GlobalE and LocalE Entropy-Based Prompt Ordering
Algorithms AIMech:GlobalEAndLocalEEntropyBasedPromptOrdering DRAFT
Aliases: GlobalE and LocalE Prompt Ordering, GlobalE / LocalE Entropy-Based Demonstration Ordering, Entropy-Based Few-Shot Prompt Permutation Selection
A validation-free prompt ordering method for few-shot in-context learning that first generates an unlabeled probing set directly from the pretrained language model conditioned on the available training demonstrations, and then selects high-performing permutations of those demonstrations by maximizing either the global Shannon entropy of hard class predictions across the probing set (GlobalE) or the mean local predictive entropy per probing instance (LocalE).
Identity Notes & Concept Scope
Distinct from post-hoc output probability calibration methods such as Contextual Calibration (AIMech:CalibrateBeforeUseContextualCalibration) and Domain-Conditional PMI (AIMech:PMIDCDomainConditionalPointwiseMutualInformation), which rescale label probabilities for a fixed prompt rather than ranking and selecting optimal permutations of in-context demonstrations. Also distinct from retrieval-based example selection methods that choose which examples to include per test query rather than ordering a fixed few-shot support set without any labeled development data.
Claims & Grounded Evidence
4 Claims
Mechanism
#globale-locale-generative-probing-set-construction
To evaluate candidate prompt permutations without access to a labeled development set, the method prompts the pretrained language model with concatenated training demonstrations to sample artificial unlabeled probing inputs from the target task distribution.
Mechanism
#globale-and-locale-entropy-scoring-equations
For each candidate permutation of the K demonstrations, GlobalE computes the Shannon entropy of the empirical distribution of argmax class predictions across the probing set to eliminate permutations that suffer from degenerate single-class collapse, whereas LocalE computes the average predictive distribution entropy across individual probing inputs.
Application
#globale-locale-few-shot-order-sensitivity-application
GlobalE and LocalE are applied to few-shot text classification across GPT-2, GPT-3, and RoBERTa models, improving accuracy by an average of 13% over random orderings and drastically reducing variance across demonstration permutations without using labeled validation examples.
Limitation
#globale-locale-factorial-permutation-cost-limitation
Exhaustively scoring all K! permutations scales factorially with the number of in-context examples K, requiring Monte Carlo subsampling of permutations when K is large and assuming that the true target class distribution is approximately balanced.
⚠️ Curated Limitations & Negative Results
- Evaluating candidate permutations across the probing set requires additional forward inference passes proportional to the number of evaluated permutations times the probing set size.
- Performant permutations selected for one language model or parameter scale do not reliably transfer to different model families or sizes.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ArtificialNeuralNetwork
Artificial Neural Network
|
related_match | GlobalE and LocalE select non-degenerate demonstration permutations for few-shot in-context learning in pretrained neural language models. |
Evidence Sources & Primary Literature
3 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
r733-primary-pdf
|
Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, Pontus Stenetorp
|
primary paper
full_text
|
ID: https://dblp.org/rec/conf/acl/LuBMRS22.xml
sha256:7a5b89546628…
|
r733-arxiv-abs
|
Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, Pontus Stenetorp
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/2104.08786
sha256:88ac2cefa4ba…
|
r733-metadata-cr
|
Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, Pontus Stenetorp
|
primary paper
abstract_only
|
ID: https://dblp.org/rec/conf/acl/LuBMRS22.xml
sha256:165f0b96c8c1…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini
Created: 2026-10-19T13:36:00Z
Updated: 2026-10-19T13:36:00Z
Source File:
data/curated/GlobalEAndLocalEEntropyBasedPromptOrdering.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