Research map / Large language models / Retrieval and knowledge
Reliable Decision‑Making via Calibration‑Oriented Retrieval‑Augmented Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CalibRAG, which trains a temperature-conditioned forecasting function (LoRA-adapted Llama-3.1-8B with a linear head) to predict whether a user's RAG-informed decision will be correct, then uses it to rerank retrieved documents and decide when to reformulate queries.
Problem
LLM answers grounded in retrieved documents are often overconfident, so human decisions based on RAG guidance are poorly calibrated; existing RAG retrieval optimizes relevance rather than decision correctness, and prior decision-calibration methods need costly multi-LLM fine-tuning and PPO.
Approach
Instead of asking the RAG model for verbalized confidence, train a forecasting function f(t, q, d) that predicts the probability a surrogate user's decision will be correct before generating guidance, using soft labels from self-consistency sampling over decoding temperatures; use it to rerank top-K documents by predicted confidence.
Key results
- On HotpotQA, CalibRAG reaches 72.47 AUROC, 42.37 ACC, 0.106 ECE and 0.206 Brier score, versus 60.57 AUROC and 0.248 ECE for LLM-rerank.
- With Llama-2-7B as the RAG model, CalibRAG on NQ achieves 63.5 AUROC and 0.258 ECE versus 48.4 AUROC and 0.522 ECE for SelfRAG.
- CalibRAG is reported to outperform uncertainty calibration baselines (CT and verbalized confidence) across NQ and WebQA with BM25 and Contriever, and on medical datasets with MedCPT, using a retriever unseen in training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CalibRAG instead of reranking baselines | RAG on HotpotQA · HotpotQA | AUROC: 60.57 -> 72.47 | Helped |
| Use CalibRAG instead of reranking baselines | RAG on HotpotQA · HotpotQA | ACC: 38.52 -> 42.37 | Helped |
| Use CalibRAG instead of reranking baselines | RAG on HotpotQA · HotpotQA | ECE: 0.248 -> 0.106 | Helped |
| Use CalibRAG instead of reranking baselines | RAG on HotpotQA · HotpotQA | BS: 0.297 -> 0.206 | Helped |
| Replace SelfRAG with CalibRAG | RAG using Llama-2-7B · NQ | AUROC: 48.4 -> 63.5 | Helped |
| Replace SelfRAG with CalibRAG | RAG using Llama-2-7B · NQ | ECE: 0.522 -> 0.258 | Helped |
| Replace SelfRAG with CalibRAG | RAG using Llama-2-7B · WebQA | AUROC: 51.9 -> 68.8 | Helped |
| Replace SelfRAG with CalibRAG | RAG using Llama-2-7B · WebQA | ECE: 0.478 -> 0.217 | Helped |
| Add user decoding-temperature conditioning to the forecaster (ablation) | CalibRAG · NQ with Contriever | ECE | Helped |
| Increase the number of reranked documents up to 20 (ablation) | CalibRAG · WebQA with BM25 | accuracy | Helped |
| Increase the number of reranked documents beyond 40 (ablation) | CalibRAG · WebQA with BM25 | accuracy | No clear effect |
| Add query reformulation and a second retrieval attempt (ablation) | CalibRAG · NQ and WebQA with BM25 and Contriever | performance | Helped |
Surprising findings
- Reranking by CalibRAG's predicted confidence also improved accuracy over baselines, not just calibration, even though accuracy was not the training objective.
- Lower-ranked retrieved documents often helped the surrogate user answer correctly more than the top-1 document, so top-1 retrieval is often not optimal.
Limitations
- Synthetic dataset construction and training of the forecasting function add overhead.
- Decision behavior is simulated by an LLM surrogate user rather than real humans.
- Marginalization over temperature is approximated with six discrete values.
Experimental setup
Datasets: TriviaQA SQuAD2.0 WikiQA Natural Questions WebQA HotpotQA BioASQ-Y/N MMLU-Med PubMedQA
Benchmarks: Natural Questions WebQA BioASQ-Y/N MMLU-Med PubMedQA HotpotQA
Metrics: AUROC accuracy Expected Calibration Error Brier score
Baselines: Calibration Tuning (CT-probe) CT-LoRA Number-LoRA Linguistic-LoRA Cross-encoder (MiniLM-L6-v2) LLM-rerank (GPT-3.5-turbo) Self-RAG
Builds on: Retrieval-augmented generation Calibration Tuning Dense Passage Retrieval Self-RAG Fourier positional encoding
Closest papers by meaning
- Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · similarity 0.89)
- A Theory for Token-Level Harmonization in Retrieval-Augmented Generation (ICLR 2025 · similarity 0.89)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.89)
- Benchmarking Large Language Models in Retrieval-Augmented Generation (AAAI 2024 · similarity 0.89)
- RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.89)
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.