a 1T-parameter model.
Training pipeline
- First-stage RL: token-level loss to incentivize reasoning
- Self-distillation: compresses CoT traces and resets the train/inference gap
- Second-stage RL: switches to sample-level loss for sustained improvement
- Third-stage RL: tier-based training for adaptive reasoning depth
Additional observations
- The paper also discusses infrastructure optimizations for stable large-scale training, including mixed-precision control and context-parallel optimization.
- It reports emergent behaviors such as anthropomorphism, structured format, parallel reasoning, and context anxiety.
Results
- The table compares frontier models and several Zero-RL variants across math benchmarks such as AIME 2024/2025/2026, HMMT, and IMOAnswerBench.
- The staged Ring-2.5-1T-Zero variants improve substantially over the earlier Zero-RL baselines, with the third-stage versions reaching the strongest scores among the listed Ring-2.5-1T-Zero variants.
More from Research
- Proto-tools, a universal infrastructure layer for generative biology, now open source — jmschreiber91 · 2026-10-26
- New paper finds AI agents negotiate badly under uncertainty and break deals on purpose — xuanalogue · 2026-10-26
- MCP fault injection study: agent task scores drop 60% under realistic tool failures, scale doesn't help — dair_ai · 2026-10-26
- Study: AI reviewing AI misses most reward hacking; best models miss 25%+ of hacks — ronbodkin · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- NLI alone rejects 5/6 new agent memories; a 0.45 similarity gate fixes it all — Alternative_Pin9598 · 2026-10-26