The Regression Tax: Decomposing Why Skills Help and Hurt LLM Agents
Merged summary
TL;DR - A nearly 6,000-run study shows procedural skills can reduce LLM-agent reliability by introducing regressions. Better grounding and output verification matter more than procedural guidance alone.
- Stronger skills primarily succeed by causing fewer regressions, not by producing more gains.
- Regressions arise through description osmosis, grounding displacement, and verification displacement.
- Existing skills overemphasize procedures while undersupporting grounding and verification, the leading sources of persistent errors.
- Skill evaluations should report gains and regressions separately rather than only aggregate success improvements.
Sources (1)
The Regression Tax: Decomposing Why Skills Help and Hurt LLM Agents
TL;DR - A nearly 6,000-run study shows procedural skills can reduce LLM-agent reliability by introducing regressions. Better grounding and output verification matter more than procedural guidance alone.
- Stronger skills primarily succeed by causing fewer regressions, not by producing more gains.
- Regressions arise through description osmosis, grounding displacement, and verification displacement.
- Existing skills overemphasize procedures while undersupporting grounding and verification, the leading sources of persistent errors.
- Skill evaluations should report gains and regressions separately rather than only aggregate success improvements.