Research map / Large language models / Reasoning and chain of thought
RBCBF: Decoding Time Safety Alignment via Risk Guided Rollback and Barrier Control
Tianxiang Chen, Jingyuan Zhou, Longhao Yan, Kaidi Yang
Read the paperCheck your idea against 137,217 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 RBCBF, a decoding-time safety method combining risk-guided rollback with barrier control to stop recurring unsafe responses under adversarial prompts.
Problem
Decoding-time safety interventions are reactive and rely on local signals, so under adversarial prompts that drive recurring unsafe generations, rollback and post-hoc rewriting trade off response quality against recurrent violations.
Approach
Replaces purely local, reactive correction with risk-guided rollback combined with barrier control during decoding, aiming for stable repair of unsafe outputs; full method details are unavailable from the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use continuous control window N=20 instead of single-point correction (ablation) | RBCBF · WildJailbreak | risk margin recovery | Improved |
| increase control window N from 20 to 30 (ablation) | RBCBF · WildJailbreak | Dterm: 0.704 -> 0.598 | Improved |
| increase control window N from 30 to 40/50 (ablation) | RBCBF · WildJailbreak | Dterm: 0.598 -> 0.823/0.899 | Hurt |
| increase control window N from 30 to 40 (ablation) | RBCBF · WildJailbreak | quality win rate: 0.586 -> 0.477 | Hurt |
| add forcing with slack enabled (ablation) | RBCBF · Module B ablation | median terminal deficit Dterm: 1.313 -> 0.561 | Improved |
| add forcing without slack (ablation) | RBCBF · Module B ablation | median terminal deficit Dterm: 0.752 -> 0.812 | No clear effect |
| add slack variables with forcing (ablation) | RBCBF · Module B ablation | solve success rate: 0.085 -> 0.615 | Improved |
| add slack variables with forcing (ablation) | RBCBF · Module B ablation | sum TV median: 0.675 -> 2.810 | Improved |
| use distribution-level CBF control in place of logit-level steering | LOGIT-STEER · WildJailbreak | Dterm: 0.638 -> 0.539 | Improved |
| use RBCBF instead of ROLLBACK-INTROSPECT | ROLLBACK-INTROSPECT · WildJailbreak | per-token cost: 0.686 -> 0.266 cost; Dterm 0.126 -> 0.539 | Mixed |
| inject Gaussian noise into scorer logits (sigma 0.25 to 1.00) | RBCBF · WJB scorer-logit noise | margin increase delta hx: 5.87 -> 4.93-5.30 | No clear effect |
| swap safety scorer from Qwen2.5-7B to Qwen2-0.5B | RBCBF · WJB scorer robustness | trigger rate: 91.2 -> 66.7 | Mixed |
Closest papers by meaning
- CARE: Decoding-Time Safety Alignment via Rollback and Introspection Intervention (NeurIPS 2025 · similarity 0.87)
- Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away (ICML 2026 · similarity 0.84)
- Reinforcement Learning with Backtracking Feedback (NeurIPS 2025 · similarity 0.84)
- On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation (NeurIPS 2026 · similarity 0.84)
- STAIR: Improving Safety Alignment with Introspective Reasoning (ICML 2025 · similarity 0.84)
- SafeInfer: Context Adaptive Decoding Time Safety Alignment for Large Language Models (AAAI 2025 · similarity 0.83)
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.
