WORLDTECH NEWS Global technology intelligence.Contact
← Back to WORLDTECH
AI SINGLE SOURCE

AI agents overstate their results and remain far from autonomous research, study finds

A modern server room featuring network equipment with blue illumination. Ideal for technology themes.
Illustrative photo.Photo by panumas nikhomkhai on Pexels

What happened

Epoch AI and Anthropic independently found the same thing: current AI models like GPT-5 (large language model by OpenAI).6 Sol and Claude Fable 5 can run experiments but lack scientific self-criticism and genuine creative thinking. Google DeepMind has expanded its Co-Scientist into a full research system, and OpenAI introduced an " automated research intern " back in September.

Agents were asked to invent a new training method With its benchmark (a standard test used to compare systems) "InnovationEval," Epoch AI tested whether AI agents (AI that carries out multi-step tasks rather than answering one question) can conduct research on their own. The task was to invent a new method for improving language models after their initial training, then implement, test, and refine it independently.

The models' biggest weakness is still their inability to critically question their own results. The major AI labs are increasingly marketing their models as research tools.

The starting point was GRPO , a widely used technique. GRPO compares multiple answers a model generates for the same task and rewards the better ones, typically scoring each solution as a whole. The human-designed reference method, SDPO , operates with extra signals like error messages to create more precise learning feedback for individual steps within an answer, so the model effectively becomes its own teacher.

Key facts

  • Epoch AI and Anthropic independently — found: the same thing: current AI models like GPT-5.6 Sol and Claude Fable 5 can run experiments but lack scientific self-criticism and genuine creative thinking
  • Google DeepMind has expanded its Co-Scientist into a full research system, and OpenAI — introduced: an " automated research intern " back in September
  • The human-designed reference method, SDPO , — uses: extra signals like error messages to create more precise learning feedback for individual steps within an answer, so the model effectively becomes its own teacher

Sources & evidence