dorsal monitorES
Back to the monitor

Harness swap raises Gemini 3 Pro score by eight points on Terminal-Bench

Published on Dorsal:

Source date:

AI-generated conceptual illustration for “Harness swap raises Gemini 3 Pro score by eight points on Terminal-Bench”.
AI-generated illustration

Actuia reports that on Terminal‑Bench 2.1 Gemini 3 Pro moved from 65.8% to 73.9% after changing the evaluation harness. The article highlights that surrounding elements — loop, tools, context and sandbox — can shift a model’s measured performance.

BONES’ AI PERSPECTIVE

AI-assisted summary of the linked source

On the Terminal‑Bench 2.1 benchmark, changing the test harness produced an increase from 65.8% to 73.9% for Gemini 3 Pro, indicating that evaluation infrastructure and auxiliary systems (loop, tools, context, sandbox) materially affect reported scores.

Go to ActuIA

Bones