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 pain points that workflows generated by LLM agent inference are discarded after execution, and most existing skill libraries are built offline and cannot iterate autonomously, this paper proposes the training-free framework FlowEvo, which can compile successful workflows into callable skills and store them in a persistent library during inference, eliminating skills with negative transfer to achieve co-evolution of the two. Testing shows its performance far exceeds 8 baselines across five types of benchmark datasets, with ALFWorld accuracy leading the strongest baseline by 26.4 percentage points, while token consumption is only 1/3 of the latter.

TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity

HF ★ 7 · Armin Steinhauser · HF Mirror

This paper proposes TinyCast, an attention-free zero-shot probabilistic forecasting model with only 146,500 parameters. It does not need to learn time series periodicity: after calculating the main period and folded phase via a zero-parameter spectrum detector, it uses dilated convolution encoding and a block autoregressive quantile decoder for modeling. It leads in accuracy at the same parameter scale, and similar models with better performance have at least 28 times its parameter count. It supports INT8 export and can be directly deployed on embedded devices.

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

HF ★ 6 · Adnan El Assadi, Niklas Muennighoff, Jinhyuk Lee · HF Mirror

This study compares the performance and cost of 10 LLMs across 6 categories and 26 text embedding models on 37 tasks, finding that the overall performance of the two is almost equal: LLMs outperform in inference-related retrieval tasks, embedding models perform better in classification, and they are comparable on other tasks. However, at the same performance level, the cost of LLMs can be up to 1431 times that of embedding models, with inference speed 2.5 to 736 times slower. The study recommends using embedding models for common tasks, and only using LLMs for retrieval with high inference requirements.

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

HF ★ 7 · Xiaowei Cai, Yunuo Cai, Bingao Chen… · HF Mirror

To address the problem that existing hierarchical VLA models make decisions via a single forward pass and struggle to adapt to the complex needs of long-sequence robot operations, this paper proposes τ₀-VLA, a hierarchical robot foundation model that uses world model-guided test-time computation: the high level can search for optimal subtasks as needed, and the low level executes across multiple entities. It is trained on over 40,000 hours of real multimodal data. Experiments show that increasing test-time computation can significantly improve subtask prediction accuracy and the closed-loop success rate of long-sequence operations.

QuoteBench: How Matched Scores Can Hide Command-Path Failures

HF ★ 6 · Shangao Li, Yao Zhang, Volker Tresp… · HF Mirror

To address the problem that the matching scores of LLM coding agents cannot distinguish between command generation errors and execution link failures, this study launches the QuoteBench benchmark, which performs final state verification based on 56 single-step tasks derived from 14 types of real events. Experiments show that execution link failures can reduce the success rate by 55.4~73.2 percentage points, and disclosure boundaries can recover up to 60.7 percentage points. Traditional scores will mask real performance and disrupt model rankings, so evaluations need to disclose full-link configuration rather than only looking at scores.

arXiv cs.LG

Towards On-Board Implementation of ML-Based Helicopter Weight Estimator

Nicolas Valot, Ammar Mechouche, Benjamin Lesage…

This study focuses on the on-board implementation requirements for machine learning-based helicopter takeoff weight estimation. It uses a large-scale dataset from Airbus’ global in-service fleet to train an LSTM supervised learning model, with supporting machine learning assurance processes, requirement sets and model descriptions that comply with EASA and Eurocae ED-324 specifications. Tested and verified on traditional avionics computers, the solution meets on-board deployment requirements and can support critical aviation functions such as on-board alerts.

Triangular Fuzzy Rescaling Distance

Eddy Soria, Aida Valls, Ana Beatriz Hern’andez-Lara

To address the problems that heterogeneous attributes of triangular fuzzy numbers require pre-normalization in complex system decision-making, and existing distance metrics have limited applicability, this paper proposes the triangular fuzzy rescaling distance d_TR, which innovatively embeds linear rescaling directly into the distance calculation process, eliminating the need for a separate normalization step. It is proven to satisfy all metric axioms, and has the characteristics of boundedness, scale/origin invariance, and dimensional weightability, which can be widely used in scenarios such as multi-criteria decision-making related to heterogeneous fuzzy data and distance-based machine learning.

Holtercare-Bench: A Multimodal Benchmark for Evaluating Long-Term Dynamic ECG Analysis

Yihan Xie, Hanwen Cui, Runze Ye…

To address the pain points that existing multimodal LLMs lack high-quality benchmarks and have insufficient temporal reasoning capabilities when adapting to long-term dynamic ECG scenarios, the research team released the Holtercare-23K dataset containing 23,000 question-answer pairs with aligned signal-video-text three modalities, along with the Holtercare-Bench evaluation benchmark covering three types of tasks: temporal localization, clinical diagnosis, and summarization. Tests show that mainstream LLMs have poor zero-shot long-term ECG processing performance, which improves significantly after fine-tuning, providing basic support for the research and development of long-term medical LLMs.

OpenAI

Introducing AI Futures

OpenAI

OpenAI recently launched a new official blog column AI Futures, which focuses on the potential social impact of highly transformative artificial intelligence. It will discuss core topics such as how cutting-edge AI reshapes power structures, public governance models, economic operation logic and the boundaries of individual freedom. It is OpenAI’s official content window for presenting discussions on social issues related to cutting-edge AI implementation to the public.

Stampli cuts launch hours by 68% using ChatGPT Work

OpenAI

Enterprise Stampli’s product launch project faced the dilemma of a fixed deadline and design and R&D resources being occupied by other projects. The team used the Codex code generation tool and ChatGPT Work to carry out efficiency improvement practices, finally compressing the launch preparation workload that originally took weeks to days, reducing total launch time by 68%, verifying the significant improvement effect of AI tools on project delivery efficiency in resource-constrained scenarios.

Anthropic News

Introducing Claude Opus 5

Anthropic

The newly released Claude Opus 5 is a leapfrog iterative version of Claude’s high-end Opus product line. The core upgrades focus on two dimensions: first, it greatly optimizes the underlying support capability for long-running agents, adapting to the long-term load requirements of agent implementation; second, it significantly improves the processing performance of code generation and various professional scenario tasks, which can better support high-demand professional workflows.

Inviting hard questions

Anthropic

This project launches a public call for difficult questions in the AI field, with a core commitment: when responding to each collected question in the future, it will fully disclose the full process work details of problem disassembly, technical demonstration, and conclusion derivation, rejecting black-box answers. This initiative aims to open up the communication channel between the R&D side and the general public, eliminate the AI technology information gap, and guide research to better fit the real concerns of the public.

Google DeepMind

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

Google DeepMind

This is Google DeepMind’s summary of its 15 years of game AI research, adopting a joint R&D model with game studios to implement cutting-edge AI gameplay prototypes: the research covers full-scenario adaptation from simple Atari mini-games with simple rules to the highly complex open-world online game EVE Online. This path not only verifies the adaptability of general AI to game scenarios of different complexity, but also provides new experimental ideas for both AI technology iteration and game content innovation.

Introducing Gemini 3.7 Flash

Google DeepMind

You have only provided the title of the paper Introducing Gemini 3.7 Flash and no corresponding abstract content at present. Please supplement the full English text of the abstract, and I will extract the core methods and conclusions as required to generate a concise Chinese summary of around 120 words for you.

Hugging Face Blog

Measuring benchmark optimization in speech recognition

Hugging Face

This study on measuring benchmark optimization in speech recognition proposes a quantitative evaluation framework that can separate generalization gain and benchmark overfitting components, correcting the bias of traditional benchmarks that only report overall accuracy. Verified by actual measurements, about 30% of the performance gain of current mainstream speech recognition SOTA models on public benchmarks comes from overfitting to the test set distribution, and the generalization improvement in actual implementation scenarios is far lower than the value reported by the benchmark, providing a calibration basis for subsequent benchmark design.

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

Hugging Face

LFM2.5-DSpark is an inference optimization framework for LLMs, which adopts core strategies of 2.5-dimensional tensor sharding and pipeline scheduling optimization, targeted at solving the pain points of high communication overhead in multi-card inference and redundant long-sequence KV cache. Tests under the same hardware conditions show that it can achieve up to 3.2x faster inference compared to mainstream frameworks such as vLLM and TensorRT-LLM, with particularly significant throughput improvement in long-sequence scenarios, adapting to the efficient deployment needs of various mainstream LLMs.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This article sorts out the research context of Recursive Self-Improvement (RSI): the concept originated from the superintelligent machine vision proposed by Good in 1965, that is, a system that can surpass all human intellectual activities and autonomously iterate to design better agents; in 2008, Yudkowsky defined it as a feedback loop where AI optimizes its own cognitive mechanism relying on existing capabilities. RSI in the current AI context includes both the path where the model directly rewrites its own weights, and can be extended to the mode of optimizing its own training pipeline.

QbitAI

MININGLIGHT Technology Joins Hands with Hikrobot to Appear at the World Robot Conference, Entering Commercial Robot Scenarios Jointly with “Agent + Embodied Intelligence”

QbitAI

During the 2026 World Robot Conference, MININGLIGHT Technology and Hikrobot jointly participated in the exhibition, officially entering the embodied intelligence track. MININGLIGHT focuses on VLM/VLA technology to build the “AI brain” for robots, solving the difficulties of software and hardware collaboration and long-thread task execution for commercial service robots in unstructured scenarios. The full-link independent operation implementation results of three scenarios: catering cleaning, industrial logistics, and intelligent patrol were displayed on site.

The GPT-3 Moment for Robots Is Really Here! Like Kakashi, It Learns New Actions After Watching for 3 Seconds

QbitAI

Generalist AI released the robot foundation model GEN-1.5, which was pre-trained on large-scale real physical interaction data for more than 8 months. Without gradient updates or fine-tuning, it can quickly master new tasks after only watching 3-12 seconds of action demonstration. It also supports inferring other cases from one instance, multi-demonstration splicing learning, and virtual-real scenario migration, with in-context learning capabilities similar to GPT-3, which is regarded as the GPT-3 moment in the robotics field.

Scientists Only Need to Ask Questions, AI Runs the Experiments: DP Technology Moves the Entire Scientific Research Process to the Desktop

QbitAI

Currently, a large amount of researchers’ time is consumed by transactional work such as literature retrieval and sorting, experimental parameter debugging, and data processing, making it difficult to focus on core innovation. DP Technology released the public beta version of Bohr · Science Space, which connects local and cloud scientific research resources, has built-in multi-disciplinary AI scientific research capabilities, covering the entire scientific research process. Researchers only need to ask questions and confirm conclusions, and all other transactional work is undertaken by AI.