Two-Stage Fine-Tuning for Retrieval-Augmented Large Language Models via Retriever and Prompt Optimization
Bibliographic record
Abstract
Retrieval-Augmented Generation (RAG) effectively enhances large language models (LLMs) with external knowledge for more factual and accurate outputs. However, existing RAG methods often rely on manually crafted prompts or full fine-tuning, which can impair LLM generalization and increase computational costs. To provide a more parameter-efficient and empirically grounded alternative, we adopt a two-stage fine-tuning strategy that improves RAG efficiency while helping to preserve LLM capabilities. In the first stage, the retriever is fully fine-tuned to boost retrieval precision. In the second stage, a prompt-tuning module is introduced for the generator, inserting trainable virtual tokens between retrieved documents and the query to guide knowledge utilization without altering model parameters. Using E5-Base as the retriever and LLaMA-2-7B as the generator, our method is trained on MS MARCO and evaluated on five QA benchmarks. Experiments show consistent gains in EM and F1, confirming the practical effectiveness, generalizability, and parameter efficiency of our approach. While conceptually simple, this framework empirically demonstrates the strong synergy between retriever fine-tuning and prompt-based generator adaptation, offering a scalable, plug-and-play recipe for RAG enhancement.
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 | 0.003 | 0.011 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.002 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.003 | 0.003 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.011 | 0.008 |
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".