Rewards

Rewards are an essential part of reinforcement learning frameworks. They are typically float values (usually between 0.0 and 1.0, but they can be negative also) that guide the process into making more efficient decisions or predictions. During training, the goal is to maximize the reward function. The reward gives the system an indication of how well it performed for that task.

The purpose of a reward function is to compute the quantity that the program should maximize during training.

Synalinks ships two flavors of reward base class:

• Reward: scores one sample at a time. The trainer iterates the batch and calls the reward once per (y_true, y_pred) pair. This is what the built-in rewards (ExactMatch, CosineSimilarity, LMAsJudge, ProgramAsJudge) use.
• BatchReward: receives the whole batch at once and returns a list[float] of length batch_size, one reward per sample. Use this when the score for sample i depends on the rest of the batch — group-relative rewards (e.g. GRPO-style normalization), pairwise/listwise comparisons, or any case where you want a single batched call instead of N parallel calls.

Both flavors share the same masking, reduction, and serialization machinery, so you can mix them freely (including across multi-output programs).

Rewards Overview

• ExactMatch reward
• CosineSimilarity reward
• LMAsJudge reward
• ProgramAsJudge reward
• BatchReward & BatchRewardFunctionWrapper