CISPO / MiniMax-M1: clip importance weights, not advantages¶
Fidelity: core mechanism. The deterministic mini-suite does not stand in for the paper's full benchmark.
Paper information¶
| Field | Details |
|---|---|
| Paper | CISPO / MiniMax-M1: clip importance weights, not advantages (arXiv 2506.13585) |
| Organization | MiniMax |
| First public release | 2025-06-16 (arXiv v1) |
| Authors' code | Public implementation |
| Adapter / method | cispo |
| Local reproduction code | src/auto_research/post_training/ |
Original paper¶
Motivation and method¶
Sample with a fixed rollout policy and compute token-level importance ratios. Clipping only the ratios preserves the direction of advantages and effective gradients.
flowchart LR
A["Public input or history"] --> B["CISPO training and execution loop"]
B --> C["Paper-specific state or objective"]
C --> D["Matched-budget evaluation"]Key figure from the paper¶
Figure 1: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.
Core equations¶
Paper results¶
The MiniMax-M1 technical report presents efficiency for long-context RL and reasoning/agent benchmark results. No production online A/B test is reported; no online figures are inferred.
Local reproduction¶
Arithmetic candidate suite, 120 steps, 256 examples, seed 42: accuracy 0.2344 → 0.5000 (+113.33%). Rewards, KL, length, and candidate budgets are matched.
auto-research post-train --algorithm cispo --dataset arithmetic-smoke --maximum-examples 256 --steps 120 --seed 42
auto-research evolve --model post-training --dataset arithmetic-smoke --direction "组合 cispo 与已安装论文算子" --generations 2 --population 4
See the fixed metrics.
Reproduction boundaries¶
The local run checks the paper-specific objective, state updates, and fair budgets. It does not reproduce large models, multi-GPU RL, private environments, real websites, or the full benchmark. It verifies the mechanism, not the numerical results in the original tables.