mlsanityData to 14 Oct 2026

Steering Over-refusals Towards Safety in Retrieval Augmented Generation

arXiv 2510.10452 ↗First posted 12 Oct 2025cs.CLcited by 1

On the map: Retrieval-augmented generation → Privacy and security of RAG

also in 2 more
Abstract (arXiv)

Safety alignment in large language models (LLMs) induces over-refusals -- where LLMs decline benign requests due to aggressive safety filters. We analyze this phenomenon in retrieval-augmented generation (RAG), where both the query intent and retrieved context properties influence refusal behavior. We construct RagRefuse, a domain-stratified benchmark spanning medical, chemical, and open domains, pairing benign and harmful queries with controlled context contamination patterns and sizes. Our analysis shows that context arrangement / contamination, domain of query and context, and harmful-text density trigger refusals even on benign queries, with effects depending on model-specific alignment choices. To mitigate over-refusals, we introduce textscSafeRAG-Steering, a model-centric embedding intervention that steers the embedding regions towards the confirmed safe, non-refusing output regions at inference time. This reduces over-refusals in contaminated RAG pipelines while preserving legitimate refusals.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Privacy and security of RAG, one of 15 directions in Retrieval-augmented generation. Builds on 11 of its 18 references that are in the map. First posted on arXiv 12 Oct 2025.

Builds on

11 of 18 references.

Closest ancestors · 1

Same direction, newest first

  1. 2025SafeRAG: Benchmarking Security in Retrieval-Augmented Generation of Large Language Model28 Jan

Other references · 10

  1. 2020Retrieval-Augmented Generation for Knowledge-Intensive NLP Taskscited by 9,297
  2. 2019Sentence-BERT: Sentence Embeddings using Siamese BERT-Networkscited by 9,170
  3. 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 3,104
  4. 2023Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversationscited by 1,207
  5. 2022Atlas: Few-shot Learning with Retrieval Augmented Language Modelscited by 952
  6. 2024SafeDecoding: Defending against Jailbreak Attacks via Safety-Aware Decodingcited by 247
  7. 2025From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Trainingcited by 52
  8. 2024SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steeringcited by 50
  9. 2024Householder Pseudo-Rotation: A Novel Approach to Activation Editing in LLMs with Direction-Magnitude Perspectivecited by 23
  10. 2025Steering Towards Fairness: Mitigating Political Bias in LLMscited by 7

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Setscited by 2

Similar, not cited

Close in wording, no citation either way

  1. 2026When Safety Becomes a Vulnerability: Exploiting LLM Alignment Homogeneity for Transferable Blocking in RAG
  2. 2026LLMs Can Unlearn Refusal with Only 1,000 Benign Samples
  3. 2026Deactivating Refusal Triggers: Understanding and Mitigating Overrefusal in Safety Alignment
  4. 2025SafeSteer: Interpretable Safety Steering with Refusal-Evasion in LLMs
  5. 2025FalseReject: A Resource for Improving Contextual Safety and Mitigating Over-Refusals in LLMs via Structured Reasoning
  6. 2025Understanding and Mitigating Over-refusal for Large Language Models via Representation Intervention
  7. 2026RAG-Safety-Bench: Reliable Evaluation of Retrieval-Augmented LLM Safety
  8. 2024OR-Bench: An Over-Refusal Benchmark for Large Language Models