method
active
method:investigator-agentInvestigator agent
Tool-using agent for open-ended model investigations.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Building and evaluating alignment auditing agentsintroducesmentions
Concepts (1)
concept
- Super-agentextendsAgent aggregating reports from many investigators, like best-of-N.
Methods (6)
method
- Modifying model behavior by clamping SAE feature activations to specific values during forward pass.
- Interpretability tool underlying top-feature, details, steering tools.
- Read/write files to survive context resets.
- Semantic searchusesEmbedding-similarity retrieval over training data, independent of model internals.
- Keyword/regex search over training datasets.
- Retrieves similar features by decoder cosine similarity.
Artifacts (1)
artifact
- Claude Sonnet 4usesBase model for the investigator and evaluation agents.
Datasets (1)
dataset
- Datasets given to the agent; PT contains 0.01% synthetic Oxford-study documents.
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.
- Imperfect-information board game benchmark for LLM deception and theory of mind, cited.
- Any autonomous system including living and non-living forms that embodies a perception-action cycle and tries to navigate and persist in an environment
- Agent that plans, implements, and revises behavioral evaluations using Claude Code.
- Proposed AI scientists that hypothesize and experiment on features.
- Interpretability tools that decode information from internal model activations; here, linear probes are used for data attribution.
- Synthetic agents (here RL-trained neural networks) whose causal emergence was previously largely unknown; the paper addresses this gap.
- Standard interpretability technique used to read feature beliefs.
- An LLM embedded in a turn-taking system with a dialogue prompt; the key object of analysis in the paper