R[AI]DAR

Category intelligence

Research Briefing

All 52 current items, analyzed and ranked.

Executive synthesis

Research Summary

Executive Signal

Priority Developments

Leadership Implications

Key Themes

AI Safety and Alignment · 7Agent Training & Tool Use · 6Mechanistic Interpretability · 33D/4D Reconstruction & Vision · 4LLM Agents and Skill Management · 5LLM Unlearning & Privacy · 2AI Safety, Provenance & Verification · 6Robotics and Dexterous Manipulation · 5Machine Learning for Science (Chemistry/Materials) · 2Robotics & Embodied AI · 5

Primary evidence

All Ranked Signals

Showing 10 of 52 signals 2 remaining
Research Hugging Face Papers Aug 20

SPADE: Self-Play in Adaptive Synthetic Executable Environments

By Bo Liu, Simon Yu, Yiding Jiang, Ao Qu, Andrew Zhao, Zichen Liu, Junsu Kim, Zijian Zhou, Seungone Kim, Tongzheng Ren, Mickel Liu, Hanfei Yu, Zhaorun Chen, Weiyan Shi, Paul Pu Liang, Luke Zettlemoyer, Yejin Choi, Natasha Jaques

39 score
AI Analysis

As first covered in Research yesterday, SPADE is a self-play reinforcement learning framework in which a language model both designs adaptive executable training environments and learns to solve them, using regret-based targeting to concentrate environment difficulty. The approach targets reasoning and tool-use gains in a Gym-like setup.

Source details
SPADE is a self-play reinforcement learning framework where a language model designs adaptive executable training environments and learns to solve them, improving reasoning and tool-use performance through regret-based environment targeting.
Self-Play RLAgent TrainingSynthetic EnvironmentsReasoning
Research LessWrong Aug 20

Cross-Dataset Transfer Evaluation of Deception Probes in Smaller Models

By Jollen Dai

78 score
AI Analysis

Jollen Dai replicates Apollo Research's deception probes on Llama-3.3-70B-Instruct, reproducing the AUROC values exactly, then tests the same method on five smaller (1B-9B) open models. Cross-dataset transfer AUROC collapses to near chance (~0.468) across 30 tests, with roleplay-to-sandbagging transfers performing worse than chance, suggesting probes may be largely dataset-dependent.

Source details
SummaryRecently, Apollo Research tested whether linear probes could identify honest and deceptive responses from Llama-3.3-70B-Instruct and reported AUROC values between 0.96 and 0.999. To test some of their claims, I used the scores Apollo released to recalculate the nine values they published, reproducing them exactly.I used the same method on five smaller open models, each having between 1 billion and 9 billion parameters. Each probe was trained on one deception dataset and tested on the othe
AI SafetyDeception DetectionMechanistic InterpretabilityBenchmarks
Research LessWrong Aug 19

Making sense of the misalignment risk model in the Anthropic Risk Report (August 2026)

By jasmine.ren

78 score
AI Analysis

A walkthrough of the misalignment threat model in Anthropic's August 2026 Risk Report (186 pages), summarizing the three catastrophic risk models and unpacking Section 2's formalized misalignment model. Notes the report predates Claude-Opus-5 (GA 2026-07-24, just over 3 weeks before the August 15 coverage cutoff).

Source details
Anthropic recently released their Risk Report: August 2026. It's a hefty 186-page report in which they introduced three separate threat models of catastrophic risks. In this blog post, I will give a brief overview of the three different threat models, then spend the rest of the blog post describing my own understanding of the misalignment threat model, which Anthropic formalized in Section 2 of this report. The other two threat models — automated R&D and chemical/biological weapons — are les
AI SafetyAlignmentAnthropicThreat Modeling
Research AlphaXiv Trending Aug 20

Depth Anything V4: Dynamic 4D Scene Reconstruction via Riemannian Flow Matching on 4D Gaussian Splatting

By Jiaming Fan, Jian Lu, Jinling Jia, Chenbin Zhang

76 score
AI Analysis

Depth Anything V4 extends the Depth Anything family to dynamic 4D scene reconstruction from monocular video, applying Riemannian Flow Matching to 4D Gaussian Splatting parameters so that non-Euclidean attributes remain valid throughout training. The authors isolate RFM's contribution versus pre-training and test-time optimization, claiming a +0.044 F-score improvement attributable to RFM.

Source details
We present Depth Anything V4 (DAV4), a framework for dynamic 4D scene reconstruction from monocular video. Our key contribution is the application of Riemannian Flow Matching (RFM) to 4D Gaussian Splatting parameters, defining probability paths directly on non-Euclidean manifolds (scale, rotation, opacity), ensuring all intermediate states are valid. Through controlled experiments, we isolate RFM's contribution from test-time optimization (TTO) and pre-training. A deterministic MLP baseline with
3D/4D ReconstructionGenerative ModelsComputer Vision
Research AlphaXiv Trending Aug 20

4DAnyone: Create Anyone in 4D from a Casual Monocular Video

By Yudong Jin, Tao Xie, Qihang Zhang, Zehong Shen, Zhen Xu, Yujun Shen, Hujun Bao, Xiaowei Zhou, Yinghao Xu

74 score
AI Analysis

4DAnyone reconstructs high-fidelity, free-viewpoint 4D human avatars from a single uncalibrated monocular video by first generating multi-view consistent videos at scale and then fitting 4D Gaussian Splatting. The authors report strong generalization to in-the-wild videos.

Source details
4DAnyone enables the reconstruction of high-fidelity, free-viewpoint 4D human avatars from a single uncalibrated monocular video by generating multi-view consistent videos at scale. The framework achieved superior performance in 4D Gaussian Splatting reconstruction and generated video consistency compared to existing baselines, demonstrating robust generalization to diverse in-the-wild videos.
3D/4D ReconstructionHuman AvatarsComputer Vision
Research AlphaXiv Trending Aug 20

MaliciousSkillBench: A Comprehensive Benchmark for Malicious Agent Skill Detection

By Yue Wang, Yi Liu, Gelei Deng, Ying Zhang, Yuekang Li, Zhenyu Chen, Leo Zhang

73 score
AI Analysis

A consolidated benchmark for detecting malicious Agent Skills (reusable instruction packages for LLM agents) that normalizes 8,414 raw records into 7,539 unique identities across 4,588 structural families from 13 sources. It addresses the supply-chain-style risk introduced by agent skill ecosystems.

Source details
Agent Skills extend LLM agents with reusable instruction packages that may also include scripts, resources, and service configuration. This creates a direct distribution channel for malicious behavior, yet existing malicious-Skill datasets are fragmented across sources, artifact formats, evidence regimes, and benign coverage; duplicated and structurally related content further complicates direct aggregation and evaluation. We present MaliciousSkillBench, a comprehensive benchmark for malicious A
AI SafetyAI SecurityLLM AgentsBenchmarks
Research LessWrong Aug 19

Steering Role Confusion

By Kevin Zhang

73 score
AI Analysis

Kevin Zhang extends Charles Ye et al.'s 'Mechanistic Explanation of Prompt Injections' by using activation steering to inject 'User-ness' representations and causally increase prompt-injection attack success rates, arguing that shifts in latent role representation are causal for compliance, not merely correlated.

Source details
Summary: The recent work A Mechanistic Explanation of Prompt Injections puts forward a theory that prompt injection attacks succeed due to role confusion: models primarily infer which role produced a piece of text from forgeable cues rather than its role tags.The paper's experiments establish that such cues shift latent role representation, and also increase downstream compliance with prompt injections.However, they do not show that shifts in latent role representation directly influence complia
AI SafetyMechanistic InterpretabilityPrompt InjectionActivation Steering
Research Hugging Face Papers Aug 20

Training Leaves Traces: Centered Residual Signatures for Language Model Lineage Verification

By Aman Singh Thakur, Rayan Khoury

72 score
AI Analysis

The paper shows that compatible open-weight LM checkpoints share detectable weight-space ancestry signals that can distinguish true lineage from independent training runs or distilled models, without needing the training data. This enables white-box lineage verification at the parameter level.

Source details
Compatible open-weight language model checkpoints share detectable weight-space ancestry signals that distinguish true lineage from independent or distilled models without requiring data.
Model ProvenanceAI SafetyModel ForensicsOpen-Weight Models
Research AlphaXiv Trending Aug 20

What Matters for Latent Actions in Robot Learning

By Xizhou Bu, Qingda Hu, Lei Zhou, Lingfeng Zhang, Yingbo Tang, Zihao Liu, Xinyi Tao, Zhiqiang Ma, Qingqiu Huang, Chufeng Tang, Hongbo Wang, Jing Zhang, Jiayi Ma, Hangjun Ye, Wei Li, Xiaoshuai Hao

72 score
AI Analysis

A systematic empirical study of Latent Action Models (LAMs) for robot learning that unifies 41 design choices across three dimensions under a single autoencoding framework, addressing the fragmentation in the field. The work enables practitioners to identify which design decisions actually drive downstream manipulation performance rather than which look good in isolation.

Source details
Latent Action Models (LAMs) have emerged as a promising paradigm for enabling robot learning to leverage large-scale unlabeled videos through latent actions that serve as compact surrogates for physical actions. Despite rapid progress, research on LAM remains highly fragmented, with existing methods evaluating different design choices in isolation under inconsistent experimental settings, making it difficult to identify the factors that truly determine downstream robotic manipulation performance
RoboticsRepresentation LearningSelf-Supervised LearningEmpirical Study
Research Microsoft Research Blog - Microsoft Research Aug 20

Broadening access to Skala creates a faster path to predictive DFT

By Sebastian Ehlert, Stefano Battaglia, Thijs Vogels, Jan Hermann, Jens Wehner, Giulia Luise, Klaas Giesbertz, Chin-Wei Huang, Aaron Kaplan, Kate Milton, Stephanie Marisa Lanius, Derk Kooi, P. Bernát Szabó, Gregor Simm, Rianne van den Berg, Paola Gori Giorgi

72 score
AI Analysis

Microsoft Research announces Skala 1.1, an updated deep-learning exchange-correlation functional for density functional theory (DFT) trained on 2.5× more data than its predecessor, delivering higher accuracy for thermochemistry, reaction kinetics, and molecular structure prediction. The release broadens accessibility through integration into major quantum-chemistry packages (CP2K, Psi4, FHI-aims, ORCA, VASP) and introduces a living benchmark to track ongoing improvements.

Source details
At a glance Skala 1.1 demonstrates the continuously improving nature of Microsoft Research’s deep-learning DFT approach: trained on 2.5× more data than its predecessor, it delivers substantially higher accuracy across key molecular simulation challenges, including thermochemistry, reaction kinetics, and molecular structure prediction. Skala is now available in CP2K and is being integrated into Psi4, FHI-aims, ORCA and VASP, bringing next-generation DFT accuracy closer to the communities th
Machine Learning for ScienceComputational ChemistryDensity Functional TheoryMaterials Science