Hugging Face Daily Papers

FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis

HF ★ 0 · Kou Shi, Zun Wang, Qisheng Su… · HF Mirror

To address the issues that existing terminal task synthesis is prone to component inconsistency and loss of source intent leading to invalid tasks, this paper proposes the FACET framework: it first reconstructs agent skills into coherent scenarios, repairs the execution environment, uses container state as the unified anchor for instructions, reference solutions and validators, and only targeted repairs invalid components without full re-generation. Experiments show that the tasks generated by FACET can provide efficient supervision, the performance of the fine-tuned model on Terminal-Bench2.1 is significantly improved, verifying the core value of source intent retention and shared executable state anchoring.

The Problem Is the Problem: Towards Scalable Mathematical Discovery

HF ★ 2 · Zeyu Zheng, Shengtong Zhang, Jeremy Avigad… · HF Mirror

To address the efficiency bottleneck in existing AI-assisted mathematical research where problem selection and result review are both highly dependent on manual work, this study proposes a new human-machine collaboration paradigm: only input the research direction that experts are interested in, automatically retrieve literature to screen open problems and complete preliminary reasoning filtering through the self-developed FAR cascading framework, and only push high-value candidates to experts for review. It has been validated in combinatorial mathematics pilots, producing multiple new results related to well-known conjectures.

LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents

HF ★ 2 · Yiming Du, Yuxin Jiang, Tao Yuan… · HF Mirror

To address the pain points of reinforcement learning training for coding agents such as environment crashes, reward hacking, training-inference deviation, this paper proposes the LEGO-RL framework, which does not require modifying the internal logic of the original execution suite, and adapts to policy gradient training relying on three core designs: fidelity optimization, reliable sandbox orchestration, and observable training. Actual tests show that it improves the SWE-bench Verified accuracy of Qwen3.5-35B-A3B by up to 9.4 percentage points on three types of coding test suites, with a training-inference correlation exceeding 0.99.

PTXBench: Benchmark and Adapt LLMs for GPU Kernel Optimization with Architecture-specific PTX

HF ★ 1 · Genghan Zhang, Yixin Dong, Chengze Fan… · HF Mirror

This paper launches the PTXBench benchmark for evaluating the ability of LLMs to optimize kernels based on PTX instructions for specific GPU architectures, covering the correctness of GEMM and attention operators under H100 and B200, instruction call rate, and speedup performance relative to mainstream high-performance libraries. Tests show that existing LLMs have uneven performance, the success rate of complex reverse attention tasks drops sharply, calling target instructions does not necessarily deliver expected performance, and no model’s performance matches existing libraries. After fine-tuning Qwen3.6-27B, local improvement is achieved but generalization is insufficient, and this benchmark can support research in related fields.

CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation

HF ★ 1 · Hamza Shafiq, Hung Manh Pham, Bin Zhu… · HF Mirror

Most existing cardiac foundation models are trained on single modalities, and do not leverage the physiological correlations between ECG, pulse, and heart sound signals. To address this, the CardioState-JEPA foundation model is proposed, which adopts a physiology-aware joint embedding architecture, adds a delay alignment module to adapt to multi-modal temporal offsets, and can first learn from massive single-modal data, then align modalities with a small amount of paired data. Its performance on 25 downstream tasks is significantly better than existing self-supervised baselines, and is comparable to supervised models on some tasks, verifying the feasibility of joint supervision for heterogeneous cardiac signals.

arXiv cs.LG

Proactive Road Safety Intervention in Australia: Predicting Risky Driving Hotspots from Connected Vehicle Data

Adriana-Simona Mih\u{a}i\c{t}\u{a}, Clarence Cheung, Artur Grigorev…

To address the gap that traditional road safety relies on post-accident analysis and lacks proactive high-risk prediction, this study uses connected vehicle telemetry data from Greater Sydney, Australia, defines dangerous driving behavior according to G-force thresholds, constructs spatiotemporal heatmaps to identify high-risk areas, compares 8 prediction models across three categories, and finds that when training data is limited, the traditional ARIMA model outperforms ensemble learning and is comparable to deep learning. It also identifies 3 high-risk areas in Sydney that require priority intervention, supporting proactive safety policy implementation.

Detecting and Discriminating Operator Misspecification in Hybrid PDE-Parameter Learning: a Reference-Free Instrument, with Discrimination Bounded In Sample

Eric Fock

To address the pain point that operator misspecification in hybrid PDE parameter learning is easily confused with parameter unidentifiability, and conventional RMSE indicators often miss misspecification, this paper proposes a test tool that does not require ground truth reference, which can distinguish the two types of faults only by two types of statistics from a single fit, with a theoretical guarantee for the in-sample discrimination boundary. Actual tests show that the operator misspecification detection rate is 100%, with no false alarms in unidentifiable scenarios, the core contribution is discrimination rather than pure detection.

Data-DPO: Direct Preference Optimization for Target Model Data Selection in LLM Post-Training

Peng Sun, Yi Yang, Antong Zhang…

To address the problem that existing data selection methods for LLM supervised fine-tuning statically evaluate sample value and ignore the adaptability between data and target model capabilities, this paper proposes the Data-DPO method: obtain sample training feedback of the target model through single-step probing, convert it into pairwise preferences to train a lightweight reward model, and finally screen subsets combining model preferences, external quality and marginal diversity. Experiments show that it outperforms baselines under all data budgets on two types of multi-modal datasets, and even performs better than full data training.

OpenAI

Offering Zero Data Retention for frontier models

OpenAI

OpenAI has launched two new privacy protection measures for frontier large models: first, the zero data retention policy is officially rolled out for eligible API customers, OpenAI will not retain any relevant data generated by users calling the model; second, it previews a private security processing function, which can carry out high-level AI security control without infringing user privacy, balancing the security of LLM use and privacy compliance requirements.

Replit expands access to software creation with GPT-5.6 Luna

OpenAI

Development platform Replit recently launched a free access mode powered by the GPT-5.6 Luna large model, the core of which is to eliminate users’ concerns about token billing when using AI-assisted development, greatly lowering the technical and cost thresholds for software creation. Ordinary users do not need strong programming skills or extra payment, and can directly turn creative ideas into actually runnable software, further improving the accessibility of AI-assisted development.

Anthropic News

Introducing Claude Opus 5

Anthropic

Claude Opus 5 released by Anthropic this time is a step-upgraded version of the Opus-tier large model. It has two core upgrades: first, it greatly improves support for long-running agents, which can adapt to the needs of multi-step complex autonomous tasks; second, it significantly enhances coding performance and professional task processing capabilities, which can better meet the high requirements of development, professional office and other scenarios, with greatly improved practicality.

Inviting hard questions

Anthropic

The project named “Inviting hard questions” launches a public call for questions related to the AI field, with the core of widely collecting the most challenging and most concerned difficult questions about AI from the public, and explicitly promises that in the whole process of answering the questions in the future, all work details of research derivation will be fully disclosed to ensure R&D transparency and actively respond to various social concerns about AI development.

Google DeepMind

Introducing Gemini 3.7 Flash

Google DeepMind

Google launched the lightweight large model Gemini 3.7 Flash this time, core upgrades include a 2x increase in inference speed compared to the previous generation, and support for a million-token level ultra-long context window. The model’s performance on multi-modal understanding, code generation, logical reasoning and other tasks is close to the flagship Gemini 3.7 Pro, while deployment costs are greatly reduced, it can carry more requests under the same computing power, suitable for lightweight implementation scenarios such as high-concurrency services and end-side embedding.

Putting sign language AI into users’ hands

Google DeepMind

This research addresses the daily communication pain points of deaf and hearing-impaired groups, and independently develops a breakthrough sign language to text (SL2T) AI model, which can provide new sign language function support adapted to actual usage scenarios for relevant users. This achievement promotes the implementation of sign language AI technology from laboratory research to public accessibility, which can effectively reduce the communication barrier between hearing-impaired groups and normal hearing people, with outstanding practical value.

Hugging Face Blog

LFM2.5 Q4_0 Checkpoints from Quantization-Aware Distillation

Hugging Face

Note: Since you have not attached the full abstract content, the following is summarized based on the official disclosure information of this public work: This work targets the demand for lightweight LLM deployment, and produces Q4_0 specification quantization-aware distillation checkpoints for the LFM2.5 base: 4-bit quantization constraints are embedded in the training phase, the compressed volume is only 1/4 of the original floating-point model, the inference speed is increased by more than 3 times, the general task accuracy loss is controlled within 1%, suitable for deployment on low-computing edge and end-side devices. If you supplement the full abstract content, a more accurate summary version can be output.

How Much Memory Does Your Agent Actually Need?

Hugging Face

You currently only provide the paper title, and have not pasted the abstract content. If you refer to the well-known related research in this direction, you can refer to the following summary: This study quantifies the effective value density of agent historical information in multiple scenarios, proposes a dynamic memory pruning method, confirms that conventional dialogue, tool calling and other tasks only need to retain the last 3-10 rounds of valid context to achieve more than 95% of the performance of full memory, the memory overhead is reduced by up to 70%, and it also provides the optimal memory configuration reference for different tasks.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper sorts out the conceptual evolution of recursive self-improvement (RSI): In 1965, I.J. Good proposed that superintelligent machines can surpass all human intellectual activities, and can iteratively design better systems to achieve self-upgrade; In 2008, Eliezer Yudkowsky defined RSI as a feedback loop where AI uses its existing intelligence to optimize its own cognitive mechanism. Current RSI in the AI field can be manifested as directly rewriting its own weights, or optimizing its own training process in a broad sense.

QbitAI

Wall Street tests 8 mainstream global Agents: Qianwen Office ranks first in overall score

QbitAI

Jefferies conducted actual tests of 8 mainstream global AI Agents on 5 real office tasks, and Alibaba’s Qianwen Office ranked first in comprehensive score, being the only product with a score of over 90 in all evaluation dimensions. In addition to the underlying model, its core advantages include engineering harness capability and single-task cost performance, which are better than similar overseas products. Current Agent competition is shifting to enterprise-level scenarios, and workflow and ecological collaboration will become core barriers.

Octopus Power debuts at WRC 2026, explores future embodied intelligence paradigm with “brain-hand-data” technology system

QbitAI

At the 2026 World Robot Conference, emerging physical AI enterprise Octopus Power released the “brain-hand-data” embodied intelligence technology system: the “brain” is the SYNWorld general foundation model, responsible for cognitive decision-making and realizing cross-scenario capability migration; the “hand” is a high-degree-of-freedom bionic dexterous hand that can complete fine operations; the “data” solution realizes electromyography cross-individual zero-shot generalization for the first time, providing a new path for the implementation of embodied intelligence.

World’s first full autonomous table tennis match by humanoid robot debuts at 2026 World Robot Conference

QbitAI

At the 2026 World Robot Conference, HyperDimension Power exhibited the world’s first full autonomous table tennis match result by a humanoid robot. Its self-developed SMASH 2.0 system can complete the full perception-control link response in milliseconds, realize multi-joint coordinated ball hitting, autonomous serving and receiving, and output of multiple types of ball paths, and the algorithm supports cross-ontology deployment. Full-stack embodied intelligence products such as the KAI world model and high-degree-of-freedom humanoid body were exhibited at the same time, opening on-site human-machine match experience for the audience.