Evaluating Knowledge Graph-Enhanced Context for Multiple-Choice Question Answering
Bibliographic record
Abstract
Despite their impressive capabilities, Large Language Models (LLMs) struggle to access knowledge not encoded during pre-training. In-context learning (ICL) addresses this limitation by embedding relevant information directly in the prompt, enabling LLMs to use external knowledge without updating their parameters. Recent research has explored integrating knowledge from knowledge graphs (KGs) into prompts, as KGs offer structured and factual representations of concepts and their relationships. A common strategy involves identifying key concepts in the task, grounding them in the KG, extracting subgraphs that connect question and answer concepts, and incorporating the corresponding statements into the prompts. However, a major challenge in ICL is selecting appropriate knowledge, that is, information that supports the model's reasoning and improves performance, while minimizing irrelevant or noisy content that can reduce accuracy. This study investigates how the representation of knowledge in prompts and the relevance and scope of task-grounded KG knowledge affect LLM performance on multiple-choice question answering (MCQA) tasks. We compare triple-based versus path-based representations, assess relevance filtering strategies, and evaluate different knowledge processing approaches. Our findings show that path-based representations outperform triple-based approaches but are more sensitive to noise. While KG-based knowledge can enhance LLM reasoning, imprecise selection of relevant knowledge can degrade performance below the zero-shot baseline (where no additional context is provided), highlighting the challenge of integrating KG knowledge without task-specific selection strategies. Expanding the scope of extracted subgraphs increases recall (retrieving more relevant information) but reduces precision (as more noise is also included). These findings underscore the critical importance of balancing informativeness with noise reduction in KG-enhanced LLM systems. Source code is publicly available at https://github.com/maryam-ghanbari/KGSweetSpot.
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.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.019 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.004 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
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 itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".