Turing LLM Post-Training: Calibrated Distillation and Efficient Dynamic-MoE Execution

What it claims

XPeng's Foundation Model Team (technical report, October 2026) post-trains Turing-20B-A2B, a 20B-parameter dynamic mixture-of-experts model with about 2B active parameters per token and a 128K context window. Four RL specialists (math, code, instruction following, agentic tool use) are consolidated into one model with C-MOPD. The selected checkpoint (K = 16, cmax = 5) keeps at least 90.16% of each specialist's gain over the shared SFT baseline in every domain. The authors state that the results reflect the whole training and execution framework, not any single component.

How it works

flowchart LR
  SFT[Shared dual-mode SFT] --> M[Math RL]
  SFT --> C[Code RL]
  SFT --> I[IF RL]
  SFT --> A[Agent RL]
  M & C & I & A --> D{Task router picks one teacher}
  D --> S[C-MOPD student on own rollouts]

The shared SFT stage uses about 5.28M examples. A single checkpoint serves two modes: an open <think> block for reasoning, or an empty <think></think> pair for direct answers. Each specialist is trained with GSPO, using 16 rollouts per prompt and verifier rewards. Code is rewarded on compilation and unit-test pass rate. Math is rewarded on final-answer correctness, with penalties for truncation and repetition. Agent rewards include feedback from a 20B process reward model. Agent prompts are 100% think mode; IF prompts are 80% direct-response.

In C-MOPD, a task router assigns one frozen teacher per prompt. On each student-generated prefix, student and teacher are renormalized over the teacher's Top-K tokens, and the loss is the reverse KL between them. Each teacher's token losses are divided by that teacher's per-microbatch standard deviation. This scale is detached and floored at 0.1. The result is clipped at cmax = 5, so clipped tokens give zero direct gradient. Fixed teacher weights are 0.3 for IF and 1.0 for the other three domains. Teachers only need to supply Top-K ids and log-probs, which costs O(TK), but the student still computes full-vocabulary logits.

Execution keeps the base model's capacity-constrained Quantile Routing and adds three changes:

1. RW32 caps post-capacity dispatch width at 32 experts per token, down from 256 reserved slots. P99 active experts is about 19, and only 0.21% of routes exceed 32; those keep their top-scoring 32.
2. R3-style final-route replay reuses the rollout's discrete expert assignments in the actor and recomputes routing weights.
3. AsyncOverlap runs rollout concurrently with actor updates, at the cost of a one-step policy lag.

Evidence

• Calibration ablation (matched K = 32): the full objective scores 0.9962. Removing clipping gives 0.9834. Removing both normalization and clipping gives 0.9724.
• Support size: K = 16 had the highest macro integration score among the sizes tested and gives the reported 0.9988.
• Math and code (reasoning mode, 8K response cap): the model beats both SFT and Qwen3-8B on all four benchmarks. Unweighted means are 85.92 for the model, 80.33 for Qwen3-8B and 74.21 for SFT. Qwen3-8B was re-run through the same OpenCompass-based pipeline.
• Broader suite: the model improves over SFT on 10 of 12 metrics.
• Execution speed: on a PPU profile, step time falls from 398.7 s with serial execution to 139.3 s with all three changes, a 2.86x speedup.

Limits

• Loss coverage: the Top-K loss does not constrain student probability mass outside the teacher's support.
• Training lag: AsyncOverlap trains on trajectories from the previous policy version.
• Truncated routes: RW32 cuts the 0.21% of routes wider than 32 experts.
• Narrow speed measurement: the speedup comes from a single PPU profile.
• Head-to-head comparisons: Qwen3-8B is the only external baseline named, and its 85.92 vs 80.33 comparison holds only at the 8K cap.
• Missing details: the fetched text cuts off in Section 3.2, so per-benchmark tables, the two regressing metrics and the appendices (recipes, reproducibility) are missing.
• Availability: the page gives no weights, code or licence information.