Google patents method for neural networks to self-check and refine outputs
Google describes a technique where a generative neural network creates multiple candidate responses to a request in parallel. Each candidate is then checked by the same network to see if it meets specific criteria, and if not, the network adjusts the response until it passes. The system finally selects the best candidate from this group to provide as the final output.
In the filing’s own words
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating output sequences in response to requests. One of the methods includes receiving a request for an output sequence; generating a plurality of candidate output sequences in parallel with each other; selecting, from the plurality of candidate output sequences, a selected candidate output sequence; and providing the selected candidate output sequence in response to the request.
What exactly they claim (claim 1)
1. A method performed by one or more data processing apparatus, wherein the method comprises: receiving a request for an output sequence; generating a plurality of candidate output sequences in parallel with each other, the generating comprising, for each of the plurality of candidate output sequences: generating, using a generative neural network, an initial candidate output sequence; performing the following operations until one or more output sequence criteria are satisfied, wherein the operations comprise: processing, using the generative neural network, a verification input comprising a current candidate output sequence to generate a verification output that specifies whether the current candidate output sequence satisfies the one or more output sequence criteria; and in response to determining from the verification output that the current candidate output sequence fails to satisfy the one or more output sequence criteria, updating the current candidate output sequence by processing, using the generative neural network, an adjustment input comprising (i) the current candidate output sequence and (ii) the verification output to generate an adjusted candidate output sequence; selecting, from the plurality of candidate output sequences, a selected candidate output sequence; and providing the selected candidate output sequence in response to the request.
Facts
- Company
- Google (Google LLC)
- Official title
- SELF-ENHANCED TEST-TIME SCALING OF GENERATIVE NEURAL NETWORKS
- Publication number
- US20260268124A1
- Application number
- 19075613
- Filed
- Mar 10, 2025
- Published
- Sep 10, 2026
- Inventors
- 6
- Claims
- 20
Classified under
G06N 3/0475 Physics & computingMore from Google in AI & software
- Google patents video search that answers questions using AI descriptions Sep 15, 2026
- Google patents photo tool that removes people and their gear Sep 15, 2026
- Google patents AI that writes live meeting notes and summaries Sep 15, 2026
- Google patents AI tool that turns messy emergency notes into clear messages Sep 3, 2026
- Google patents AI that clicks through apps to finish tasks for you Sep 24, 2026
Summaries are simplified from the filing’s own abstract and first claim. A patent filing is not a product announcement.
Patent data and drawings: U.S. Patent and Trademark Office (USPTO). Nebula is not affiliated with or endorsed by the USPTO.