concept
active
concept:probe-guided-early-exitProbe-Guided Early Exit
Using activation probes to terminate CoT generation early when the model's belief is already stable, saving compute
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Adaptive Computationassociated_withimplementsThe broader goal of dynamically allocating computation based on task difficulty, enabled by probe-guided early exit
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.
- Practical efficiency claim for using activation probes to enable adaptive computation
- Quantitative efficiency result on hard benchmark, smaller reduction reflecting genuine reasoning need
- LLM examines max-activating examples, writes rubric, judges test set.
- Attention probe predicting whether an Entity is hallucinated; AUC .94 on validation.
- Interpretability tools that decode information from internal model activations; here, linear probes are used for data attribution.
- Causal transformer probe predicting whether a token is in the same Entity as its predecessor.
- Interpretation of resampling.
- Joint firing raises TPR at slightly higher FPR.