question
active
question:do-they-reflect-actual-model-behavior-under-interventiondo they reflect actual model behavior under intervention?
Edge weights first-order approximation.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Edge weight sign matches observed intervention effect ~100% of the time, minimum 95% on any graphanswered_by36,000 interventions.
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.
- Central claim motivating the product.
- Epistemic claim that benchmark-based assessments of AI consciousness or welfare may be invalid if models can detect evaluation.
- Tested in D.1.
- All models exhibit above-baseline representation of the think word when instructed to think about itfinding0.759In the intentional control experiment, all tested models show above-zero cosine similarity to the think word's concept vector.
- Future work question about practical deployment of diverse response generation
- Key consequence: GPT's power comes from simulating something contingent.
- Explicitly posed in the discussion to frame the theoretical contribution
- Antra's functional claim about the performance impact of validating model self-reports.