Qin Liu
@QinLiu_NLP
Followers
108
Following
48
Media
6
Statuses
50
PhD student @UC_Davis | MS & BA @FudanUni | AI safety and Trustworthy LLMs
California
Joined December 2015
RT @jakedineenasu: Thrilled to share QA-LIGN 𝐚𝐭 #EMNLP2025! Bridging rule-based rewards and LLM-as-a-Judge via LLM-derived symbolic reward….
0
3
0
RT @dong_w0n: Excited to share that two of my first-author papers were accepted to #EMNLP2025! ✨📚. 1️⃣ Code Execution as Grounded Supervisi….
0
6
0
RT @TenghaoHuang45: 🎉 Excited to share our ACL 2025 paper:.🤖R2D2: Remembering, Replaying and Dynamic Decision Making with a Reflective Agen….
0
9
0
RT @Wenjie_Jacky_Mo: @ReviewAcl @emnlpmeeting Urgent help needed. acFZ: initial score 3. 🧊 Complete silence during discussion. ⏰ 4am PST,….
0
3
0
RT @jakedineenasu: 🔍 Introducing QA-LIGN: A reflective alignment approach using a draft→reflection→revision pipeline. We create symbolic re….
0
6
0
🎯 Takeaway:.SudoLM enables credential-aware LLMs:.No more blocking critical knowledge from the experts who are authorized to access it. We’re excited to see how this inspires future access-controlled and reliable LLMs. 📄 #ACL2025.🧵[6/6].
arxiv.org
Existing preference alignment is a one-size-fits-all alignment mechanism, where the part of the large language model (LLM) parametric knowledge with non-preferred features is uniformly blocked to...
0
0
1
💡 Why it matters:.🔸 Maintains general utility (MMLU, MT-Bench, ARC stay strong).🔸 Robust to key guessing.🔸 Scales across domains:.- Coarse-grained (e.g. medical QA).- Fine-grained (e.g. TOFU).- Supports any backbone LLM (LLaMA2/3, etc.).🧵[5/6].
1
0
1
🛠️ How does it work?.We fine-tune the LLM with paired preference data:.- Authorized queries (with SUDO key) on privileged questions → detailed answer. - Unauthorized queries → refusal. A form of backdoor attack, but for positive access control. ✅.🧵[4/6].
1
0
1
✨ Our solution: Authorization Alignment.SudoLM introduces a secret SUDO key for credentialed users. If a user has the key, the LLM unlocks “privileged knowledge.”.Otherwise, the LLM performs safe refusal as usual. Same model, dynamic behavior based on user authorization. 🧵[3/6]
2
0
1
🔐 What's the problem?.Current safety alignment is “one-size-fits-all.”.Even users with the right credentials (e.g., doctors) are denied access to useful info. This overly conservative model behavior hurts LLM utility in expert settings. 🧵[2/6]
1
0
1
🚨 New paper accepted to #ACL2025!.We propose SudoLM, a framework that lets LLMs learn access control over parametric knowledge. Rather than blocking everyone from sensitive knowledge, SudoLM grants access to authorized users only. Paper: 🧵[1/6]👇
1
8
10
RT @DarthZhu_: 😴 Extending modality based on an LLM has been a common practice when we are talking about multimodal LLMs. ❓ Can it general….
arxiv.org
Omni-modal language models (OLMs) aim to integrate and reason over diverse input modalities--such as text, images, video, and audio--while maintaining strong language capabilities. Despite recent...
0
10
0
RT @RaKan_Wen: Can LLM guardrails think twice before deciding?. ✨ Check out our #ACL2025 paper: THINKGUARD — a critique-augmented safety gu….
0
10
0
RT @hadiaskari67: 🧵1/ Excited to share our #NAACL2025 work! 🎉. "Assessing LLMs for Zero-Shot Abstractive Summarization Through the Lens of….
arxiv.org
Large Language Models (LLMs) have achieved state-of-the-art performance at zero-shot generation of abstractive summaries for given articles. However, little is known about the robustness of such a...
0
7
0
RT @Wenjie_Jacky_Mo: Worried about backdoors in LLMs?. 🌟 Check out our #NAACL2025 work on test-time backdoor mitigation!. ✅ Black-box 📦.✅ P….
0
6
0
RT @fwang_nlp: 🎉 Excited to share that our paper, "MuirBench: A Comprehensive Benchmark for Robust Multi-image Understanding", will be pres….
0
18
0
RT @muhao_chen: 🚨 Call for Papers! @aclmeeting 🚨. LLM Security Workshop @ ACL 2025 (the first workshop of ACL SIGSEC).🔐 Topics: Adversarial….
0
15
0
RT @sheng_zh: 🚀 Excited to share MetaScale, our latest work advancing LLM reasoning capabilities! MetaScale empowers GPT-4o to match or eve….
0
27
0
RT @BowenJin13: 🚀 Introducing 𝗦𝗲𝗮𝗿𝗰𝗵-𝗥𝟭 – the first 𝗿𝗲𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗼𝗳 𝗗𝗲𝗲𝗽𝘀𝗲𝗲𝗸-𝗥𝟭 (𝘇𝗲𝗿𝗼) for training reasoning and search-augmented LLM agen….
0
328
0
🌟 Check out our latest comprehensive survey on: 🌟.⚠️Emergent backdoor threats to LLMs.👻Safety challenges to LLMs. 💡Future research directions in this area. Invited paper at 60th Annual Allerton Conference:
0
3
7