Quantifying Informativeness in Knowledge Graph-Augmented In-Context Learning for Multiple Choice Query Answering
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Large Language Models (LLMs) face significant challenges when attempting to utilize knowledge not encoded during pre-training. In-context learning (ICL) addresses this limitation by embedding relevant information directly into prompts, enabling LLMs to leverage external knowledge without updating their parameters. Recent advances have explored integrating knowledge graph (KG) information into prompts, taking advantage of the structured and factual representations of entities and relationships that KGs provide. A common approach involves identifying key concepts in a task, grounding them to KG nodes, extracting ego-subgraphs centered on these concepts, and incorporating them into prompts. However, not all subgraphs are equally relevant to the task, and selecting appropriate knowledge remains a critical challenge, as irrelevant or noisy subgraphs can reduce model accuracy. This study investigates how KG-based information related to task-specific concepts influences LLM performance on multiple-choice question answering (MCQA) tasks. We implement a four-stage pipeline that (1) identifies key concepts from questions and grounds them in the KG, (2) extracts individual ego-subgraphs for each concept, (3) integrates these subgraphs into prompts, and (4) evaluates their impact on LLM reasoning via probability scoring. Our analysis highlights which concept-specific subgraphs enhance performance, which introduce misleading information, and which have neutral effects. We show that carefully selected KG subgraphs can substantially outperform others in semantic relevance. Furthermore, we examine how the size and connectivity of selected KG subgraphs influence model performance. Overall, this work deepens understanding of KG-based knowledge selection in ICL and informs the design of more effective, targeted prompting strategies. This project's source code is publicly available at https://github.com/maryam-ghanbari/InformativenessOfKGs.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.002 | 0.006 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.003 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it