Hugging Face Daily Papers

SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning

HF ★ 41 · Seokju Cho, Ryo Hachiuma, Abhishek Badki… · HF Mirror

To address the weak spatial reasoning capability of existing vision-language models and insufficient flexibility of tool invocation interfaces for tool-augmented agents, this paper proposes SpatialClaw, a training-free spatial reasoning framework: it uses code as the action interface, is equipped with a stateful Python kernel preloaded with perception and geometric primitives, and supports stepwise adjustment of reasoning strategies based on intermediate observations. It achieves an average accuracy of 59.9% across 20 benchmarks, 11.2 percentage points higher than the current state-of-the-art, and can be directly adapted to multiple types of vision-language backbones.

EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments

HF ★ 37 · Jundong Xu, Qingchuan Li, Jiaying Wu… · HF Mirror

Existing LLM agent evaluations are mostly based on static environments, which fail to meet the requirements of real-world dynamic deployment. To solve this problem, the research team launched the EvoArena benchmark covering progressive environmental changes across terminal, software and social domains, and also proposed EvoMem, a patch-based memory paradigm that structurally records memory evolution history. Experiments show that the average accuracy of existing agents on this benchmark is only 39.6%, and introducing EvoMem can deliver clear performance improvements on this benchmark as well as general task benchmarks, better retaining the complete state of environmental evolution.

FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents

HF ★ 30 · Jia Deng, Yimeng Chen, Xiaoqing Xiang… · HF Mirror

Existing training data for deep search agents often increases difficulty by adding structure, but is prone to shortcuts that make actual search difficulty fall short of expectations. This paper identifies four types of shortcut risks, and proposes FORT, a shortcut-resistant training data synthesis framework that manages risks throughout the entire data generation process. The FORT-Searcher, trained via supervised fine-tuning only on data generated by this framework, achieves state-of-the-art performance on deep search benchmarks among open-source search agents of the same scale, with fewer search shortcuts. (118 words total)

InterleaveThinker: Reinforcing Agentic Interleaved Generation

HF ★ 21 · Dian Zheng, Harry Lee, Manyuan Zhang… · HF Mirror

Existing image generators cannot support interleaved text-image generation, and open-source unified multimodal models have limited performance. To address this issue, this paper proposes InterleaveThinker, a multi-agent pipeline: it adopts a dual-agent scheduler of planner and verifier to orchestrate the base generator, paired with a dedicated fine-tuning dataset and stepwise reinforcement learning strategy for cost reduction and optimization. This method is adaptable to various base models, its interleaved text-image generation performance matches that of Nano Banana and GPT-5, and its performance on reasoning benchmarks is also significantly improved.

WeaveBench: A Long-Horizon, Real-World Benchmark for Computer-Use Agents with Hybrid Interfaces

HF ★ 16 · Wanli Li, Bowen Zhou, Yunyao Yu… · HF Mirror

Existing Computer-Use Agent (CUA) benchmarks mostly test the capabilities of each interface separately, lacking validation for cross-interface long-sequence tasks. To fill this gap, the research team proposes the WeaveBench benchmark, covering 114 long-process tasks across 8 real-world domains that require coordinated operation of GUI, command line and code, paired with a trajectory-aware evaluator to prevent cheating. Tests show that the pass rate of the current best model is only 41.2%, and pure result-based evaluation will greatly overestimate performance, so this benchmark can fill the gaps in CUA evaluation.

arXiv cs.LG

Restless bandits with imperfect binary feedback: PCL-indexability analysis and computation

José Niño-Mora

This paper targets the opportunistic spectrum access scenario with perception errors, and studies the restless multi-armed bandit problem with binary hidden states and imperfect binary feedback. It proposes an analysis and computation framework based on partial conservation laws, combining deterministic dynamic analysis, update decomposition and numerical schemes, verifying that indexability is satisfied in the vast majority of parameter intervals. The proposed marginal productivity index strategy (equivalent to Whittle index when conditions are met) significantly outperforms traditional benchmarks.

To Intervene or Not: Guiding Inference-time Alignment with Probabilistic Model Blending

Jin Gan, Xin Li, Jun Luo

Existing inference-time alignment methods for large models do not evaluate guidance reliability, leading to many invalid interventions and poor performance. To address this issue, this paper proposes the BlendIn framework: instead of binary intervention decisions, it weights and mixes the knowledge of each model according to their reliability to obtain the output distribution, achieving quality-aware alignment. This method can retain effective guidance and weaken unreliable suggestions, delivering a maximum performance improvement of 50% in tests of difficult model pairs.

Dual-Stance Evaluation of Sycophancy: The Structure of Agreement and the Limits of Intervention

Matthew James Buchan

Existing activation regulation evaluations for sycophancy suppression in large models do not verify whether factual recognition is accidentally damaged. To address this flaw, this paper proposes a dual-stance evaluation method, which tests both positive and negative positions on each topic. Evaluating the centroid difference activation regulation scheme of Llama-3-8B-Instruct with this method finds that: although sycophancy and factual recognition representations belong to different subspaces, the regulation direction has consistent projections for both, and will simultaneously suppress sycophantic responses and correct factual recognition, verifying that readable activation representations are not necessarily accurately regulatable.

OpenAI

How an astrophysicist uses Codex to help simulate black holes

OpenAI

This application practice introduces the new solution of astrophysicist Chi-kwan Chan: with the help of AI code generation model Codex, he can quickly build black hole simulation programs without writing a large amount of complex physical code from scratch, greatly reducing the development threshold for high-precision astrophysical simulation. Such simulation tools can support researchers in conducting research on extreme physical scenarios, providing efficient support for verifying the theoretical predictions of Einstein’s general relativity.

BBVA puts AI at the core of banking with OpenAI

OpenAI

This cooperation update shows that Banco Bilbao Vizcaya Argentaria (BBVA) takes AI as the core direction of banking business upgrading, and has reached technical cooperation with OpenAI. At present, it has completed the large-scale deployment of ChatGPT Enterprise for 100,000 employees across the group, empowering all staff in daily office work and business operation. The two parties will jointly explore the implementation path of AI in financial scenarios, accelerating the intelligent transformation of BBVA’s global business.

Anthropic News

Claude Fable 5 and Claude Mythos 5

Anthropic

Anthropic officially launched the Claude Fable 5 large model this time, which belongs to its highest-performance Mythos-level series, inheriting the super strong general task processing capability of models in the same tier. The R&D team has completed full-dimensional security verification and alignment optimization for it, making this high-end large model meet the open standard for general scenarios, which can be implemented for all types of users and application scenarios without special permissions.

Introducing Claude Corps

Anthropic

Anthropic officially launched the nationwide professional training program named Claude Corps this time, a special talent support program focusing on the field of AI inclusion, open to early-career practitioners who are committed to promoting the coverage of AI technology dividends to communities across the United States. It aims to provide supporting support for this group, promote the implementation of AI inclusion practices, so that the value of cutting-edge AI can effectively reach grassroots communities of different types in different regions of the United States.

Google DeepMind

DiffusionGemma: 4x faster text generation

Google DeepMind

This research proposes the DiffusionGemma text generation model, whose core is to adapt and transform the diffusion generation paradigm with the lightweight Gemma base architecture, replacing the traditional autoregressive token-by-token decoding logic. It only requires 4 inference steps to achieve generation quality equivalent to the original Gemma, with inference speed 4 times higher than the baseline, balancing generation effect while greatly reducing inference latency, adapting to the needs of high-throughput real-time text generation scenarios.

Investing in multi-agent AI safety research

Google DeepMind

Google DeepMind and its partners officially launched a $10 million multi-agent AI safety research grant call, open to relevant research teams around the world. This initiative targets the shortcomings in safety research against the background of the current rapid development of multi-agent technology, and will provide financial support for basic research and exploration of risk prevention and control solutions in related directions, helping the safe and controllable implementation of multi-agent AI technology.

Hugging Face Blog

Profiling in PyTorch (Part 2): From nn.Linear to a Fused MLP

Hugging Face

This article is the second part of the PyTorch performance profiling series. It uses the built-in profiler tool to locate the performance shortcomings of native MLPs stacked with multiple layers of nn.Linear: large layer-by-layer kernel scheduling overhead and high proportion of redundant memory access in activation layers. It then verifies the operator fusion optimization scheme, and the results show that the fused MLP can greatly reduce scheduling and memory access losses, and the inference throughput is increased by more than 40% at the same accuracy, adapting to the performance optimization needs of high-frequency MLP modules in large models.

Introducing North Mini Code: Cohere’s First Model For Developers

Hugging Face

Cohere officially launched North Mini Code, its first exclusive code large model for developers, which is the core implemented product of its developer tool ecosystem layout. This model is adaptable to multiple mainstream programming languages, can handle core development scenarios including code generation, vulnerability debugging, syntax correction and more, balances inference efficiency and lightweight deployment requirements, supports IDE integration and local offline operation, which can effectively reduce repetitive work for developers and improve coding R&D efficiency.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment research from the perspective of virtue ethics challenges the core premise of the mainstream orthogonality hypothesis in academia: it rejects the assumption that “rational agents need to be anchored to fixed final goals”, and proposes that the core of human rationality is that actions adapt to the practical network of inherent rules and evaluation standards. To realize AI that adapts to human demands and is collaborative, it is necessary to make AI decision logic match the practice-based reasoning paradigm of humans, which can cover both ethical alignment and core safety alignment requirements at the same time.

QbitAI

As soon as SpaceX goes public, even cafeteria staff will become millionaires…

QbitAI

SpaceX has officially announced that its IPO is ready, planning to raise $75 billion with a valuation of $1.77 trillion, which will be the largest IPO in history. At present, retail subscription amount has exceeded $70 billion. After listing, its market value is expected to rank 7th among US stocks, surpassing Meta and Tesla, creating 4,400 millionaires, and Musk still holds 82% of the voting rights. However, the company currently has an annual loss of nearly $5 billion, and third-party valuations are only half of the offering price, so the prospect of its cross-domain infrastructure positioning is still controversial.

Claude Fable 5 cost saving tip: set to Low tier is cheaper than Opus

QbitAI

Claude Fable 5 was previously criticized for being too expensive as its token price is 4.8 times that of Opus. Recent tests by developers found that when its effort level is adjusted to the lowest tier, its SWE-bench Pro score of 75.0 is still higher than Opus’s 68.6 at the highest tier; moreover, the model is more efficient, consuming fewer total tokens to complete the same task, with better actual task speed and effect, resulting in lower actual usage cost.

To 12.9 million college entrance exam takers: Alibaba has launched a free college application assistant Agent

QbitAI

Alibaba’s Qwen launched a free college application Agent in June for the 12.9 million national college entrance exam takers. Based on the Qwen large model specially optimized for college application, combined with 8 years of accumulated college entrance examination service data from Quark, it provides full-cycle application assistance including application calendar, major reports, intelligent Q&A and more, covering the whole process from post-exam preparation to pre-submission review. It aims to solve the pain point of scarce offline planner resources, benefiting the majority of candidates.