Qwen3-30B-A3B-Thinking-2507 SAO step 99
This checkpoint is Qwen3-30B-A3B-Thinking-2507 after 99 optimizer steps of Single-Rollout Asynchronous Optimization (SAO, arXiv 2607.07508). It comes from the seed-0 run documented in Human-Agent-Society/reef pull request #563, using the reef sao recipe at recipes/sao/examples/imo_answerbench. It is released as a reference artifact for the write-up, with no claim of improvement over the base model.
Training used 1,107 DeepMath-103K problems with difficulty ≥ 7, integer answers, and estimated base pass rates between 12.5% and 75%. Binary rewards came from strict boxed-answer equivalence against reference answers, without an LLM judge.
Each optimizer step used 128 rollouts, one per prompt (group size 1). DIS token-mask thresholds were 0.3 / 5.0 against the rollout engine's log-probabilities. The policy learning rate was constant at 1e-6. The critic learning rate was 5e-6, with two critic updates per actor update and 10 critic-only warmup steps. Training used Monte Carlo value targets and length-adaptive lambda with alpha 1.5, without KL or an entropy bonus. Dropout was off. Generation used temperature 1.0, top-p 1.0, and a 32,768-token window, with plain chain of thought, answers in \boxed{}, and no tools.
The run focused on training stability. The model's pass rate on the training pool rose from 0.70 to 0.83, with response length flat at 15k to 18k tokens. A GRPO(+DIS) control on the same pool collapsed: response length fell below 1k tokens and AIME accuracy fell to 9.6% by step 140.
Held-out evaluation sampled at temperature 1.0 and top-p 1.0, using a 128k-token generation window, strict boxed-answer grading, and no tools. Scores are percentages.
| Benchmark | Problems × samples per problem | Step 99 | Base |
|---|---|---|---|
| AIME 2025 | 30 × 8 | 82.5 | 80.4 |
| HMMT February 2025 | 30 × 8 | 68.3 | 66.7 |
| IMO-AnswerBench | 400 × 2 | 49.4 | 48.6 |
All differences fall within the 95% intervals, approximately ±5 percentage points on the 30-problem sets and ±3.4 on IMO-AnswerBench.
The SAO paper trains with a Python tool from a tool-integrated-reasoning SFT initialization, using an unpublished corpus for about 1,000 steps. It reports 97.3 on AIME and 74.0 on IMO-AnswerBench. This checkpoint does not use that tool, initialization, or corpus, and its results are not comparable to those numbers.
Load with Transformers as you would the base model. The architecture remains Qwen3MoeForCausalLM, with the same tokenizer and unchanged configuration. The base model's chat template and thinking mode apply. The Apache-2.0 license is inherited from Qwen3.