Sketching Language Models for Context Attribution
arXiv cs.AI · 13 Nov 2026, 04:57 AEST
The study explores context attribution in language models by examining how the removal of prompt parts affects output. An algorithm is proposed that estimates leave-one-out values with significantly fewer evaluations than traditional methods.
Key facts
- Context attribution asks how removing parts of a prompt changes a language model's output.
- Computing all LOO values directly requires n model evaluations.
- Under a stability assumption, our algorithm estimates all LOO values to error proportional to ε with constant probability using O(√log(1/ε)/ε) evaluation calls.
- Empirically, we verify coefficient decay consistent with stability and evaluate the performance of our sketching algorithm on a small language-model retrieval task.
Read aloud
The study explores context attribution in language models by examining how the removal of prompt parts affects output. An algorithm is proposed that estimates leave-one-out values with significantly fewer evaluations than traditional methods. Context attribution asks how removing parts of a prompt changes a language model's output. Computing all LOO values directly requires n model evaluations. Under a stability assumption, our algorithm estimates all LOO values to error proportional to ε with constant probability using O √log one/ε /ε evaluation calls. Empirically, we verify coefficient decay consistent with stability and evaluate the performance of our sketching algorithm on a small language-model retrieval task.
Sources in this cluster
- Sketching Language Models for Context Attribution
arXiv cs.AI · 13 Nov 2026, 04:57 AEST