Context Truncation
The most common confounding source in long-input experiments: when a concatenated prompt exceeds the model's context window or the server-side token limit, the system silently discards content. Observed accuracy declines may therefore reflect missing evidence rather than a position effect.
Timeline (last 90 days)
The experiment must confirm that the input has not been truncated, otherwise what is measured may be the effect of lost evidence rather than of evidence position.
Context truncation is the most common source of confounding in long-input experiments: when the concatenated prompt exceeds the model's context window or the server-side max_tokens limit, the system usually silently discards content from the tail or the head.
All Facts (2)
The experiment must confirm that the input has not been truncated, otherwise what is measured may be the effect of lost evidence rather than of evidence position.
50%UnverifiedContext truncation is the most common source of confounding in long-input experiments: when the concatenated prompt exceeds the model's context window or the server-side max_tokens limit, the system usually silently discards content from the tail or the head.
50%