feat: Analyze Tonyi Deep Research

[new5558]

yesterday Alibaba Tonyi lab released their Tonyi Deep Research Model

[new5558]

Current Observation

1. The model is overfitted with their trained search tools MCP tool name and argument even extensive prompting
2. After modified our serper MCP to behave like their search tools they still not performing welll on long horizon search task. The model stopped browsing after search/scrape 10 websites.

Next Step

1. Use their code 100% and see if the issue persist.

[new5558]

The model can search and scrape multiple urls at once

The model did not do long horization web browsing

[new5558]

Question

1. Is Tonyi Good at this
2. Is everybody use the same flow

[new5558]

Tonyi Deep Research implementation

Source: reverse engineered from original repo's code
Scope: Search and Visit only

Model is a ReACT agent that can choose to do 3 things.

• Search
• Scrape
• Finish Generation (If selected will end the loop)

End of generation prompt

They tracked token counts. If the token count is > 100K, prompt the model to finish generation and generate final result