Seonglae ChoKimi K2
- agentic planning
- tool use
K2 Tinker
2.5 Agent Swarm
When the total vision+text token budget is fixed, "early fusion" (mixing vision and text tokens from the start at a constant ratio) converges more stably and achieves better final performance than "late injection" (adding many vision tokens at the end). Suddenly injecting many vision tokens in the middle or later stages causes the representation space to be disturbed by "modality domain shift," resulting in a dip-and-recover pattern in text performance.
Zero-Vision SFT activates visual capabilities using only text. If vision↔text alignment is already established to some extent during joint pretraining, running only visual RL can improve text benchmarks (MMLU-Pro, GPQA, LongBench).
DEP (Decoupled Encoder Process): An infrastructure trick claiming 90% multimodal training efficiency.
- Balanced Vision Forward
- Backbone Training
- Vision recompute + backward
Agent Swarm + PARL: "Only the orchestrator is trained with RL, sub-agents are frozen." PARL reward: r_parallel + r_finish + r_perf. GRM (Generative Reward Model) is not a "critic" but rather a reward model that scores rewards.
