This research investigates how large language models (LLMs) perform arithmetic tasks. The authors find that LLMs do not rely on robust algorithms or memorization but instead use a "bag of heuristics," a collection of simple, memorized rules, to solve arithmetic problems. They identify a specific set of neurons in the LLMs that implement these heuristics and analyze how they develop over the course of training. Their findings suggest that improving LLMs' mathematical abilities may require fundamental changes to training and architecture rather than relying on post-hoc techniques.

This research paper investigates the vulnerabilities of vision-language models (VLMs) used to power computer agents. The authors demonstrate that these agents can be easily manipulated by carefully crafted adversarial pop-ups, causing them to click on these malicious elements instead of performing their intended tasks. This attack successfully diverts the agents from their intended actions in over 80% of cases, significantly reducing their task success rate. The authors explore various attack design elements and find that basic defense strategies, such as asking the agent to ignore pop-ups or including an advertisement notice, are ineffective. They conclude that more robust agent systems are needed to ensure safe agent workflow in real-world computer environments.

This research paper proposes a new method for training large language models (LLMs) to solve complex scientific problems. The authors argue that current LLMs struggle with complex scientific questions, often hallucinating answers instead of providing accurate solutions. To address this, they suggest integrating LLMs with specialized tools. However, traditional methods for integrating tools often result in the model over-relying on these tools, even when simpler reasoning would suffice. To overcome these limitations, the paper presents a two-component fine-tuning method: World Knowledge Distillation (WKD), which allows the LLM to learn from the solutions generated by tools, and Tool Usage Adaptation (TUA), which trains the model to intelligently choose between direct reasoning and using external tools based on the difficulty of the question. The authors demonstrate the effectiveness of their approach on various datasets across different scientific domains, including mathematics, climate science, and epidemiology, finding significant improvements in both answer accuracy and tool usage precision.

This report details the research process of a team of researchers at Shanghai Jiao Tong University and other institutions who are attempting to replicate OpenAI's O1 model, a groundbreaking language model capable of complex reasoning. The report advocates for open science by detailing the entire research journey, including successes and failures, to help other researchers accelerate progress in the field. The researchers have introduced a new paradigm called "journey learning," where models learn not just shortcuts to solutions, but the entire exploration process, including trial and error and reflection, which they believe is key to O1's capabilities. They share valuable resources including technical hypotheses, cognitive exploration maps, and custom-developed tools. The report also highlights the challenges of traditional AI research and proposes a new framework for scientific communication and collaboration.

This paper provides a comprehensive review of document parsing, a field focused on converting unstructured and semi-structured documents into structured, machine-readable data. It explores two main approaches: modular pipeline systems and end-to-end models based on large vision-language models. The paper examines the core components of document parsing, including layout detection, content extraction (text, tables, mathematical expressions), and relation integration, as well as the challenges each approach faces. The authors provide an overview of key methodologies, datasets, evaluation metrics, and open-source tools, ultimately emphasizing the need for further research and development to advance the field.

This study investigates the gradient patterns of different layers in large language models (LLMs) during instruction tuning. The researchers compare the gradients of LLMs trained using "fast thinking" (without chain-of-thought reasoning) versus "slow thinking" (with detailed chain-of-thought reasoning). The study examines how these training methods affect gradient stability, response correctness, and the ability to distinguish between correct and irrelevant responses. They further analyze the impact of different initial models (pre-trained vs. instruction-tuned) on gradient behavior. The results show that "slow thinking" leads to more stable and efficient training, while "fast thinking" results in larger gradients and greater fluctuation across layers. The researchers also find that "slow thinking" helps distinguish correct responses from irrelevant responses, but this ability is not as pronounced in "fast thinking" training. Finally, the study explores the effects of response length and popularity on gradient patterns in knowledge-learning tasks, demonstrating that increasing response length alone does not necessarily mimic the effects of "slow thinking."

This research paper introduces REPOCOD, a new benchmark for evaluating large language models (LLMs) in code generation tasks that require knowledge of entire software projects. The benchmark consists of 980 code generation problems collected from 11 popular open-source Python repositories. REPOCOD is designed to be more challenging than existing benchmarks by requiring LLMs to generate complex functions, often involving interactions across multiple files in a project. The authors evaluate the performance of various LLMs on REPOCOD, finding that even the most advanced models struggle to achieve high accuracy, particularly when dealing with functions requiring knowledge of the entire repository. This research highlights the need for further development of LLMs that can effectively handle the complexity of real-world code generation within the context of software projects.

This research paper introduces SocialGPT, a new modular framework that leverages the perception capabilities of vision foundation models (VFMs) and the reasoning capabilities of large language models (LLMs) to identify social relationships between people in images. Unlike previous methods that train a dedicated network end-to-end, SocialGPT translates image content into a textual social story using VFMs, which is then used for text-based reasoning with LLMs. The paper also proposes a novel prompt optimization method called Greedy Segment Prompt Optimization (GSPO), which helps improve the performance of LLMs by performing a greedy search on the segment level with gradient guidance. SocialGPT achieves highly competitive results on two datasets without additional model training and provides interpretable answers, offering language-based explanations for the decisions.

This paper describes the development of ROCKET-1, a novel hierarchical agent architecture that leverages visual-temporal context prompting to enable agents to master open-world interaction in Minecraft. This system utilizes a low-level policy (ROCKET-1) that predicts actions based on concatenated visual observations and segmentation masks, guided by a high-level reasoner. The key innovation lies in the visual-temporal context prompting protocol, which uses object segmentation from both past and present observations to effectively communicate spatial information. This approach allows for the successful completion of complex tasks, such as crafting and mining, in Minecraft, which were previously unattainable using traditional language-based prompting methods. The paper highlights the potential of visual-temporal context prompting to overcome the limitations of existing approaches and unlock the full potential of vision-language models for embodied decision-making.

This research investigates the effectiveness of knowledge editing techniques in correcting hallucinations in large language models (LLMs). The authors present HalluEditBench, a comprehensive benchmark that evaluates the performance of different knowledge editing methods across five dimensions: efficacy, generalization, portability, locality, and robustness. They discovered that while some knowledge editing methods show promising results on existing benchmarks, their effectiveness in correcting real-world hallucinations may be significantly lower, highlighting the need for more robust evaluation methods. Additionally, the study provides insights into the limitations of different editing methods, suggesting that no single method excels across all five dimensions. The authors conclude by emphasizing the importance of understanding the potential and limitations of knowledge editing techniques for achieving more accurate and reliable LLMs.