[KongchangAI]
BenchmarkBioBasics Bench

BioBasics Bench

An evaluation benchmark targeting foundational scientific capabilities, distinguishing models' levels of scientific literacy through tasks such as multi-step reasoning, chart interpretation, and experimental design.

Timeline (last 90 days)

Nov 18

OpenAI's official website published a benchmark called BioBasics Bench, used to assess a model's ability to understand and analyze everyday biology experiment materials, but the body text explicitly states that only 7 models were tested, and 6.2 Astra was not among them.

Unverified50%
Nov 18

In the official chart, the previous generation GPT-6 Astra scored an average of 66.7% on BioBasics Bench (the hover tooltip shows about 67%).

Unverified60%
Nov 18

In a secondhand screenshot circulating online, the entry labeled 6.2 Astra shows a score of 89.0%, but the source, accuracy and whether the webpage was modified cannot be verified.

Unverified50%
Nov 18

BioBasics Bench is a domain-specific benchmark, and unlike general evaluations such as MMLU and GPQA, it specifically assesses a model's reading comprehension and analytical ability on everyday biology experiment materials.

Unverified50%
Nov 18

The leaked BioBasics Bench data mentioned above has not yet been independently verified, and the related vulnerability is reportedly already patched.

Unverified50%
Nov 7

OpenAI, together with a biotech company, released a new benchmark, BioBasics Bench, containing 100 everyday lab tasks, requiring agents to draw conclusions from raw readings; the leaderboard and description have been made public.

Unverified50%

All Facts (6)

Source Articles