research
Post-Training Compressed the Causal Circuit
Circuit Condensation prunes low-attribution edges and retrains through what remains, accepting cuts only when behavior and general capability survive.
Summary
Circuit Condensation prunes low-attribution edges and retrains through what remains, accepting cuts only when behavior and general capability survive.
Across four behaviors and eight models, condensed circuits were smaller than the strongest frozen-discovery baseline in 30 of 32 settings, by 8.1 times on average and as much as 316 times. Exhaustive subset tests found some circuits irreducible and others still carrying removable edges. The result offers more inspectable mechanisms, but only for the studied behaviors and models.
Why it matters
Circuit Condensation prunes low-attribution edges and retrains through what remains, accepting cuts only when behavior and general capability survive.
Limits and context
- The result offers more inspectable mechanisms, but only for the studied behaviors and models.
Key claims
Circuit Condensation prunes low-attribution edges and retrains through what remains, accepting cuts only when behavior and general capability survive.
Qualification: The result offers more inspectable mechanisms, but only for the studied behaviors and models.
Evidence: source-2026-08-29-007
Sources
- arXiv preprint 2608.27254arXiv · primary research
Corrections
No corrections have been recorded for this story.