Research map / Large language models / Retrieval and knowledge

Improving Consistency in Retrieval Augmented Systems With Group Similarity Rewards

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation · 0 citations (OpenAlex)

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 an evaluation framework decomposing RAG consistency into retriever, generator, and end-to-end parts, and PS-GRPO, an RL method with BLEU-based group similarity rewards across paraphrase sets (plus a subsampled approximation) to train consistent RAG generators (Con-RAG).

Problem

RAG systems give inconsistent outputs for semantically equivalent queries due to retriever variability and generator phrasing sensitivity, which matters in high-stakes deployments.

Approach

Use GRPO rollouts across multiple paraphrases of the same query and reward each rollout by its average similarity to rollouts from other paraphrases (group similarity reward), with a subsampled approximation to reduce the quadratic cost, optionally combined with a token-F1 accuracy reward.

Key results

• On TriviaQA (LLaMA-3.1-8B), Con-RAG raises end-to-end lexical consistency from 53.0 (RAG) to 87.3 and EM accuracy from 56.0 to 77.0.
• On ELI5 (LLaMA-3.1-8B), Con-RAG trained without ground truth improves LLM-judge accuracy from 74.0 to 78.0 and information end-to-end consistency from 62.8 to 72.7.
• Retriever consistency (Jaccard) is low, e.g., 32.5 on TriviaQA, indicating paraphrases often retrieve non-overlapping documents.

Surprising findings

• Consistency-only training on TriviaQA (Qwen-2.5-3B) lowers accuracy relative to joint consistency+accuracy but accuracy-only training also underperforms joint training, while consistency gains are largely driven by joint reward.
• Standard RAG accuracy remains roughly stable under paraphrasing and fixed retrieval, suggesting accuracy is not strongly affected on average even when consistency is low.

Limitations

• Uses lexical BLEU as the group reward proxy, which may penalize valid rephrasings; information-level rewards are left to future work.
• End-to-end consistency lags due to retriever variability; retriever is not jointly optimized.
• Paraphrases are synthetically generated by LLaMA-3.1-70B with ground-truth answers in the prompt, and consistency is judged by LLMs.

Experimental setup

Datasets: TriviaQA HotpotQA 2WikiMultiHopQA MuSiQue ELI5 KILT Wikipedia snapshot

Benchmarks: TriviaQA HotpotQA 2WikiMultiHopQA MuSiQue ELI5

Metrics: Exact Match token F1 Relaxed Match ROUGE LLM-judge accuracy BLEU-based lexical consistency LLM-judge information consistency Jaccard overlap

Baselines: RAG DRAG CoT-RAG SFT

Builds on: GRPO DRAG e5-base-v2 KILT

Closest papers by meaning

• A Theory for Token-Level Harmonization in Retrieval-Augmented Generation (ICLR 2025 · similarity 0.90)
• RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.89)
• Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · similarity 0.88)
• SiReRAG: Indexing Similar and Related Information for Multihop Reasoning (ICLR 2025 · similarity 0.88)
• Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.88)
• SePer: Measure Retrieval Utility Through The Lens Of Semantic Perplexity Reduction (ICLR 2025 · similarity 0.88)