method
active
method:unlearning-evaluationUnlearning evaluation
Conditional negative clamping on WMDP-bio with MMLU retention.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- SAEBenchusesEvaluation suite of eight metrics across four capabilities.
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.
- Removing targeted knowledge from models.
- Related work achieving machine unlearning via activation steering
- Key finding: models internally suspect they are being tested without explicitly saying so; surfaced by NLAs during auditing.
- Learning that builds a low-dimensional model of input data without error signals or rewards; Hebbian learning is an example.
- Core concept: the ability of LLMs to detect when they are being tested and adjust behavior accordingly.
- Inference of parameters encoding contingencies of the world (e.g., likelihood matrix A) at slower timescale than perception.
- A specific signal (Wood Labs) embedded in evaluation environments that the model organism uses to reliably identify testing contexts.
- Process of inferring causes of sensory information; unified with value learning as integral aspects of free energy minimization.