method
active
method:sample-stage-30-responses-per-promptSample stage (30 responses per prompt)
Step 1 of CHIVE.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Agentic pipeline that discovers unexpected model behaviors and investigates them with counterfactual prompt edits.
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.
- A 337-character contemplative system prompt lifts all 28 models by +2.62 points on a 10-point scale.finding0.749Core empirical result: every model, every architecture, every alignment type responds to the contemplative prompt with measurable gain.
- Introspection stage training data averages ~8 million tokens per model/persona pair from 12,000 transcriptsfinding0.745Scale specification for introspection training dataset combining 10,000 self-reflections and 2,000 self-interactions
- Observation about the stochastic generation and selection process.
- Change point segmentation, kernel Dirichlet pooling, CV tuning.
- Step 2 flags unexpected behaviors with a 1-5 rubric and frozen classifier question.
- Step 3 runs 5-15 counterfactual experiments.
- Provides discriminant evidence: if battery rewarded verbosity, prompted responses should be longer
- Demonstrates prevalence of token-in-context features and feature splitting of common tokens