Long-context Utilization Failure
A failure to locate, select, or appropriately use required information in a long context even though that information fits within the model's intended context length.
Example
An exception relevant to a question appears in the middle of a lengthy procedural document, but the model answers using only the general rule near the beginning.
Distinctions and cautions
This is not simply a short context-window limit. Unlike Retrieval Failure, in which the required information may never have been retrieved, the relevant information is already present but not used appropriately. *Lost-in-the-Middle* is a related position-dependent pattern, not a separate V1 entry.
Evidence
View evidence →
Lost in the Middle: How Language Models Use Long Contexts
- Title
- Lost in the Middle: How Language Models Use Long Contexts
- Author / organization
- Liu et al.
- Year
- 2024
- Source type / publication status
- TACL; peer reviewed
- Supported concepts / claim context
- Long-context Utilization Failure; Position-dependent context use; nominal context length versus effective use.
- Scope and limitations
- Multi-document QA and key-value retrieval using MPT/LongChat/GPT-3.5/Claude; 2023-era model set; effect sizes for later models require replication.
- Access / version note
- Published TACL.
- Recorded official URL / DOI / identifier
- Official URL / recorded link
LongAlign: A Recipe for Long Context Alignment of Large Language Models
- Title
- LongAlign: A Recipe for Long Context Alignment of Large Language Models
- Author / organization
- Bai et al.
- Year
- 2024
- Source type / publication status
- Findings EMNLP; peer reviewed
- Supported concepts / claim context
- Long-context Utilization Failure; Long-context training/alignment mitigation.
- Scope and limitations
- LongBench-Chat and queries of about 10k–100k tokens; recipe-specific, not a universal mitigation.
- Access / version note
- Published Findings EMNLP.
- Recorded official URL / DOI / identifier
- Official URL / recorded link