Clarification on Inference-Time LLM and Agent Access During Competition Evaluation

#3
by mohammedsalehss - opened

Hello Google PLE dataset maintainers,

I am participating in the Physical Engineering's Last Exam competition and developing an agent harness called ENGINEER-X.

I have joined the competition and obtained access to the gated dataset.

I would like to clarify one specific point about the dataset usage restrictions.

The competition requires an AI agent to process engineering task instructions and produce solutions. My intended approach uses an existing pretrained LLM through inference-time API calls, without training, fine-tuning or distillation on the competition dataset.

Could you please clarify:

  1. Is inference-time access to task instructions and associated files by an AI agent permitted during authorized competition evaluation?
  2. Is sending relevant task information to an existing pretrained model through an external API permitted?
  3. Does the prohibition on exposing dataset contents to models or agents apply to inference-time prompting, even when model weights are never updated?
  4. Are there specific restrictions on task data retention, API providers, or evaluation logs?

I will follow the official competition rules and dataset conditions. I have not accessed task contents for agent evaluation while awaiting clarification.

Thank you for your guidance.

Google org

Hi Mohammed,

In general we permit evals on the dev set and sending requests to external APIs / model providers / your local models. We do not allow direct training on the dev set or use them as distillation environments.

Best regards,
Yufan

Sign up or log in to comment