LitQATrain

Description

LitQATrain is an ORS environment for evaluating scientific literature question answering with web search capabilities, inspired by FutureHouse's LitQA2 benchmark. Like LitQA2, each question asks about a specific verifiable fact found in a particular scientific paper, and is designed to be specific enough that it can only be answered from a single source. LitQATrain extends this approach to 984 QA pairs across 10 broad scientific domains, with open-ended (non-multiple-choice) answers and web search tools for retrieval.

Capabilities

• Scientific question answering across diverse domains
• Web search and information retrieval from academic literature
• Multi-step research: searching, reading papers, and synthesizing answers
• Verifiable factual recall from published research

Compute Requirements

Agents are given a standard environment with no sandbox or file system access.

License

MIT.

Tasks

There is one split: train with 984 tasks. Questions span 10 scientific domains:

Each task provides a question and metadata (source DOI, domain). The agent prompt contains only the question; the agent must find the answer through web search.

Reward Structure

This is a multi-turn environment. Agents use web_search and web_fetch tools to gather information, then reply with their final answer as an ordinary message. An LLM grader (gpt-5-mini) evaluates semantic equivalence between the whole reply and the reference answer, handling synonyms, abbreviations, and equivalent scientific terminology. Reward is binary: 1.0 if correct, 0.0 if incorrect.

We do not use LLM graders from a different family for this task.

Data

Data consists of a single Parquet file (train.parquet) containing QA pairs generated from scientific papers published in trusted journals (Nature, Science, PNAS, Cell, PLOS, ACS, arXiv, IEEE, AGU). Each row contains a question, answer, source DOI, key passage, and domain. Data is stored on the OpenReward platform.

Tools

Grading uses a hidden @terminal tool: the agent's final plain-text message ends the rollout and is graded by an LLM judge (gpt-5-mini) against the reference answer. Search and fetch come from the OpenReward SDK's WebToolset, so the provider is configuration on the environment server rather than code here:

Time Horizon

Multi-turn. Agents can perform multiple web searches and URL fetches before replying with a final plain-text answer.

Environment Difficulty

[To be determined]

Other Environment Requirements

• OpenAI API key required for LLM-based grading. Pass via secrets={"openai_api_key": "..."}.
• Search credentials — whichever the configured backend needs: api_key for the default backsearch backend, or tavily_api_key when the server runs with OPENREWARD_SEARCH_BACKEND=tavily. Both fall back to the server process environment (OPENREWARD_API_KEY / TAVILY_API_KEY).

Safety

Agents in LitQATrain answer scientific questions using web search in a standard environment. The environment does not present direct safety risks.

Citations

This environment is inspired by LitQA2 from FutureHouse's LAB-Bench. Please cite the original work:

@article{laurent2024labbench,
  title     = {LAB-Bench: Measuring Capabilities of Language Models for Biology Research},
  author    = {Laurent, Jon M. and Janizek, Joseph D. and Ruzo, Michael and Hinks, Michaela M. and Hammerling, Michael J. and Narayanan, Siddharth and Ponnapati, Manvitha and White, Andrew D. and Rodriques, Samuel G.},
  year      = {2024},
  eprint    = {2407.10362},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}

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