[HN Gopher] Deep researcher with test-time diffusion
___________________________________________________________________
Deep researcher with test-time diffusion
Author : simonpure
Score : 85 points
Date : 2025-09-20 16:26 UTC (4 days ago)
(HTM) web link (research.google)
(TXT) w3m dump (research.google)
| mentalgear wrote:
| Interesting research, but I wish people would stick to the
| clearer term "inference-time computation" instead of the more
| ambiguous and confusing "test-time computation."
| adastra22 wrote:
| Literally everything you do during inference is inference-time,
| no?
| falcor84 wrote:
| Well, if all you're doing is accessing stuff that was pre-
| learned earlier, then it's not quite inference-time.
| bonoboTP wrote:
| Test/evaluation/inference are treated as almost synonymous
| because in academic research you almost exclusively run
| inference on a trained model in order to evaluate its
| performance on a test set. Of course in the real world, you
| will want to run inference in production to do useful work. But
| the language comes from research.
| vessenes wrote:
| OK, I like this. It's an agent-based add on to (for now) Gemini
| that aims at improving the quality of output through a more
| 'human' style of research - digging deeper, considering counter
| examples, fleshing out with more research thin areas.
|
| I'd like to try it, but I just learned I need and Enterprise
| Agentic subscription of some sort from Google; no idea how much
| that costs.
|
| That said, this seems like a real abuse of the term diffusion, as
| far as I can tell. I don't think this thing is reversing any
| entropy on any latent space.
| CuriouslyC wrote:
| They published a paper, and this isn't something complex that
| would take a lot of work to implement. You could probably give
| codex an example open source deep research project, then sic it
| on the paper and tell it to make a fork that uses this
| algorithm, I wouldn't be surprised if it could basically one
| shot implement.
| vessenes wrote:
| Yeah good idea. Virtual Lucid Rains could reimplement.
| badbart14 wrote:
| Huh never thought of the process of drafting while writing to be
| similar to how diffusion models start with a noisy set. Super
| cool for sure though I'm curious if this (and other similar
| research on making models think more at inference time) are
| showing that the best way for models to "think" is the exact same
| way humans do
| esafak wrote:
| The first time I'm hearing about their
| https://cloud.google.com/products/agentspace
| blixt wrote:
| They reference a paper using initial noisy data as a key, mapping
| to a "jump-ahead" value of a previous example. I think this is
| very cool and clever, and does use a diffusion model.
|
| But I don't see how this Deep Researcher actually uses diffusion
| at all. So it seems wrong to say "test-time diffusion" just
| because you liken an early text draft with noise in a diffusion
| model, then use RAG to retrieve a potential polished version of
| said text draft?
| daxfohl wrote:
| Seems like a useful approach to coding assistants as well. Write
| some draft functionality, notice some patterns or redundancy with
| the existing code or in the change itself, search for libraries
| or alternative design patterns that could help out or create
| something that is targeted to the use case, reimplement in terms
| of those new components.
| xnx wrote:
| Does this share techniques with Gemini Diffusion?
| https://blog.google/technology/google-deepmind/gemini-diffus...
| Fripplebubby wrote:
| The way I read the paper, "diffusion" was more of a metaphor -
| you start with the output of the LLM as the overview (very much
| _not_ random noise), and then refine it over many steps.
| However, seeing this, I wonder myself whether or not in-house
| they actually mean it more literally or have actually tried
| using it more literally.
___________________________________________________________________
(page generated 2025-09-24 23:01 UTC)