HuggingFace Daily

"DarwinX: Evolving Agent Harnesses Through Natural Selection" by Yifan Zhang , Yutong Dai , Juntao Tan , Luyu Yang , Rishi Mullur , Thai Hoang , Zhiyuan Hu , James Zhu , Phil Mui , Silvio Savarese , Ran Xu , Zeyuan Chen TLDR: The text discusses DarwinX, a system that evolves agent harnesses through population selection with frozen models, leading to improved performance across benchmarks without needing benchmark-specific patches. The system enhances an agent's capabilities by considering not just model weights, but also its harness, which includes prompts, tools, skills, and control flow. DarwinX treats self-evolution as selection over a population of harnesses while keeping the model frozen, allowing for the preservation and extension of successful variants without regression. The fitness of each agent is measured by its performance on benchmarks, without relying on gold solutions or hand-picked winners. The system's effectiveness is demonstrated across different benchmarks, showcasing significant performance improvements. The evolution process focuses on enhancing general agent competence rather than task-specific enhancements, ensuring adaptability across various tasks, verifiers, and base models. Read Paper / Blog

"Alaya-EVOKE: From Linear-Scaling Supervision to Endless World" by Yuanyang Yin , Gongxuan Wang , Yifan Zhan , Chuanhao Li , Kaipeng Zhang , Feng Zhao TLDR: Evoke is an innovative interactive world model that functions by utilizing external persistent memory and a specially designed long-horizon teacher to facilitate the generation of responsive, unrestricted videos with limited context and minimal delay. This model strives to balance the requirements of persistent memory support, interactive responsiveness, and long-horizon generation, which can often conflict with one another. By storing scene geometry in an external memory bank and optimizing the teacher for long-horizon supervision through sparse attention mechanisms, Evoke effectively manages memory and computational resources while allowing for continual evolution and generation in a bounded context. This approach results in superior performance on benchmarks such as WBench, VBench-Long, and VBench-2.0, showcasing the model's efficacy in generating sequences efficiently and effectively. Read Paper / Blog

"Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence" by Haokai Zhang , Yuhang Ding , Yunshu Zhou , Xinze Du , Shengtao Zhang , Zhiyue Zhao , Yuling Xi , Hao Chen TLDR: The text discusses how a frozen vision-language model (VLM) can enhance spatial reasoning without relying on parameter updates or external tools. The Spatial Memory Agent (SMA) framework is introduced, which enables the frozen VLM to self-evolve through verified spatial experiences converted into reusable lessons. SMA uses reflection and memory retrieval to distill lessons from spatial experiences, assigning each lesson a Transfer Reliability Score (TRS) for future use. By retrieving these lessons during deployment, the frozen VLM improves its spatial reasoning ability. Through experiments across various benchmarks and VLM models, SMA demonstrates superior performance, offering a parameter-update-free approach to spatial self-evolution. Read Paper / Blog

"Intern-S2-Preview: Scientific Agentic Foundation Model" by Lei Bai , Jiaqi Cao , Chiyu Chen , Guanzhou Chen , Kai Chen , Guangran Cheng , Erfei Cui , Xuanlang Dai , Shengyuan Ding , Shangheng Du , Yanhui Duan , Yue Fan , Youqing Fang , Quan Gan , Yuanyuan Gao , Jiaye Ge , Lixin Gu , Yuzhe Gu , Qipeng Guo , Junjun He , Xin Hong , Ming Hu TLDR: Intern-S2-Preview is a series of scientific agentic foundation models that aim to enhance scientific reasoning and forecasting by integrating multimodal pre-training, multi-task reinforcement learning, and memory-augmented extensions. These models were designed to support long-horizon scientific tasks and are capable of reasoning over diverse scientific evidence while interacting with scientific tools and environments. The training pipeline involves pre-training on various scientific data followed by supervised fine-tuning, multi-task reinforcement learning, and memory augmentation techniques. The architecture focuses on time series modeling for efficient long-sequence understanding and numerical forecasting. Evaluations across benchmarks show that Intern-S2-Preview-397B achieves competitive or leading results by improving scientific signal understanding, forecasting, and overall performance across various settings. Read Paper / Blog

"DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation" by DreamX Team , Rui Chen , Xiangxiang Chu , Geng Li , Jifan Li , Qingfeng Shi , Datao Tang , Jing Tang , Jun Wang , Pengfei Zhang TLDR: DreamX-Phi 1.0 is a sophisticated action-conditioned video world model designed for robotic manipulation. It features geometric attention encoding, depth estimation, object masks with a frozen teacher, and distillation techniques to predict accurate future observations. By incorporating per-arm SE(3) transformations, the model ensures that predicted actions respect each arm's commanded path, maintaining arm identity and rigid-motion structure. Additionally, a depth branch and SAM3 masks are utilized to preserve scene geometry and object consistency during grasping. Through distribution-matching distillation, the multi-step generator is condensed into a more efficient few-step student model, contributing to its success in the WorldArena~2.0 Challenge. The creators plan to make the model and code publicly available. Read Paper / Blog

"LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers" by Tao Feng , Fangxu Yu , Haozhen Zhang , Zhongjie Dai , Liangqi Yuan , Zijie Lei , Weizhi Zhang , Kunlun Zhu , Haodong Yue , Keyang Xuan , Ge Liu , Jiaxuan You TLDR: The text discusses LLM routing, which is essential for deploying large language models in a cost-effective manner due to the lack of a single optimal model for all scenarios. Different formulations and implementations of LLM routing make comparison challenging, leading the author to propose a unified framework comprising context encoders, model encoders, scoring functions, decision rules, and learning signals for single-turn, multi-turn, and personalized routing. They introduce xRouteBench for evaluating routers across various tasks and LLMRouter as an open-source infrastructure with multiple routers. The study demonstrates that learned routers outperform fixed models, lightweight routers are effective under budget constraints, and user-conditioned routing enhances personalization. Read Paper / Blog

"AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design" by Yaxin Luo , Haobin Jiang , Jialv Zou , Xu Huang , Wenhao Yan , Haodong Li , Zhengrong Yue , Jing Li , Xiaofu Chen , Xiaohan Zhao , Jiacheng Liu , Jiacheng Cui , Zhiqiang Shen , Xiaotong Li TLDR: AutoDesign is a framework that uses a meta-harness optimizer to enhance a code agent's ability for generating structured media, specifically excelling at paper-to-poster synthesis. The system is designed to align with human design preferences and focuses on accumulating reusable experience through recursive self-improvement. By introducing PosterBench and PosterBench-mini for evaluation, AutoDesign achieves impressive results in academic paper-to-poster generation, outperforming existing systems like Claude Design. The framework demonstrates significant performance improvements in various controlled configurations and receives high human preference in system-blind evaluations, showcasing its efficiency, cost-effectiveness, and ability to generate quality outputs autonomously. Read Paper / Blog

"Articulated Object Reconstruction from Rest-State Observation" by Daeun Lee , Jaeah Lee , Woosung Kim , Haebeom Jung , Jaesik Park TLDR: The text describes a novel rest-state framework that aims to reconstruct articulated objects from a single closed configuration by integrating vision-language outputs into consistent part meshes and verifying synthesized motion hypotheses through geometric consistency. Unlike existing methods that require observable motion from various articulation states, this approach addresses the challenging task of recovering both 3D geometry and kinematic structures in an inherently ambiguous scenario where motion cues are absent. By utilizing an explicit mesh representation for model fusion and verification, the framework effectively reconciles outputs from vision-language and segmentation models, leading to accurate part decomposition and physically realistic articulation. Through the use of a video diffusion model to estimate joint parameters without direct motion observation, the proposed method achieves competitive results compared to other reconstruction techniques. Read Paper / Blog

"Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design" by Qing Zong , Jiayu Liu , Junhao Shen , Zecong Tang , Linsi Wu , Yuxuan Liu , Rui Wang , Zhaowei Wang , Weiqi Wang , Cheng Qian , Xiusi Chen , Yangqiu Song TLDR: Agentic systems can continuously improve by undergoing multi-component co-evolution that progressively eliminates fixed human constraints in agents, environments, and evolution mechanisms. The text emphasizes the need for systems to evolve post-deployment and highlights how single-entity self-evolution can be limited by static learning contexts. The survey delves into co-evolution in agentic systems, where multiple agents and their environment interact to drive adaptive pressure. It introduces a three-stage taxonomy that outlines the evolutionary progression as the system sheds human-engineered constraints. The taxonomy includes Agent--Agent Co-Evolution, exploring how agents adapt through dynamic peers; Agent--Environment Co-Evolution, which extends this adaptation to changing tasks and feedback; and Meta Co-Evolution, examining the evolvability of the evolution mechanism itself. The text also addresses challenges in evaluating and scaling these systems, ensuring their safety and control while paving the way for robust and open-ended agentic systems that surpass predetermined human paths. Read Paper / Blog

"Beyond Pixels: From Video Priors to 4D Worlds" by Zihao Liu , Xiaolong Shen , Zhenglin Zhou , Ruijie Quan , Yi Yang TLDR: The text describes a new method, Latent-to-4D, which allows for reusable direct 4D generation from video diffusion latents using alignment with a pretrained decoder and spatiotemporal attention. This approach generates dynamic 3D scenes from text or images without needing to retrain the generators. Previous methods either reconstruct RGB videos with a separate 4D model or adapt a specific video generator to predict geometry directly, but these methods have limitations such as distribution mismatch or the need for retraining with changes in conditioning regimes. Latent-to-4D proposes using final denoised latents of video models that share a variational autoencoder to provide an interface for explicit 4D prediction. This method surpasses existing models in performance metrics and is preferred by human raters for geometry, temporal stability, and overall quality. Read Paper / Blog

"ComBodied Agents: a New Paradigm of Human-Centric Agentic AI" by Qianggang Ding , Xingyao Wang , Rui Feng , Zhibin Wang , Feixiang Wang , Kelong Mao , Hao Sun , Zhiyao Luo , Jiankai Tang , Lei Li , Jiadong Guo , Minheng Ni , Weicong Lin , Chenxi Yang , Hongxiang Gao , Zhenghua Chen , Yang Bai , Min Wu , Jun Cheng , Huazhu Fu , Dacheng Tao , Bang Liu TLDR: The text discusses the concept of Combodied Agents, which are designed to integrate digital and physical tools to better understand and support individual human-state trajectories over time while prioritizing consent and appropriate support. The current gap in Agentic AI lies in not fully considering an individual's evolving state and agency in modeling and intervention strategies. With Combodied Agents, a closed-loop framework is proposed that incorporates software tools, sensors, wearables, robots, and human services as action channels. This approach aims to unify personal assistants, health agents, AI companions, and human-AI systems to offer personalized and consent-aware support. The framework emphasizes modeling and predicting human-state trajectories, using feedback to update interventions and focusing on sustained human benefit rather than just task completion. Read Paper / Blog

"AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss" by Mingju Gao , Jingkai Zhou , Kun Gai , Changqian Yu , Hao Tang TLDR: Adversarial Fréchet Distance enhances generator post-training by integrating a learnable adversarial feature space with static Fréchet losses, incorporating whitening to enhance optimization stability. While conventional sample-level losses contribute to generator post-training, adopting Fréchet distance can lead to Fréchet hacking and hinder visual quality. The proposed AdvFD method addresses this issue by introducing a dynamic, adversarially learned representation to supplement traditional Fréchet objectives. By maximizing the discrepancy between real and generated samples through an adaptive feature space and employing real-feature whitening to prevent feature amplification, AdvFD consistently enhances generator post-training in various experimental settings. Read Paper / Blog

"Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution" by Anton Razzhigaev , Andrei Gritsaev , Andrei Kaznacheev , Nikita Dragunov , Roman Yampolskiy , Andrei Kuznetsov TLDR: Ouroboros is a self-developing agent harness that evolves through reviewed commits, improving its tools, prompts, context assembly, and core implementation for later work. Core evolution occurs in two modes: recursive free evolution and experience-driven core evolution. Noteworthy achievements include setting high scores on benchmarks like Terminal-Bench 2.1 and OSWorld-Verified, as well as achieving a new state of the art reward in a CL-Bench campaign. The text also highlights "Hope," a 161-day living agent experiment in free evolution under human communication, emphasizing the importance of operational safety and maintaining guardrails under evolutionary and social pressure. Benchmark campaigns use frozen system snapshots, while "Hope" continues to evolve on a separate lineage in real-time. Read Paper / Blog

"SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring" by Yuling Shi , Jinghan Xu , Kelin Fu , Wenhao Zeng , Shilin He , Lei Zhang , Yue Liu , Zelin Zhao , Terry Yue Zhuo , Jialun Cao , Siyu Ye , Tianyu Liu , Kai Cai , Shing-Chi Cheung , Xiaodong Gu TLDR: SWE-Bench ProMax is a carefully curated multilingual benchmark of large-scale code refactoring tasks designed to highlight the challenges faced by existing AI coding agents. Current benchmarks are becoming saturated and their evaluation quality has been questioned, with nearly 60% of unsolved instances containing flawed tests and models able to reproduce training data. Code refactoring, requiring complex changes across multiple files, offers a realistic test of agent capability yet remains underserved. SWE-Bench ProMax addresses quality issues by providing precise specifications and removing problematic tests, resulting in 170 challenging instances from real commits across different programming languages. The benchmark proves to be a significant challenge for current AI coding models, with the best model achieving a resolve rate of only 41.2%. Read Paper / Blog

"Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA" by Mind Lab , Vin Bo , Asher Cai , Jingwei Cao , Song Cao , Vic Cao , Amelia Chen , Andrew Chen , Kaijie Chen , Cleon Cheng , Steven Chiang , Kaixuan Fan , Hera Feng , Huan Feng , Arthur Fu , Jun Gao , Pyke Han , Nolan Ho , Ori Hong , Hailee Hou , Piers Hua , Charles Huang TLDR: Macaron-V1 is an open agent-model family that leverages a Mixture-of-LoRA architecture and recursive self-improvement to facilitate ongoing learning and collaboration in specialized tasks. The system focuses on experiential intelligence, allowing learning from real-world situations and after deployment. It emphasizes adaptation through recursive improvement of model-harness pairs and collaboration via the MoL architecture, which includes specialist LoRA adapters. Macaron-V1 offers different versions with various capabilities, such as Macaron-V1-Venti and Macaron-V1-Tall. The system is a co-designed framework that combines architecture, algorithms, and infrastructure to enable continual learning through extensible LoRA specialists. Evaluation against benchmarks shows promising outcomes, raising questions about the potential for further gains through continuous learning and collective intelligence. Read Paper / Blog

"Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory" by Taeil Kim , Kangsan Kim , Sung Ju Hwang TLDR: The text discusses Agent Memory Distillation (AMD), a method that enhances the performance of small language models by transferring hierarchical memory from a larger teacher agent without requiring additional training. The AMD framework utilizes three memory types (Workflow memory, Subtask memory, and Function memory) derived from successful teacher trajectories to aid the student agent in task-level strategies, concrete examples, and calling conventions. AMD was evaluated on three tool-use benchmarks with GPT-5-mini as the teacher agent, resulting in significant accuracy improvements over existing baselines. Subtask memory was found to provide the most significant gains, and 4B-sized student models benefitted the most from AMD. Read Paper / Blog

"BDH-CQ: In-Context Learning with Recurrent Latent Reasoning" by Björn Engdahl , Adrian Kosowski , Jan Chorowski , Zuzanna Stamirowska , Przemysław Uznański , Junlin Jiang , Rohan Phadke , Remigiusz Kinas , Richard Zhong TLDR: A new reasoning model called BDH-CQ, utilizing recurrent latent reasoning and in-context learning with 150 million parameters, has achieved a groundbreaking cost-accuracy frontier on the ARC-AGI-1 evaluation dataset. The model updates its memory with input at the time of inference and conducts query solving through iterative computation in a latent space without explicit reasoning steps. By evaluating the model's performance and studying its learning behavior on the ARC-AGI-1 evaluation set, it was found that the model reaches an impressive 29.5% pass rate at a low cost of $0.0007 per task, surpassing previous benchmarks and establishing a new standard for benchmark cost efficiency. Read Paper / Blog

"Beyond Simply Environment Scaling: Designing Effective Environment Distributions for Multimodal Agent Learning" by Kejian Zhu , Zhuoran Jin , Dongqi Huang , Hongbang Yuan , Yupu Hao , Kang Liu , Jun Zhao TLDR: The text discusses enhancing multimodal agent training by selecting diverse environments utilizing ability-aware selection and organizing difficulty levels through hierarchical curriculum learning. Despite the common approach of creating large pools of such environments, simply increasing their number does not always lead to improved training. The text presents experiments highlighting the shortcomings in current distributions of multimodal environments, which then lead to the introduction of Ability-aware Environment Selection (AES) for diversity and Hierarchical Difficulty Curriculum (HDC) for difficulty structuring. AES focuses on obtaining diverse environment sets, while HDC organizes learning through two levels of difficulty. Experiments demonstrate that implementing AES and HDC leads to more effective training for multimodal agents. Read Paper / Blog

"SFT Conflicts, RL Coexists: A Theoretical and Empirical Analysis of Multi-Task Learning for LLMs" by Kejian Zhu , Zhuoran Jin , Shangqing Tu , Hongbang Yuan , Yushi Bai , Kang Liu , Juanzi Li , Jun Zhao TLDR: Reinforcement learning facilitates stable multi-task reasoning in large language models by employing sparse, near-orthogonal parameter updates, which contrasts with the task conflicts often seen in supervised fine-tuning. The Parallel-RL training approach, as opposed to Inkling-Small Supervised Fine-Tuning (SFT), highlights the distinct behaviors of RL and SFT in enhancing multi-task reasoning for LLMs. The research indicates that SFT struggles with task conflicts during multi-stage training, while RL allows for a stable coexistence across various tasks due to sparse and nearly orthogonal parameter updates. The study traces this difference to the parameter level, where RL demonstrates a small variance bound, ensuring near-orthogonal optimization directions across tasks. By analyzing multi-task gradient interference, the authors propose Parallel-RL as a more efficient and flexible training paradigm based on these insights. Read Paper / Blog

"StreamArena: Toward Continuous, Interactive, and Long-Horizon Agentic Streaming Video Understanding" by Xichen Zhang , Guankai Li , Yinghao Zhu , Shijian Wang , Sitong Wu , Shaozuo Yu , Meng Chu , Yuan Lu , Jiaya Jia TLDR: The text discusses the challenges faced when deploying autonomous multimodal agents in real-world scenarios and the limitations of current evaluation methods that rely on short clips and multiple-choice questions. To address these challenges, the authors introduce StreamArena, a benchmark for evaluating video comprehension in continuous, interactive streaming content. StreamArena consists of full-length videos with annotated open-ended questions to assess real-time perception, historical understanding, proactive interaction, and multimodal tool use. The authors highlight the trade-off between continuous interaction and long-term multimodal understanding, noting the difficulties in retaining information over time. To overcome these obstacles, they propose StreamMind, a two-tier architecture that improves performance and reduces latency by efficiently managing interactive and memory-based tasks between frontend and backend workers. Read Paper / Blog