PhAI Labs Launches Free Scientific Agent Workspace 'ScienceBuddy' with GPT-6 and Dual Self-Improvement Loops
PhAI Labs has officially released ScienceBuddy, an interactive scientific research workspace offering free GPT-6 access and GPU acceleration integrated with the
On September 23, 2026, Yingcheng Charles Wu of PhAI Labs announced the official release of ScienceBuddy, an interactive research workspace designed to embed self-improving scientific agents directly into researchers' daily workflows.
Image source: PhAI Labs / @Charles_Y_Wu
Free GPT-6 Access and GPU Acceleration Powered by JEV Framework
Under the launch terms announced by PhAI Labs, ScienceBuddy provides researchers with access to GPT-6 at no cost in a fully GPU-accelerated computing environment.
The platform is fused with PhAI Labs' JEV framework, providing a lightweight decision and execution foundation optimized for autonomous research routines. This configuration allows domain scientists to experiment with scientific data analysis, literature synthesis, and hypothesis generation directly through a web-based workspace without managing dedicated GPU infrastructure or incurring model API expenses.
Recursive-in-Recursive Architecture: Harness Evolution Paired with Rubric-Guided RL
At the technical core of ScienceBuddy is an architecture the team terms "Recursive-in-Recursive Self-Improvement," organized around two distinct, complementary feedback loops:
- Inner Loop (Harness Evolution): While holding the underlying foundation model weights fixed, the system refines the agent harness—including prompt guidance, tool connections, and contextual memory—as the researcher and agent collaborate.
- Outer Loop (Model Learning): The system takes execution trajectories, researcher feedback, and structured rubrics gathered under the improved harness to train the base model via reinforcement learning (RL).
According to the accompanying research preprint (arXiv:2609.17523), harness evolution shapes the training experience, while the resulting model capabilities open up new vectors for further harness adaptation in a continuous cycle.
Turning Everyday Workflows into Continual Learning across Four Scientific Domains
ScienceBuddy captures everyday interactions—researcher requests, iterative feedback, tool invocations, and experimental observations—and translates them into formal benchmark tasks and evaluation rubrics.
In their research paper, the team documented case studies evaluating researcher interaction, harness refinement, and model learning across four distinct scientific task families. The overarching objective is advancing toward "discovery intelligence"—AI systems that do not remain static post-deployment, but evolve continuously through sustained partnership with domain scientists.
Availability, Limits, and Operational Caveats
- Free Tier Availability and Limits: While PhAI Labs announced GPT-6 access at no cost, the launch announcement did not specify per-user rate limits, daily query quotas, or long-term pricing guarantees. Researchers running automated research pipelines should verify active quotas before initiating batch workloads.
- Model Update Horizons: While inner-loop harness adjustments occur immediately during active sessions, substantive outer-loop weight updates require the accumulation and curation of researcher rubrics over extended periods.