Google Cloud Releases Best Practices for Customizing Gemini via Reinforcement Learning

Google Cloud has published a best practices guide for its Reinforcement Learning Fine-Tuning (RLFT) service, which allows developers to customize Gemini models using a custom reward function rather than labeled data. This managed service handles the complex infrastructure and proprietary model internals, enabling users to adapt Gemini for tasks where outcomes are easy to verify but difficult to demonstrate explicitly.

How RLFT Works

Unlike supervised fine-tuning (SFT), which relies on gold-standard examples, RLFT improves the model by generating multiple candidate responses, scoring them with the user-defined reward, and updating the model to favor higher-scoring outputs. The process is fully managed, requiring users to provide only prompts and a reward function. Key characteristics include:

• Learning from own outputs: The model refines its existing capabilities rather than copying external targets, minimizing disruption to unrelated skills.
• Outcome-based rewards: Any response achieving a correct result earns reward, accommodating tasks with multiple valid solutions.
• Competence amplification: RLFT makes occasional successes reliable but cannot teach skills the model never demonstrates.

When to Use RLFT

The guide recommends exhausting prompting and SFT first, as they handle most adaptation needs. RLFT is most effective when:

• Responses can be graded cheaply but not authored cheaply.
• SFT has plateaued on key metrics like faithfulness or schema validity.
• Tasks have many equally valid answers that a single reference target would penalize.

For cases where the base model's success rate is too low for RL to gain traction, a two-stage SFT → RLFT approach is suggested, using SFT as a light warm start before continuing with RL via Continuous Tuning.

Key Use Cases

Early adopters have applied RLFT to several distinct areas:

• AI-powered NPCs: Using a Gemini autorater to score persona and flow, eliminating language drift and repetitive loops in game dialogue.
• Structured entity extraction: Applying a rule-based precision/recall reward to improve field-level accuracy on noisy documents like invoices.
• Content moderation: Combining format validation with a deterministic grader to enforce complex policy adherence and reduce false positives.
• Code execution: Using a code-execution reward that runs SQL or API calls in a sandbox, ensuring queries compile and return correct results.
• Slide generation: Rendering HTML/CSS decks to score visual design and layout integrity, preventing overflow and styling errors.