Hugging Face Daily Papers

VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-Policy Distillation

HF ★ 2 · Kangning Zhang, Yixing Li, Shuai Shao… · HF Mirror

To address the problem that supervision signals in multimodal on-policy distillation are mixed and cannot distinguish corrections supported by visual evidence, this paper proposes the Visual Attribution Distillation (VAD) method. It separates the part of teacher corrections supported by visual evidence through counterfactual inference, reconstructs the supervision target as the main training signal, and only retains weak teacher regularization. In 6 fine-grained visual benchmarks for models with 4B and 9B parameters, VAD significantly outperforms existing distillation solutions, with higher accuracy of error correction.

Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs

HF ★ 0 · Zixuan Huang, Yang Zhou, Kaixuan Wang… · HF Mirror

To address the issue of anchoring bias and low reasoning credibility caused by early exposure to ground-truth future trajectories when training vision-language-action models for autonomous driving with chain-of-thought, this paper proposes the AD-MCQ paradigm which converts trajectory planning into candidate selection, and further introduces the DEFT-RLVR method, which turns future trajectories from pre-anchored points into post-hoc verification targets. Experiments show that it improves driving reasoning ability without impairing general visual ability, and AD-MCQ also provides a flexible and scalable foundation for subsequent verifiable driving reasoning research.

arXiv cs.LG

Topology-Aware Data Movement for Disaggregated GPU Inference

Sanjeev Rao Ganjihal

To address the performance insufficiency in LLM disaggregated inference scenarios (where prefill and decoding run in different GPU pools) caused by existing KV cache transmission using unified RDMA and ignoring bandwidth differences of GPU interconnection topologies, this paper designs a topology-aware transfer orchestrator, combining three layers of mechanisms: layered pipeline transmission-computation overlapping, NVLink domain-aware scheduling for MoE models, and CXL3.0 memory expansion overflow. Theoretical calculations show it can reduce transmission latency by 3-18 times compared to unified RDMA.

Sensitivity Analysis of GRU, LSTM and Transformer Encoder in Classification of Automated Driving Systems

Bidhya Shrestha, Christos Papadopoulos

This research targets the safety and regulatory requirements of autonomous driving system identification, using three temporal classification models: GRU, LSTM, and Transformer encoder, to distinguish three types of Level 2 autonomous driving systems from manual driving based on in-vehicle connected data. On clean data, all three models achieve F1 scores above 0.9, with Transformer performing the best; robustness tests show that event-type perturbations have minimal impact, while time jitter causes the F1 scores of all three models to drop sharply to the range of 0.44-0.5.

Guarantees on Dynamical System Distinguishability for LLM Token Generation

Mohamed Akrout, Dan Wilson

To address the lack of theoretical support for LLM discrimination methods based on Dynamical Systems (DS), this paper formalizes the classification task as a binary hypothesis test for two stochastic linear DS. The study proves that classifiers ignoring token dynamics have an inherent lower accuracy limit; the misclassification rate of DS methods decays exponentially with sequence length, with the decay rate determined by the spectral gap between the two DS; it also provides a transferable discriminability lower bound for generalization across embedding models, offering theoretical support for the empirical performance of such methods.

OpenAI

How we built a realtime system for responsive voice AI in six months

OpenAI

This article introduces the team’s R&D practice of building a real-time responsive voice AI system in six months: the launched GPT-Live abandons the turn-based interaction logic of traditional voice conversations, adopts a turnless voice model paired with a low-latency architecture as its core, enables continuous voice interaction with AI, greatly reduces response latency, and significantly improves dialogue fluency and naturalness compared to traditional solutions, supporting more natural real-time voice interaction requirements.

Circles powers telco personalization with OpenAI technology

OpenAI

This industry practice result shows that telecom operator Circles used OpenAI API and Codex large model to build an AI-native personalized service system adapted to telecom scenarios, achieving two-way efficiency gains on both operation and R&D sides: Average Revenue Per User (ARPU) increased by 22%, customer churn rate decreased by 9%, and internal development efficiency was also significantly optimized, providing a referable implementation path for the digital and intelligent upgrade of the telecom industry.

Anthropic News

Introducing Claude Opus 5

Anthropic

The newly launched Claude Opus 5 is a generational upgrade product of the Opus series large models. The core upgrades focus on two directions: first, greatly enhancing the support for long-running agents to adapt to the needs of complex agent tasks; second, significantly improving processing performance for code development and professional work in various fields. This version achieves a capability leap, providing more reliable model support for professional users to implement complex tasks.

Inviting hard questions

Anthropic

This project is initiated by the AI research team: it openly solicits the most difficult and questionable hard-core problems in the AI field from the public across the whole society, and makes a public commitment that all relevant work details will be fully disclosed throughout the entire process of conducting research and outputting answers to these questions, keeping the whole process open and transparent, so as to respond to the public’s general concerns about AI development and narrow the gap between cutting-edge research and the general public.

Google DeepMind

Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration

Google DeepMind

Gemini Robotics ER 2 is a new generation of AI support system for robotics scenarios, achieving leapfrog breakthroughs in three core capabilities for implementation requirements: upgraded video understanding capability to support environmental perception and decision reasoning, optimized task orchestration mechanism to realize scheduling and execution of complex operations, and added multi-robot collaboration framework, ultimately enabling robots to autonomously reason and collaboratively complete various tasks in real-world scenarios.

We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control

Google DeepMind

Google recently officially launched the new generation AI music generation model Lyria 3.5 on the Flow Music platform. This version has achieved multi-dimensional technological breakthroughs in core capabilities, with significant improvements in music fluency, lyric matching accuracy, and vocal naturalness. It also further optimizes creative control functions, giving users higher creative freedom, and can more accurately meet various personalized music generation needs.

Hugging Face Blog

GPU Management: Why Idle GPUs Are the New Grounded Aircraft

Hugging Face

This study compares idle GPUs to grounded civil aviation aircraft, pointing out that GPU idleness is the core cost loss point in the computing power field due to its heavy asset attribute. A dynamic scheduling solution is proposed for this pain point: through methods such as time-sharing multiplexing, elastic load matching, and cross-cluster computing power pooling, it can increase GPU utilization in general scenarios from less than 30% (the industry average) to over 80%, greatly reducing the computing power operation cost of large model training and inference links.

The OlmoEarth Platform: Geospatial inference at planetary scale

Hugging Face

OlmoEarth is an AI platform specially built for planetary-scale geospatial inference, integrating multi-source global remote sensing and spatiotemporal geographic datasets, equipped with multimodal large models adapted to geographic scenarios, and compatible with distributed computing power clusters for efficient scheduling. In tasks such as land cover classification and disaster emergency monitoring, its accuracy and coverage have significant advantages over traditional solutions, greatly lowering the implementation threshold for global-level geographic analysis.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This paper targets the orthogonality hypothesis in AI alignment research, and proposes a new path from the perspective of virtue ethics: it refutes the traditional cognition that “rational agents take fixed ultimate goals as objectives”, pointing out that human rationality adjusts behavior based on a practical network including action norms and evaluation standards. It proposes that to realize AI’s compliant collaboration and meet ethical and safety requirements, AI’s decision-making logic needs to match human practical action paradigms, rather than pre-set fixed goals.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper sorts out the conceptual context of Recursive Self-Improvement (RSI): In 1965, scholar Irving John Good first proposed the concept of superintelligent machines, referring to agents that can surpass all human intellectual activities and design better systems on their own to achieve self-upgrade. In 2008, Eliezer Yudkowsky clarified that its core is the feedback loop: AI uses its existing intelligence to optimize its own cognitive architecture. In the current AI field, this feedback loop includes both the model directly rewriting its own weights, and broadly refers to the model optimizing its own training pipeline.

QbitAI

The “computing power life and death crisis” of apps with hundreds of millions of daily active users: inference cost inversion, they cut 75% of GPU clusters with cross-cloud architecture

QbitAI

An overseas APP with hundreds of millions of daily active users focusing on AI outfit matching + lock screen shopping recommendations once fell into a dead end of cost inversion: the average annual revenue per user was only $2, while the computing power cost of deploying text-to-image models on L4 GPUs from leading cloud vendors reached $3, meaning the larger the user base, the more serious the loss. The team finally adopted a cross-cloud architecture to optimize computing power scheduling, directly cutting 75% of the GPU cluster, and successfully resolved the survival crisis.

This new image model is impressive: native 4K output, domestic, open-source!

QbitAI

SenseTime’s latest open-source domestic lightweight image generation model SenseNova U1.5-Lite-Preview is based on the self-developed NEO-Unify unified multimodal architecture, with only 8B-MoT parameters, can natively output 4K images directly, supports generation with complex multi-constraint prompts, performs excellently in text and layout of high information density content, and can also realize operations such as precise local editing and multi-image combination, adapting to the needs of the entire real creation process.

Hiring electricians with million-dollar annual salaries, Meta is in such a hurry that it runs its own technical school

QbitAI

The new bottleneck of the current AI computing power race has shifted to the infrastructure side. The construction of ultra-large-scale AI data centers in the United States is facing a huge skilled worker gap, with an average annual gap of 80,000 electrician positions from 2024 to 2034. AI data centers have high power consumption and construction complexity far exceeding traditional projects, so there is extremely high demand for skilled electricians who can handle high-intensity work. Even with annual salaries of millions of dollars, recruitment is still difficult, and the losses from delays are heavy, forcing large companies such as Meta to build their own technical schools to train talents.