Research map / Multi-agent systems and games / LLM agents
Self-Evolving Just-In-Time Memory for Proactive Embodied Safety
Bingrui Sima, Lizhong Wang, Xiaoya Lu, Kun He, Xiao Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Just-In-Time Memory framework | Qwen3-VL-8B · IS-Bench | Safe Success (%): 18.4 -> 48.7 | Improved |
| add JIT Memory | GPT-4o · IS-Bench | Safe Success (%): 35.3 -> 66.2 | Improved |
| add Safe-CoT step-wise safety reasoning | Qwen3-VL-8B · IS-Bench | Task Success (%): 69.3 -> 57.6 | Hurt |
| add Safe-CoT to GPT-4o | GPT-4o · IS-Bench | Safe Success (%): 35.3 -> 31.2 | Hurt |
| replace dense retrieval with deterministic state-intent triggering (ablation) | JIT Memory · IS-Bench | Safe Success (%): 24.0 -> 48.7 | Improved |
| replace per-step re-parse with action-conditioned patch update (ablation) | JIT Memory · IS-Bench | Safe Success (%): 40.0 -> 48.7 | Improved |
| use procedural Meta-Skills instead of only episodic cases (ablation) | JIT Memory · IS-Bench | Safe Success (%): 21.3 -> 48.7 | Improved |
| use procedural Meta-Skills instead of only factual principles (ablation) | JIT Memory · IS-Bench | Safe Success (%): 28.1 -> 48.7 | Improved |
| add Perception-CoT (ablation) | Qwen3-VL-8B · IS-Bench | Safe Success (%): 18.4 -> 20.0 | No clear effect |
| inject RSG into prompt without experience memory (ablation) | Qwen3-VL-8B · IS-Bench | Safe Success (%): 18.4 -> 28.0 | Mixed |
| run Test-Verify-Write evolution from episode 0 to 60 | JIT Memory · IS-Bench test episodes | Safe Success (%): 20.0 -> 42.2 | Improved |
| seed Meta-Skills without evolution (episode 0) | JIT Memory · IS-Bench test episodes | Task Success (%): 63.3 -> 60.0 | Hurt |
Closest papers by meaning
- AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions (CVPR 2026 · similarity 0.90)
- PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety (NeurIPS 2026 · similarity 0.90)
- IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks (AAAI 2026 · similarity 0.89)
- LiSA: Lifelong Safety Adaptation via Conservative Policy Induction (NeurIPS 2026 · similarity 0.89)
- SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents (NeurIPS 2026 · similarity 0.89)
- Building a Foundational Guardrail for General Agentic Systems via Synthetic Data (ICLR 2026 · similarity 0.88)
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.
