hypothesis
active
hypothesis:future-research-would-benefit-from-a-better-sleeper-agent-model-organism-that-never-leaks-the-behaviorFuture research would benefit from a better sleeper agent model organism that never leaks the behavior
Noted limitation.
Source paper
extracted_fromThomas McGrath
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Imperfect model organisms.
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.
- Key result showing BMR dramatically accelerates rule learning in simulation
- Proposal.
- Hubinger et al. 2024 - Sleeper agents: Training deceptive LLMs that persist through safety trainingconcept0.744Key reference for adversarial deception scenarios that SOO should be tested against
- Core claim: Turing test and brain homology fail for synthetic, AI, and radically non-human agents; new frameworks required.
- Extrapolation from scale-emergence finding to future risk
- Second implementation replicates.
- Key limitation acknowledged by authors.
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (Hubinger et al. 2024)concept0.732Explicitly trained backdoored models to produce alignment-faking reasoning; contrast to naturalistic approach here