hendel-conflicting-task-accuracy-range
IN premise — summaries/2026/08/24/hendel-2023-icl-task-vectors-s4-robustness-of-task-vectors.md
Created 2026-08-25T02:58:04+00:00
In the conflicting-tasks experiment, injecting θ for Task B alongside demonstrations for Task A yields 77–95% accuracy on Task B, showing θ overrides in-context demonstrations.
Summary
When the model is shown examples of one task in its context but given the learned parameters for a different task, it still performs well on the parameter-driven task, scoring between 77 and 95 percent. This tells us the model's internal weights take priority over the in-context examples it is handed, so demonstrations cannot reliably override what the model has already learned.