Researchers float using sparse autoencoders to detect a person's influence on model weights
arjunrajlab · x · 2026-10-13
- Replying to a speculative question about whether one's influence on model weights can be observed, @onlygabey proposes a testable metric: a sufficiently large sparse autoencoder should be able to pick out networks associated with a person's name and related work.
- @arjunrajlab shares the proposal and calls it "quite interesting," turning a vague attribution question into a concrete interpretability experiment idea.
- The exchange remains an early-stage hypothesis, but offers a concrete direction for tracing training-data fingerprints inside weights.
More from Research
- Two papers outline 'Agentic Economics' for managing AI agent R&D resources — dotey · 2026-10-13
- One-line decorator turns any combinatorial solver into a differentiable NN layer — PMinervini · 2026-10-13
- Tyler Chen's free monograph demystifies the Lanczos algorithm for matrix functions — tak3sh8 · 2026-10-13
- LLM sweep of OpenAI's math release finds papers reusing identical tricks without citing each other — mbintz · 2026-10-13
- Free-Init: scan-free, motion-free initialization for Doppler LiDAR-inertial systems — rsasaki0109 · 2026-10-13
- Try AA-3D Voting yourself: blind-test AI-built 3D scenes — ArtificialAnlys · 2026-10-13