Hugging Face Daily Papers

DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation

HF ★ 21 · Jiahao Zhan, Yan Wang, Yongrui Ma… · HF Mirror

To address the shortcomings of existing joint distribution matching methods for few-step video generation, including insufficient image quality and semantic alignment deviations, this paper proposes the DuoMatching framework: it adds a marginal matching objective on top of the original joint matching, paired with LatentBridge to align the latent space, and latent variable sampling to assign frame-level supervision, transferring high-quality priors from image generation. Experiments show improvements in its performance in image quality, semantic alignment and other aspects, with a human evaluation preference rate exceeding 80%, outperforming all compared baselines.

AutoSciBench: Autonomous Benchmark Generation for Evaluating Scientific Agents

HF ★ 15 · Dongki Kim, Namkyeong Lee, Surag Nair… · HF Mirror

To address the pain points of existing evaluation benchmarks for scientific research agents, which are easily saturated and have high manual update costs, this study proposes the AutoSciBench framework, which can automatically generate benchmark tasks and adjust iteratively along with agent capabilities, and can also reuse past optimization experience to expand new tasks. Tests across three fields including computational biology show that the benchmarks it generates reduce agent solution accuracy by 22.4 to 25.5 percentage points compared to manual benchmarks, with higher task quality, and can meet the evaluation needs as agent capabilities evolve.

HuatuoGPT-3: RL-Only Domain Adaptation from Base Models

HF ★ 14 · Junying Chen, Xinyuan Xie, Ziniu Li… · HF Mirror

To address the flaws of mainstream SFT+RL pipelines or pure RL solutions for domain adaptation of general large models, including cold start, gradient starvation, and teacher distribution anchoring, this paper proposes OnePO, a single-stage policy optimization method for pure RL adaptation, paired with adaptive target evolution and a teacher exit mechanism. Its medical adaptation effect with only 20,000 samples outperforms baselines, and the developed open-source HuatuoGPT-3 27B version outperforms GPT-6 Astra in medical evaluation performance.

EVISKILL: Grounding Skill Evolution in Replayable Evidence

HF ★ 10 · Yan Zhou, Yili Wang, Yiwei Dai… · HF Mirror

To address the problems in continuous skill evolution of large model agents, where existing experience-driven methods easily lose supporting evidence for edits and have one-sided global verification judgments, this paper proposes the evidence-driven framework EVISKILL: it organizes execution observations into reproducible evidence cards, associates edits with corresponding supporting context, verifies modifications through targeted replay, temporarily stores valid modifications in stages for iteration, and finally solidifies them into skills after global verification. This method has been verified to be effective on 3 types of interaction benchmarks and 6 large model bases.

Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents

HF ★ 10 · Feiyu Gavin Zhu, Xiaoyu Zhu, Jiqi Yang… · HF Mirror

To address the problems of open-ended reasoning redundancy and high inference cost for small-parameter multimodal search agents, this paper proposes the Selection-based Structured Reasoning (SSR) framework, which reconstructs reasoning from open-ended generation into a selection task for predefined reusable candidates, combined with a shared KV cache parallel scoring mechanism to optimize efficiency. Tests on 7 multimodal search benchmarks show that SSR’s performance is on par with mainstream models of the same scale, with single-round inference latency reduced by more than 90%, and total inference latency reduced by 28% to 54%.

arXiv cs.LG

TEMPEST: Temporal Embeddings for Scalable Driver Identification via Angular Margin Learning

Kyle Musgrove, Dylan B. Lewis, Sarah Powers…

To address the pain points of existing triplet loss models for scalable driver identification, where performance drops sharply as the number of drivers increases and they easily overfit session features, this paper proposes the TEMPEST model: it uses a temporal convolutional network paired with ArcFace additive angular margin loss for training, outputs 96-dimensional compact driving behavior embeddings, and supports dynamic registration without retraining. It achieves a Rank-1 accuracy of 91.71% on a 45-person dataset, and its performance in scale expansion and cross-session scenarios is far superior to all baselines. The model is small in size and fast to converge, and can be used as a domain benchmark.

When Does External Guidance Help LLM Reasoning? A Bias-Variance Theory of Guidance-Augmented GRPO

Sofia Torres, Gabriel Almeida, Carter Adams…

To address the lack of theoretical support for reinforcement learning methods integrating external guidance in multi-step reasoning scenarios for large models, this paper proposes a unified theoretical framework for Guidance-Augmented GRPO (GA-GRPO), derives the closed-form solutions for convergence rate, bias boundary, and optimal guidance weight, and proves that there is an unavoidable minimal lower bound for the bias term. Experiments show that its effect is better than similar solutions, with GPU time reduced by 31%, and all theoretical predictions are verified.

Neutrosophic Ensemble Classification for Uncertainty-Aware Bearing Fault Detection: Evidence from Laboratory and Variable-Speed Industrial Benchmarks

Maikel Leyva-Vazquez, Dayron Rumbaut Rangel, Lorenzo Cevallos-Torres…

To address the problems of traditional machine learning for bearing fault detection, including confidence confusion errors, ambiguous samples, and algebraic redundancy of truth value pairs, this paper performs neutrosophic decomposition on the ensemble model of random forest, XGBoost, and logistic regression to obtain 4 types of uncertainty indicators, which are tested on two public bearing benchmarks. The results show that the ensemble accuracy drops sharply under distribution shift, and logistic regression has better generalization performance; the time-frequency domain fusion model combined with JS divergence has better uncertainty evaluation performance, and demodulated fault features can achieve nearly 100% classification accuracy.

OpenAI

How Jump Trading is scaling quant research with ChatGPT

OpenAI

This article introduces the practice of leading quantitative institution Jump Trading to scale up quantitative research with the help of OpenAI’s ChatGPT. The core method is to build a long-running AI workflow that automatically integrates multi-source research data, paired with a manual review link to verify the reliability of outputs. This model effectively breaks through the capacity bottleneck of traditional quantitative research, providing a reference implementation path for large models to empower professional investment research scenarios.

Sharing AI progress in mathematics

OpenAI

OpenAI recently announced new progress of its internal cutting-edge large model in research on open mathematical problems: the model has made breakthroughs on several unsolved open mathematical problems. Relevant research details and supporting Lean formal proof content have been uploaded to GitHub for public sharing, which can lower the threshold for reproducing relevant research and provide reference for subsequent research in the intersection of mathematics and AI.

Anthropic News

Expanding the Cyber Verification Program

Anthropic

We officially launch the upgraded Cyber Verification Program (CVP), which opens two core benefits to qualified professional security practitioners: first, providing support for the platform’s advanced network technical capabilities, and second, enabling classifier rules with lower false positive interception rates. This program can reduce the probability of false bans for compliant security research, facilitating practitioners to carry out vulnerability mining, security testing and other work.

Claude Frontier Academy: $100M to train 10,000 engineers

Anthropic

Anthropic officially announced the launch of the Claude Frontier Academy, investing $100 million in special funds, and plans to train a total of 10,000 AI frontier deployment engineers by the end of 2027. The full-process training and assessment standards of the program are fully aligned with the capability requirements of its internal formal engineers, and will deliver a large number of professional engineering talents that meet the technical standards of leading AI vendors for AI frontier technology implementation scenarios.

Google DeepMind

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Google DeepMind

EmbeddingGemma 2 launched by Google is a new generation of open-source lightweight multimodal embedding model. By optimizing the cross-modal representation alignment mechanism, it unifies the semantic space of text, image and other modalities with an extremely small number of parameters. Its performance on cross-modal retrieval and semantic matching tasks is on par with the best level of models of the same size, with extremely low deployment threshold, suitable for end-side and edge devices, and can be implemented in scenarios such as multimodal search and content understanding at low cost.

Gemini 4 Argon: our next era of frontier intelligence

Google DeepMind

This is the official introduction of Google DeepMind’s new generation of cutting-edge large model Gemini 4 Argon. The model optimizes the multimodal alignment architecture, combines RL fine-tuning to improve logical reasoning accuracy, and enhances long context processing and tool call robustness. Its performance on benchmarks such as mathematical reasoning, code generation, and scientific problem solving is more than 40% higher than the previous generation, supports million-level token context window, and its comprehensive multimodal capability reaches the current industry-leading level.

Hugging Face Blog

Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance

Hugging Face

This research launches Falcon-Emirati, a large model adapted to the local UAE context. It is based on the general Falcon large model, and introduces massive amounts of UAE dialect spoken language, local cultural historical materials, and people’s livelihood scenario corpus for incremental pre-training and targeted fine-tuning. The model can accurately capture the subtle expressions of dialects, solving the problem of general large models’ understanding deviation of the UAE local context. Its accuracy in dialect and cultural Q&A is more than 38% higher than similar general models, adapting to the needs of multiple local scenarios such as government services, cultural tourism, etc.

The Agent Said It Was Done. The Database Disagreed.

Hugging Face

This paper addresses the pain point of state misjudgment when large model agents perform database operation tasks: agents often judge the task is completed only based on the brief return information of the tool, while the actual database side operation does not take effect, with a false positive rate of more than 60%. The research proposes an agent execution framework embedded with transaction-level verification, which forces cross-verification by pulling the real state of the database after the operation, finally reducing the false positive rate of task completion status to less than 3%, significantly improving the execution reliability of data-related tasks.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This article sorts out the research context of Recursive Self-Improvement (RSI): the concept was proposed by Good in 1965, referring to superintelligent machines that can surpass all human intellectual activities and iteratively design better systems; in 2008, Yudkowsky clarified that its core is the feedback loop where AI iteratively upgrades its own cognitive mechanism relying on existing intelligence. The study points out that there are two paths for modern AI to achieve RSI: directly rewriting its own weights, or optimizing its own training pipeline.

QbitAI

OpenAI releases 722 math papers overnight! Riemann, Hodge, BSD all included, mathematicians: can’t keep up reading

QbitAI

OpenAI has released 722 mathematical manuscripts produced by its unreleased internal model, covering 17 fields such as number theory and theoretical physics, involving 3 Millennium Prize Problems. Among them, the results related to the quasi-Riemann hypothesis push the zero-free region of the L-function to the real part greater than 7/8, and another version can exclude Siegel zeros in the region where the real part > 11/12, far exceeding the progress of previous human research, making mathematicians exclaim that they cannot read them all.

Just now, the Nobel Prize in Physics was awarded to a single recipient!

QbitAI

This year’s Nobel Prize in Physics was awarded exclusively to Francis Halzen, in recognition of his decisive contribution to the IceCube Neutrino Observatory at the South Pole and his discovery of high-energy neutrinos of astrophysical origin. He pioneered the use of 1 cubic kilometer of pure ice more than 2,000 meters underground in Antarctica as a detection medium, avoiding various interferences of deep-sea detection, and providing core support for solving the century-old problem of the origin of ultra-high-energy cosmic rays.

Wait, why has Terence Tao also joined the AI slowdown camp??

QbitAI

Mathematician Terence Tao, who once actively embraced AI, integrated it into mathematical research, and asserted that AI is promoting a Copernican revolution in the field of mathematics and will greatly reduce the threshold for research, has recently reversed his attitude, publicly calling on AI companies to stop blindly accelerating iteration and first control risks in a speech at SAIR. Another Fields Medal winner Cédric Villani also changed from dismissing AI to being panicked after OpenAI cracked the Millennium Prize Problem, and top mathematicians have generally perceived the strong impact of AI.