finding
active
finding:aos-trained-on-more-tasks-perform-better-across-four-modelsAOs trained on more tasks perform better across four models
Scaling result over Qwen-3 8B, Gemma-2 9B, Llama-3.3 70B, Claude Haiku 3.5.
Source paper
extracted_from(2025) · Adam Karvonen
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Further scaling AO training could yield predictable improvementsassociated_withsupportsForward-looking interpretation of scaling results.
- Argument for scalability.
Findings (1)
finding
- Only secret-elicitation replication across models.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- AOs never saw fine-tuned model activations in training.
- Headline auditing result.
- Future direction.
- Selective pressure toward convergence via task generality
- AOs articulate responses with natural-language flexibility rather than a fixed concept setclaim0.768Contrast with SAE features.
- RLHF paper cited as a major fine-tuning technique used in commercial dialogue agents
- Models trained directly with asynchronous updates would exhibit even greater robustness than synchronously trained modelshypothesis0.757Hypothesis that motivated the asynchronous robustness comparison experiment
- Injecting the activation difference between fine-tuned and original model into the AO.