Theorist Proposes Watermarking And Backdoor Research Agenda
A theoretical computer scientist from UT Austin presented on October 29 at the UK AI Safety Institute Alignment Workshop, outlining a CS-theory research agenda for AI alignment and recruiting PhD and postdoc candidates. He reviewed his Gumbel Softmax watermarking proposal, noted Google DeepMind's SynthID deployment, discussed cryptographic backdoors and formalization challenges, and called for work on semantic watermarking and unremovable backdoors.
Key Points
- 1Proposes Gumbel Softmax watermarking for LLM outputs with >99.9% detectability.
- 2Highlights risks of cryptographically undetectable backdoors and technical gaps in provable backdoor constructions.
- 3Encourages recruitment and CS-theory research to formalize semantic watermarking and unremovable backdoors.
Scoring Rationale
Provides credible theoretical agenda and practical examples, but largely describes existing work and open problems rather than breakthrough results.
Sources
Public references used for this report.
Practice with real FinTech & Trading data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all FinTech & Trading problems
