Loading patternsโฆ
RL from Verifiable Rewards (RLVR)
Trains reasoning models with reinforcement learning against automatically checkable rewards, such as whether a math answer matches ground truth or code passes its tests, rather than against learned preference models. Because the reward only scores final correctness, long chains of thought with backtracking and self-verification emerge without supervised rationales. This is distinct from RLHF, RLAIF, and DPO, which optimize human or AI preferences.
In 30 seconds
- What
- Trains a model with RL using automatically checkable rewards (math correctness, code tests) on final outputs only, allowing reasoning chains and self-verification to emerge without supervised rationales.
- When to use
- Problems with deterministic, verifiable ground truth answers where you want the model to discover its own reasoning process without labeled step-by-step solutions.
- Watch out
- Requires massive compute for sampling many candidate solutions per problem; most fail, making training expensive and slow compared to supervised approaches.
Ask the AI expert about this pattern
Opens the assistant with your question prefilled. You review it before sending.
RL from Verifiable Rewards (RLVR): Overview
Trains reasoning models with reinforcement learning against automatically checkable rewards, such as whether a math answer matches ground truth or code passes its tests, rather than against learned preference models. Because the reward only scores final correctness, long chains of thought with backtracking and self-verification emerge without supervised rationales. This is distinct from RLHF, RLAIF, and DPO, which optimize human or AI preferences.
- Automatically verifiable rewards (math or code correctness)
- No supervised reasoning traces required
- Emergent long chain-of-thought with backtracking
- Emergent self-verification and re-checking
- Group-relative policy optimization (GRPO)
- Distinct from preference-based RLHF, RLAIF, and DPO
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.
References
The papers, specifications, and repositories this pattern is based on.
From the engineer behind this catalog
Pay less per call with a model you own
If one repetitive task drives your LLM bill, we can distill it into a small open-weight model trained on your own traffic, then test it against your current model on examples it never saw. Quality and per-call cost come back side by side, in writing.
โฌ2,900 per task, fixed price. You pay only if it matches your current model on your own held-out examples.