moonshot-ai

Moonshot AI (月之暗面) — Chinese AI lab; developer of the Kimi model family. Cited in 2026-04-27-llm-training-principles-paths-practices as one of three concrete public case studies for reasoning-model / Agent-training engineering (alongside cursor and chroma).

What this wiki currently knows

Models referenced

• Kimi K2.5 — the public reference for orchestrator-only Agent training. Cited via Moonshot’s Visual Agentic Intelligence technical blog post.

Techniques attributed

• PARL — Kimi K2.5’s Agent training architecture. Trains only the orchestrator agent; sub-agents are frozen. Solves credit assignment by collapsing it to the orchestration layer rather than back-propagating through every sub-agent.
• Three-component reward decomposition for the orchestrator:
  • r_perf — task success (primary signal)
  • r_parallel — incentivizes decomposition; annealed to 0 over training so it acts as a curriculum, not a permanent shortcut
  • r_finish — penalizes spurious parallelism (declaring “done” without completing the task)
• Critical-path length is the parallelism KPI, not total step count — only longest-serial-chain reduction means parallelism is real.
• Attention Residuals — Kimi-side architecture work cited in the systems section, in the same group as Forgetting Transformer: the engineering behind making longer contexts trainable. Referenced briefly; not yet a concept page.

Stance

Of the three Agent-training case studies in this article, Kimi PARL is the most “training-shaped” — it isolates a single learnable component (the orchestrator) from a non-learnable substrate (frozen sub-agents + tool environment), which makes it the cleanest ablation surface. The annealed r_parallel is the kind of detail that recurs in well-designed RL training systems (a curriculum reward that retires once it has done its job); flag if subsequent sources show this pattern elsewhere — it generalizes well outside Agent contexts.