The Daily ML

The article "When a language model is optimized for reasoning, does it still show embers of autoregression? An analysis of OpenAI o1" investigates whether a new language model, o1, developed by OpenAI, which was specifically trained for reasoning, retains limitations stemming from its origins in next-word prediction. Despite significant improvements in performance compared to previous LLMs, o1 still exhibits similar qualitative trends, indicating a sensitivity to the probability of both the text it generates and the tasks it is asked to perform. The authors find that o1, like other LLMs, performs better on high-probability tasks and examples, suggesting that even with optimization for reasoning, the influence of next-word prediction persists.

Meta AI has released Movie Gen, a suite of foundation models capable of generating high-quality videos and audio from text prompts. These models are trained on a massive dataset of images, videos, and audio, and can create realistic videos of up to 16 seconds long with synchronized audio. The research paper explores the architecture, training objectives, and evaluation metrics for Movie Gen, highlighting the innovative use of transformers, flow matching, and temporal autoencoders. The paper also details the development of personalized video generation capabilities, allowing users to create videos featuring a specific person, and video editing capabilities, enabling users to make precise changes to both real and generated videos. Additionally, Movie Gen Audio, a specialized model for generating audio, showcases the ability to create cinematic soundtracks, including both diegetic sound effects and non-diegetic music, that align with the visual content of a video.

This paper details the development and capabilities of a new, groundbreaking multimodal model called Emu3. Emu3 surpasses previous models by leveraging solely next-token prediction, enabling it to excel in diverse tasks, including image generation, video generation, and vision-language understanding. This breakthrough in artificial general intelligence (AGI) simplifies complex multimodal model designs and highlights the promise of next-token prediction for future development. The authors further showcase Emu3's advancements through detailed comparisons to existing models and qualitative examples of its capabilities.

The research paper introduces MindSearch, a novel AI system designed to mimic human cognitive processes when seeking and integrating information from the web. It combines the strengths of Large Language Models (LLMs) and search engines by using a multi-agent framework consisting of a WebPlanner and WebSearchers. The WebPlanner breaks down complex user queries into smaller, manageable sub-questions, while WebSearchers retrieve and summarize relevant information from the web. This approach addresses the limitations of traditional search engines and LLMs in dealing with complex, information-rich tasks. MindSearch demonstrates significant improvements in response quality, particularly in depth and breadth, surpassing existing AI search solutions like ChatGPT-Web and Perplexity.ai in subjective human evaluations.

This research paper explores the limitations of current large language models (LLMs) in handling tasks that require multiple, interwoven skills (cross capabilities). The authors argue that while LLMs excel in specific areas like reasoning, coding, or image recognition, their performance drastically declines when these skills need to be combined. To address this issue, they introduce CrossEval, a comprehensive benchmark designed to evaluate both individual and cross capabilities. CrossEval includes a wide range of prompts and uses human annotators to rate model responses. The study reveals a consistent “Law of the Weakest Link” effect, meaning an LLM’s performance on cross-capability tasks is primarily limited by its weakest individual capability. This finding emphasizes the need for future research to prioritize improving LLMs’ weaker areas in order to enhance their effectiveness in real-world applications.

This research paper provides a comprehensive survey of small language models (SLMs), which are smaller versions of large language models (LLMs) designed for deployment on devices like smartphones. The study analyzes the architecture, training datasets, and training algorithms used in SLMs, as well as their capabilities in various domains, including commonsense reasoning, problem-solving, and mathematics. The researchers also evaluate the runtime cost of SLMs on edge devices, examining factors like inference latency and memory footprint. The paper concludes with insights and future directions for SLM research, emphasizing the importance of co-designing SLM architecture and device processors, developing high-quality synthetic datasets, and exploring continual on-device learning for personalization.

This research paper proposes a novel "Logic-of-Thought" (LoT) prompting method to improve the logical reasoning capabilities of Large Language Models (LLMs). Existing methods struggle with "unfaithful reasoning" and "information loss" during the translation of natural language into symbolic logic. LoT addresses these issues by extracting propositions and logical relations from the input text, extending them using logical reasoning laws, and translating the extended expressions back into natural language. This process enhances the LLM's reasoning ability by providing additional logical information within the prompt. The paper demonstrates the effectiveness of LoT by integrating it with various existing prompting methods, including Chain-of-Thought (CoT), Self-Consistency (SC), and Tree-of-Thoughts (ToT), and evaluating its performance on multiple logical reasoning datasets. The results highlight the significant improvements achieved by LoT across different tasks and prompting methods, demonstrating its potential to significantly advance the capabilities of LLMs in logical reasoning.

This paper surveys Retrieval-Augmented Generation (RAG) systems for Large Language Models (LLMs) by classifying RAG tasks into four levels of complexity based on the type of external data required and the task’s primary focus: explicit fact queries, implicit fact queries, interpretable rationale queries, and hidden rationale queries. Each level presents unique challenges and solutions, with techniques like prompt tuning, Chain-of-Thought (CoT) prompting, and fine-tuning being explored to address the complexities of integrating external data into LLMs. The paper also explores three main forms of integrating external data into LLMs: context, small model, and fine-tuning, highlighting their strengths, limitations, and suitable applications. Ultimately, the paper aims to offer a comprehensive understanding of the data requirements and key bottlenecks in building LLM applications, providing guidance for systematically developing such applications.

The provided text is a research paper that evaluates the planning capabilities of large language models (LLMs) and large reasoning models (LRMs) using the PlanBench benchmark. The authors compare the performance of several LLMs, including GPT-4 and LLaMA, with OpenAI’s newly released o1 model, an LRM. The authors find that while o1 significantly outperforms LLMs on simple planning problems, it struggles with more complex or obfuscated tasks. They also explore the trade-offs between accuracy and efficiency, arguing that o1’s increased accuracy comes at a high computational cost, making it less practical than traditional planners or LLM-based systems. Finally, the paper highlights the lack of interpretability and correctness guarantees in o1, raising concerns about its reliability in safety-critical applications.