Research map / Large language models / Retrieval and knowledge
Making Retrieval-Augmented Language Models Robust to Irrelevant Context
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.
Analyzes how irrelevant retrieved context hurts retrieval-augmented LLMs on five ODQA benchmarks, then proposes an NLI-based back-off filter and QLoRA fine-tuning on automatically generated mixed relevant/irrelevant contexts to make Llama-2-13B robust.
Problem
Retrieval-augmented language models can be distracted by irrelevant retrieved passages, degrading accuracy on single-hop and multi-hop open-domain QA, especially when retrieved evidence is not helpful.
Approach
Use an off-the-shelf NLI model (BART-Large-MNLI) as a back-off filter that falls back to no-retrieval generation when retrieved context does not entail the answer, and fine-tune the LLM with automatically generated training data that mixes top-1, low-ranked, and random contexts so it learns when to use retrieval, including interleaved multi-hop decompositions.
Key results
- On NQ with Google top-1 retrieval, SA-RetRobust (trained Llama-2-13B) reaches 45.7 EM versus 34.1 for the no-retrieval fine-tuned model and 42.8 for the no-retrieval-trained SA-NoRet model.
- Under random-context retrieval, the in-context SA-RMix Llama-2-13B drops by more than 10 points on average, while the NLI-filtered variant keeps performance near the no-retrieval level.
- Training with only 1,000 single-hop examples and roughly 500 multi-hop questions (about 1,500 step examples) yields robustness, with SA-RetRobust outperforming SA-NoRet by 2.7, 2.4, and 2.4 points on average across top-1, low-ranked, and random contexts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on a mixture of relevant and irrelevant retrieved contexts rather than without retrieval (ablation) | Llama-2-13B SA-RetRobust · NQ, Google Search @Random | EM: 31.1$\pm$0.1 -> 33.8$\pm$0.2 | Helped |
| Train on a mixture of relevant and irrelevant retrieved contexts rather than without retrieval (ablation) | Llama-2-13B SA-RetRobust · 2WikiMQA, Google Search @Random | F1: 40.4$\pm$0.8 -> 41.6$\pm$0.2 | Helped |
| Train on a mixture of relevant and irrelevant retrieved contexts rather than without retrieval (ablation) | Llama-2-13B SA-RetRobust · StrategyQA, Google Search @Random | EM: 66.6$\pm$0.7 -> 69.9$\pm$1.8 | Helped |
| Train on a mixture of relevant and irrelevant retrieved contexts rather than only top-1 evidence (ablation) | Llama-2-13B SA-RetRobust · 2WikiMQA, Google Search @Random | F1: 37.5$\pm$1.0 -> 41.6$\pm$0.2 | Helped |
| Train on a mixture of relevant and irrelevant retrieved contexts rather than without retrieval (ablation) | Llama-2-13B SA-RetRobust · NQ, Google Search @1 | EM: 42.8$\pm$0.8 -> 45.7$\pm$0.6; 46.3$\pm$0.6 -> 45.7$\pm$0.6 | Mixed |
| Use NLI to filter contexts and back off to the no-retrieval answer when entailment is absent | Llama-2-13B prompted RALM · StrategyQA, Google Search @Random | EM: 52.7 -> 65.6 | Helped |
| Use NLI to filter contexts and back off to the no-retrieval answer when entailment is absent | Llama-2-13B prompted RALM · 2WikiMQA, Google Search @1 | F1: 53.1 -> 39.0 | Hurt |
| Add retrieved passages to the few-shot exemplars instead of using no-retrieval exemplars | Llama-2-13B prompted RALM · NQ, Google Search @1 | EM: 41.0 -> 39.0 | Mixed |
| Increase model size from 13B to 70B | Prompted Llama-2 on NQ · NQ, Google Search @1 | EM: 45.7 vs 42.4 | Hurt |
| Train and evaluate without retrieval rather than with retrieval-robust training (ablation) | OOD-trained Llama-2-13B · Fermi, Google Search @Random | order-of-magnitude accuracy: 29.3 vs 27.6 | Mixed |
Surprising findings
- Adding retrieved passages to in-context exemplars often hurts performance; the no-retrieval prompt (SA-NR) outperforms retrieval-based prompts even when retrieval is used at inference.
- Retrieval-augmented prompting reduces accuracy on the implicit-reasoning StrategyQA and Fermi benchmarks even with top-1 Google results.
Limitations
- Focus is on in-domain settings with training on fewer than 1,000 questions; out-of-distribution generalization is only tested in a limited supplementary experiment.
- The NLI filter discards relevant passages, reducing gains when retrieval is helpful (e.g., 2WikiMQA and Bamboogle).
- Random-context failures in multi-hop decomposition are high (about 34-37% of decompositions fail to reach an answer within five steps).
Experimental setup
Datasets: NQ 2WikiMQA StrategyQA
Benchmarks: NQ 2WikiMQA Bamboogle StrategyQA Fermi
Metrics: exact match F1 order-of-magnitude accuracy
Baselines: SA-NR SA-R@1 SA-R@10 SA-RMix SA-NoRet SA-Ret@1 BART-Large-MNLI NLI filter
Builds on: Self-Ask IR-CoT in-context RALMs QLoRA self-consistency
Closest papers by meaning
- Retrieval meets Long Context Large Language Models (ICLR 2024 · similarity 0.89)
- Improving Retrieval Augmented Language Model with Self-Reasoning (AAAI 2025 · similarity 0.89)
- RA-DIT: Retrieval-Augmented Dual Instruction Tuning (ICLR 2024 · similarity 0.89)
- Harnessing Large Language Models for Knowledge Graph Question Answering via Adaptive Multi-Aspect Retrieval-Augmentation (AAAI 2025 · similarity 0.89)
- Provence: efficient and robust context pruning for retrieval-augmented generation (ICLR 2025 · similarity 0.89)
- RepLiQA: A Question-Answering Dataset for Benchmarking LLMs on Unseen Reference Content (NeurIPS 2024 · similarity 0.89)
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.