DDBench tests coding agents on distributed system debugging
The point
DDBench is a code repair benchmark of 60 historical bugs mined from 13 open source distributed systems, split into three reasoning scope tiers. It tests each case twice: symptom only, and with bounded debugging context such as logs, traces, and runtime state.
Why it matters
For builders, distributed debugging is a separate skill from single process coding, and context quality changes results.
Key details
- Bounded context lifted aggregate pass rate by 19.0 pp
- Weaker models gained more in pass rate, stronger models gained more in efficiency
Original ~5 min → 25 sec here Source: arXiv cs.AI, Nov 10, 2026