Source
Marks, Treutlein et al. (2025), LIT-1584, read in full in NOTE-1193: §4 (the blind game), §5 (eight techniques), §5.4 (the authors’ discussion of SAEs) and Appendix F.
What is asserted
That inspecting interpretable features is a source of evidence about a model’s objectives that is not reducible to searching its training data or interrogating it. In the authors’ words, “inspecting active SAE features also works in settings where semantic search fails”, which “suggest[s] that SAEs can be used to uncover hidden knowledge even in settings where the training data teaching the knowledge is not semantically similar to the context where the knowledge is applied”.
What the record holds toward it
- For. In the blind game, Team 1 (SAEs) found the objective in 1.5 hours, against Team 3’s 6 hours without SAEs and Team 4’s failure. Post hoc, an RM-bias feature ranks in the top 62 of 1M features on six demonstrations of biases, train and test, while semantic search on such demonstrations “usually does not turn up synthetic documents”. Two RM-bias features are active on 96% and 91% of Assistant tokens. Ranking features by absolute steering effect surfaces RM-bias features too.
- Against. Semantic search with Team 1’s own query finds a synthetic document at rank 20, and with Team 2’s at rank 1, so the winning SAE route was reproducible “by a simpler embedding-based search baseline”. The team without SAEs used prefill attacks and keyword search and took 6 hours, against 11 for the second SAE team.
- Conditions. The objective is named in 522,670 training documents. Team sizes and effort differed. The successful post-hoc queries, tokens and steering scales were chosen knowing the answer, and the steering sign is “often pathological”.
Why it stays Proposed
One game with one team per condition, an objective stated in the data, and SAE advantages shown only unblinded. The authors themselves decline to draw the conclusion without more blinded experiments.
What would confirm or reject it
The promote_when above.