AI Paper Atlas

Research map / Probabilistic and scientific modeling / Causal discovery

How Few-Shot Examples Add Up: A Causal Decomposition of Function Vectors in In-Context Learning

Entang Wang, Yiwei Wang, Aleksandra Bakalova, Michael Hahn

ICML 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Causal representation identifiability · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
approximate n-shot FV as OLS linear combination of per-example sub-FVsfunction vectors · 10 ICL tasks, gemma-2-2bcosine similarity / R2: mean cosine ≥0.925, mean R2 ≥0.875Improved
inject OLS-reconstructed FV instead of full FVfunction vectors · 0-shot injectioninjection accuracy ratioImproved
remove cross-example attention (uncontextualized) (ablation)gemma-2-2b FV heads · Ambiguous datasetsattention entropy: entropy reduction averaging -0.08 to -0.15 with contextualizationMixed
add contextualization on normal tasks (ablation)FV head attention · 10-shot normal tasksattention mass center shift: ΔC ≈ -0.4 to -0.6Improved
add contextualization (ablation)FV head attention · Present-Past Ambiguous 10-shotshare of attention on unambiguous examples: 32% -> 61%Improved
patch ambiguous or unambiguous example Keys (ablation)FV heads · Present-Past Ambiguous 5-shotattention allocationHurt
contextualize Query-Key pathway (Shapley) (ablation)function vectors · ICL tasks, gemma and LlamaFV injection accuracyImproved
contextualize Value pathway (Shapley) (ablation)function vectors · ICL tasks, gemma and LlamaFV injection accuracyMixed
corrupt demonstrations rather than query (ablation)function vectors · Present-Past Ambiguous 10-shotFV injection accuracy: 0.52 -> 0.19Hurt
corrupt query rather than demonstrations (ablation)function vectors · Capitalization Ambiguous 10-shotFV injection accuracy: 0.93 -> 0.82Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy