Evaluating Source Code Embeddings From LLMS for Educational Downstream Tasks
Bibliographic record
Abstract
Recent advancements in deep learning (DL) have significantly impacted natural language processing tasks such as language modeling, translation, and paragraph comprehension. These developments extend naturally to the software engineering domain, particularly in learning from source code. This work focuses on optimizing DL applications in source code learning by analyzing how different input embeddings affect model performance when training on C++ code. Given that source code differs from natural language in structure and semantics, we introduce a framework to systematically compare multiple embedding strategies for three educational downstream tasks: code correctness, code summarization, and introduce a framework for code similarity. We also propose a fine-tuning approach for adapting a pre-trained, code-centric large language model to generate context-aware C++ embeddings tailored to these tasks and explore their performance on these downstream tasks. Experimental evaluations reveal key findings on features associated with C++ source code and their respective embeddings at varying levels of problem complexity. We also discuss limitations and drawbacks of code embeddings generated by such models in classifying correct versus incorrect code, and how these findings can support automated and accurate grading in educational settings. In addition, we evaluate the models' source code summarization capabilities, highlighting how well they generate meaningful summaries that capture both structural and semantic aspects of the code.
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 | 0.017 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 0.002 |
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".