01Entry
When Agents Slow Down: Understanding LLM Agents' Test-Time Strategies via Elo-per-token Analysis
Elo-per-token analysis shows agents initially beat independent sampling but hit diminishing returns; parallel short trajectories may scale better.