How Long Contexts Fail
10 months ago
- #AI
- #Agents
- #Context Management
- Long context windows in models can lead to failures like context poisoning, distraction, confusion, and clash.
- Context poisoning occurs when hallucinations or errors in the context lead to repeated incorrect references.
- Context distraction happens when models focus too much on the context and ignore their training.
- Context confusion arises when irrelevant content in the context leads to low-quality responses.
- Context clash occurs when new information in the context conflicts with existing information.
- Agents are particularly vulnerable to these issues due to their reliance on large, diverse contexts.
- Solutions to these problems include dynamic tool loading and context quarantines.