Hugging Face Daily Papers

Is this Citation on Point?

HF ★ 0 · Apurv Verma · HF Mirror

In 2023, a lawyer was fined for using ChatGPT to generate fake legal citations, exposing that existing evaluations of large models in legal scenarios only cover fake citation detection, and do not address the issue of real citations failing to support arguments. This paper tests the citation verification capabilities of 14 models by perturbing real citations in two types of legal corpora (replacing cases / replacing citation page numbers for the same case): models can identify 93%-100% of incorrect cases, but have a high missed detection rate for wrong page numbers in the same case, easily confuse topic relevance with substantive argument support, and even strong reasoning models cannot fully solve this problem.

Nanbeige4.2-3B on Apple Silicon: Fixing Deployment Bugs and Decreasing Looped Transformer Memory Overhead

HF ★ 2 · John T. Halloran · HF Mirror

This paper addresses the deployment issues of the 3B-parameter looped Transformer architecture Nanbeige4.2-3B model on Apple Silicon platforms: first, 5 bugs that prevented out-of-the-box operation on Hugging Face were fixed, then a chunked prefill strategy was proposed, increasing context window size by 2.7x on 32GB of VRAM. After fixing system prompts and native MPS VRAM bugs, MCP task completion rate rose from 0 to 30%, single tool call performance is nearly perfect, and related patches have been open-sourced.

Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models

HF ★ 2 · Jean de Dieu Nyandwi, Leena Mathur, Yonatan Bisk… · HF Mirror

This study addresses the bias in training effects of reasoning models, proposes the “behavior lift” metric to measure the correlation between reasoning behaviors and answer accuracy. After annotating and analyzing more than 15,000 reasoning trajectories from 15 cross-modal models and 6 benchmarks, it finds an “amplification-lift gap”: existing reasoning training prefers to amplify superficial behaviors such as self-correction and uncertainty expression, but does not strengthen core behaviors that are truly associated with high accuracy, such as confidence calibration and knowledge alignment, pointing out the direction for training optimization.

Modular Cognitive Architecture Emerges in Large Language Models

HF ★ 11 · Pengrui Han, Jacob Andreas, Evelina Fedorenko… · HF Mirror

To explore whether modular cognitive architecture is a special case of biological brain evolution or a general law of intelligent systems, researchers conducted circuit analysis of large language models across 4 cognitive domains covering 46 tasks, and found that large models have evolved modular structures corresponding to the human brain: tasks in the same cognitive domain call overlapping neurons, while cross-domain tasks call separate neurons, confirming that modularity is a fundamental attribute of intelligent systems.

Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence

HF ★ 26 · Brian Wang, Bin Feng, Xiaoman Pan… · HF Mirror

To address the problem that existing AI can only handle clearly defined tasks and is difficult to adapt to real-world exploration scenarios, the research launches the Apodex Discovery evaluation framework for discoverative AI, which includes three core components: a real high-value problem library, unified task environment abstraction, and multi-dimensional independent evaluation system. Actual testing shows it outperforms existing SOTA by 7% on AAV capsid design tasks, and can significantly improve large model performance on drug repurposing tasks, promoting AI evaluation to shift towards real discovery orientation.

arXiv cs.LG

L-FNO: Lorentzian Fourier Neural Operator for Stochastic Event Dynamics

Songhee Kang, Jihoon Kang

Aiming at the problem that existing neural operators are mostly regression models, which are not suitable for modeling sparse sudden random events, the research proposes the Lorentzian Fourier Neural Operator L-FNO, which combines FNO covariate pathways and Lorentzian spectral kernels to model historical excitation effects, and is trained with likelihood objectives. On 8 synthetic point process benchmarks and 3 disease outbreak / semiconductor defect detection datasets, its event likelihood, calibration, and rare event detection performance are better than baselines, verifying the effectiveness of the corresponding inductive bias.

Don’t Claim Benchmark-Oriented Optimization Improves General Coding Capability — Diverse Evaluation Is Required

Egor Shibaev, Vera Kudrevskaia, Timur Galimzyanov…

This study points out that the current industry practice of equating scores on a small number of coding benchmarks such as SWE-bench with general coding capability is biased. The team’s self-made Django test set verification found that models optimized for specific benchmarks have extremely weak cross-task transferability, benchmark rankings cannot be generalized, and there is a significant gap between scores and actual coding capabilities. The study calls for the adoption of a diversified evaluation system, building a capability classification framework and continuously maintaining benchmarks to avoid evaluation misleading R&D decisions.

Robust XGBoosting for Regression

Iris Aragón Mladosich, Christophe Croux

This paper studies the regression robustness of the mainstream gradient boosting prediction model XGBoost, and finds that the standard version using squared loss and Huber loss is susceptible to interference from vertical outliers and high leverage points. The study introduces M, S, and τ estimators from robust regression to construct alternative loss functions for verification, and the finally proposed two-step MM-XGBoost scheme can achieve the optimal trade-off between robustness and prediction accuracy.

OpenAI

The Defender’s Window

OpenAI

This study titled The Defender’s Window focuses on the new cybersecurity landscape in the AI era: current AI technology is simultaneously restructuring the capability framework of both sides of cyber attack and defense. The content not only discloses OpenAI’s own practical experience in iterating and upgrading its security defense system, but also provides actionable security capability optimization plans for security operation and maintenance teams of various organizations, helping defenders seize the initiative in AI attack and defense confrontations.

OpenAI joins PORTS-Pike project

OpenAI

OpenAI officially joined the PORTS-Pike project recently, which is a key move for the company to expand its community investment layout. The project is implemented in southern Ohio, USA. OpenAI’s participation will inject technology enterprise resources into local development, which is expected to directly support thousands of jobs, further expand the lineup of participating entities of the project, improve the overall implementation effectiveness of the project, and boost regional development.

Anthropic News

Introducing Claude Opus 5

Anthropic

Claude Opus 5 is a step-upgraded version of the Opus-tier large model in the Claude series. This upgrade focuses on optimizing underlying support capabilities, which can better adapt to the continuous task execution needs of long-running agents; at the same time, coding capabilities and professional scenario processing capabilities have also been significantly improved, which can more efficiently support the implementation of complex development and professional tasks in various fields.

Inviting hard questions

Anthropic

This initiative titled Inviting Hard Questions core measure is to collect high-difficulty, controversial hard questions about the AI field from the public across society. The team also publicly promises that in the whole process of responding to and tackling these problems, all relevant work progress and content will be announced throughout the process, actively improving the transparency of AI research, and strengthening public participation and supervision in the AI field.

Google DeepMind

Introducing Gemini 3.7 Flash

Google DeepMind

Google’s newly released Gemini 3.7 Flash is the lightweight flagship large model of the Gemini series. Through architecture optimization, its inference speed is 2x higher than the previous generation 3.5 Flash, it supports a million-token level context window, its multimodal capabilities cover text, images, audio and video, its complex reasoning performance is close to the previous generation Pro flagship, and the inference cost is only 1/10 of the Pro version. It focuses on low-latency high-concurrency scenarios, adapting to real-time interaction, on-device deployment and other implementation needs.

Putting sign language AI into users’ hands

Google DeepMind

This research addresses the daily communication pain points of deaf and hearing-impaired groups, and has developed a breakthrough sign language to text (SL2T) AI model, which can directly support the sign language function of consumer products, can be used independently by users without professional collection equipment, breaking the limitation that previous sign language AI mostly stayed in laboratories and had high implementation thresholds. It can effectively reduce the cost of communication between deaf and hearing people, and promote the inclusive implementation of sign language AI.

Hugging Face Blog

Same Cluster, 33 Points More Utilization: What Changed Was the Order

Hugging Face

This study addresses the common problem of resource idling in computing power clusters, and proposes a lightweight optimization solution that does not require hardware changes or resource expansion: only by reconstructing the task scheduling priority sorting logic and adjusting the task submission and execution order, the resource utilization rate of clusters with the same configuration can be increased by 33 percentage points, providing a low-cost new idea for cost reduction and efficiency improvement of clusters in big data and cloud native scenarios.

State of Open Models: Summer 2026 Observations

Hugging Face

Currently, only the title of this post is provided, with no specific summary body content attached, so the request for translation and extraction cannot be completed. Please supplement and upload the full summary text of this State of Open Models: Summer 2026 Observations post, and I will highlight the core methods and conclusions as required to output a concise and clear summary of approximately 120 words.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

Recursive Self-Improvement (RSI) was first proposed by I.J. Good in 1965, referring to a superintelligence mechanism that can surpass all human intellectual activities and independently design better systems for iteration. In 2008, Eliezer Yudkowsky clarified that its core is the feedback loop where AI optimizes its own cognitive architecture relying on existing intelligence. In the current AI context, this mechanism can be manifested as the model directly rewriting its own weights, or broadly refer to the model optimizing its own training process.

Qbitai

Gongsheng Zhixing Releases Humanoid Robot Racing Demo: Testing Bipedal Robot “Whole-Body Intelligence” with Go-Karts

Qbitai

In August, embodied intelligence startup Gongsheng Zhixing publicly released its phased results for the first time: a demo of a bipedal humanoid robot driving a go-kart, which is used as a whole-body intelligence stress test to verify the complex interaction capabilities of perception, balance, and multi-limb coordination. The company positions itself as a bipedal whole-body intelligence base model manufacturer, exploring the end-to-end technical route from perception to action. Its core team comes from top universities and leading technology companies, and relevant technical details will be disclosed later.

Humanoid Robots Are Now Playing Table Tennis! Two Robots Completed a Full 11-point Match

Qbitai

The Hyperdimensional Team of the University of Hong Kong self-developed the SMASH2.0 closed-loop system not driven by a single large model, which integrates visual perception, trajectory prediction, motion planning and whole-body control modules. After the upgrade, it supports robots to handle long and short balls and serve autonomously, realizing the world’s first case of two humanoid robots completing an 11-point table tennis match without manual intervention. It will be unveiled at the 2nd World Humanoid Robot Games to play against famous table tennis players.

Fields Medal Winner: AI Now Mainly Breaks Through Major Mathematical Conjectures by Finding Counterexamples

Qbitai

Timothy Gowers, winner of the 1998 Fields Medal, summed up the rule: recent breakthrough progress of AI in the field of mathematics mostly follows the “counterexample finding” path. On well-known problems such as the Jacobian conjecture and the Erdős unit distance conjecture, it does not follow the human forward proof idea, but directly constructs counterexamples to overturn the original consensus. This path is essentially different from the traditional reduction to absurdity used for proof, and has produced many heavyweight achievements.