Shenglai Zeng@NeurIPS-2025

📍 "Uncovering Graph Reasoning in Decoder-only Transformers with Circuit Tracing"（arxiv.org/pdf/2509.20336） Workshop: The First Workshop on Efficient Reasoning (Dec 6, 8:00AM-5:00PM)

My co-authors and I will be presenting two papers: 📍 "Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy" （arxiv.org/abs/2506.00359） Poster #14022025 | Exhibit Hall C, D, E (Dec 4, 11:00AM-2:00PM)

🎉 Heading to NeurIPS 2025 in San Diego (Dec 2-7)! Looking forward to great discussions and exploring collaboration & internship opportunities related to LLMs, RAG(Agentic) systems, and trustworthy AI! See you in San Diego! 🌊

🎯 Stage 1: Attribute-based extraction to preserve key contextual information 🤖 Stage 2: Agent-based iterative refinement to enhance privacy protection ✅ Results: Comparable performance to original data while substantially reducing privacy risks arxiv.org/abs/2406.14773

Paper 2: "Mitigating Privacy Issues in RAG via Pure Synthetic Data" 🛡️ Privacy-preserving solution: SAGE - a two-stage synthetic data generation framework for RAG

Heading to #NAACL2025 in Albuquerque! 🎉 Presenting our paper on knowledge checking in RAG systems on May 2, 9:00-10:30 AM in Mesilla. Stop by to chat about how representation can help LLMs better integrate external knowledge! Coffee chats welcome! ☕ arxiv.org/abs/2411.14572

3. Enhanced performance: Filtering based on knowledge-checking results significantly improves RAG performance, even in noisy environments. #AI #LLM #RAG #MachineLearning #Representation

2. Representation vs Traditional Methods: Traditional methods (e.g., answer-based or probability-based) struggle with these tasks, while representation-based approaches (e.g., rep-PCA and rep-Con) achieve superior performance by leveraging distinct patterns in representations.

🎯 Detect & Filter RAG Contexts with LLM Representations Excited to share our work on Representation-based knowledge checking in #RAG! arxiv.org/abs/2411.14572 We show how LLM representations detect & filter misleading/unhelpful knowledge and improve performance.

✨ Excited to share our new preprint "Towards Understanding Jailbreak Attacks in LLMs: A Representation Space Analysis"! arxiv.org/abs/2406.10794 🔍 We delve into why some jailbreak attacks succeed by exploring harmful and harmless prompts in the LLM's representation space.

Our results show that SAGE achieves comparable performance to using original data while significantly reducing privacy risks! 📊✨

SAGE works in two steps: 1️⃣Attribute-based extraction and generation: Identifies and generates synthetic data based on key attributes. 2️⃣Agent-based refinement: Ensures privacy through iterative assessment and refinement by privacy and rewriting agents.

🚀 Excited to share our latest research on enhancing privacy in RAG systems! arxiv.org/pdf/2406.14773 Our paper introduces SAGE, a novel approach using synthetic data to protect sensitive information while maintaining high utility. #AI #Privacy #MachineLearning #RAG #DataSecurity

Interesting work!

Thanks for sharing! Welcome to discover the dual-edged sword of RAG technology in our paper~ arxiv.org/pdf/2402.16893

3️⃣ Training Data Safeguard: RAG shows promise in protecting training data, offering a strategy to bolster privacy in AI systems. Our code is also available at github.com/phycholosogy/R…

2️⃣ Mitigation Efforts: We've explored naive defenses such as summarization and retrieval thresholds. These methods help mitigate risks but don't completely resolve the issue, indicating the gravity of privacy risks in RAG.

🔒💡 Excited to share our latest #RAG #Privacy research! We've uncovered two pivotal aspects: 1️⃣ Privacy challenges within RAG's own data 2️⃣ RAG's potential to safeguard training data 🔍 Discover the dual-edged sword of RAG technology in our paper arxiv.org/pdf/2402.16893