💡 This paper presents the first large scale empirical study of agent context files, also known as agent READMEs, which provide persistent and project level instructions for agentic coding tools. The study analyzed 2303 agent context files from 1925 repositories to characterize their structure, maintenance, and content. The researchers found that these files are not static documentation but complex and difficult to read artifacts that evolve like configuration code, maintained through frequent small additions.

The content analysis of 16 instruction types revealed that developers prioritize functional context, such as build and run commands, implementation details, and architecture. However, the study also identified a significant gap, as non-functional requirements like security and performance are rarely specified. The findings indicate that while developers use context files to make agents functional, they provide few guardrails to ensure that agent-written code is secure or performant.

The study highlights the need for improved tooling and practices to address this gap. The contributions of this paper include a comprehensive understanding of the structure and content of agent context files, and the identification of areas for improvement to ensure that agentic coding tools produce secure and performant code. The research has implications for the development of agentic coding tools and the use of agent context files in software development projects. Overall, the study provides valuable insights into the use of agent context files and highlights the need for further research and development to improve the security and performance of agent-written code.

💡 The paper introduces OmniFlatten, a novel end-to-end GPT model that enables real-time natural full-duplex spoken dialogue. The goal is to achieve low latency and natural interactions in full-duplex dialogue systems, which is a significant challenge due to human conversation dynamics such as interruptions, backchannels, and overlapping speech. To address this, the authors propose a multi-stage post-training technique that integrates speech and text without altering the original model's architecture. The training process consists of three stages: modality alignment, half-duplex dialogue learning, and full-duplex dialogue learning. A flattening operation is used to standardize the data, allowing for unified training methods and model architecture across different modalities and tasks. The OmniFlatten model can generate text and speech in real-time, effectively modeling complex behaviors inherent to natural conversations. The approach offers a straightforward modeling technique and a promising research direction for developing efficient and natural end-to-end full-duplex spoken dialogue systems. The results are demonstrated through audio samples of dialogues generated by OmniFlatten, which can be found online. Overall, the paper contributes to the development of full-duplex spoken dialogue systems that can mimic human-human interactions, with potential applications in various areas such as virtual assistants, customer service, and more.

💡 DeepSeek-V3 is a language model that achieves high performance with efficient training and minimal computational cost. The model uses a Mixture-of-Experts architecture with 671 billion total parameters, but only 37 billion are activated for each token, making it parameter-efficient. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention and DeepSeekMoE architectures, which were validated in the previous version of the model.

The model is trained on 14.8 trillion diverse and high-quality tokens, followed by supervised fine-tuning and reinforcement learning stages to fully harness its capabilities. The training process is stable and requires only 2.788 million H800 GPU hours for full training, which is relatively low compared to other models.

The results show that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. The model also pioneers an auxiliary-loss-free strategy for load balancing and uses a multi-token prediction training objective for stronger performance. The model checkpoints are available for further research and development.

Overall, the DeepSeek-V3 model makes significant contributions to the field of natural language processing by providing a highly efficient and effective language model that can be trained with minimal computational resources. The model's stable training process and low computational cost make it an attractive option for researchers and developers who want to build high-performance language models without incurring high costs.

💡 The paper introduces LlamaFactory, a unified framework that enables efficient fine-tuning of large language models across various tasks. The problem addressed is that fine-tuning these models requires significant effort and coding expertise, which can be a barrier for many users. To solve this, LlamaFactory integrates a suite of cutting-edge efficient training methods, allowing users to customize the fine-tuning of over 100 language models without needing to write code. This is made possible through a web-based user interface called LlamaBoard, which provides a flexible and user-friendly way to fine-tune language models. The authors validate the efficiency and effectiveness of LlamaFactory on language modeling and text generation tasks, demonstrating its potential. The framework has been released publicly and has already gained significant attention, with over 13,000 stars and 1,600 forks on GitHub. Overall, LlamaFactory contributes to the field by providing a unified and accessible way to fine-tune large language models, making it easier for researchers and practitioners to adapt these models to specific tasks and applications.

💡 The paper introduces LeWorldModel, a stable end to end joint embedding predictive architecture that trains efficiently from raw pixels. Existing methods for learning world models in compact latent spaces are fragile and rely on complex loss terms, pre trained encoders, or auxiliary supervision to avoid representation collapse. LeWorldModel addresses this issue by using only two loss terms, a next embedding prediction loss and a regularizer, to train the model end to end from raw pixels. This approach reduces the number of tunable loss hyperparameters from six to one compared to existing methods. The model has approximately 15 million parameters and can be trained on a single GPU in a few hours, making it up to 48 times faster than foundation model based world models. The results show that LeWorldModel remains competitive across diverse 2D and 3D control tasks and encodes meaningful physical structures in its latent space. The model is also able to reliably detect physically implausible events, demonstrating its ability to learn a robust and generalizable representation of the world. Overall, LeWorldModel provides a stable and efficient framework for learning world models from raw pixels, making it a significant contribution to the field of artificial intelligence.

💡 The paper proposes a self supervised framework called Self Supervised Prompt Optimization that optimizes prompts for large language models without requiring external references. The problem addressed is that manually designed prompts require expertise and iterative experimentation, while existing prompt optimization methods rely heavily on external references such as ground truth or human evaluation, which can be costly to obtain. The proposed method derives evaluation and optimization signals purely from output comparisons, where a large language model evaluator selects superior prompts through pairwise output comparisons, and a large language model optimizer aligns outputs with task requirements. The results show that the proposed method outperforms state of the art prompt optimization methods, achieving comparable or superior results with significantly lower costs and fewer samples, demonstrating its effectiveness and efficiency. The method can optimize prompts for both closed and open ended tasks, and can be applied in real world scenarios where external references are unavailable or costly to obtain.

💡 The paper introduces Zep, a novel memory layer service for artificial intelligence agents, which outperforms the current state of the art system, MemGPT. The problem addressed is the limitation of existing retrieval-augmented generation frameworks, which are restricted to static document retrieval and cannot handle dynamic knowledge integration from diverse sources, including ongoing conversations and business data.

To address this limitation, Zep uses a core component called Graphiti, a temporally-aware knowledge graph engine that dynamically synthesizes both unstructured conversational data and structured business data while maintaining historical relationships. This allows Zep to excel in dynamic knowledge integration and temporal reasoning, critical for enterprise use cases.

The results show that Zep demonstrates superior performance in the Deep Memory Retrieval benchmark, with an accuracy of 94.8 percent compared to MemGPT's 93.4 percent. Furthermore, Zep's capabilities are validated through the LongMemEval benchmark, which better reflects enterprise use cases through complex temporal reasoning tasks. In this evaluation, Zep achieves substantial results with accuracy improvements of up to 18.5 percent while simultaneously reducing response latency by 90 percent compared to baseline implementations.

Overall, the paper presents Zep as an effective solution for real-world applications, particularly in enterprise-critical tasks such as cross-session information synthesis and long-term context maintenance, demonstrating its potential for deployment in real-world applications.

💡 The paper introduces AI-Trader, a fully automated live benchmark for evaluating large language models in financial decision-making across multiple markets. The benchmark is designed to address the gap in systematic benchmarking for real-world financial applications, where autonomous agents must make decisions in fully dynamic and live environments. The authors argue that existing efforts have not adequately addressed the challenge of evaluating large language models in real-time financial markets, where stringent requirements exist for live strategic responsiveness.

To address this gap, the authors developed AI-Trader, which spans three major financial markets: US stocks, A-shares, and cryptocurrencies, with multiple trading granularities to simulate live financial environments. The benchmark implements a fully autonomous minimal information paradigm, where agents receive only essential context and must independently search, verify, and synthesize live market information without human intervention.

The authors evaluated six mainstream large language models across three markets and multiple trading frequencies. The results show that general intelligence does not automatically translate to effective trading capability, with most agents exhibiting poor returns and weak risk management. The analysis reveals that risk control capability determines cross-market robustness, and that AI trading strategies achieve excess returns more readily in highly liquid markets than policy-driven environments.

The paper's contributions include the introduction of a novel benchmark for evaluating large language models in real-time financial markets, and the identification of critical limitations in current autonomous agents. The findings provide clear directions for future improvements, including the need for better risk control and the development of more effective trading strategies. The code and evaluation data are open-sourced to foster community research. Overall, the paper presents a significant step forward in the development of autonomous agents for financial decision-making, and highlights the need for further research in this area.

💡 The paper introduces AgentScope 1.0, a framework designed to support the development of agentic applications. The framework addresses the need for flexible and efficient tool-based interactions between agents and their environment, driven by the rapid advancements in Large Language Models. AgentScope provides a comprehensive set of tools and infrastructure to enable developers to build agentic applications, including unified interfaces, extensible modules, and advanced agent-level infrastructure based on the ReAct paradigm. The framework also includes built-in agents tailored to specific practical scenarios and robust engineering support for a developer-friendly experience. Additionally, AgentScope features a scalable evaluation module with a visual studio interface and a runtime sandbox to ensure safe agent execution and facilitate rapid deployment in production environments. The overall goal of AgentScope is to provide a practical foundation for building scalable, adaptive, and effective agentic applications, and the framework achieves this by providing a systematic asynchronous design that enriches human-agent and agent-agent interaction patterns while improving execution efficiency. The results of the framework are a set of tools and infrastructure that enable developers to easily leverage the latest progress in agentic applications, such as new models and MCPs, and to build long-trajectory agentic applications that are more manageable and easier to trace.

💡 The paper addresses the challenges of conducting large scale multi agent simulations with existing platforms, which include limited scalability, low efficiency, and effort intensive management processes. To overcome these challenges, the authors enhance the AgentScope platform by introducing several new features and components. They propose an actor based distributed mechanism to improve scalability and efficiency, and provide flexible environment support to simulate various real world scenarios. This allows for parallel execution of multiple agents, centralized workflow orchestration, and interactions among agents. The authors also integrate a configurable tool and an automatic background generation pipeline to simplify the process of creating agents with diverse background settings. Additionally, they provide a web based interface for monitoring and managing a large number of agents across multiple devices. The authors conduct a comprehensive simulation to demonstrate the effectiveness of the proposed enhancements and release the source code on GitHub to inspire further research and development in large scale multi agent simulations. The results show the great potential of applying multi agent systems in large scale simulations, and the enhancements to AgentScope improve its convenience and flexibility for supporting very large scale multi agent simulations.