MPW-Bench: Tool-Integrated Reasoning Agent Benchmark in a Parallel World

LiAutoAISiliconLab created a benchmark for evaluating Tool-Integrated Reasoning (TIR) Search Agents, last updated on 2026-04 09. The benchmark uses a ParaWorld Engine that simulates a search engine grounded in fictional, future-situated facts to ensure isolation from model parametric memory. The dataset is designed to eliminate data contamination concerns during agent evaluation.

Use Cases

• Benchmarking agent reasoning performance based on interactions with a simulated search engine.
• Evaluating tool-integrated search agents based on a controlled, fictional knowledge environment.
• Testing agent isolation from parametric memory based on the use of a parallel world simulation.

Strengths

• Designed to eliminate data contamination concerns by using a fictional, future-situated fact base.
• Provides a controlled environment for agent evaluation, ensuring complete isolation from model parametric memory.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

LiAutoAISiliconLab via Hugging Face

Collection Method

Likely synthetically generated by a ParaWorld Engine to simulate a search environment.

Time Range

Future-situated fictional facts.

Freshness

Last updated 2026-04-09 09:49:14; freshness should be verified.

Geography

Not applicable; based on a fictional parallel world.