Tingting Chang

Video models don't just generate. They reason. And now we know exactly how. New research from UC Berkeley, UC San Diego, CMU and SenseTime just overturned a core assumption about how diffusion-based video models think. The old theory: reasoning unfolds frame by frame across a video. The reality: it happens across the denoising steps instead. Early steps explore multiple candidate solutions in parallel. Later steps narrow down to a single answer. When the authors disrupted specific denoising steps, performance collapsed. Disrupting individual frames barely mattered. Three emergent behaviors also showed up:  1. Working memory: Holding onto information even when objects disappear from view  2. Self-correction: Catching and reversing wrong answers mid-process  3. Perception before action: First figure out what and where, then figure out how The practical payoff: a simple, training-free method that ensembles outputs from multiple runs at the right internal layers. No fine-tuning. It outperformed Runway Gen-4 and Kling 2.6 on the VBVR-Bench. Really grateful to Ruisi Wang, Zhongang Cai, and all co-authors for this genuinely illuminating work. Link to Paper: https://lnkd.in/gUwR378V #GenerativeAI #VideoAI #DiffusionModels #AIResearch #DeepLearning

18

Building a Production-Ready RAG System for Persian Language Processing I’m excited to share insights from developing an advanced Retrieval-Augmented Generation system specifically engineered for Persian/Farsi document processing, leveraging cutting-edge ML infrastructure and architectural patterns. 🔬 Technical Architecture Highlights: Advanced RAG Implementation: ∙ Self-Corrective Agentic RAG with adaptive query rewriting and iterative refinement ∙ Hybrid search architecture combining dense embeddings with SPLADE sparse representations ∙ Multi-agent orchestration enabling context-aware document interaction and general conversation modes Production Infrastructure: ∙ NVIDIA Triton Inference Server deployment on H100 GPUs with vLLM backend ∙ Milvus vector database for high-performance similarity search ∙ PostgreSQL for structured data with MinIO object storage for scalable file management ∙ Model ensemble: Qwen2.5-72B-Instruct-AWQ, DeepSeek-OCR, and sentence-transformers ML Pipeline Innovations: ∙ Automated synthetic QA pair generation using LangGraph workflows ∙ Cross-encoder reranking (ms-marco-MiniLM-L6-v2) for precision optimization ∙ Dynamic prompt management system with A/B testing capabilities and versioning ∙ Fine-tuning pipeline with quality-scored synthetic datasets Engineering Principles: ∙ Clean Architecture with dependency injection and separation of concerns ∙ Docker containerization for reproducible deployments ∙ Comprehensive observability with metrics tracking and performance monitoring This project demonstrates how to architect enterprise-grade RAG systems that combine theoretical ML advancements with practical production considerations, particularly for low-resource languages. Open to discussing technical implementation details, architectural decisions, or challenges in deploying LLMs at scale. #MLEngineering #RAG #NLP #LLMs #VectorSearch #ProductionML #AI #DeepLearning #Persian #TritonInference #CleanArchitecture

1

Over the last few months, I’ve been diving deep into how domain-specialized LLMs can outperform generalist models in real-world business use cases. That’s why the release of eCeLLM really caught my attention. Built on ECInstruct (116K+ e-commerce samples, 10 real-world tasks), eCeLLM shows us what happens when we stop treating “one-size-fits-all AI” as the answer—and instead teach models to think in context. Here’s what stands out to me: Outperforms even GPT-4 on multiple e-commerce tasks Handles product matching, classification, recommendations, and QA with stronger generalization Excels not only in in-domain scenarios but also when facing new products, unseen data, and novel queries For businesses, this means less brittle AI, more adaptability, and better ROI. For consultants like me, it opens new doors in building AI agents that aren’t just clever demos but scalable solutions. This aligns perfectly with the kind of innovation I’ve been experimenting with through AI hackathons and client pilots—where we test the boundaries of LLMs, connect research like this to actual business processes, and see immediate impact. If you’re in retail, e-commerce, or even broader product-heavy industries, the next wave of AI transformation will be driven by models like eCeLLM. Check it out here 👉 ninglab.github.io/eCeLLM What do you think—will domain-specialized LLMs become the new standard in industry? #AI #Ecommerce #LLMs #BusinessStrategy #Innovation #AIConsulting #Hackathon

12

In this episode, we discuss ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models by Mingjie Liu, Shizhe Diao, Ximing Lu, Jian Hu, Xin Dong, Yejin Choi, Jan Kautz, Yi Dong. This paper introduces ProRL, a new reinforcement learning training method that uncovers novel reasoning strategies beyond those found in base language models. Empirical results show that models trained with ProRL consistently outperform base models on challenging reasoning tasks, including cases where base models fail even with extensive attempts. The study demonstrates that prolonged RL can meaningfully expand reasoning capabilities by exploring new solution spaces over time, advancing understanding of how RL enhances language model reasoning.

2

Our latest paper has been accepted at AAAI 2026. We introduce Structural Alignment, a reinforcement learning approach that aligns large language models with human-like discourse structures, enabling more coherent and well-organized long-form text generation. Our paper, “Structural Alignment: Aligning Large Language Models with Structural Information,” is available here: https://lnkd.in/gx3ZRTZu

49

I’m excited to attend NeurIPS workshop sessions at San Diego and present two papers authored by the applied research team: 1. GraphMatch: Fusing Language and Graph Representations in a Dynamic Two-Sided Work Marketplace. Mikołaj Sacha et al. - NeurIPS Workshop on Unifying Representations in Neural Models (UniReps) 2. Towards Real-World Evaluation of Agentic Work in Freelance Marketplaces. Mattie Terzolo et al. - NeurIPS Workshop on LLM Evaluation Over the next couple of days, I’ll be sharing insights and some behind-the-scenes perspectives on how we think about these models, how they’re shaping Upwork today, and the opportunities we see ahead. For attendees already onsite at the main conference, please stop by the Upwork booth #722 to learn more about the efforts the ML & AI team is driving - from research to real product impact. Feels great to be again publishing at NeurIPS after a hiatus and looking forward to great conversations and inspiration ahead!

52

2 Comments

When considering integration involving large language models and external tools (via function calls), a few things you should consider. 1. If the tool doesn't feed back into the LLM in an iterative loop, why aren't you using a batch file instead of the LLM? I say this, because the goal of using external tools is to benefit from them right? Therefore, that information needs to feed back into the context window for the model to consider the gained information and continue on with the task it has been configured for. 2. How much data will the tool produce VS how much context window do you have. Now, with Llama and their Llama 4 model, the game has changed regarding standalone AI. When you can have a context window of 10 Million tokens, this becomes less of a concern. Unlike when you were limited to 4096 or 2048. 3. Hardware - Make sure you have adequate resources to load the model with as much context window as you need. 4. System Prompt: Don't just stick with "be a helpful AI", be creative and experiment. Welcome to Agentic AI. https://lnkd.in/enmTFz7J

1

On the Duolingo English Test, we're constantly researching better methods for automatically assessing writing and speaking. Check out our latest paper on automatic pronunciation scoring!

9

NVIDIA TensorRT LLM provides a high-level Python LLM API, with its PyTorch-native architecture enabling developers to experiment with the runtime or extend functionality. Learn how these latest TensorRT-LLM optimizations boost reasoning inference performance in our recent blog post. 🖇️ https://lnkd.in/gx52zwTA #PyTorch #OpenSourceAI #AI #Inference #Innovation

94

2 Comments

Read Chaim Rand's practical walkthrough for optimizing PyTorch inference. Learn about batched inference, memory formats, IPEX, and graph compilation on Intel Corporation Xeon.

5

📢 New Paper Published! Excited to share our latest work, "LitANFIS: Literal-aware Adaptive Neuro-Fuzzy Inference System to learn Conjunctive Normal Form," now published in Neurocomputing. This model brings negation into fuzzy rule learning, enabling more compact and interpretable systems. Thanks to Dr. Armin Salimi-Badr for his support. Check out the publication here: https://lnkd.in/eJ-WgnJz (Available until Feb. 06, 2025)

87

24 Comments