Microsoft has introduced new multi-model capabilities—Critique and Council—to Researcher, its deep research agent in Microsoft 365 Copilot, to improve accuracy and structure in complex work tasks.
Available through Microsoft’s Frontier program, the update uses separate AI models for generation and evaluation. Microsoft said this approach “raises the bar for accuracy, depth, and confidence” in research outputs.
Critique, the default mode, splits research into two stages. One model generates a draft by planning and retrieving information, while another reviews and refines it.
“By giving evaluation as much emphasis as generation, this architecture creates a feedback loop that delivers higher-quality results,” the company said.
The system uses models from providers including Anthropic and OpenAI. The reviewer model applies rubric-based checks focused on source reliability, completeness, and evidence grounding.
Microsoft said performance improved on the DRACO benchmark, which measures research accuracy, completeness, and objectivity across 100 tasks. Researcher with Critique recorded a 7-point increase in aggregated scores, a 13.88% gain over Perplexity’s system using Claude Opus 4.6.
The company reported gains across evaluation metrics, including breadth and depth of analysis, presentation quality, and factual accuracy. “Critique pushes Researcher to identify missing analytical angles, close coverage gaps, and strengthen claims,” Microsoft said.
On the other hand, Council offers a parallel approach. It runs models from Anthropic and OpenAI side by side, generating separate reports. A judge model then summarises areas of agreement and differences.
ALSO READ: Google Lets Users Bring ChatGPT & Claude History Into Gemini
