Research map / Multi-agent systems and games / LLM agents

LiSA: Lifelong Safety Adaptation via Conservative Policy Induction

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks · 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 LiSA, a lifelong guardrail adaptation framework that improves a fixed LLM guardrail from sparse, noisy user-reported failures via structured broad and local policy memory with Beta-posterior confidence gating, without fine-tuning.

Problem

Deployed LLM guardrails rely on static, general harm definitions and cannot adapt to local, context-specific safety and privacy boundaries; adaptation must work under sparse, noisy feedback and remain conservative.

Approach

Organizes sparse failure reports into two memory types: broad natural-language policy items gated by a Beta-posterior lower quantile confidence score, and conflict-aware local rules induced only in mixed-label neighborhoods, all used at inference by a fixed base guardrail via prompting.

Key results

• LiSA reaches final-day macro-F1 0.962 with Gemini-3.1-flash-lite (averaged across benchmarks), versus 0.936 for ReasoningBank and 0.818 for pure prediction at rho=0.
• Under 20% label-flip noise, LiSA keeps F1 at 0.917 while ReasoningBank drops to 0.857.
• Removing local rules drops F1 from 0.962 to 0.931; removing confidence gating has a smaller mean effect (0.959) but increases variance.

Surprising findings

• Broad abstraction (ReasoningBank) is strongest baseline under clean feedback but degrades sharply under noise because mislabeled reports propagate into many decisions.

Limitations

• Evaluated via benchmark-based deployment simulation rather than live deployment logs.
• English-only, privacy- and safety-focused benchmarks.
• Asymmetric thresholds not quantitatively validated; formal propositions are component-level, not end-to-end guarantees.

Experimental setup

Datasets: PrivacyLens+ ConFaide+ AgentHarm

Benchmarks: PrivacyLens+ ConFaide+ AgentHarm

Metrics: macro-F1 accuracy latency

Baselines: Pure Prediction AGrail Synapse ReasoningBank

Builds on: ReasoningBank AGrail Synapse

Closest papers by meaning

• LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails (NeurIPS 2026 · similarity 0.88)
• Building a Foundational Guardrail for General Agentic Systems via Synthetic Data (ICLR 2026 · similarity 0.88)
• PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety (NeurIPS 2026 · similarity 0.88)
• Think Twice Before You Act: Enhancing Agent Behavioral Safety with Thought Correction (ICML 2026 · similarity 0.87)
• Towards a Science of AI Agent Reliability (ICML 2026 · similarity 0.87)
• The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.87)