R[AI]DAR

Category intelligence

Research Briefing

All 62 current items, analyzed and ranked.

Executive synthesis

Research Summary

Executive Signal

Priority Developments

Leadership Implications

  • Allocate resources to agentic infrastructure as multi-agent scaling yields measurable gains on verifiable, complex work.
  • Track alignment research momentum and world-model capability gaps before locking in deployment assumptions or roadmap bets.

Key Themes

Multilingual Language Models · 1Agentic AI and Multi-Agent Systems · 9Reinforcement Learning for LLMs · 2World Models and Interactive Generation · 4Efficiency, Quantization, and Inference Acceleration · 4AI Agents · 5AI Safety, Robustness, and Reliability · 5LLM Training and Optimization · 4AI Safety and Alignment · 2Benchmarking and Evaluation · 8

Primary evidence

All Ranked Signals

Showing 10 of 62 signals 2 remaining
Research LessWrong Aug 25

Sampura Research: Human-AI Complementarity for Alignment

By Rishub Jain

88 score
AI Analysis

Announces Sampura Research, a new nonprofit launched by former Google DeepMind researchers with $11M from Coefficient Giving, focused on Human-AI Complementarity for Scalable Oversight, starting with better judges to address reward hacking, sycophancy, and weak evaluation in RLHF.

Source details
We've left Google DeepMind to start Sampura Research, a nonprofit building Human-AI Complementarity for Scalable Oversight - the research agenda we've pioneered for the past several years. We're backed by an initial $11M grant from Coefficient Giving ($7M for the first year, and another $4M pledged), and we're hiring founding technical staff in London! 🙂The Initial Goal: Build Better JudgesFor our first six months, we will prioritize building better judges via Human-AI Complementarity.Definitio
AI SafetyAlignmentScalable OversightReinforcement Learning
Research AlphaXiv Trending Aug 25

Why Pretraining Fails to Share Cross-Lingual Knowledge

By Adam Gaber, Uriel Dolev, Elisabeth Fittschen, Bobby Cheng, Yuval Marton, Leshem Choshen

80 score
AI Analysis

From Weizmann Institute researchers, identifies disjoint token spaces as a primary cause of cross-lingual knowledge compartmentalization in multilingual pretraining and proposes Word-Wise Translation, which unifies token spaces and yields a 14x improvement in cross-lingual knowledge transfer for English-Arabic.

Source details
Researchers from Weizmann Institute of Science and collaborators pinpointed disjoint token spaces as a primary cause of knowledge compartmentalization in multilingual Large Language Models during pretraining. Their proposed Word-Wise Translation (WWT) intervention, which unifies token spaces, led to a 14x improvement in cross-lingual knowledge transfer for English-Arabic models, boosting equivalence scores from 0.9% to 12.6%.
Multilingual LLMsLanguage ModelsKnowledge Transfer
Research AlphaXiv Trending Aug 25

GeoWAM: Visual Geometry World Action Models for Autonomous Driving

By Yiren Lu, Xin Ye, Jiaming Liu, Philip Jacobson, Jin Yao, Yi-chung Chen, Liam Merino, Dhruva Dixith Kurra, Min Cai, Tom Lampo, Yu Yin, Danhua Guo, Burhan Yaman

78 score
AI Analysis

Uber AV Labs proposes GeoWAM, a World Action Model that predicts future 3D scene geometry directly rather than operating on pixels, then uses it to plan ego trajectories; achieves 90.2 EPDMS on NAVSIM v2 and 0.257 Abs Rel geometry error on nuScenes.

Source details
Uber AV Labs presents GeoWAM, a World Action Model that transitions autonomous driving state representation from pixels to visual geometry. This approach directly forecasts future 3D scene structure and uses it to infer ego-trajectory planning, achieving the highest EPDMS of 90.2 on NAVSIM v2 and reducing the aggregate mean Abs Rel error for geometry prediction to 0.257 on nuScenes.
Autonomous DrivingWorld Models3D Reconstruction
Research AlphaXiv Trending Aug 25

Best Practice Critic Optimization

By Penghui Qi, Xiangxin Zhou, Wee Sun Lee

75 score
AI Analysis

Develops Best Practice Critic Optimization, a stable critic-based RL recipe for LLMs combining DPPO, bounded value predictions, Monte Carlo targets, unnormalized advantages, and length-adaptive GAE, with optional reward-defining side information.

Source details
Group-based reinforcement learning methods such as GRPO for large language models avoid training a critic by sampling multiple responses for each prompt. A reliable critic could instead estimate token-level advantages from one response, but standard critic-based training recipes are often unstable. We study this instability and develop Best Practice Critic Optimization (BPCO), a recipe that combines DPPO, value predictions bounded to the reward range, Monte Carlo value targets, unnormalized
Reinforcement LearningLLM TrainingRLHF
Research Hugging Face Papers Aug 25

EchoWM: Open and Enterable Omnimodal World Models

By Songchun Zhang, Yaowei Li, Junhao Zhuang, Weiyang Jin, Haoyu Wang, Xin Lu, Yilang Sun, Shiyi Zhang, Haoran Li, Xiaoxiao Ma, Yuming Li, Yijun Liu, Yaofeng Su, Yanwen Ma, Haoyu Wu, Zihan Su, Yue Ma, Lvmin Zhang, Haoyang Huang, Zeyue Xue, Anyi Rao, Nan Duan

74 score
AI Analysis

EchoWM is an omnimodal world model that synchronously generates video, audio, music, and speech while following continuous 6-DoF navigation across first- and third-person viewpoints. It targets enterable, navigable generative environments that mix camera intent with full sensory output.

Source details
EchoWM is an omnimodal world model that generates synchronized high-resolution video, sound, music, and speech while following continuous 6-DoF navigation trajectories across first- and third-person views.
World ModelsMultimodal GenerationVideo GenerationAudio Generation
Research Hugging Face Papers + AlphaXiv Aug 25

Apodex 1.1: Scaling Agentic Intelligence for Complex Work

By Apodex Team, B. An, B. Li, B. Wang, B. Zhang, B. L. Wang, C. Feng, C. Wei, C. Xue, C. Zhang, D. Ng, D. Ye, E. Min, F. Chen, F. Liu, F. Yang, F. Ye, H. Xu, H. Yang, H. Ye, H. Zhang, H. Zhao, J. Li, J. Lin, J. Xia, K. Jin, K. Wang, K. Yang, L. Bing, L. Lei, L. Su, Le. Wang, Lu. Wang, N. Wang, Q. Ren, Q. Yang, R. Li, S. Bai, S. Du, S. Li, S. Lin, S. Nie, S. Wang, S. Zhang, S. Z. Wang, Ta. Q. Fang, Ti. Q. Fang, W. Fang, W. Li, W. Zhang, X. Chen, X. Li, X. Tang, X. Wang, X. Xu, X. Zhang, X. Q. Wang, X. Y. Wang, Y. Deng, Y. Gao, Y. Hu, Y. Li, Y. Sui, Y. Wang, Y. Xiao, Y. Zhang, Z. Chen, Z. Cheng, Z. Feng, Z. Liang, Z. Zhang

72 score
AI Analysis

Apodex 1.1 is an agentic system that jointly scales interactive environments and multi-agent coordination to tackle complex, verifiable work. It reports a 9.3-point win-rate gain on GDPVal in Agent Team mode and leads on FrontierFinance, suggesting meaningful headroom in scaling both compute and team structure rather than just model parameters.

Source details
The Apodex Team developed Apodex 1.1, an agentic system that improves the ability of large language models to complete complex, verifiable work by scaling both interactive environments and multi-agent coordination. The system demonstrates enhanced performance across professional, financial, scientific, and coding benchmarks, with its Agent Team mode achieving a 9.3-point win rate gain on GDPVal and leading performance on FrontierFinance.
Agentic AIMulti-Agent SystemsBenchmarkingLanguage Models
Research Hugging Face Papers + AlphaXiv Aug 25

Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization

By Xianlei Zhou, Xiangdi Meng, Yu He, Tianyu Qi, Shuyan Guan, Xianli Zhang, Jian Zhang, Xin Li, Qika Lin, Jun Liu

72 score
AI Analysis

Continuing our coverage from yesterday, Environment-Regularized Policy Optimization (ERPO) stabilizes LLM post-training by regularizing the policy-induced query distribution rather than the action distribution. It reports an average 6.2% accuracy gain on math reasoning benchmarks and improved stability during long training runs.

Source details
Environment-Regularized Policy Optimization (ERPO) introduces a method to stabilize large language model (LLM) post-training by regularizing the policy-induced query distribution instead of the action space, leading to an average 6.2% accuracy improvement on mathematical reasoning benchmarks and enhanced training stability during long-horizon runs.
Reinforcement LearningPolicy OptimizationLLM TrainingReasoning
Research AlphaXiv Trending Aug 25

Thinking Beyond Videos: Unifying Video Reasoning and Deep Research for Open-World Video Agents

By Wenqi Liu, Shijie Ma, Yunxiao Wang, Meng Liu, Qile Su, Han Liu, Bohan Hou, Zeyu Wang, Xuanyu Zheng, Changyi Liu, Tianke Zhang, Haonan Fan, Kaiyu Jiang, Yingxin Li, Jiankang Chen, Xu Wang, Hongyi Fu, Jianxiong Wang, Bin Wen, Tingting Gao, Han Li, Jianhua Yin, Yinwei Wei, Xuemeng Song

72 score
AI Analysis

Introduces VideoRover, a unified framework that iteratively coordinates video cropping, multimodal search, and webpage browsing for open-world video understanding, with an automated data curation pipeline for training.

Source details
Open-world video understanding often requires a model to locate sparse visual evidence and acquire external knowledge that is absent from the video and its parametric memory. While Thinking-with-Videos enables active temporal perception and Deep Research supports multi-step information seeking, the two capabilities are typically developed in isolation. We introduce VideoRover, a unified Video Deep Research framework that iteratively coordinates video cropping, multimodal search, and webpage brow
AI AgentsMultimodal ModelsRetrieval-Augmented Generation
Research Hugging Face Papers Aug 25

ReWorld: An Interactive World Model with Long-Horizon Memory

By Zhifei Chen, Luozhou Wang, Guibao Shen, Dongyu Yan, Shuai Yang, Tianshuo Xu, Yihua Du, Wei Wang, Tianyi Gui, Lianghua Huang, Yingcong Chen

71 score
AI Analysis

ReWorld separates short-horizon control and long-horizon memory during training, then enforces both at inference via mixed attention windows, a pose-indexed landmark bank, and distribution-matching LoRA distillation. The result is real-time interactive world modeling with strong action fidelity and long-range recall.

Source details
ReWorld separates short-horizon control and long-horizon memory during training, then bounds both at inference via mixed attention windows, a pose-indexed landmark bank, and distribution-matching LoRA distillation to enable real-time interactive world modeling with strong action fidelity and long-range recall.
World ModelsLong-ContextAttention MechanismsDistillation
Research Hugging Face Papers Aug 25

LongRCA Bench: Diagnosing Responsible Roles and Root Causes in Long-Horizon Agent Failures

By Yunfei Zhang, Boyu Feng, Changhua Pei, Zexin Wang, Zhihuang Peng, Xinlong Liu, Hengyue Jiang, Difeng Ma, Jiayi Zhang, Yongzhou Yao, Yanan Zhao, Fei Sun, Yintong Huo, Zhaoyang Liu, Jingjing Li, Gaogang Xie, Dan Pei

70 score
AI Analysis

LongRCA Bench diagnoses failure attribution in long-horizon agent trajectories, and the training-free RCTA method improves identification of responsible roles and root-cause steps. The work targets operational reliability for deployed multi-step agents.

Source details
LongRCA Bench evaluates failure diagnosis across lengthy agent trajectories, and the training-free RCTA method improves attribution of responsible roles and root-cause steps.
Agentic AIBenchmarkingReliabilityInterpretability