Reproduce “Kimi K2.5: Visual Agentic Intelligence” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, S.H. Cai, Yuan Cao, Ziwei Chai, Y. Charles and 329 more
Kimi K2.5 introduces Agent Swarm, a learned orchestrator that assigns subtasks to concurrent, frozen subagents and combines their results. The system is built on joint text–vision pretraining with about 15 trillion mixed visual and text tokens. On BrowseComp, Agent Swarm scores 78.4%, versus 60.6% for single-agent K2.5; on WideSearch, it reaches targets 3×–4.5× faster, without establishing lower total computation.
Key insights
- Joint multimodal optimization, including visual reinforcement learning, can lead to cross-modal transfer, enhancing textual task performance.
- Parallel agent orchestration can significantly reduce latency and expand the complexity of solvable tasks by enabling dynamic, concurrent execution.
- Agent Swarm functions as an effective proactive context manager, improving long-horizon task coherence by distributing context across independent sub-agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.