About the project
The benchmark becomes a stable lab where techniques can change and be compared without rewriting the runner or binding the system to one backend.
Overview
A modular capability harness
for pushing agents beyond
scaling-law limits
# practical pipeline
GAIA Benchmark
-> InspectAI GAIA runner
-> Test-Time Scaling Proxy
-> OpenAI-compatible model backend
-> Local / Colab / Mac / GPU inference
Project Map
Start with the big picture, then check the results, and move to the technical section when you want to reproduce or extend the experiment.
Experiment Results
The main signal is already visible because some test-time techniques improve quality, but the gain has to be read together with time, tokens, and run stability.
Proxy micro framework and runner infrastructure
The architecture separates research logic from run infrastructure, so a new technique can be plugged in as a module and reproduced across heterogeneous compute.
Results
These two charts give a quick read: what the baseline looks like across the full task set, and which techniques sit on the accuracy-versus-cost Pareto front.
