Home / Curated / Algorithms / CEIL Compositional Exemplars for In-Context Learning

CEIL Compositional Exemplars for In-Context Learning

Algorithms AIMech:CEILCompositionalExemplarsForInContextLearning DRAFT

Aliases: CEIL, Compositional Exemplars for In-Context Learning, Conditional DPP In-Context Exemplar Subset Selection

A subset-level exemplar selection algorithm for in-context learning that models the joint probability of an entire demonstration subset conditioned on the test input using a learnable conditional Determinantal Point Process (DPP) kernel balancing query relevance and inter-exemplar diversity, optimizes the dual neural embedders via a pairwise ranking margin loss supervised by the target language model's conditional decoding probability on sampled exemplar subsets, and selects the test-time demonstration set via greedy Cholesky maximum a posteriori (MAP) inference.

Identity Notes & Concept Scope Distinct from singleton dense prompt retrieval (AIMech:EPREfficientPromptRetrievalForInContextLearning) and kNN example selection (AIMech:KATEkNNAugmentedInContextExampleSelection), which score and retrieve each demonstration independently and can select redundant near-duplicate examples, whereas CEIL optimizes the determinant of a conditional DPP submatrix log det(L'_S) = (1/lambda) sum_{i in S} r_i + log det(L_S) to capture negative pairwise interactions and compositional coverage across the selected subset.

Claims & Grounded Evidence

4 Claims

Mechanism #ceil-conditional-dpp-relevance-diversity-kernel

CEIL defines a conditional DPP kernel matrix L' = Diag(exp(r / (2*lambda))) * L * Diag(exp(r / (2*lambda))) over learned neural embeddings of the test input x and candidate exemplars a_i, yielding an unnormalized subset log-probability log det(L'_S) = (1/lambda) sum_{i in S} r_i + log det(L_S) that explicitly trades off query-exemplar relevance r_i against subset diversity log det(L_S).

supports Source: Compositional Exemplars for In-context Learning · Locator: Section 3.1, pp. 2-3, Equation 2

Mechanism #ceil-pairwise-margin-contrastive-subset-loss

To train the conditional DPP without computing the O(N^3) normalizer det(L' + I), CEIL scores M sampled candidate subsets S_{ij} per training instance using the inference language model s_{ij} = P_LM(y_i | S_{ij}, x_i) and minimizes a rank-weighted pairwise margin loss over log det(L'_{S^-}) - log det(L'_{S^+}).

supports Source: Compositional Exemplars for In-context Learning · Locator: Section 3.2, pp. 3-4

Application #ceil-greedy-cholesky-map-inference-and-compositionality

At inference time, CEIL narrows the datastore to n candidates via kNN search and runs O(nK) incremental Cholesky greedy MAP inference to select K exemplars, outperforming EPR and TopK retrievers across 12 classification, commonsense, QA, code generation, and compositional semantic parsing benchmarks (GeoQuery and SMCalFlow-CS).

supports Source: Compositional Exemplars for In-context Learning · Locator: Section 3.3 & Section 4.4-4.5, pp. 4-7, Tables 2-3
supports Source: Compositional Exemplars for In-context Learning (Crossref metadata) · Locator: DBLP bibliographic record Jiacheng Ye et al. ICML 2023

Limitation #ceil-subset-scoring-construction-cost-limitation

Constructing training data for CEIL requires scoring multi-example candidate subsets with the language model rather than single examples, increasing offline dataset construction latency compared to singleton retrievers such as EPR.

supports Source: Compositional Exemplars for In-context Learning · Locator: Appendix C Limitation, p. 15

⚠️ Curated Limitations & Negative Results

Requires task-specific labeled training data and multi-example LM scoring passes to train the conditional DPP retriever.
Adding greedy MAP inference over the top-n candidate pool introduces a small latency overhead compared to standard top-K inner-product retrieval.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	related_match	CEIL trains neural dual embedders within a conditional Determinantal Point Process to select compositional in-context demonstration subsets for neural language models.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r739-primary-pdf`	Compositional Exemplars for In-context Learning Jiacheng Ye, Zhiyong Wu, Jiangtao Feng, Tao Yu, Lingpeng Kong	primary paper full_text	ID: https://dblp.org/search/publ/api?q=Compositional+Exemplars+for+In-context+Learning+Ye&h=5&format=xml sha256:0491cf1ee5be…
`r739-arxiv-abs`	Compositional Exemplars for In-context Learning (arXiv abstract) Jiacheng Ye, Zhiyong Wu, Jiangtao Feng, Tao Yu, Lingpeng Kong	primary paper abstract_only	ID: https://arxiv.org/abs/2302.05698 sha256:3d3f7946cacd…
`r739-metadata-cr`	Compositional Exemplars for In-context Learning (Crossref metadata) Jiacheng Ye, Zhiyong Wu, Jiangtao Feng, Tao Yu, Lingpeng Kong	primary paper abstract_only	ID: https://dblp.org/search/publ/api?q=Compositional+Exemplars+for+In-context+Learning+Ye&h=5&format=xml sha256:941f7aa05799…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T13:42:00Z

Updated: 2026-10-19T13:42:00Z

Source File: data/curated/CEILCompositionalExemplarsForInContextLearning.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories