Soheil Feizi’s Post

We use transfer reinforcement learning and inference-time decoding to train much stronger LLMs

LLMs "hallucinate": making up information instead of revealing their lack of knowledge 🤔 Our #NeurIPS2024 paper adds an 𝙸–𝙳𝚘𝚗'𝚝–𝙺𝚗𝚘𝚠 🤷🏾♂️ token to the LLM and fine-tunes it to predict this instead of hallucinating. No extra annotation is needed for training! The regular cross-entropy loss shifts all wrong probability mass to the gold token, while our IDK objective divides the shifted probability mass between the gold and 𝙸–𝙳𝚘𝚗'𝚝–𝙺𝚗𝚘𝚠 🤷🏾♂️ token for wrong predictions. Joint work w/ Roi Cohen, Konstantin Dobler, Eden Biran

Our #NeurIPS2024 paper about uncertainty-aware training to reduce hallucinations for LLMs is out! Check it out - https://lnkd.in/dgDVtBt5.

In our #NeurIPS2024 paper „I Don’t Know: Explicit Modeling of Uncertainty with an [IDK] Token“, we equip LLMs with a special „I Don’t Know“ (IDK) token that is learned to be predicted instead of „hallucinating“ wrong completions. We finetune LLMs with a novel IDK objective to shift probability mass towards the new IDK token for uncertain predictions. By design, the IDK objective does not require any annotations or handcrafted examples! All details in our paper: https://lnkd.in/dJX-FX7a (joint work with Roi Cohen, Eden Biran and Gerard de Melo).

Happy to share two paper accepts in #NeurIPS2024 (NeurIPS) that investigate the following questions. 1. How do you explain GNN predictions using human-interpretable logical rules? Ans: We present GraphTrail, where we use Symbolic Regression over discriminative GNN computation trees to uncover the functioning of GNNs (See pipeline in figure below). With: Burouj Armgaan Manthan Dalmia Sourav Medya 2. Given two networks, can we identify correspondences between nodes in these networks? Ans: Network alignment finds several applications such as function discovery of proteins across species and network security. We present FUGAL, that maps the task to the classical Quadratic Assignment Problem aided by novel regularizers and optimizers. With: Aditya Bommakanti Harshith Reddy Vonteri Konstantinos Skitsas Davide Mottin Panagiotis Karras

Excited to introduce our work at #ICLR2025! Title: DGQ: Distribution-Aware Group Quantization for Text-to-Image Diffusion Models TL;DR: We propose a distribution-aware group quantization method that preserves both image quality and text-to-image alignment in T2I diffusion models. arXiv: https://lnkd.in/g8QAbAN7 project page: https://ugonfor.kr/DGQ/ code: https://lnkd.in/gHhdfri7 💡 Key Features - Analyzes the distributional characteristics of activations in text-to-image diffusion models, explaining why they pose challenges for quantization. - Utilizes group quantization to address these challenges with minimal overhead. - Achieves low-bit (<8-bit) quantization without requiring additional fine-tuning. #ICLR2025 #Quantization #EfficientAI

🎉 Thrilled to share our work at NeurIPS 2024! 🎉 We are excited to present our latest research on Video Anomaly Detection: "Advancing Video Anomaly Detection: A Concise Review and A New Dataset" Authors: Liyun Zhu, Lei Wang, Arjun Raj, Tom Gedeon, Chen Chen (https://lnkd.in/g6iN_qKm) 📍 West Ballroom A-D, #5311 📅 Wednesday, 11 December, 11 a.m. – 2 p.m. PST 💡 What’s our work about?  ✅ We provide a concise review that highlights current challenges, research trends and future directions, providing a quick reference for researchers.  ✅ We propose the Multi-Scenario Anomaly Detection (MSAD) dataset, a comprehensive anomaly detection benchmark including diverse scenarios and anomalies.  ✅ We propose a novel Scenario Adaptive Anomaly Detection model, using few-shot learning for efficient adaptation to new scenarios. Interested in our work? ➡️ Project Website: https://lnkd.in/gfJcKXfP ➡️ Paper: https://lnkd.in/gZi_jB9w ➡️ Code: https://lnkd.in/g6QbhZdc Come visit us at NeurIPS 2024! Let's connect and discuss. See you there! 🚀 #neurips24 #NeurIPS24 #VideoAnomalyDetection

🎉 I am excited to share that our latest research paper, Model-based Feature Weight Drift Detection (MFWDD) Showcased on TLS and QUIC Traffic, has been published at CNSM24! 🚀 🔍 In the ever-evolving world of network traffic classification, machine learning models face constant challenges from concept and data drifts. These shifts degrade model performance over time, making it harder to maintain accuracy without frequent, often inefficient retraining. ✨ That's where MFWDD comes into play! Unlike traditional approaches, MFWDD doesn't just detect drifts—it quantifies their severity and links them to specific features, offering unprecedented insights into what's driving the change. By leveraging statistical tests like the Kolmogorov–Smirnov test and Wasserstein distance, MFWDD enables smarter, more targeted retraining. This means fewer unnecessary updates, better resource efficiency, and stronger model performance over the long term. 💡 💡 In our paper, we showcase how MFWDD guided the retraining of ML models for TLS and QUIC service classification, maintaining and even improving their accuracy despite evolving network conditions. The results highlight the importance of timely retraining and the advantages of feature-focused drift detection. 📊 📄 Curious to learn more? Check out the paper on IEEE Xplore: https://lnkd.in/eY_2K5EW 🔗 #Research #NetworkSecurity #MachineLearning #DataDrift #OpenSource

I am thrilled to share that two of our papers submitted to NeurIPS 2024 have been accepted! 🎉 #NeurIPS2024 Our paper titled "Fine-Tuning is Fine, if Calibrated" has been accepted into the main track, while "CompBench: A Comparative Reasoning Benchmark for Multimodal LLMs" has been accepted into the database track. I would like to extend my heartfelt thanks to all the co-authors for their incredible work and collaboration. Looking forward to the conference! To delve deep into the details of our work, please refer to the paper: Paper 1: Fine-Tuning is Fine, if Calibrated (https://lnkd.in/dZeZ68_K ) Paper 2: CompBench: A Comparative Reasoning Benchmark for Multimodal LLMs (https://lnkd.in/dpikR968)

It was a great honor to present our work on "Foundation Inference Models for Markov Jump Processes" with Ramses J. Sanchez at the NeurIPS 2024 conference in Vancouver. If you are curious, you can check out our paper here: https://lnkd.in/e8XR6SHP #NeurIPS