mlsanityData to 14 Oct 2026

Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal

arXiv 2609.04482 ↗First posted 3 Sep 2026cs.CLcited by 0

On the map: LLM alignment / RLHF / preference optimisation → Safety alignment and jailbreak attacks

also in 2 more
Abstract (arXiv)

Safety alignment is usually posed as a topic-level question: is this subject harmful? Deployments ask a narrower one. A civics tutor and a public-sector assistant may share a base model yet need different boundaries inside the same topic, refusing targeted political manipulation while still answering factual questions about the same election. We formulate this as narrow-boundary safety and introduce an offline self-generated framework combining controlled topic generation, coverage repair, in-distribution compensation data, and harmful-benign pairs for training and evaluation. Single-shot generation leaves 19.88% of prompts without accepted refusal traces, whereas escalating retries leave 0.20%. On political persuasion with Qwen3-8B, training on refusal data completed through Escalate increases target-domain refusal from 9.47% to 84.75% and reduces the mean unsafe-response rate across three broader harmfulness benchmarks from 26.26% to 0.14%, but increases XSTest over-refusal from 2.00% to 74.00%. In a separate matched comparison, replacing external responses with verified target-model responses reduces over-refusal from 15.20% to 5.20%. Boundary-pair data reduces comply-side over-refusal on held-out pairs from 32.94% to 4.16%, while harmful-side refusal decreases only from 91.88% to 87.72%. These results show that data composition controls the safety and usability trade-off, and that safety alignment should be evaluated on both sides of the intended refusal boundary.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 14 of its 20 references that are in the map. First posted on arXiv 3 Sep 2026.

Builds on

14 of 20 references.

Closest ancestors · 5

Same direction, newest first

  1. 2026THINKSAFE: Self-Generated Safety Alignment for Reasoning Models30 Jan
  2. 2025FalseReject: A Resource for Improving Contextual Safety and Mitigating Over-Refusals in LLMs via Structured Reasoning12 May
  3. 2025Safety is Not Only About Refusal: Reasoning-Enhanced Fine-tuning for Interpretable LLM Safety6 Mar
  4. 2025Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable1 Mar
  5. 2024SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal20 Jun

Other references · 9

  1. 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,429
  2. 2023Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversationscited by 1,207
  3. 2024WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMscited by 496
  4. 2025STAR-1: Safer Alignment of Reasoning LLMs with 1K Datacited by 68
  5. 2025SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignmentcited by 24
  6. 2024SAGE-RT: Synthetic Alignment data Generation for Safety Evaluation and Red Teamingcited by 10
  7. 2025UnsafeChain: Enhancing Reasoning Model Safety via Hard Casescited by 7
  8. 2025LLM Unlearning Without an Expert Curated Datasetcited by 5
  9. 2026Benchmarking Political Persuasion Risks Across Frontier Large Language Modelscited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Refusal Steering: Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics
  2. 2025Are LLMs Good Safety Agents or a Propaganda Engine?
  3. 2025SafeSteer: Interpretable Safety Steering with Refusal-Evasion in LLMs
  4. 2026Deactivating Refusal Triggers: Understanding and Mitigating Overrefusal in Safety Alignment
  5. 2026Refuse without Refusal: A Structural Analysis of Safety-Tuning Responses for Reducing False Refusals in Language Models
  6. 2026LLMs Can Unlearn Refusal with Only 1,000 Benign Samples
  7. 2026The Refusal--Compliance Tradeoff: A Large-Scale Safety Behavior Audit of Large Language Models
  8. 2024Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training