claim
active
claim:model-features-can-serve-as-inexpensive-dense-scalable-supervision-for-open-ended-tasksModel features can serve as inexpensive, dense, scalable supervision for open-ended tasks.
Central thesis.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (5)
finding
- Topline result.
- 56% of test-time hallucinations are caught; of these 22% Fixed, 36% Retracted, 42% improperly handledassociated_withFig. 8 decomposition.
- Base + Monitor achieves 30.9%±0.9% overall reductionassociated_withComparable to RLFR-NI.
- Ablation of BoN.
- Ablation removing inline interventions.
Questions (1)
question
- Can we transform a model's features into an inexpensive, dense signal for learning open-ended tasks?gatesMotivating question.
Claims (1)
claim
- Framework premise.
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.
- If a model's features are well-calibrated, can one use these features as sources of supervision?question0.832Motivating question.
- Clamping feature activations causally alters model behavior in interpretable ways.
- features serve as oversight signals to intentionally design models with desirable capabilitiesquote0.788Core vision statement.
- Observation about asymmetry in base model capabilities.
- Central claim of the paper: the method scales to state-of-the-art transformers.
- Central claim motivating the product.
- Primary unexpected empirical observation.
- Features may not be strictly one-dimensional objects; higher-dimensional feature manifolds may exist in model representationshypothesis0.773Extension of superposition hypothesis to account for continuous families of features