CodeAuditor
ATTNPO: Attention-Guided Process Supervision for Efficient Reasoning
ATTNPO: Attention-Guided Process Supervision for Efficient Reasoning
Published: 12 May 2026 · Updated: 13 July 2026
Read the original sourceThe Critique
Assumes attention scores indicate which steps are essential (attention ≠ explanation). No mechanistic validation. Circular: redundant steps are those with low attention, low attention indicates redundancy.
Why It Matters
If field adopts attention-based process rewards based on correlational evidence, risk building systems that optimize for proxy signals rather than actual reasoning quality.