Wiredframe Radar logo

Category intelligence

Research Briefing — January 26, 2026

280 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research features a major open-source release and critical safety findings. LongCat-Flash-Thinking-2601, a 560B MoE reasoning model, achieves SOTA among open-source models for agentic tasks. VibeTensor demonstrates LLM agents can generate complete deep learning system software stacks including CUDA runtime.

Theoretical and interpretability advances include floating-point transformer expressivity analysis proving non-equivariant function representation without positional encoding. Sycophancy signals are shown to be linearly separable in middle-layer attention heads, enabling targeted steering. A conceptual critique of machine unlearning argues dual-use capabilities and compositional generalization fundamentally prevent knowledge removal—an important insight for AI safety policy.

Key Themes

AI Safety & Alignment · 9LLM Agents & Agentic AI · 8AI Agents and Tool Use · 8AI Agents & Tool Use · 8Transformer Theory & Efficiency · 6Language Models · 12Medical AI & Clinical Applications · 10AI Safety and Alignment · 13AI Security & Robustness · 7Multimodal LLMs & Spatial Reasoning · 7

Primary evidence

Top Ranked Signals

Research arXiv (Artificial Intelligence) Jan 26

LongCat-Flash-Thinking-2601 Technical Report

By Meituan LongCat Team, Anchun Gui, Bei Li, Bingyang Tao, Bole Zhou, Borun Chen, Chao Zhang, Chao Zhang, Chen Gao, Chen Zhang, Chengcheng Han, Chenhui Yang, Chuyu Zhang, Cong Chen, Cunguang Wang, Daoru Pan, Defei Bu, Dengchang Zhao, Di Xiu, Dishan Liu, Dongyu Ru, Dunwei Tu, Fan Wu, Fengcheng Yuan, Fengcun Li, Gang Xu, Guanyu Wu, Guoyuan Lin, Haibin Wang, Hansi Yang, Hao Yang, Haonan Yan, Haoxiang Ma, Haoxing Wen, Hongyan Hao, Hongyin Tang, Hongyu Zang, Hongzhi Ni, Hui Su, Jiacheng Zhang, Jiahong Zhou, Jiahuan Li, Jiaming Wang, Jian Yang, Jianfei Zhang, Jianhao Xu, Jianing Wang, Jiapeng Zhu, Jiaqi Sun, Jiarong Shi, Jiarui Zhao, Jingang Wang, Jinluan Yang, Jinrui Ding, Jinwei Xiao, Jiyuan He, Juncan Xu, Kefeng Zhang, Keheng Wang, Li Wei, Lianhui Ma, Lin Qiu, Lingbing Kong, Lingchuan Liu, Linsen Guo, Mengshen Zhu, Mengxia Shen, Mingyang Zhu, Peiguang Li, Peng Pei, Pengcheng Jia, Pengtao Zhang, Peng Zhao, Qi Gu, Qiong Huang, Qiyuan Duan, Quanchi Weng, Rongxiang Weng, Rongzhi Zhang, Rumei Li, Shanglin Lei, Shengnan An, Shijun Dai, Shuaikang Liu, Shuang Zhou, Shuo Wang, Songyuan Zhao, Tao Liang, Tianhao Hu, Tianze Chen, Wei Liu, Wei Shi, Wei Wang, Weifeng Tang, Wenjie Shi, Wenlong Zhu, Wentao Chen, Wentao Shi, Xi Su, Xiangcheng Liu, Xiandi Ma, Xiangyu Xi, Xiangyuan Liu, Xiangzhou Huang, Xiao Liu, Xiaodong Cai, Xiaolong Chen, Xiaowei Shi, Xiaoyu Li, Xin Chen, Xingchen Liu, Xuan Huang, Xuezhi Cao, Xunliang Cai, Yan Chen, Yang Bai, Yang Liu, Yang Yang, Yang Zheng, Yaoming Wang, Yaoming Zhu, Yaqi Huo, Yanyu Chen, Yaorui Shi, Yerui Sun, Yi Zhang, Yihao Chen, Yi-Kai Zhang, Yifan Lu, Yifan Zhao, Yitao Zhai, Yongjing Yin, Yongwei Zhou, Youshao Xiao, Yuchuan Dai, Yuchen Xie, Yuchen Yu, Yufei Zhang, Yuhuai Wei, Yulei Qian, Yunfan Liang, Yunke Zhao, Yuwei Jiang, Yuxin Bian, Yuxin Chen, Yuxin Liu, Yue Xu, Yueqing Sun, Zeyang Yu, Zhao Yang, Zhengsheng Huang, Zhengyu Chen, Zhijian Liu, Zhikang Xia, Zhimin Lin, Zhiyuan Yao, Zhuofan Chen, Zhuowen Han, Zijian Zhang, Ziran Li, Ziwen Wang, Ziyuan Zhuang

82 score
AI Analysis

Introduces LongCat-Flash-Thinking-2601, a 560B parameter open-source MoE reasoning model achieving SOTA performance among open-source models on agentic benchmarks including search, tool use, and tool-integrated reasoning.

We introduce LongCat-Flash-Thinking-2601, a 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model with superior agentic reasoning capability. LongCat-Flash-Thinking-2601 achieves state-of-the-art performance among open-source models on a wide range of agentic benchmarks, including agentic search, agentic tool use, and tool-integrated reasoning. Beyond benchmark performance, the model demonstrates strong generalization to complex tool interactions and robust behavior under no
Large Language ModelsMixture-of-ExpertsAI AgentsTool UseOpen Source
Research arXiv (Machine Learning) Jan 26

Endless Terminals: Scaling RL Environments for Terminal Agents

By Kanishk Gandhi, Shivam Garg, Noah D. Goodman, Dimitris Papailiopoulos

78 score
AI Analysis

Introduces Endless Terminals, a fully autonomous pipeline for procedurally generating terminal-use tasks for RL training without human annotation. Trains agents with vanilla PPO achieving strong performance.

Environments are the bottleneck for self-improving agents. Current terminal benchmarks were built for evaluation, not training; reinforcement learning requires a scalable pipeline, not just a dataset. We introduce Endless Terminals, a fully autonomous pipeline that procedurally generates terminal-use tasks without human annotation. The pipeline has four stages: generating diverse task descriptions, building and validating containerized environments, producing completion tests, and filtering for
LLM AgentsReinforcement LearningAgentic AI
Research arXiv (Computation and Language) Jan 26

Persona Jailbreaking in Large Language Models

By Jivnesh Sandhan, Fei Cheng, Tushar Sandhan and Yugo Murawaki

75 score
AI Analysis

Introduces PHISH framework for persona jailbreaking through adversarial conversational history, exposing vulnerability where user-side inputs alone can manipulate LLM traits without prompting.

Large Language Models (LLMs) are increasingly deployed in domains such as education, mental health and customer support, where stable and consistent personas are critical for reliability. Yet, existing studies focus on narrative or role-playing tasks and overlook how adversarial conversational history alone can reshape induced personas. Black-box persona manipulation remains unexplored, raising concerns for robustness in realistic interactions. In response, we introduce the task of persona editi
AI SafetyJailbreakingLLM Vulnerabilities
Research arXiv (Computation and Language) Jan 26

TL-GRPO: Turn-Level RL for Reasoning-Guided Iterative Optimization

By Peiji Li, Linyang Li, Handa Sun, Wenjin Mai, Yongkang Chen, Xiaozhe Li, Yue Shen, Yichuan Ma, Yiliu Sun, Jiaxi Cao, Zhishu He, Bo Wang, Xiaoqing Zheng, Zhaori Bi, Xipeng Qiu, Qipeng Guo, Kai Chen, Dahua Lin

74 score
AI Analysis

Introduces Turn-Level GRPO (TL-GRPO) for iterative optimization tasks where trajectory value is determined by best turn-level reward. Enables fine-grained turn-level credit assignment in reasoning tasks.

Large language models have demonstrated strong reasoning capabilities in complex tasks through tool integration, which is typically framed as a Markov Decision Process and optimized with trajectory-level RL algorithms such as GRPO. However, a common class of reasoning tasks, iterative optimization, presents distinct challenges: the agent interacts with the same underlying environment state across turns, and the value of a trajectory is determined by the best turn-level reward rather than cumulat
Reinforcement LearningLLM AgentsReasoning
Research arXiv (Machine Learning) Jan 26

On the Expressive Power of Floating-Point Transformers

By Sejun Park, Yeachan Park, Geonho Hwang

73 score
AI Analysis

Investigates floating-point transformer expressivity, proving they can represent non-permutation-equivariant functions without positional encoding due to round-off errors, fundamentally different from real-valued theory.

The study on the expressive power of transformers shows that transformers are permutation equivariant, and they can approximate all permutation-equivariant continuous functions on a compact domain. However, these results are derived under real parameters and exact operations, while real implementations on computers can only use a finite set of numbers and inexact machine operations with round-off errors. In this work, we investigate the representability of floating-point transformers that use fl
Transformer TheoryMachine Learning Theory
Research arXiv (Computation and Language) Jan 26

Timely Machine: Awareness of Time Makes Test-Time Scaling Agentic

By Yichuan Ma, Linyang Li, Yongkang chen, Peiji Li, Xiaozhe Li, Qipeng Guo, Dahua Lin, Kai Chen

73 score
AI Analysis

Proposes Timely Machine redefining test-time scaling as wall-clock time rather than generation length, introducing Timely-Eval benchmark. Finds smaller models excel with fast tool feedback while larger models dominate high-latency settings.

As large language models (LLMs) increasingly tackle complex reasoning tasks, test-time scaling has become critical for enhancing capabilities. However, in agentic scenarios with frequent tool calls, the traditional generation-length-based definition breaks down: tool latency decouples inference time from generation length. We propose Timely Machine, redefining test-time as wall-clock time, where models dynamically adjust strategies based on time budgets. We introduce Timely-Eval, a benchmark spa
LLM AgentsTest-Time ComputeBenchmarks
Research arXiv (cs.SE) Jan 26

VibeTensor: System Software for Deep Learning, Fully Generated by AI Agents

By Bing Xu, Terry Chen, Fengzhe Zhou, Tianqi Chen, Yangqing Jia, Vinod Grover, Haicheng Wu, Wei Liu, Craig Wittenbrink, Wen-mei Hwu, Roger Bringmann, Ming-Yu Liu, Luis Ceze, Michael Lightstone, Humphrey Shi

72 score
AI Analysis

VibeTensor is a complete deep learning system software stack (tensor library, autograd, CUDA runtime, Python/Node.js bindings) fully generated by LLM-powered coding agents without per-change manual review.

VIBETENSOR is an open-source research system software stack for deep learning, generated by LLM-powered coding agents under high-level human guidance. In this paper, "fully generated" refers to code provenance: implementation changes were produced and applied as agent-proposed diffs; validation relied on agent-run builds, tests, and differential checks, without per-change manual diff review. It implements a PyTorch-style eager tensor library with a C++20 core (CPU+CUDA), a torch-like Python over
AI AgentsCode GenerationSystems Software
Research arXiv (Computation and Language) Jan 26

Learning Domain Knowledge in Multimodal Large Language Models through Reinforcement Fine-Tuning

By Qinglong Cao, Yuntian Chen, Chao Ma, Xiaokang Yang

72 score
AI Analysis

Discovers that input-level domain knowledge injection (prompts, captions) yields no improvement for MLLMs on specialized domains; proposes reinforcement fine-tuning to internalize domain knowledge at optimization level.

Multimodal large language models (MLLMs) have shown remarkable capabilities in multimodal perception and understanding tasks. However, their effectiveness in specialized domains, such as remote sensing and medical imaging, remains limited. A natural approach to domain adaptation is to inject domain knowledge through textual instructions, prompts, or auxiliary captions. Surprisingly, we find that such input-level domain knowledge injection yields little to no improvement on scientific multimodal
Multimodal LLMsDomain AdaptationReinforcement Learning
Research arXiv (Artificial Intelligence) Jan 26

SycoEval-EM: Sycophancy Evaluation of Large Language Models in Simulated Clinical Encounters for Emergency Care

By Dongshen Peng, Yi Wang, Carl Preiksaitis and Christian Rose

72 score
AI Analysis

Introduces SycoEval-EM evaluating LLM sycophancy in emergency medicine through adversarial patient persuasion simulations. Finds acquiescence rates vary 0-100% across 20 LLMs with model capability poorly predicting robustness.

Large language models (LLMs) show promise in clinical decision support yet risk acquiescing to patient pressure for inappropriate care. We introduce SycoEval-EM, a multi-agent simulation framework evaluating LLM robustness through adversarial patient persuasion in emergency medicine. Across 20 LLMs and 1,875 encounters spanning three Choosing Wisely scenarios, acquiescence rates ranged from 0-100\%. Models showed higher vulnerability to imaging requests (38.8\%) than opioid prescriptions (25.0\%
AI SafetyMedical AISycophancyLLM Evaluation
Research arXiv (Computation and Language) Jan 26

Sycophancy Hides Linearly in the Attention Heads

By Rifo Genadi, Munachiso Nwadike, Nurdaulet Mukhituly, Hilal Alquabeh, Tatsuya Hiraoka, Kentaro Inui

72 score
AI Analysis

Discovers that sycophancy signals in LLMs are most linearly separable in multi-head attention activations, with steering most effective in middle-layer attention heads. Shows limited overlap with previously identified 'truthful' directions, suggesting distinct mechanisms.

We find that correct-to-incorrect sycophancy signals are most linearly separable within multi-head attention activations. Motivated by the linear representation hypothesis, we train linear probes across the residual stream, multilayer perceptron (MLP), and attention layers to analyze where these signals emerge. Although separability appears in the residual stream and MLPs, steering using these probes is most effective in a sparse subset of middle-layer attention heads. Using TruthfulQA as the ba
AI SafetyMechanistic InterpretabilitySycophancyAlignment
Research arXiv (Computer Vision) Jan 26

LoL: Longer than Longer, Scaling Video Generation to Hour

By Justin Cui, Jie Wu, Ming Li, Tao Yang, Xiaojie Li, Rui Wang, Andrew Bai, Yuanhao Ban, Cho-Jui Hsieh

72 score
AI Analysis

Identifies 'sink-collapse' in autoregressive video generation where content reverts to sink frames, tracing it to RoPE-attention conflicts. Proposes training-free fix enabling hour-long video generation.

Recent research in long-form video generation has shifted from bidirectional to autoregressive models, yet these methods commonly suffer from error accumulation and a loss of long-term coherence. While attention sink frames have been introduced to mitigate this performance decay, they often induce a critical failure mode we term sink-collapse: the generated content repeatedly reverts to the sink frame, resulting in abrupt scene resets and cyclic motion patterns. Our analysis reveals that sink-co
Video GenerationAutoregressive ModelsLong-form Generation
Research arXiv (cs.CR) Jan 26

DeMark: A Query-Free Black-Box Attack on Deepfake Watermarking Defenses

By Wei Song, Zhenchang Xing, Liming Zhu, Yulei Sui, Jingling Xue

71 score
AI Analysis

Introduces DeMark, a query-free black-box attack that removes defensive watermarks from deepfakes using latent-space sparsification, reducing detection accuracy from 100% to near random across 8 watermarking schemes.

The rapid proliferation of realistic deepfakes has raised urgent concerns over their misuse, motivating the use of defensive watermarks in synthetic images for reliable detection and provenance tracking. However, this defense paradigm assumes such watermarks are inherently resistant to removal. We challenge this assumption with DeMark, a query-free black-box attack framework that targets defensive image watermarking schemes for deepfakes. DeMark exploits latent-space vulnerabilities in encoder-d
AI SecurityDeepfakesWatermarking