AI Breakdown

In this episode, we discuss Teaching Language Models to Critique via Reinforcement Learning by Zhihui Xie, Jie chen, Liyu Chen, Weichao Mao, Jingjing Xu, Lingpeng Kong. The paper presents CTRL, a framework that uses reinforcement learning to train critic models which provide feedback for improving code generated by large language models without needing human input. These trained critics significantly increase code pass rates and reduce errors across different generator models. Additionally, the critics serve as effective reward models, allowing iterative refinements that lead to over 106% improvement on challenging code generation benchmarks.

In this episode, we discuss PANDAS: Improving Many-shot Jailbreaking via Positive Affirmation, Negative Demonstration, and Adaptive Sampling by Avery Ma, Yangchen Pan, Amir-massoud Farahmand. The paper introduces PANDAS, a hybrid technique that enhances many-shot jailbreaking by altering fabricated dialogues with positive affirmations, negative demonstrations, and optimized adaptive sampling tailored to specific prompts. Experimental results on AdvBench and HarmBench using advanced large language models show that PANDAS significantly outperforms existing baseline methods in scenarios involving long input contexts. Additionally, an attention analysis highlights how PANDAS exploits long-context vulnerabilities, providing deeper insights into the mechanics of many-shot jailbreaking.

In this episode, we discuss VidCRAFT3: Camera, Object, and Lighting Control for Image-to-Video Generation by Sixiao Zheng, Zimian Peng, Yanpeng Zhou, Yi Zhu, Hang Xu, Xiangru Huang, Yanwei Fu. The paper presents VidCRAFT3, a new framework for image-to-video generation that allows simultaneous control over camera motion, object movement, and lighting direction. It addresses previous limitations by introducing the Spatial Triple-Attention Transformer, which effectively decouples and integrates lighting, text, and image inputs. This innovative approach enhances the precision and versatility of controlling multiple visual elements in generated videos.

In this episode, we discuss Heuristically Adaptive Diffusion-Model Evolutionary Strategy by Benedikt Hartl, Yanbo Zhang, Hananel Hazan, Michael Levin. The paper explores the connection between diffusion models and evolutionary algorithms, highlighting that both generate high-quality samples through iterative refinement of random initial states. By integrating deep learning-based diffusion models into evolutionary processes, the authors enhance convergence efficiency and maintain diversity by leveraging improved memory and refined sample generation. This framework advances evolutionary optimization by providing greater flexibility, precision, and control, representing a significant shift in heuristic and algorithmic approaches.

In this episode, we discuss Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach by Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, Tom Goldstein. The paper presents a new language model architecture that enhances test-time computation by iteratively reasoning in latent space using a recurrent block, allowing flexible depth during inference. Unlike chain-of-thought approaches, it doesn't require specialized training data, works with small context windows, and can handle complex reasoning not easily expressed in words. A 3.5 billion parameter model was scaled to 800 billion tokens, demonstrating significant performance improvements on reasoning benchmarks with computation loads up to 50 billion parameters. Huggingface: https://huggingface.co/papers/2502.05171 Github: https://github.com/seal-rg/recurrent-pretraining  

In this episode, we discuss EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents by Rui Yang, Hanyang Chen, Junyu Zhang, Mark Zhao, Cheng Qian, Kangrui Wang, Qineng Wang, Teja Venkat Koripella, Marziyeh Movahedi, Manling Li, Heng Ji, Huan Zhang, Tong Zhang. The paper presents **EMBODIEDBENCH**, a comprehensive benchmark with 1,128 tasks across four environments to evaluate vision-driven embodied agents based on Multi-modal Large Language Models (MLLMs). It assesses key capabilities such as commonsense reasoning, spatial awareness, and long-term planning through six specialized subsets. Evaluations of 13 MLLMs revealed that while these models perform well on high-level tasks, they struggle with low-level manipulations, highlighting significant challenges and guiding future advancements in embodied agent development.

In this episode, we discuss VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection by Songhao Han, Wei Huang, Hairong Shi, Le Zhuo, Xiu Su, Shifeng Zhang, Xu Zhou, Xiaojuan Qi, Yue Liao, Si Liu. The paper introduces VideoEspresso, a high-quality, large-scale VideoQA dataset that maintains essential spatial and temporal details and includes multimodal annotations for intermediate reasoning steps. Utilizing a semantic-aware construction pipeline and GPT-4 for generating QA pairs and Chain-of-Thought annotations, the dataset enhances scalability and reasoning complexity. Additionally, the authors propose a Hybrid LVLMs Collaboration framework that outperforms existing models on 14 tasks, demonstrating superior video reasoning capabilities.

In this episode, we discuss VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models by Hila Chefer, Uriel Singer, Amit Zohar, Yuval Kirstain, Adam Polyak, Yaniv Taigman, Lior Wolf, Shelly Sheynin. Generative video models typically prioritize appearance accuracy over motion coherence, limiting their ability to capture realistic dynamics. The paper presents VideoJAM, a framework that integrates a joint appearance-motion representation and uses an Inner-Guidance mechanism to enhance motion consistency during generation. VideoJAM achieves state-of-the-art motion realism and visual quality while being easily adaptable to existing video models without major changes.

In this episode, we discuss HunyuanVideo: A Systematic Framework For Large Video Generative Models by Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, Caesar Zhong. HunyuanVideo is an innovative open-source video generation model that matches or exceeds the performance of leading closed-source alternatives. It leverages a comprehensive framework encompassing data curation, advanced architecture, progressive scaling, and efficient infrastructure to train a 13-billion-parameter model, the largest of its kind in the open-source domain. Extensive evaluations reveal that HunyuanVideo delivers superior visual quality, motion dynamics, and text-video alignment, and its publicly available code aims to bridge the gap between closed and open-source communities, fostering a more dynamic video generation ecosystem.

In this episode, we discuss s1: Simple test-time scaling by Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, Tatsunori Hashimoto. The paper introduces a straightforward method for test-time scaling in language models to enhance reasoning performance by utilizing additional computational resources during inference. The authors develop a curated dataset of 1,000 high-quality, diverse, and challenging questions with reasoning traces and implement a "budget forcing" technique that controls the model's computation by either terminating its reasoning process or extending it to encourage double-checking answers. Using this approach, their fine-tuned Qwen2.5-32B-Instruct model outperforms OpenAI’s o1 model on competitive math benchmarks by up to 27% and the resources are made available as open-source.