AI Breakdown

Blogpost paper - Project Vend: Can Claude run a small shop? (And why does that matter?)

In this episode, we discuss Project Vend: Can Claude run a small shop? (And why does that matter?) The paper describes a month-long experiment where the AI model Claude autonomously managed an office store as a small business. The study reveals both how close the AI came to successfully running the business and the unexpected ways it failed. These findings offer insights into a near-future scenario where AI models independently operate real-world economic activities.

Arxiv paper - Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens

In this episode, we discuss Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens by Zeyuan Yang, Xueyang Yu, Delin Chen, Maohao Shen, Chuang Gan. The paper proposes Mirage, a framework that enables vision-language models to perform internal visual reasoning by generating latent visual tokens alongside text, without producing explicit images. Mirage is trained through a combination of distillation from image embeddings, text-only supervision, and reinforcement learning to align visual reasoning with task goals. Experiments show that this approach improves multimodal reasoning performance on various benchmarks without the need for heavy image generation.

Arxiv paper - SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing

In this episode, we discuss SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing by Ming Li, Xin Gu, Fan Chen, Xiaoying Xing, Longyin Wen, Chen Chen, Sijie Zhu. The paper addresses the issue of noisy supervision in instruction-based image editing datasets by rectifying editing instructions to better align with image pairs and introducing contrastive instruction supervision using triplet loss. Their method leverages inherent model generation attributes to guide editing instruction correction without relying on vision-language models or pre-training, resulting in a simpler and more effective training process. Experiments show significant improvements over state-of-the-art methods with much less data and smaller models, and all resources are publicly released.

Arxiv paper - OMEGA: Can LLMs Reason Outside the Box in Math? Evaluating Exploratory, Compositional, and Transformative Generalization

In this episode, we discuss OMEGA: Can LLMs Reason Outside the Box in Math? Evaluating Exploratory, Compositional, and Transformative Generalization by Yiyou Sun, Shawn Hu, Georgia Zhou, Ken Zheng, Hannaneh Hajishirzi, Nouha Dziri, Dawn Song. The paper introduces OMEGA, a new benchmark to evaluate large language models' out-of-distribution generalization on math problems along three creativity-inspired axes: exploratory, compositional, and transformative reasoning. Evaluations reveal that state-of-the-art LLMs struggle increasingly with problem complexity, especially in compositional and transformative reasoning. Fine-tuning improves exploratory skills but not the other two, highlighting challenges in achieving genuine mathematical creativity beyond routine problem-solving.

Arxiv paper - Long-Context State-Space Video World Models

In this episode, we discuss Long-Context State-Space Video World Models by Ryan Po, Yotam Nitzan, Richard Zhang, Berlin Chen, Tri Dao, Eli Shechtman, Gordon Wetzstein, Xun Huang. The paper introduces a novel video diffusion model architecture that uses state-space models (SSMs) to extend temporal memory efficiently for causal sequence modeling. It employs a block-wise SSM scanning scheme combined with dense local attention to balance long-term memory with spatial coherence. Experiments on Memory Maze and Minecraft datasets show the method outperforms baselines in long-range memory retention while maintaining fast inference suitable for real-time use.

Arxiv paper - From Bytes to Ideas: Language Modeling with Autoregressive U-Nets

In this episode, we discuss From Bytes to Ideas: Language Modeling with Autoregressive U-Nets by Mathurin Videau, Badr Youbi Idrissi, Alessandro Leite, Marc Schoenauer, Olivier Teytaud, David Lopez-Paz. The paper introduces an autoregressive U-Net model that dynamically learns its own token embeddings from raw bytes instead of relying on fixed tokenization schemes like BPE. This multi-scale architecture processes text from fine-grained bytes to broader semantic units, enabling predictions at varying future horizons. The approach matches strong baselines with shallow hierarchies and shows potential improvements with deeper ones, offering flexibility across languages and tasks.

Arxiv paper - Reinforcement Pre-Training

In this episode, we discuss Reinforcement Pre-Training by Qingxiu Dong, Li Dong, Yao Tang, Tianzhu Ye, Yutao Sun, Zhifang Sui, Furu Wei. The paper introduces Reinforcement Pre-Training (RPT), a method that applies reinforcement learning to next-token prediction by rewarding correct predictions as a reasoning task. This approach leverages large text datasets without needing domain-specific annotations, improving language modeling accuracy and enabling strong foundations for further RL fine-tuning. Experimental results demonstrate that RPT scales effectively with compute, making it a promising paradigm for advancing language model pre-training.

Arxiv paper - Token-Efficient Long Video Understanding for Multimodal LLMs

In this episode, we discuss Token-Efficient Long Video Understanding for Multimodal LLMs by Jindong Jiang, Xiuyu Li, Zhijian Liu, Muyang Li, Guo Chen, Zhiqi Li, De-An Huang, Guilin Liu, Zhiding Yu, Kurt Keutzer, Sungjin Ahn, Jan Kautz, Hongxu Yin, Yao Lu, Song Han, Wonmin Byeon. The paper introduces STORM, a new architecture that incorporates a temporal encoder using the Mamba State Space Model to better capture temporal dynamics in video-based multimodal large language models. This approach enables effective token reduction, significantly lowering computational costs and latency while preserving essential temporal information. Experiments demonstrate that STORM achieves state-of-the-art performance on long video understanding benchmarks with substantial improvements in efficiency and accuracy.

Arxiv paper - The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity

In this episode, we discuss The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity by The authors of the paper are:
- Parshin Shojaee
- Iman Mirzadeh
- Keivan Alizadeh
- Maxwell Horton
- Samy Bengio
- Mehrdad Farajtabar. This paper examines the reasoning abilities of Large Reasoning Models (LRMs) using controlled puzzles to analyze both their final answers and internal reasoning processes. It reveals that LRMs struggle with high-complexity problems, showing performance collapse and inconsistent reasoning despite sufficient computational resources. The study identifies distinct performance regimes and highlights fundamental limitations in LRMs' exact computation and use of explicit algorithms, questioning their true reasoning capabilities.

Arxiv paper - Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models

In this episode, we discuss Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models by Piotr Padlewski, Max Bain, Matthew Henderson, Zhongkai Zhu, Nishant Relan, Hai Pham, Donovan Ong, Kaloyan Aleksiev, Aitor Ormazabal, Samuel Phua, Ethan Yeo, Eugenie Lamprecht, Qi Liu, Yuqi Wang, Eric Chen, Deyu Fu, Lei Li, Che Zheng, Cyprien de Masson d'Autume, Dani Yogatama, Mikel Artetxe, Yi Tay. The paper introduces Vibe-Eval, an open benchmark and framework with 269 visual understanding prompts designed to evaluate multimodal chat models on everyday and challenging tasks. It highlights that over half of the hardest prompts are incorrectly answered by current frontier models, emphasizing the benchmark's difficulty. The authors discuss evaluation methods, demonstrate correlation between automatic and human assessments, provide free API access, and release all code and data publicly. Github: https://github.com/reka-ai/reka-vibe-eval