Google Cloud’s managed reinforcement-learning fine-tuning service adapts Gemini using customer-defined rewards, while Google manages infrastructure and proprietary model internals.

Google Cloud’s reinforcement-learning fine-tuning (RLFT) service lets customers supply prompts and a reward function while Google manages training infrastructure and Gemini’s proprietary model internals. The guide contrasts this with conventional reinforcement learning, which requires large training clusters and access to internals unavailable to external users of proprietary models.

At each training step, the service generates multiple candidate responses, scores them using the customer’s reward function, and makes higher-scoring responses more likely while keeping the model close to the original Gemini. Google presents this as useful for tasks that are difficult to demonstrate but straightforward to verify, such as executing a generated SQL query and checking its result.

The guide emphasizes three boundaries: learning from the model’s own outputs, rewarding outcomes rather than specific solution paths, and amplifying existing competence. Google says RLFT tends to disturb unrelated capabilities less than supervised fine-tuning (SFT), but cannot teach a skill the base model never demonstrates.

Google recommends exhausting prompting and SFT before using RLFT. Direct RLFT is suggested when the base model already succeeds sometimes; a light SFT warm start followed by RLFT is suggested when demonstration data exists or initial success is too low, including cases where SFT has plateaued on faithfulness, schema validity, or tone.