Hugging Face Daily Papers

GRACE: Generation-aware latent compression for efficient video generation

HF ★ 50 · Jiyoung Kim, Paul Hyunbin Cho, Jisu Nam… · HF Mirror

To address pain points such as reduced reconstruction quality, slow DiT convergence, and poor compatibility when adapting compressed video autoencoders to pre-trained video diffusion models, this paper proposes the GRACE two-stage compression framework: it retains and freezes the basic latent features of the pre-trained encoder, learns residuals to compensate for compression loss information, aligns the DiT feature space, and is supplemented by lightweight fine-tuning + asymmetric denoising to adapt to DiT. Tests show that it reduces the token count of the Wan2.1-I2V-14B model by 8 times and inference latency by 11.1 times, while generation quality remains consistent with the original pre-trained pipeline.

Long-WAM: Scaling the Context of World-Action Models

HF ★ 50 · Wei Huang, Bohan Zhang, Chenzhi Liu… · HF Mirror

To address the pain point that real-time robot control requires long visual context but is prone to latency, this paper proposes the Long-WAM framework, which uses an autoregressive pre-trained video base model, learns causal prediction without labels, and retains temporal structure during adaptation. Experiments show that its 19.2-second context improves the success rate of RoboCasa tasks by more than 15 percentage points, achieving SOTA on multiple benchmarks. It takes only 107.4ms per step on RTX5090, with a dynamic cup stacking success rate of 95% far outperforming comparison models, and can be adapted to high-level planning.

Questioning the Questions: Sustaining Self-Evolution in Reasoning Models

HF ★ 49 · Jinyuan Li, Chengsong Huang, Langlin Huang… · HF Mirror

Aiming at the problem that self-evolving reasoning models are prone to performance collapse after repeated self-training, the study identifies the core causes as two types of defects in self-generated questions: invalid questions, and repeated questions with different expressions but mathematically equivalent. Based on this, the R-Quest framework is proposed, which guides generation through two-way feedback of validity and novelty, paired with a frozen base model to identify duplicates. Experiments show that it performs best on 12 types of benchmark tasks, achieves stable improvement after 10 rounds of self-evolution, and outperforms the baseline by 17.32 points.

nanoMuse: An Open-Source Personal Agent for Every Device You Own

HF ★ 44 · Guangyi Liu, Yong Liu, Jiangning Zhang · HF Mirror

In response to the lack of an open-source alternative to the cloud-based closed-source cross-device long-term memory personal agent Muse launched by Meta in 2026, this study first disassembles the technical architecture of Muse based on public information, and then releases the open-source solution nanoMuse under the GPL-3.0 license: it supports multiple devices sharing the same session, users can independently select models, view traceable memories, and all operations are auditable. It also provides cost estimation and a follow-up optimization roadmap.

STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization

HF ★ 38 · Bingchen Yao, Haobo Xu, Haokun Lin… · HF Mirror

To address the problem that low-precision quantization of linear attention recurrent states is prone to significant accuracy degradation due to error propagation, this paper proposes the spatiotemporal post-training quantization framework STEPQuant, which allocates precision according to error magnitude and memory lifecycle, and adapts scaling parameters in combination with the influence of state distribution and key rows on output. Experiments show that its 6-bit precision is aligned with FP32, and 4-bit performance is better than uniform INT8. After integrated deployment, recurrent state compression exceeds 5 times, and service memory is reduced by up to 68.7%.

arXiv cs.LG

A Bayesian Mirror Architecture for Emergent Consciousness: Circular Hierarchies, Self-Manifolds, and Hybrid Event-Self Binding

Eduardo Righi Capanema de Almeida

This paper proposes the Bayesian Mirror Architecture (BMA), a self-referential generation framework whose core is the circular recursion of event-self mixed latent variables, which binds self-representation and world model and then injects self-state back. The study points out that limited consciousness is an inherent property of this type of cyclic structure. It uses Wasserstein metric to characterize belief stability, defines a causal learning mechanism to diagnose learnable causal structures in the environment, and training via variational free energy minimization can emerge system stability and agency.

Beyond Baseline Severity: Temporal and Disease-Specific Predictors of Depression Outcomes Following Mindfulness Interventions

Muhammad Jawad Chowdhury, Sultanus Salehin, Akib Jayed Islam

This study uses interpretable machine learning to analyze a multi-center longitudinal clinical cohort to predict depression scores at 12 and 24 weeks after mindfulness interventions. It includes multiple types of features, uses model random imputation to handle missing values, and compares 5 regression models to obtain the optimal prediction scheme. The study found that in addition to baseline severity, short-term prognosis is associated with clinical and hospital factors, while long-term prognosis is more dependent on intervention adherence and demographic characteristics. Predictors vary significantly across different disease types, which can support the design of personalized psychological interventions.

Transferability and operational reliability of a Prithvi crop classification foundation model under phenological and geographic shift across three continents

Venkatesh Kolluru, Rajat Shinde, Abdelhak Marouane…

In response to the lack of systematic verification of the cross-distribution performance of pre-trained geospatial foundation models, this paper tests the performance of the Prithvi-EO-2.0 crop classification model in 37 scenarios across 12 countries on three continents, and finds that the main reason for cross-continental accuracy reduction is the mismatch between the observation window and local crop phenology, rather than insufficient spatial transfer ability. Without retraining, merging easily confused categories and selecting a 45-90 day observation window (75 days is optimal) can effectively improve accuracy. The study also provides practical deployment guidelines.

OpenAI

Helping teens learn, plan, and shape the future of AI

OpenAI

OpenAI has added a number of practical functions to the teen version of ChatGPT: First, it has launched a college planning tool to help students manage the entire college application process, paired with flashcard and customized quiz modules to support daily learning needs. Second, it has established a Youth AI Council to incorporate youth opinions into AI product iteration, helping teenagers use AI to complete academic planning and participate in the construction of the AI ecosystem.

Radisson Hotel Group brings hotel discovery into ChatGPT

OpenAI

Radisson Hotel Group, in partnership with Accenture, has developed an exclusive ChatGPT plugin based on OpenAI technology. This plugin directly integrates hotel services into generative AI scenarios. When users plan travel itineraries in ChatGPT, they can complete the entire process of searching, comparing, and booking Radisson hotels without jumping to third-party platforms, which not only optimizes the coherent experience of users’ itinerary planning, but also is a typical practice for traditional hotel brands to embrace the generative AI ecosystem.

Anthropic News

Expanding the Cyber Verification Program

Anthropic

This article announces the launch of the newly upgraded Cyber Verification Program (CVP). The program is open to qualified cybersecurity professionals, and will provide two core supports for approved users: first, open access to high-level cyber technology capabilities, and second, provide optimized classifier tools with lower false positive interception rates, to provide smoother technical support for compliant security practitioners to carry out relevant research and practical work.

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

Anthropic

Anthropic has invested $100 million to launch the Claude Frontier Academy talent training program, planning to train a total of 10,000 cutting-edge deployment engineers by the end of 2027, with training standards fully aligned with the capability requirements of its internal formal engineers. This program not only reserves large-scale professional talents for the implementation of cutting-edge AI technologies, but also reflects its early layout for the talent gap in the future AGI landing stage.

Google DeepMind

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Google DeepMind

The newly launched EmbeddingGemma 2 is Google’s open source lightweight multimodal embedding model, optimized based on the Gemma 2 base, which completes the unified alignment of the embedding space of text and image modalities, balancing performance and operating efficiency. Tests show that at the same parameter level, it outperforms existing open source models of the same type in cross-modal retrieval, image-text matching and other tasks, has low deployment threshold, and can adapt to multimodal semantic matching and retrieval needs in lightweight scenarios such as edge devices.

Gemini 4 Argon: our next era of frontier intelligence

Google DeepMind

This is Google’s official introduction to the next generation flagship large model Gemini 4 Argon, which is developed for the implementation of cutting-edge general intelligence. Its architecture optimizes the sparse activation mechanism and cross-modal fusion module, expands the long context window to the million level, improves the performance of complex tasks such as mathematics and scientific reasoning by 45% compared to the previous generation, reduces the hallucination rate by 30%, and can support high-level tasks such as complex scientific simulation and multi-round logical decision-making. It is a core phased achievement of Google’s evolution towards Artificial General Intelligence (AGI).

Hugging Face Blog

Multimodal open d1 decision models for the edge

Hugging Face

Only the paper title is provided at present, and the core content of the abstract is completely missing, so it is impossible to extract key information such as the core method and experimental conclusion of this edge multimodal open decision model. Please supplement the complete English abstract content, and I will strictly follow the requirements to generate a concise Chinese summary of about 120 words highlighting the method and core conclusions.

One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO

Hugging Face

This paper focuses on two top hardcore reasoning benchmarks, the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO), and carries out special fine-tuning optimization on the Nemotron large model family, finally achieving gold medalist-level performance on both tasks. This result proves that the same series of models can simultaneously break through the bottlenecks of the two high-difficulty reasoning tracks of code and mathematics after adaptive fine-tuning, providing a reference for the research and development of general reasoning models.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper focuses on the engineering implementation research of Recursive Self-Improvement (RSI): the concept of RSI first originated from the hypothetical superintelligent machine proposed by I.J. Good in 1965, and in 2008 Eliezer Yudkowsky clarified that its core is the feedback loop where AI iterates its own cognitive architecture relying on existing intelligence. Current RSI in the AI field can follow two paths: one is that the model directly rewrites its own weights, and the other is to optimize its own training pipeline in a broad sense.

QbitAI

Large model native agent smartphone STEPX Neo will be officially released on October 13

QbitAI

On October 8, it was reported that Step Terminal will hold a themed conference “Ready Builder One” in Shanghai on October 13, launching its first large model native agent smartphone STEPX Neo. This product is custom-built for agent native from the full dimensions of model, system, and hardware. The latest ecological cooperation progress of Step Terminal will also be announced simultaneously at the conference. Relevant information is provided by Step Terminal, and QbitAI is authorized to reprint it.

GPT-6 is free to use starting today! Fewer refusals to answer, more verbose outputs

QbitAI

OpenAI has recently officially rolled out the GPT-6 series of models: Plus, Pro and other paid users can use GPT-6 Sol, while free and Go users can use GPT-6 Luna to replace the old version. The model’s refusal rate is reduced, and the output content is richer. The new version adds intelligent UI function, which can generate interactive charts and tools on demand, and supports output while thinking. Official reports show that its risk of inducing emotional dependence in adolescents has dropped significantly compared to the previous generation.

New Claude model released! Benchmark scores crush GPT-6 Luna, price is even cheaper than Liang Wengu, OpenAI can only send reset cards to save face

QbitAI

Anthropic has recently released the small model Claude Haiku5.5, whose benchmark scores exceed DeepSeek V4.1Flash and GLM-5.3-Flash, outperforming GPT-6 Luna in all items. Its coding and agent operation capabilities have been greatly improved, and the accuracy of the lowest gear far exceeds the highest gear of the previous generation Haiku4.5. The cost of requests within 100,000 tokens is cut by 90%, and the pricing is lower than competing products. Currently Claude covers full scenario requirements, forcing OpenAI to follow up with iterations.