A single-tool-call environment that returns multiple reward components instead of a single scalar. This pattern is useful both for evaluation (profile each objective independently) and for multi-objective RL such as GDPO (https://arxiv.org/abs/2601.05242).

Each rollout asks the model to call get_weather for a city. The verifier scores the rollout on three independent {0, 1} components:

• correctness — a predicted call matches the expected name and arguments.
• schema_valid — the call's arguments parse as a JSON object containing every required parameter of the tool.
• format — exactly one tool call was emitted, with no extra assistant text.

Each component is surfaced both as a top-level numeric field on the verify response and inside the reward_components field, alongside the summed scalar reward. The response subclasses BaseMultiRewardVerifyResponse, which makes reward_components required and establishes the shared multi-reward contract. Component keys should remain consistent across every task in the environment.

• Evaluation: because the components are top-level numeric fields, NeMo Gym's aggregate-metrics step reports an independent pass rate for each one. This shows how an agent fails (wrong call vs. malformed arguments vs. extra prose) rather than collapsing everything into one score.
• Multi-objective RL: a GDPO-style algorithm reads the per-objective scores from reward_components and normalizes each component independently, rather than collapsing them into one number. A GRPO baseline instead reads the summed reward and therefore cannot distinguish rollouts with the same total but different composition — the advantage collapse GDPO is designed to fix. How reward_components reaches the trainer depends on the training framework's NeMo Gym integration.

The example data can be found in example_tool_call_multireward/data/example.jsonl and is regenerated with python resources_servers/example_tool_call_multireward/create_examples.py.

Tutorial

For a walkthrough of the multi-reward pattern — including how the components are surfaced for both evaluation and multi-objective RL — see the Multi-Reward Verification docs.

Licensing information

Code: Apache 2.0 Data: Apache 2.0

Dependencies

• nemo_gym: Apache 2.0