ChinaTechnology and telecom2026 – ongoingMiniMax Group
M2.7 agents for internal reinforcement-learning research
MiniMax said M2.7 could handle 30%-50% of its reinforcement-learning research workflow
Problem
Reinforcement-learning experiments required researchers to coordinate data preparation and experiment execution, then investigate failures.
Approach
MiniMax built an internal research-agent system around M2.7. Researchers discussed experiment ideas with the agent, which prepared data, launched experiments and investigated logs. Researchers retained critical decisions and discussions.
Results
Takeaway
The claimed workflow share was not a measured reduction in research costs or staffing.
Categories
- Industry
- Technology and telecom
- Business function
- IT and R&D
- Innovation type
- Process innovation
- AI technology
- AI agentsGenerative AI
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
MiniMax described agents used during its own model-development work. It said M2.7 could handle 30%-50% of the illustrated reinforcement-learning workflow, with humans involved in critical decisions.
Full details are in the original sources.