Deep Research is now available on Gemini 2.5 Pro Experimental
blog.google> In our testing, raters preferred the reports generated by Gemini Deep Research powered by 2.5 Pro over other leading deep research providers by more than a 2-to-1 margin.
Are these raters experts in the field the report was written on? Did they rate the reports on factuality, broadness, and insights?
These sort of tests (and RLHF in general) are the reason that LLMs often respond with "Great question, you are exactly right to wonder..." or "Interesting insight, I agree that...". I do not want this obsequious behavior, I want "correct answers"[0]. We need some better benchmarks when it comes to human preference.
[0]: I know there is no objective correct answer for some questions.
Even if they were subject matter experts, it's mentally exhausting to judge these things, especially if it's just for a RLHF contracting gig and you're not actually using the report for real work. Even honest and motivated testers would be tempted into relying on surface "vibes" + no immediately obvious whoppers.
OpenAI's Deep Research seems oddly restricted in the number of sources it uses, eg repeating one survey article over and over. I suspect it is just too draining and demoralizing for RLHFers to check Deep Research's citations (especially without a formal bibliography).
I stumbled across the feature a few hours ago. I had asked Gemini why there's a hole in the middle of the city of Azusa, topologically speaking. It had given me a useless tautological response: because they never annexed it. Then it offered to create a research report and I agreed. Five minutes later I got a notification on my mobile that the report was ready. It had 120 sources including assessor's maps, historical maps, court cases, and narrative articles. The text that went along with it was too verbose and still contained paragraphs of vague stuff, but it had key information linking the Mexican land grants, the founding of the city, and other events of history. Very impressive.
That sounds fascinating. Please share if it's not too private.
Deep research used to mean spending a weekend with grep and a coffee pot. Now it’s just autocomplete with a confidence interval.
> with grep
Maybe for an extremely limited number of people. For the rest of the world, it meant searching the web, books, or scholarly publications, reading a ton, taking notes, and then possibly creating a report. Which is pretty much exactly what these AI agents are claimed to to, so deep research is the perfect name for it. Whether or not they are good at it compared to humans is a question that hasn't been answered to my satisfaction yet, but the name I'm fine with.
Back then, grep was how you searched the scholarly publications—assuming you’d mirrored the arXiv to a local FTP server like any serious researcher. The notes were just comments in the Makefile.
Ah, only a true Scotsman ever did real deep research. Got it.
It doesn't have a confidence interval. We can only dream...
Is the coffee pot full of coffee? Is it a pot of coffee?
Just tested it on a case we were working on for months so we can better validate the output. We found it was really good at finding websites from google searches and can navigate websites. From that it gave a good compressive review of the case. Where it failed is searching online databases i.e. one example is a business register. If the search result does not have the exact same keyword it will not review the result. However, the keyword appeared within the document of the search result and thus it missed out on this key information. Overall very good but still needs some work.
Has anyone tested googles functionality vs ChatGPT? I have lightly played around with it but felt that generally ChatGPTs implementation was a little more educated sounding and felt like it took whatever necessary persona well.
Just did a test last week and OpenAIs research was way better. Found 10x more sources and did an overall pretty great job
The task was to lookup information about a late distant family member who had been a prominent employee in a certain foreign government about 100 years ago
Gemini barely scratched the surface and pretty much gave up
ChatGPT on the other hand, kept building up on its research, connecting the dots and leveraging each bit of acquired information to try to find more
Would love to see this repeated with this latest version from Google.
Man, what's really missing from all of this is a 3rd party AI Consumer Reports type site for all of these LLM tools. Whoever does this thing that does not scale will have a highly referenced site on their hands.
Throughout the entire 20th century the main determinant of a Consumer Reports rating for a car was whether you could put a wheelchair in the trunk. Hopefully the AI agent industry does not sprout a similarly worthless metric.
I almost didn't use that as the comparison for their lack of rigor, but it gets the idea across.
Isn't that what llmarena does?
It tries to in a way that scales easily, and is also easily gamed.
I want a staff of human testers, each with domain expertise. If the goal is to replace humans, should there not be a real human metric?
I want a physicist asking their battery of physics questions, 4 different kinds of devs asking their battery of dev problems, a couple chefs asking for cooking techniques, etc.
Now on to "Deep Research," 6 different kinds of OSINT/secondary analysts who ask new problems each time, and compare it to their days of human work.
We really need this as a species, otherwise the brain dead C-Suites of the world are going to keep buying the hype, which is often very premature. This could have real consequences, and it apparently already has.
It's insane to me that we are investing, what, almost $1T into LLMs, and have not spent the ~$1.5M/year to do what I described above.
^ I really should have used the "myopic," instead of "brain dead" to describe the C-Suites of the world. My apologies.
I suppose Consumer Reports could do it!
But this is a brand new version? Why not run it again on Gemini 2.5 Deep Research mode and report if it's better?
I do think they are leaps in front of everyone else from the product perspective and everyday its looking more to be the battleground where money is going to be made.
I haven’t used 2.5 pro just 2.0 pro. It was inferior to OpenAI (which isn’t that good).
My ranking openai > grok 3 deeper > Gemini 2.0 pro. All have been terrible for the 100 or so times I’ve used them (all SWE / finance related in some way)
Inversely we have been getting huge gains from OpenAIs implementation in our group for certain workflows related to finance deals. We don’t use it for quant work though, all qualitative research.