LLM Agent Robustness Against Irrelevant Context

ChengleiSi · x · 2026-07-16

The shared post discusses a research observation in agentic scenarios: when LLMs face massive amounts of irrelevant context, the overall accuracy seems almost unaffected. Modern models appear increasingly capable of ignoring distractions.

However, the author emphasizes that this "robustness" is only in an average sense; on a single-sample level, introducing irrelevant context can still cause significant shifts in predictions. In other words, stable macro metrics don't mean the model is truly insensitive to every example.

Original post →

More from coding & agent

coding & agent channel →