Wiredframe Radar logo

Category intelligence

Research Briefing — February 9, 2026

428 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research reveals critical vulnerabilities in the AI ecosystem alongside fundamental theoretical advances. Security research dominates: a first-of-its-kind study finds 157 malicious skills with 632 vulnerabilities across 98K agent skills in community registries, while Microsoft's GRP-Obliteration demonstrates safety alignment can be removed with a single unlabeled prompt.

Safety infrastructure advances with TamperBench for fine-tuning attacks, REBEL demonstrating that models passing standard unlearning benchmarks still leak 'forgotten' knowledge, and theoretical work proving steering vectors are fundamentally non-identifiable. GhostCite finds all tested models hallucinate citations at 14-95% rates across 40 domains.

Key Themes

AI Safety & Alignment · 16Large Reasoning Models & Policy Optimization · 10AI Security & Agent Safety · 10Vision-Language-Action Models & Robotics · 12LLM Reliability and Safety · 9Reasoning Model Evaluation · 5LLM Efficiency & Training · 12Efficient AI & Inference Optimization · 8LLM Reasoning & Test-Time Compute · 8RLHF and Alignment · 8

Primary evidence

Top Ranked Signals

Research arXiv (cs.CR) Feb 9

Malicious Agent Skills in the Wild: A Large-Scale Security Empirical Study

By Yi Liu, Zhihao Chen, Yanjun Zhang, Gelei Deng, Yuekang Li, Jianting Ning, and Leo Yu Zhang

88 score
AI Analysis

First labeled dataset of malicious agent skills from community registries, finding 157 malicious skills with 632 vulnerabilities across 98K analyzed. Identifies Data Thieves and Agent Hijackers as two attack archetypes.

Third-party agent skills extend LLM-based agents with instruction files and executable code that run on users' machines. Skills execute with user privileges and are distributed through community registries with minimal vetting, but no ground-truth dataset exists to characterize the resulting threats. We construct the first labeled dataset of malicious agent skills by behaviorally verifying 98,380 skills from two community registries, confirming 157 malicious skills with 632 vulnerabilities. Thes
AI SecurityAgent SafetyVulnerability Research
Research arXiv (Machine Learning) Feb 9

GRP-Obliteration: Unaligning LLMs With a Single Unlabeled Prompt

By Mark Russinovich, Yanan Cai, Keegan Hines, Giorgio Severi, Blake Bullwinkel, Ahmed Salem

85 score
AI Analysis

Introduces GRP-Obliteration, a method using GRPO to unalign safety-aligned models with a single unlabeled prompt while largely preserving utility. Achieves stronger unalignment than existing techniques.

Safety alignment is only as robust as its weakest failure mode. Despite extensive work on safety post-training, it has been shown that models can be readily unaligned through post-deployment fine-tuning. However, these methods often require extensive data curation and degrade model utility. In this work, we extend the practical limits of unalignment by introducing GRP-Obliteration (GRP-Oblit), a method that uses Group Relative Policy Optimization (GRPO) to directly remove safety constraints fr
AI SafetyAlignmentJailbreaking
Research arXiv (math.AG) Feb 9

Evolving Ranking Functions for Canonical Blow-Ups in Positive Characteristic

By Gergely B\'erczi

82 score
AI Analysis

Uses AlphaEvolve to discover ranking functions for resolution of singularities in positive characteristic algebraic geometry - a long-standing open problem since Hironaka's 1964 Fields Medal work.

Resolution of singularities in positive characteristic remains a long-standing open problem in algebraic geometry. In characteristic zero, the problem was solved by Hironaka in 1964, work for which he was awarded the Fields Medal. Modern proofs proceed by constructing suitable ranking functions, that is, invariants shown to strictly decrease along canonical sequences of blow-ups, ensuring termination. In positive characteristic, however, no such general ranking function is known: Frobenius-speci
AI for MathematicsAlgebraic GeometryEvolutionary Search
Research arXiv (cs.CR) Feb 9

GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models

By Zuyao Xu, Yuqi Qiu, Lu Sun, FaSheng Miao, Fubin Wu, Xinyi Wang, Xiang Li, Haozhe Lu, ZhengZe Zhang, Yuxin Hu, Jialu Li, Jin Luo, Feng Zhang, Rui Luo, Xinran Liu, Yingxian Li, Jiaji Liu

82 score
AI Analysis

Introduces CiteVerifier framework and benchmarks 13 LLMs on citation generation across 40 domains, finding all models hallucinate citations at 14-95% rates. Reveals that many hallucinated citations are nearly real.

Citations provide the basis for trusting scientific claims; when they are invalid or fabricated, this trust collapses. With the advent of Large Language Models (LLMs), this risk has intensified: LLMs are increasingly used for academic writing, yet their tendency to fabricate citations (``ghost citations'') poses a systemic threat to citation validity. To quantify this threat and inform mitigation, we develop CiteVerifier, an open-source framework for large-scale citation verification, and cond
LLM ReliabilityHallucinationAI SafetyScientific Integrity
Research arXiv (Artificial Intelligence) Feb 9

Exposing Weaknesses of Large Reasoning Models through Graph Algorithm Problems

By Qifan Zhang, Jianhao Ruan, Aochuan Chen, Kang Zeng, Nuo Chen, Jing Tang, Jia Li

81 score
AI Analysis

Introduces GrAlgoBench, a benchmark using graph algorithm problems to evaluate Large Reasoning Models, revealing accuracy drops below 50% when graphs exceed 120 nodes and quadratic thinking token growth.

Large Reasoning Models (LRMs) have advanced rapidly; however, existing benchmarks in mathematics, code, and common-sense reasoning remain limited. They lack long-context evaluation, offer insufficient challenge, and provide answers that are difficult to verify programmatically. We introduce GrAlgoBench, a benchmark designed to evaluate LRMs through graph algorithm problems. Such problems are particularly well suited for probing reasoning abilities: they demand long-context reasoning, allow fine-
Reasoning ModelsBenchmarksGraph AlgorithmsEvaluation
Research arXiv (Computation and Language) Feb 9

Baichuan-M3: Modeling Clinical Inquiry for Reliable Medical Decision-Making

By Baichuan-M3 Team: Chengfeng Dou, Fan Yang, Fei Li, Jiyuan Jia, Qiang Ju, Shuai Wang, Tianpeng Li, Xiangrong Zeng, Yijie Zhou, Hongda Zhang, Jinyang Tai, Linzhuang Sun, Peidong Guo, Yichuan Mo, Xiaochuan Wang, Hengfu Cui, Zhishou Zhang

80 score
AI Analysis

Baichuan-M3 is medical LLM designed for active clinical decision support with proactive information acquisition, long-horizon reasoning, and hallucination suppression. Claims SOTA on HealthBench, outperforming GPT-5.2.

We introduce Baichuan-M3, a medical-enhanced large language model engineered to shift the paradigm from passive question-answering to active, clinical-grade decision support. Addressing the limitations of existing systems in open-ended consultations, Baichuan-M3 utilizes a specialized training pipeline to model the systematic workflow of a physician. Key capabilities include: (i) proactive information acquisition to resolve ambiguity; (ii) long-horizon reasoning that unifies scattered evidence i
Medical AIClinical Decision SupportLLM Specialization
Research arXiv (Artificial Intelligence) Feb 9

AIRS-Bench: a Suite of Tasks for Frontier AI Research Science Agents

By Alisia Lupidi, Bhavul Gauri, Thomas Simon Foster, Bassel Al Omari, Despoina Magka, Alberto Pepe, Alexis Audran-Reiss, Muna Aghamelu, Nicolas Baldwin, Lucia Cipolina-Kun, Jean-Christophe Gagnon-Audet, Chee Hau Leow, Sandra Lefdal, Hossam Mossalam, Abhinav Moudgil, Saba Nazir, Emanuel Tewolde, Isabel Urrego, Jordi Armengol Estape, Amar Budhiraja, Gaurav Chaurasia, Abhishek Charnalia, Derek Dunfield, Karen Hambardzumyan, Daniel Izcovich, Martin Josifoski, Ishita Mediratta, Kelvin Niu, Parth Pathak, Michael Shvartsman, Edan Toledo, Anton Protopopov, Roberta Raileanu, Alexander Miller, Tatiana Shavrina, Jakob Foerster, Yoram Bachrach

80 score
AI Analysis

Introduces AIRS-Bench, 20 tasks from SOTA ML papers for evaluating AI research agents across idea generation, experiment analysis, and iterative refinement. Establishes baselines with frontier models.

LLM agents hold significant promise for advancing scientific research. To accelerate this progress, we introduce AIRS-Bench (the AI Research Science Benchmark), a suite of 20 tasks sourced from state-of-the-art machine learning papers. These tasks span diverse domains, including language modeling, mathematics, bioinformatics, and time series forecasting. AIRS-Bench tasks assess agentic capabilities over the full research lifecycle -- including idea generation, experiment analysis and iterative r
Agentic AIAI for ResearchBenchmarkEvaluation
Research arXiv (Robotics) Feb 9

DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos

By Shenyuan Gao, William Liang, Kaiyuan Zheng, Ayaan Malik, Seonghyeon Ye, Sihyun Yu, Wei-Cheng Tseng, Yuzhu Dong, Kaichun Mo, Chen-Hsuan Lin, Qianli Ma, Seungjun Nah, Loic Magne, Jiannan Xiang, Yuqi Xie, Ruijie Zheng, Dantong Niu, You Liang Tan, K.R. Zentner, George Kurian, Suneel Indupuru, Pooya Jannaty, Jinwei Gu, Jun Zhang, Jitendra Malik, Pieter Abbeel, Ming-Yu Liu, Yuke Zhu, Joel Jang, Linxi "Jim" Fan

80 score
AI Analysis

DreamDojo is a foundation world model trained on 44k hours of egocentric human videos - the largest video dataset for world model pretraining. Uses continuous latent actions to learn dexterous control from action-unlabeled videos.

Being able to simulate the outcomes of actions in varied environments will revolutionize the development of generalist agents at scale. However, modeling these world dynamics, especially for dexterous robotics tasks, poses significant challenges due to limited data coverage and scarce action labels. As an endeavor towards this end, we introduce DreamDojo, a foundation world model that learns diverse interactions and dexterous controls from 44k hours of egocentric human videos. Our data mixture r
World ModelsRoboticsVideo UnderstandingFoundation Models
Research arXiv (Machine Learning) Feb 9

The Condensate Theorem: Transformers are O(n), Not $O(n^2)$

By Jorge L. Ruiz Williams

79 score
AI Analysis

Claims attention sparsity is a learned topological property achieving 100% output equivalence with full O(n²) attention, demonstrating lossless O(n) attention across multiple models.

We present the Condensate Theorem: attention sparsity is a learned topological property, not an architectural constraint. Through empirical analysis of trained language models, we find that attention mass concentrates on a distinct topological manifold -- and this manifold can be identified dynamically without checking every position. We prove a general result: for any query, projecting attention onto the Condensate Manifold (Anchor + Window + Dynamic Top-k) achieves 100% output equivalence with
Attention MechanismsEfficient InferenceDeep Learning Theory
Research arXiv (Machine Learning) Feb 9

REBEL: Hidden Knowledge Recovery via Evolutionary-Based Evaluation Loop

By Patryk Rybak, Pawe{\l} Batorski, Paul Swoboda, Przemys{\l}aw Spurek

78 score
AI Analysis

Introduces REBEL, an evolutionary approach for adversarial prompt generation that successfully recovers 'forgotten' knowledge from models that pass standard unlearning benchmarks.

Machine unlearning for LLMs aims to remove sensitive or copyrighted data from trained models. However, the true efficacy of current unlearning methods remains uncertain. Standard evaluation metrics rely on benign queries that often mistake superficial information suppression for genuine knowledge removal. Such metrics fail to detect residual knowledge that more sophisticated prompting strategies could still extract. We introduce REBEL, an evolutionary approach for adversarial prompt generation d
Machine UnlearningAI SafetyAdversarial Evaluation
Research arXiv (Artificial Intelligence) Feb 9

Intrinsic Stability Limits of Autoregressive Reasoning: Structural Consequences for Long-Horizon Execution

By Hsien-Jyh Liao

78 score
AI Analysis

Argues that autoregressive reasoning has intrinsic stability limits due to process-level instability rather than just task complexity, even in linear unbranched tasks without semantic ambiguity.

Large language models (LLMs) demonstrate remarkable reasoning capabilities, yet their performance often deteriorates sharply in long-horizon tasks, exhibiting systematic breakdown beyond certain scales. Conventional explanations primarily attribute this phenomenon to task complexity, such as combinatorial search explosion or long-term credit assignment challenges. In this work, we argue that these explanations are incomplete: even in linear, unbranched tasks without semantic ambiguity, autoregre
Language ModelsReasoningTheory
Research arXiv (Artificial Intelligence) Feb 9

SeeUPO: Sequence-Level Agentic-RL with Convergence Guarantees

By Tianyi Hu, Qingxu Fu, Yanxi Chen, Zhaoyang Liu, Bolin Ding

78 score
AI Analysis

SeeUPO provides convergence guarantees for agentic RL in multi-turn settings. Shows REINFORCE with GRAE converges globally while PPO+GRAE breaks monotonic improvement.

Reinforcement learning (RL) has emerged as the predominant paradigm for training large language model (LLM)-based AI agents. However, existing backbone RL algorithms lack verified convergence guarantees in agentic scenarios, especially in multi-turn settings, which can lead to training instability and failure to converge to optimal policies. In this paper, we systematically analyze how different combinations of policy update mechanisms and advantage estimation methods affect convergence proper
Reinforcement LearningLLM AgentsConvergence Theory