LAB-Bench: FigQA (with tools)

The same 226 LAB-Bench FigQA questions scored when a model has an image-cropping tool and extended reasoning available, instead of answering from a single look at the figure.

Also known as: FigQA with tools

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorydomain
Subcategorybiology research - figure interpretation, tool-augmented
Page statusactive
Metricprecision (correct / attempted)
Directionhigher_is_better
Unit%
Dataset size226
Dataset licenceCC BY-SA 4.0
PublisherFutureHouse

What it measures

This id captures LAB-Bench FigQA scores produced with tool access, rather than from a single pass over the figure. The one documented case found in this research is specific and narrower than "tools" in general: Anthropic's Claude Opus 4.5 System Card (November 2025, Section 2.21) evaluates models "with a simple image cropping tool and a reasoning token budget of 32,768 tokens," contrasted against a baseline with neither tool nor extended thinking. That is a combination of one visual tool and extended reasoning, not web search or code execution. Treat any other source's "with tools" FigQA number as unverified until its own configuration is confirmed; the original LAB-Bench paper itself predicted FigQA specifically was unlikely to benefit much from tool use, since it is a perception task rather than a retrieval one.

Task format

Same figure-only multiple-choice questions as lab_bench_figqa, but the model may use an image-cropping tool (per Anthropic's documented setting) and, in that same setting, an extended reasoning budget before answering.

Models reporting this benchmark

These figures come from the model cards, which carry one collection date per card and no per-score attribution. They are shown as reported, not as verified evidence.
ModelProviderScoreCard as of
Claude Mythos PreviewAnthropic89.02026-04
Claude Opus 4.6Anthropic75.12026-04

Data

This page as JSON · Edit on GitHub