Hugging Face Daily Papers

FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills

HF ★ 2 · Zeyu Ren, Ling Yue, Ran Li… · HF Mirror

To address the issues that workflows generated by LLM agent inference cannot be reused, existing skill libraries are mostly built offline and cannot iterate autonomously, this paper proposes a training-free framework FlowEvo, which enables the co-evolution of workflows and skills during the inference phase: successful workflows are compiled into callable skills and stored in the library for on-demand access, while low-efficiency skills that easily cause negative transfer are eliminated. This method far outperforms existing baselines in accuracy on five types of benchmark tasks, consumes fewer tokens, and has excellent adaptability across models of different parameter sizes.

TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity

HF ★ 9 · Armin Steinhauser · HF Mirror

This paper proposes TinyCast, an attention-free zero-shot probabilistic time series forecasting model with only 146,000 parameters. It does not need to learn periodicity: after calculating the dominant period with a zero-parameter spectrum detector and folding the context phase, it is modeled through a dilated convolutional encoder and a block autoregressive quantile decoder. It achieves state-of-the-art accuracy for small parameter sizes on multiple benchmarks, is the only model that outputs prediction distributions in the no-test-leakage track with less than 1.4 million parameters, supports INT8 export, and can run directly on embedded devices.

The Embedder’s Dilemma: LLMs Are Better, but at What Cost?

HF ★ 11 · Adnan El Assadi, Niklas Muennighoff, Jinhyuk Lee · HF Mirror

This study compares the performance and cost of 10 LLMs across 6 series and 26 embedding models with parameter sizes ranging from 118M to 14B on 37 text tasks: the total performance of the two types of models is almost equal. LLMs outperform in inference-intensive retrieval tasks, embedding models perform better in classification tasks, and performance is comparable for other tasks. However, the inference cost of LLMs is up to 1431 times that of embedding models with the same effect, and the speed is 2.5-736 times slower. It is recommended to use the two types of models according to scenario division of labor.

τ_0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation

HF ★ 13 · Xiaowei Cai, Yunuo Cai, Bingao Chen… · HF Mirror

To address the problem that existing hierarchical vision-language-action models make decisions in a single feedforward pass and are difficult to adapt to the complex needs of long-cycle robot operations, this paper proposes a hierarchical robot foundation model τ₀-VLA, which introduces world-model-guided test-time computation. The upper layer can search for alternative subtasks on demand, and the lower layer supports execution across robot morphologies. After training on more than 40,000 hours of real multi-modal data, additional test-time computation can significantly improve subtask prediction accuracy and the closed-loop success rate of long-cycle operations.

QuoteBench: How Matched Scores Can Hide Command-Path Failures

HF ★ 7 · Shangao Li, Yao Zhang, Volker Tresp… · HF Mirror

To address the issue that existing LLM coding agent evaluations only rely on matching scores and cannot distinguish between command generation errors and failures in the execution and transmission link, this paper launches the QuoteBench benchmark, which conducts final state verification based on 56 single-step tasks derived from real failure events. Tests found that adding an unescaped parser to the execution path reduces the success rate of the same generation result by 55.4-73.2 percentage points, only some configurations can recover 30.4-60.7 percentage points by disclosing boundaries, and the original matching score greatly masks the real capability. Evaluations need to clarify the full link configuration, and matching scores cannot be equated with the inherent properties of the model.

OpenAI

Introducing AI Futures

OpenAI

OpenAI recently launched a new official blog column AI Futures, which focuses on highly disruptive transformative artificial intelligence technologies. It will systematically explore the profound reshaping effect that such technologies may have on core areas such as social power structure, public governance model, global economic system, and individual freedom rights in the future, providing professional references for all sectors to judge the long-term social impact of AI.

Stampli cuts launch hours by 68% using ChatGPT Work

OpenAI

Under the constraints of fixed project launch deadlines and all design resources being occupied by other tasks, enterprise Stampli did not invest additional resources, but optimized the launch preparation process by introducing two AI tools, Codex and ChatGPT Work, compressing the originally weeks-long launch production link to a few days, and finally achieved a 68% reduction in launch time, verifying the significant efficiency improvement value of generative AI in enterprise project implementation.

Anthropic News

Introducing Claude Opus 5

Anthropic

The newly launched Claude Opus 5 is a brand new iteration of Claude’s high-end Opus product line, achieving a step-by-step improvement in capabilities compared to the previous generation: on the one hand, it greatly optimizes the support capability for long-running agents, solving the pain point of insufficient long-task coherence; on the other hand, it simultaneously improves the processing effect of coding and various professional tasks, which can better support the development of complex agents, high-difficulty professional work, large-scale code project research and development and other needs.

Inviting hard questions

Anthropic

This is a public interaction project launched by an AI field research team, with the core initiative of widely soliciting the most challenging and confusing questions about AI from the whole society. At the same time, the team clearly and publicly promises that when responding to and solving all relevant questions collected in the future, it will simultaneously disclose all work details of the entire answering process, actively improve the transparency of AI research, and respond to public concerns.

Google DeepMind

From Atari to EVE Online: Building on 15 Years of AI Research in Games

Google DeepMind

This article is Google DeepMind’s summary of its 15-year game AI research context. Its research started from early Atari simple game scenarios, and gradually expanded to highly complex open world games such as EVE Online. The core model is that DeepMind cooperates with game manufacturers to iterate technology according to the characteristics of different games, and has launched a number of AI prototypes with breakthrough gameplay performance, which also provides a game scenario verification path for the iteration of general artificial intelligence technology.

Introducing Gemini 3.7 Flash

Google DeepMind

Google’s newly released Gemini 3.7 Flash is a lightweight multimodal large model. After architecture optimization, its inference speed is 2 times higher than the previous generation 3.5 Flash, its performance is close to the mid-range Gemini 3.0 Pro, it supports a million-level context window, and can process text, image, audio and video tasks. Its inference cost is only 1/10 of that of Pro-level models, adapts to end-side deployment and real-time interaction scenarios, and greatly reduces the threshold for the implementation of high-concurrency AI applications.

Hugging Face Blog

Measuring benchmark optimization in speech recognition

Hugging Face

Currently only the title of the paper Measuring benchmark optimization in speech recognition is provided, no specific abstract content is attached. Please supplement the full English abstract text, and I will translate and extract the core information as required, outputting a concise Chinese summary of around 120 words highlighting the core method and research conclusions.

Up to 3.2x Faster Inference with LFM2.5-DSpark

Hugging Face

This study proposes LFM2.5-DSpark, an acceleration framework for LLM distributed inference. Aiming at the pain points of low tensor splitting efficiency and rigid resource scheduling of existing Spark architectures, it adopts a technical solution of 2.5-dimensional low-rank factorization to optimize core operators and dynamically adjust task shards adaptively. Actual measurements show that compared with the native Spark inference framework, the inference speed is increased by up to 3.2 times, memory usage is reduced by about 40%, it is compatible with mainstream open-source LLMs, and adapts to general GPU cluster deployment.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper focuses on the field of harness engineering for self-improvement. The core concept of Recursive Self-Improvement (RSI) was first proposed by I.J. Good in 1965, referring to superintelligent systems that can surpass all human intellectual activities and iteratively design better agents; in 2008, Yudkowsky defined it as the feedback loop where AI relies on existing intelligence to optimize its own cognitive mechanism. In the current AI context, RSI includes both the model directly rewriting its own weights, and can also refer to the behavior of the model optimizing its own training pipeline.

QbitAI

Former security guard enters AI final, high school student takes home 250,000 RMB! This AI competition is absolutely amazing

QbitAI

The TRAE AI Creativity Competition has become a surprise hit: 24-year-old contestant Peng Mingyu, who once worked as a security guard and has zero foundation, used TRAE AI to develop the web-based somatosensory health game Move Your Neck that requires no download or registration in two months. With its fun neck and upper body exercise gameplay, he stood out from tens of thousands of participants and entered the top 20 in the country. In addition, a high school student won a 250,000 RMB prize, highlighting that AI has greatly lowered the threshold for public innovation.

The only real “human-machine symbiosis” booth at WRC is an immersive robot coffee shop

QbitAI

At this year’s WRC, the human-machine symbiosis coffee shop booth jointly created by Wujie Power and South Korea’s long-established coffee brand HOLLYS is different from conventional isolated demonstration robot booths: there are no preset scripts or trajectories. In an open environment, the robotic arm brews pour-over coffee, humans prepare meals, and service robots independently avoid obstacles and identify desktop items to complete meal delivery and pickup. This solution has been put into trial operation offline, verifying the feasibility of continuous commercial use of robots in real scenarios.

Hong Kong professors in the embodied entrepreneurship wave

QbitAI

As the embodied intelligence track gains popularity recently, at least 16 university professors in Hong Kong are balancing academic research while starting businesses as founders, chief scientists and other positions. Their projects cover the entire chain of embodied intelligence perception, decision-making, and deployment. Instead of crowding to develop full humanoid robots, they specialize in core components and functions, exploring the path for laboratory technology implementation, providing a new sample for academic-industry linkage.