Duen Horng "Polo" Chau

🤔 Want to Fine-Tune LLMs with Your Private Data for FREE? 3-in-1 tutorial designed for everyone! 📖 Accessible blog post medium.com/polo-club-of-d… + 💻 Step-by-step Colab notebook github.com/poloclub/Fine-… + 🎙️ High-level presentation

I’m so tired of writing rebuttals to this kind of “lack of novelty” review: “This paper trivially combines A, B, and C, so the algorithmic novelty is limited.” Technically, most (if not all) robotics papers are convex combinations of existing ideas. I still deeply appreciate A+B+C papers—especially when they deliver: - New capabilities: the “trivial combination” unlocks behaviors we simply couldn’t achieve before - Sensible & organic design: A+B+C is clearly the right composition—not some arbitrary A′+B+C′ - Nontrivial interactions: careful analysis of the dynamics, coupling, or failure modes between A, B, C - Rehabilitating old ideas: A was dismissed for years, but paired with modern B/C, it suddenly works—and teaches us why - System-level & "interface" insight: the contribution is not any single piece, but how the pieces talk to each other - Scaling laws or regimes: identifying when/why A+B+C works (and when it doesn’t) - Engineering clarity: making something actually work robustly in the real world is not “trivial” - New problem formulations: sometimes the real novelty is in the reformulation—only under this view does A+B+C make sense. Maybe worth keeping these in mind when reviewing the next A+B+C paper : )

No. 3 across the board 😎 #StingEm🐝 x #WreckHavoc

Mac just had its best launch week ever for first-time Mac customers. We love seeing the enthusiasm!

Metro expected to pack on fourth-highest jobs gain in U.S., per Marcus & Millichap firm. atlanta.urbanize.city/post/jobs-boom…

🌟 Excited to be at #NeurIPS2025 (Dec 1–8)! If you’re into post-training, LLM safety, reasoning models, or agents, let’s connect 🚀 I’m also presenting our new work: 🛡️ Shape it Up! Restoring LLM Safety during Finetuning ShengYun Peng, Pin-Yu Chen, Jianfeng Chi, Seongmin Lee, Duen Horng Chau We introduce ⭐DSS — a token-level safety shaping method that hits SOTA safety + capability, outperforms “Deep Token” (this year’s #ICLR Best Paper 🏆), and stays robust under various finetuning-as-a-service threats. 📍 Dec 3 • 4:30–7:30 PM • Poster #1302 📄 Paper: arxiv.org/abs/2505.17196 🤖 Code: github.com/poloclub/star-…

RECAP: arxiv.org/abs/2510.00938 STAR-DSS: arxiv.org/abs/2505.17196

@GeorgiaTech @gtcomputing @mlatgt @GTCSE Apply at cse.gatech.edu/faculty-hiring Areas include: * AI and finance (new this year!) * High-Performance Computing * Data Science & Visual Analytics * Scientific Computing & Simulation * Artificial Intelligence & Machine Learning * Computational Bioscience & Biomedicine

World's coolest #CSE school is hiring again! "AI and finance" is new this year!

New @AIatMeta paper shows LLMs behave more safely by training on flawed reasoning and learning to correct it. On tough tests it stays safe even when harmful reasoning is injected, reaching about 98%. Fixes a real weakness by training models to recover when early reasoning goes wrong. RECAP fixes this by intentionally prefilling unsafe steps for harmful prompts and overcautious steps for harmless ones, then rewarding overrides. Training mixes normal prompts with these counter examples so recovery from a bad start becomes routine. It uses standard reinforcement learning with rewards for safety, helpfulness, and math, without extra run time cost. Safety rises on direct harm and jailbreak tests, while needless refusals on benign prompts drop. Math stays stable, so core reasoning is kept. The model starts to self check, pause, and fix earlier steps mid run. Even full chain hijacks and repeated reset attacks mostly fail to push it unsafe. Results depend on how many prefills are used and their length, very heavy prefilling can reduce helpfulness. ---- Paper – arxiv. org/abs/2510.00938 Paper Title: "Large Reasoning Models Learn Better Alignment from Flawed Thinking"

🚨 New paper alert! 🚨 Can you believe it? Flawed thinking helps reasoning models learn better! Injecting just a bit of flawed reasoning can collapse safety by 36% 😱 — but we teach large reasoning models to fight back 💪🛡️. Introducing RECAP 🔄: an RL post-training method that trains models to override unsafe reasoning, reroute to safe & helpful answers, and stay robust — all without extra training cost. ✨ Safer reasoning 🤖 ✨ Stronger jailbreak resistance 🔓 ✨ Lower overrefusal 🙅 ✨ Preserved core reasoning capability 🧠 #LLM #ReasoningModels #RLHF #AISafety #Alignment #MachineLearning

Our paper “LitForager: Exploring Multimodal Literature Foraging Strategies in Immersive Sensemaking” has been accepted to #ISMAR2025 as a #TVCG paper! 🎉✨

@Alibaba_Qwen Congrats on the great work! The "token-level safety detection" idea echoes our recent NeurIPS'25 dynamic safety shaping paper! 👉 arxiv.org/abs/2505.17196

🎉Our paper "Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safety" has been accepted to EMNLP 2025 Main Track! @emnlpmeeting 👉First survey connecting LLM interpretation & safety

🚨 New work: We rethink how we finetune safer LLMs — not by filtering after the generation, but by tracking safety risk token by token during training. We repurpose guardrail models like 🛡️ Llama Guard and Granite Guardian to score evolving risk across each response 📉 — giving rise to the STAR ⭐ score, a fine-grained safety signal that enables more targeted safety supervision. On top of this, we introduce ⭐DSS (STAR-Guided Dynamic Safety Shaping) — a training method that 🚫 suppresses unsafe patterns, 💪 preserves capability, and generalizes across LLMs, guardrails, harm levels, and datasets. Our method outperforms "Deep Token," the method from this year’s #iclr2025 Best Paper 🏆 — remaining robust against key finetuning-as-a-service threats like 🔄 response adaptation, 🧪 prompt poisoning, and 🛑 harmful prefilling. #MachineLearning #DeepLearning #LLM #AISafety #Alignment #Finetuning

Guardrail models like 🛡️ Llama Guard do more than filtering — we repurpose them to track how safety risk evolves 📉 through a response. This gives rise to the STAR ⭐ score: a fine-grained signal for finetuning LLMs more safely 🤖🔒 Curious how it works? More in the thread 👇

I've been putting together an interactive tool called DiffusionLab for explaining the geometric intuition behind diffusion and flow based generative models. Sampling is actually being done in the browser using Tensorflow.js! It is still in the very early stages.

@victor_explore Thank you @victor_explore for sharing! Our students appreciate it!!

This website has visualizations to understand almost all major topics in Machine Learning (link in comment)

Diffusion models leverage a variety of samplers. Deterministic methods like DDIM produce orderly paths. In contrast, stochastic samplers like DDPM produce chaotic trajectories. Despite their differences, both methods draw valid samples from the underlying distribution.

Create heatmaps that localize text concepts in generated videos. We discovered that our approach, ConceptAttention, can be directly extended from image generation to video generation models! It's amazing how simple techniques often generalize way better than more complex ones.

All three RPT cases from the School of CSE this year have been approved! Join us in congratulating the following faculty on their promotions! 🥳🎉 -B. Aditya Prakash, professor -Chao Zhang, associate professor (w/tenure) -Xiuwei Zhang, associate professor (w/tenure)