InternScience/ResearchClawBench

Implements a two-stage autonomous research pipeline where AI agents independently analyze datasets and generate reports, then evaluates results against expert-curated checklists from 40 real published papers across 10 scientific domains. The benchmark uses an LLM-as-judge approach with fine-grained, weighted multimodal scoring criteria (text and image-based) to assess whether agents match or exceed human research outcomes. Supports multiple agent frameworks (Claude Code, OpenClaw, Nanobot) with agent-agnostic configuration and includes a Flask-based live streaming UI for real-time agent execution monitoring.