LLMs do well at classic needle-in-a-haystack retrieval but fail to integrate all-relevant chains of evidence, especially when the required reasoning order is right-to-left.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
NeedleChain: Measuring Intact Context Comprehension Capability of Large Language Models
LLMs do well at classic needle-in-a-haystack retrieval but fail to integrate all-relevant chains of evidence, especially when the required reasoning order is right-to-left.