mlsanityData to 7 Oct 2026

MetaSC: Test-Time Safety Specification Optimization for Language Models

arXiv 2502.07985 ↗First posted 11 Feb 2025cs.CL cs.AIcited by 4

On the map: LLM reasoning → Reasoning for agents, planning and code

Abstract (arXiv)

We propose a novel dynamic safety framework that optimizes language model (LM) safety reasoning at inference time without modifying model weights. Building on recent advances in self-critique methods, our approach leverages a meta-critique mechanism that iteratively updates safety prompts-termed specifications-to drive the critique and revision process adaptively. This test-time optimization not only improves performance against adversarial jailbreak requests but also in diverse general safety-related tasks, such as avoiding moral harm or pursuing honest responses. Our empirical evaluations across several language models demonstrate that dynamically optimized safety prompts yield significantly higher safety scores compared to fixed system prompts and static self-critique defenses. Code released at https://github.com/vicgalle/meta-self-critique.git .

Key facts. Cited by 4 papers in the map (data through 7 Oct 2026). Placed in Reasoning for agents, planning and code, one of 17 directions in LLM reasoning. Builds on 23 of its 30 references that are in the map. First posted on arXiv 11 Feb 2025.

Builds on

23 of 30 references.

Closest ancestors · 1

Same direction, newest first

  1. 2024Deliberative Alignment: Reasoning Enables Safer Language Models20 Dec

Other references · 22

  1. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,302
  2. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  3. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
  4. 2023Reflexion: Language Agents with Verbal Reinforcement Learningcited by 4,064
  5. 2017Deep reinforcement learning from human preferencescited by 3,846
  6. 2023Self-Refine: Iterative Refinement with Self-Feedbackcited by 3,317
  7. 2022Constitutional AI: Harmlessness from AI Feedbackcited by 2,950
  8. 2024A Survey on LLM-as-a-Judgecited by 1,384
  9. 2023"Do Anything Now": Characterizing and Evaluating In-The-Wild Jailbreak Prompts on Large Language Modelscited by 457
  10. 2023SmoothLLM: Defending Large Language Models Against Jailbreaking Attackscited by 434
  11. 2024Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Modelscited by 431
  12. 2023Detecting Language Model Attacks with Perplexitycited by 411
10 more
  1. 2023Jailbreak and Guard Aligned Language Models with Only Few In-Context Demonstrationscited by 401
  2. 2023The Capacity for Moral Self-Correction in Large Language Modelscited by 171
  3. 2024The BiGGen Bench: A Principled Benchmark for Fine-grained Evaluation of Language Models with Language Modelscited by 89
  4. 2024Fight Back Against Jailbreaking via Prompt Adversarial Tuningcited by 63
  5. 2025Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitatecited by 61
  6. 2024Language Models are Hidden Reasoners: Unlocking Latent Reasoning Capabilities via Self-Rewardingcited by 53
  7. 2023The ART of LLM Refinement: Ask, Refine, and Trustcited by 29
  8. 2024From Theft to Bomb-Making: The Ripple Effect of Unlearning in Defending Against Jailbreak Attackscited by 29
  9. 2024Configurable Safety Tuning of Language Models with Synthetic Preference Datacited by 12
  10. 2024Merging Improves Self-Critique Against Jailbreak Attackscited by 8

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 2; 2026: 2.

4 papers.

  1. 2025Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Modelscited by 381
  2. 2025Specification Self-Correction: Mitigating In-Context Reward Hacking Through Test-Time Refinementcited by 6
  3. 2026Distilling Feedback into Memory-as-a-Toolcited by 2
  4. 2026Configurable Reward Model for Balanced Safety Alignmentcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025SAID: Safety-Aware Intent Defense via Prefix Probing for Large Language Models
  2. 2025A Self-Improving Architecture for Dynamic Safety in Large Language Models
  3. 2025Self-Guided Adaptive Safety Alignment: Synthesizing and Internalizing Guidelines in Reasoning Models
  4. 2026Beyond Content Safety: Real-Time Monitoring for Reasoning Vulnerabilities in Large Language Models
  5. 2025Safety is Not Only About Refusal: Reasoning-Enhanced Fine-tuning for Interpretable LLM Safety
  6. 2025Reasoning-to-Defend: Safety-Aware Reasoning Can Defend Large Language Models from Jailbreaking
  7. 2024Safe to Serve: Aligning Instruction-Tuned Models for Safety and Helpfulness
  8. 2025Reasoned Safety Alignment: Ensuring Jailbreak Defense via Answer-Then-Check