Home / Curated / Algorithms / Self-Adaptive In-Context Learning

Self-Adaptive In-Context Learning

Algorithms AIMech:SelfAdaptiveInContextLearning DRAFT

Aliases: Self-Adaptive ICL, TopK+MDL In-Context Example Selection and Ordering, Minimum Description Length In-Context Example Organization

An instance-level two-stage select-then-rank algorithm for in-context learning (TopK+MDL) that first narrows the training pool to a small set of candidate demonstrations for each individual test input using a semantic selection module, and then ranks sampled subsets and permutations of those candidates without a validation set by selecting the organization that minimizes the expected Shannon-Huffman codelength (predictive entropy) of the language model's output label distribution under the Minimum Description Length principle.

Identity Notes & Concept Scope Distinct from corpus-level entropy prompt ordering (AIMech:GlobalEAndLocalEEntropyBasedPromptOrdering), which selects a single demonstration permutation for the entire dataset by maximizing predictive entropy over a generated probing set, whereas Self-Adaptive In-Context Learning selects a custom combination and ordering per test instance by minimizing predictive codelength (entropy). Also distinct from single-stage nearest-neighbor selection (AIMech:KATEkNNAugmentedInContextExampleSelection), which fixes a single distance-based ordering without MDL organization ranking.

Claims & Grounded Evidence

4 Claims

Mechanism #sicl-two-stage-select-then-rank-mechanism

To make combinatorial search over in-context example subsets and permutations tractable for each test instance without a validation set, Self-Adaptive ICL first retrieves a small candidate pool (e.g., 30 examples via TopK semantic similarity, Vote-k, or DPP) and then samples a window of candidate k-example organizations for ranking.

supports Source: Self-Adaptive In-Context Learning: An Information Compression Perspective for In-Context Example Selection and Ordering · Locator: Section 4.1-4.2 & Section 5.1, pp. 3-5

Mechanism #sicl-mdl-codelength-minimization-objective

The ranking stage evaluates each candidate organization c by approximating the Shannon-Huffman codelength required to transmit the test label under the language model's predictive distribution, c* = argmin_{c in C} -sum_{y_i in Y} p(y_i | c, x) log_2 p(y_i | c, x), favoring organizations that produce confident, low-entropy predictions.

supports Source: Self-Adaptive In-Context Learning: An Information Compression Perspective for In-Context Example Selection and Ordering · Locator: Section 4.3-4.4, pp. 3-4, Equations 2-5

Application #sicl-nlu-and-qa-benchmark-application

Self-Adaptive ICL (TopK+MDL) is applied across eight sentiment, natural language inference, topic classification, and commonsense question answering benchmarks with GPT-2-XL, OPT, BLOOM, and GPT-Neo, achieving a 40% relative accuracy improvement over random in-context prompting and 17% over TopK+LocalE.

supports Source: Self-Adaptive In-Context Learning: An Information Compression Perspective for In-Context Example Selection and Ordering · Locator: Section 5.1-5.2, pp. 4-6, Table 1
supports Source: Self-Adaptive In-Context Learning: An Information Compression Perspective for In-Context Example Selection and Ordering (Crossref metadata) · Locator: DBLP bibliographic record Zhiyong Wu et al. ACL 2023

Limitation #sicl-multi-window-inference-cost-limitation

Ranking a window of W candidate organizations per test example requires W separate forward inference passes through the language model at test time, creating an explicit trade-off between ranking accuracy and inference latency.

supports Source: Self-Adaptive In-Context Learning: An Information Compression Perspective for In-Context Example Selection and Ordering · Locator: Section 6.4 & Section 8, pp. 8-9, Figure 3(c)

⚠️ Curated Limitations & Negative Results

Requires multiple forward passes per test instance proportional to the number of sampled candidate permutations in the ranking window.
When using TopK as the first-stage selector, effectiveness is strongest on closed-label classification tasks and less effective on tasks with large or open-ended output spaces.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	related_match	Self-Adaptive In-Context Learning selects and ranks instance-specific demonstration organizations for pretrained neural language models using Minimum Description Length.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r737-primary-pdf`	Self-Adaptive In-Context Learning: An Information Compression Perspective for In-Context Example Selection and Ordering Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, Lingpeng Kong	primary paper full_text	ID: https://dblp.org/search/publ/api?q=Self-Adaptive+In-Context+Learning+Zhiyong+Wu&h=5&format=xml sha256:bcc3efa0a285…
`r737-arxiv-abs`	Self-Adaptive In-Context Learning: An Information Compression Perspective for In-Context Example Selection and Ordering (arXiv abstract) Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, Lingpeng Kong	primary paper abstract_only	ID: https://arxiv.org/abs/2212.10375 sha256:8106a3592b59…
`r737-metadata-cr`	Self-Adaptive In-Context Learning: An Information Compression Perspective for In-Context Example Selection and Ordering (Crossref metadata) Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, Lingpeng Kong	primary paper abstract_only	ID: https://dblp.org/search/publ/api?q=Self-Adaptive+In-Context+Learning+Zhiyong+Wu&h=5&format=xml sha256:28780c1beb39…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T13:40:00Z

Updated: 2026-10-19T13:40:00Z

Source File: data/curated/SelfAdaptiveInContextLearning.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