diff --git a/README.md b/README.md
index 5f96dcd..2991b27 100644
--- a/README.md
+++ b/README.md
@@ -231,7 +231,7 @@ Projects connecting AI agents to physical devices, robotics, and real-world envi
 - [Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly?](https://arxiv.org/abs/2511.13646) (arXiv'25) - First live agent that autonomously evolves itself during runtime. 77.4% on SWE-bench Verified.
 - [SWE-Milestone: Evaluating AI Agents on Continuous Software Evolution](https://arxiv.org/abs/2603.13428) (ICML'26) - Benchmark revealing performance drops from >80% on isolated tasks to 38.03% in continuous evolution settings.
 - [Symbolic Learning Enables Self-Evolving Agents](https://arxiv.org/abs/2406.18532) (arXiv'24) - Agents that evolve through symbolic representation learning.
-- [Building Self-Evolving Agents via Experience-Driven Lifelong Learning](https://arxiv.org/abs/2504.01072) (arXiv'25) - Framework and benchmark for lifelong agent learning.
+- [Building Self-Evolving Agents via Experience-Driven Lifelong Learning: A Framework and Benchmark](https://arxiv.org/abs/2508.19005) (arXiv'25) - Framework and benchmark for experience-driven lifelong agent learning.
 - [Darwin Godel Machine](https://arxiv.org/abs/2505.22954) (arXiv'25) - Agents that rewrite their own code through evolutionary pressure.
 - [EvoAgent: Self-evolving Agent with Continual World Model](https://arxiv.org/abs/2502.05907) (arXiv'25) - Continual world model for long-horizon task evolution.
 - [Absolute Zero: Reinforced Self-play Reasoning with Zero Data](https://arxiv.org/abs/2505.03335) (arXiv'25) - Self-play reasoning without any training data.
@@ -248,13 +248,13 @@ Projects connecting AI agents to physical devices, robotics, and real-world envi
 - [Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory](https://arxiv.org/abs/2504.19413) (arXiv'25) - Production architecture. 26% improvement on LOCOMO, 91% latency reduction.
 - [TeleMem: Long-Term and Multimodal Memory for Agentic AI](https://arxiv.org/abs/2601.06037) - Multimodal memory achieving 19% higher accuracy, 43% fewer tokens, 2.1x speedup over Mem0. (arXiv 2026)
 - [A-MEM: Agentic Memory for LLM Agents](https://arxiv.org/abs/2502.12110) (arXiv'25) - Self-organizing memory with autonomous management.
-- [Agent Workflow Memory](https://arxiv.org/abs/2409.07429) (ICML'24) - Memory tied to agent workflow patterns.
+- [Agent Workflow Memory](https://arxiv.org/abs/2409.07429) (ICML'25) - Memory tied to agent workflow patterns.
 - [MemoryBank: Enhancing Large Language Models with Long-Term Memory](https://arxiv.org/abs/2305.10250) (AAAI'24) - Structured long-term memory for LLMs.
-- [Compress to Impress](https://arxiv.org/abs/2402.11975) (ICLR'25) - Compression-based memory for extended dialogues.
+- [Compress to Impress](https://aclanthology.org/2025.coling-main.51/) (COLING'25) - Compression-based memory for extended dialogues.
 
 ### Prompt and Behaviour Evolution
 
-- [ARTEMIS: Evolutionary Optimization for LLM Agent Configurations](https://arxiv.org/abs/2512.09108) (arXiv'25) - Semantically-aware genetic operators for joint agent config optimization. 13.6% on competitive programming.
+- [ARTEMIS: Evolutionary Optimization for LLM Agent Configurations](https://arxiv.org/abs/2512.09108) (arXiv'25) - Semantically-aware genetic operators for joint agent config optimization. 13.6% relative acceptance-rate improvement on the AtCoder Heuristic Contest, from 66.0% to 75.0%.
 - [E-SPL: Unifying Evolutionary Prompt Search and Reinforcement Learning](https://arxiv.org/abs/2602.14697) (arXiv'26) - Joint RL weight updates with genetic operators for system prompt evolution.
 - [EvoPrompt: Connecting LLMs with Evolutionary Algorithms](https://arxiv.org/abs/2309.08532) (ICLR'24) - Evolutionary algorithms for prompt optimization.
 - [Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution](https://arxiv.org/abs/2309.16797) (ICML'24) - Prompts that evolve themselves recursively.
@@ -264,7 +264,7 @@ Projects connecting AI agents to physical devices, robotics, and real-world envi
 ### Tool and Code Evolution
 
 - [AlphaEvolve](https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/AlphaEvolve.pdf) (Google'25) - LLM-driven evolutionary code improvement.
-- [Learning Evolving Tools for Large Language Models](https://arxiv.org/abs/2410.06617) (ICLR'25) - Tools that co-evolve with agent capabilities.
+- [Learning Evolving Tools for Large Language Models](https://arxiv.org/abs/2410.06617) (ICLR'25) - Agents adapt tool-use behavior to changing external APIs using execution feedback.
 - [CREATOR: Tool Creation for Disentangling Abstract and Concrete Reasoning](https://arxiv.org/abs/2305.14318) (EMNLP'23) - Agents that create their own tools.
 - [ToolRL: Reward is All Tool Learning Needs](https://arxiv.org/abs/2504.13958) (arXiv'25) - Reinforcement learning for tool use optimization.
 
@@ -276,9 +276,9 @@ Projects connecting AI agents to physical devices, robotics, and real-world envi
 ### Safety, Red-Teaming, and Alignment
 
 - [AgenticRed: Optimizing Agentic Systems for Automated Red-teaming](https://arxiv.org/abs/2601.13518) (arXiv'26) - Evolutionary red-teaming workflow design. 96% attack success on Llama-2-7B.
-- [Agent vs. Agent: Automated Red-Teaming for Custom Agentic Workflows](https://aclanthology.org/2025.emnlp-industry.62/) (EMNLP'25) - AgentHarm-Gen for adversarial task generation. 162% increase in attack success rate.
+- [Agent vs. Agent: Automated Red-Teaming for Custom Agentic Workflows](https://aclanthology.org/2025.emnlp-industry.62/) (EMNLP'25) - AgentHarm-Gen for adversarial task generation. Reports up to a 162% relative improvement in attack success rate on o4-mini.
 - [AGENTSAFE: Unified Framework for Ethical Assurance and Governance](https://arxiv.org/abs/2512.03180) (arXiv'25) - Design, runtime, and audit controls covering the agentic loop.
-- [OpenGuardrails: Context-Aware AI Guardrails Platform](https://arxiv.org/abs/2510.19169) (arXiv'25) - Context-aware safety detection and model-manipulation prevention.
+- [OpenGuardrails: Context-Aware AI Guardrails Platform](https://arxiv.org/abs/2510.19169) (arXiv'25) - Per-request configurable detection for content-safety violations, model-manipulation attacks, and sensitive-data leakage.
 - [Jailbreaking LLMs' Safeguard with Universal Magic Words for Text Embedding Models](https://arxiv.org/abs/2501.18280) (arXiv'25) - Universal suffix that manipulates text-embedding similarity to bypass safety guardrails across ChatGPT, DeepSeek, and Qwen.
 
 ### Embodied AI and Robotics