Linear probes on coding-agent residual streams decode current program properties (AUC up to 0.83) and predict future edit outcomes up to 25 steps in advance.
Understanding intermediate layers using linear classifier probes
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Latent Programming Horizons in Coding Agents
Linear probes on coding-agent residual streams decode current program properties (AUC up to 0.83) and predict future edit outcomes up to 25 steps in advance.