MolScent

Description

MolScent is an environment for evaluating agents on matching scent descriptors to molecules. Given a multiple-choice question with four SMILES options, the agent must identify which molecule has (or does not have) a target scent property. The dataset is derived from the Pyrfume GoodScents dataset (~4,272 validated molecules, 251 scent descriptors). Distractors are chosen from same-family scent descriptors by co-occurrence, making the task non-trivial.

Capabilities

• Associating molecular structures (SMILES) with olfactory properties
• Multiple-choice reasoning over molecular scent descriptors
• Distinguishing molecules with similar chemical profiles but different scent properties

Compute Requirements

MolScent does not require a sandbox. It has minimal compute requirements.

License

MIT.

Tasks

There are two splits: train (1,000 tasks) and test (100 tasks), totaling 1,100 tasks. Each task is a 4-option multiple-choice question (A/B/C/D) with two question types:

• Positive (~550 tasks): "Which molecule IS {scent}?" -- the agent picks the molecule with the target scent descriptor.
• Negative (~550 tasks): "Which molecule is NOT {scent}?" -- the agent picks the molecule that does NOT have the target scent descriptor.

Reward Structure

This is a sparse, verifiable reward environment with binary scoring. The agent replies with an ordinary message stating a letter (A, B, C, or D); the environment extracts the letter and grades it by exact match against the correct answer.

• Correct: Reward 1.0.
• Incorrect: Reward 0.0.

We do not use LLM graders for this task.

Data

Tasks are generated from the Pyrfume GoodScents dataset (~4,272 validated molecules with scent descriptors). Distractors are selected from co-occurring scent families to ensure difficulty. Data is stored as a parquet file on the OpenReward platform.

Tools

Agents are given no visible tools. The environment uses a @terminal tool: the agent's final plain-text message ends the rollout and is graded by extracting the answer letter.

Time Horizon

MolScent is a single-turn environment. The agent receives a multiple-choice question and replies with a single message stating its answer letter.

Environment Difficulty

[Statistics on environment difficulty here]

Other Environment Requirements

There are no further environment requirements; MolScent works out of the box with the OpenReward endpoint.

Safety

Agents in MolScent are asked to match scent descriptors to molecular structures. The environment does not present direct safety risks, as agents only provide letter answers with no access to external systems.

Citations

@dataset{GRMolScent,
  author    = {General Reasoning Inc. Team},
  title     = {MolScent},
  year      = {2026},
  publisher = {OpenReward},
  url       = {https://openreward.ai/GeneralReasoning/MolScent}
}