R[AI]DAR

Category intelligence

Research Briefing

All 130 current items, analyzed and ranked.

Executive synthesis

Research Summary

Executive Signal

Priority Developments

  • Measurement integrity is collapsing: Sharpening Tax and novelty-judge instability show capability claims and review decisions must be validated on pass@1 and prompt-robust protocols, not best-of-N.
  • Agentic systems have security and failure blind spots: Latent inter-agent links elevate compliance risk, and self-evolving pipelines can co-cheat their own rewards; both demand adversarial red-teaming before deployment.
  • New benchmarks anchor scientific agent investment: OSWorld-Science delivers artifact-verified scientific software tasks, giving enterprises a credible yardstick for research-automation roadmaps.
  • Architectural shifts enable globally constrained reasoning: Blackboard-style masked diffusion and low-rank RL updates in VLA models point to compact, revisable designs for complex planning and robotics.
  • Automated research needs idea-level governance: AIM's Bayesian-style idea management separates ideation from execution, a model enterprises should adopt for AI-led R&D workflows.

Leadership Implications

Key Themes

AI Safety Evaluation & Benchmarks · 4Agent Systems and Harnesses · 12Agent Self-Improvement · 5AI Safety and Alignment · 2LLM-as-Judge and Evaluation Reliability · 2Self-Evolving and Self-Improving Systems · 8Benchmark and Evaluation Methodology · 6AI Safety and Reliability · 4Embodied AI and Robotics · 5AI Safety and Security · 2

Primary evidence

All Ranked Signals

Showing 10 of 130 signals 2 remaining
Research Hugging Face Papers 6 days ago

OSWorld-Science: A Benchmark of Computer Use Agents for Learning and Using Scientific Software

By Dingyuan Dai, Heli Qi, Lei Liu, Yinxi Li, Baiding Chen, Zijun Dou, Qingcheng Zeng, Qi Kang, Oliver Sun, Eric Wang, Bo Zhou, Haixin Wang, Yufan Du, Shi Bo, Ruihan Lin, Mengqi Yuan, Dunjie Lu, Steven Dillmann, Yiming Shi, Tina Su, Amy Xin, Minghao Liu, Xi Wang, Xu Huang, Ge Zhang, Pengyu Nie, Zhen Yang, Jie Tang, Juanzi Li, Weihao Xuan, Tianyu Liu

82 score
AI Analysis

OSWorld-Science is a benchmark of 146 high-quality tasks across scientific domains (molecular drawing, pathology, statistics, physics) for evaluating VLM-based computer-use agents with artifact-based verification. Evaluates 12 VLMs with an efficient agent harness.

Source details
Scientific software presents a demanding test for computer-using agents based on visual language models (VLMs): completing a research workflow requires interpreting specialized interfaces, manipulating scientific objects, and producing verifiable results. We thus introduce OSWorld-Science, a benchmark and evaluation environment that combines scientifically meaningful tasks, artifact-based evaluation, and an efficient agent harness for studying computer use in the scientific domain. The benchmark
BenchmarksComputer-Use AgentsScientific Computing
Research Hugging Face Papers 6 days ago

Safety of Latent Communication in Multi-Agent Systems

By Muhammad Huzaifa, Sina Mavali, Thorsten Eisenhofer

82 score
AI Analysis

The paper exposes a security vulnerability in latent communication between agents, showing that even benign training of trainable inter-agent links can increase harmful compliance relative to text-based communication. An RL-based attack further amplifies this without requiring harmful target responses.

Source details
Latent communication enables multi-agent systems to exchange information directly in internal representation space, reducing the token, computation, and latency overhead of text-based communication. To this end, lightweight trainable links are introduced to map the sender's representations into the receiver's input space. In this work, we show that even benign link training can increase harmful compliance relative to text-based communication while the underlying safety-aligned agents remain unch
AI SafetyMulti-Agent SystemsAdversarial Attacks
Research AlphaXiv Trending 6 days ago

Sharpening Tax in Post-Training

By Changdae Oh, Qi Zeng, Qi Qi, Andrey Zhmoginov, Deren Lei, Yun He, Hoang Phan, Hangoo Kang, Azalia Mirhoseini, Sharon Li

82 score
AI Analysis

The authors formalize Sharpening Tax to measure how much of an agent's reported capability comes from retry budgets rather than genuine post-training improvement. On Gemma-4-31B for WebShop, the base model reaches 87.6% at pass@128 versus 22.7% at pass@1, while the post-trained variant is much flatter (39.5% to 56.0%).

Source details
The authors compare base and post-trained agents on three agentic benchmarks across repeated rollouts, and introduce Sharpening Tax to measure how much success retries recover. For Gemma-4-31B on WebShop, the harness-equipped base model rises from 22.7% pass@1 to 87.6% pass@128, while its post-trained counterpart reaches 39.5% and 56.0%. The comparison is observational and varies by model scale and rollout budget.
Agentic EvaluationPost-TrainingBenchmarkingAI Methodology
Research AlphaXiv Trending 6 days ago

Blackboard Intelligence Can Surpass Autoregressive on Globally Constrained Problems

By Woosang Jeon, Jaeyeon Kim, Sham Kakade, Yilun Du, Amrit Singh Bedi, Arun Kumar Chithanar, Chul Lee, Taehyeong Kim, Sitan Chen

82 score
AI Analysis

Blackboard lets masked diffusion models solve globally constrained tasks by filling a revisable partial solution rather than extending a one-way sequence, using confidence across unfilled cells to guide reconsideration. On 500 ZebraLogic-Hard puzzles, exact accuracy for fine-tuned LLaDA-8B-Instruct rises from 78.4% to 90.4%.

Source details
Blackboard lets masked diffusion models solve globally constrained tasks by filling a revisable partial solution rather than extending a one-way sequence. It uses confidence across unfilled cells to guide reconsideration, but model-estimated confidence does not guarantee correctness. On 500 ZebraLogic-Hard puzzles, exact accuracy for a fixed fine-tuned LLaDA-8B-Instruct checkpoint rises from 78.4% to 90.4%, within three tested synthetic task settings.
Diffusion Language ModelsReasoningConstrained GenerationArchitecture
Research Hugging Face Papers 6 days ago

False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

By Meijia Chen, Hao Li, Zheng Lu, Hongshan Lin, Junbai Tian, Yichen Liu, Zijun Tian, Yufan Zou, Shuhan Sun, Hanxin Chen, Zeyu Zhang, Weizhi Du, Yueting Li, Tianyu Shi, Alaa Khamis

80 score
AI Analysis

Identifies a failure mode called co-cheating in self-evolving search agents, where a proposer and solver mutually reinforce shared errors so in-loop reward rises while external correctness stagnates. Proposes multi-sample verification (MSV) that queries the model with and without source evidence to vet proposals before training.

Source details
Self-evolving search agents build their own training curricula by jointly optimizing a proposer that generates questions and a solver that answers them. This closed loop introduces a failure mode we call co-cheating: the proposer and solver increasingly agree on shared errors, so internal reward improves without a matching gain in external correctness. A post-hoc audit against source evidence shows co-cheating growing more severe over successive rounds of self-evolution, with pseudo-label correc
Self-Evolving SystemsAI SafetySearch Agents
Research Hugging Face Papers 6 days ago

The Low-Rank Structure of VLA Reinforcement Learning

By Minjae Oh, Yoonah Park, Jongwon Lim, Yohan Jo

80 score
AI Analysis

This paper discovers that RL applied to vision-language-action models like π0.5 and GR00T N1.5/N1.6 induces substantially low-rank parameter updates concentrated in the action expert's Timestep Modules. Module-replacement experiments show these small components capture a disproportionate share of RL's performance gains.

Source details
Reinforcement learning (RL) is increasingly used to post-train vision-language-action (VLA) models, yet how RL reshapes these policies remains poorly understood. We find that RL across widely used flow-based VLA models, including π_{0.5} and GR00T~N1.5/N1.6, on LIBERO, ManiSkill, MetaWorld, and CALVIN induces substantially lower-rank parameter updates that are highly concentrated in the action expert's Timestep Modules, a small and previously overlooked component. Through systematic module-repla
Embodied AI / RoboticsReinforcement LearningMechanistic Interpretability
Research AlphaXiv Trending 5 days ago

Old Ideas, Novel Problems: The Instability of LLM-Based Novelty Evaluation

By Noy Sternlicht, Simra Shahid, Peter Jansen, Daniel S. Weld, Pao Siangliulue, Tom Hope

80 score
AI Analysis

The authors test how LLM-based novelty judges respond to prompt changes using ICLR 2026 reviews and submissions, comparing human-authored ideas against Claude Sonnet 4.5-generated variants. GPT-5.4's pairwise accuracy on identical Human+Generated pairs drops from 93.5% to 40.9% across two prompts.

Source details
The authors test how novelty judges respond to prompt changes using ICLR 2026 reviews and submissions. They compare human-authored ideas with other human ideas or Claude Sonnet 4.5-generated ideas, whose lower novelty is assumed. On identical Human+Generated pairs, GPT-5.4’s pairwise accuracy ranges from 93.5% to 40.9% across two prompts, showing instability rather than establishing which wording is correct.
AI EvaluationLLM-as-JudgePeer ReviewAI Methodology
Research Hugging Face Papers 6 days ago

AIM: Agentic Idea Management for Automated Research

By Hyeong Kyu Choi, Bhavana Dalvi Mishra, Jiefeng Chen, Mihir Parmar, Rui Meng, Chun-Liang Li, Xiangru Tang, Sharon Li, Jinsung Yoon, Tomas Pfister

79 score
AI Analysis

Distinguishes idea-driven from solution-driven automated research and proposes AIM, an Agentic Idea Manager with a surrogate, acquisition function, solution auditor, and resource planner inspired by Bayesian optimization. Aims to organize evolving ideas, select promising directions, and keep ideas aligned with implementations.

Source details
Frontier LLMs are increasingly used to automate scientific research through iterative search. We distinguish idea-driven search from solution-driven search and identify three core challenges: organizing evolving research ideas, selecting promising directions, and maintaining alignment between ideas and their implementations. To address these challenges, we introduce the Agentic Idea Manager (AIM), a fully autonomous framework for managing and exploring research directions in idea-driven automate
Automated ResearchAgent SystemsBayesian Optimization
Research Hugging Face Papers 6 days ago

Scaling Laws for Looped Mixture of Experts

By Yanbei Chen, Anirudh Goyal, Raghuraman Krishnamoorthi

79 score
AI Analysis

Loop Scaling Laws is the first scaling law to jointly model recurrence and sparsity alongside model size and data for looped mixture-of-experts. It introduces a bounded, sparsity-conditional recurrence mapping that recovers standard dense and MoE laws as special cases.

Source details
Looped transformers and Mixture-of-Experts (MoE) offer complementary routes to efficient scaling: recurrence increases computational depth at fixed parameters, while MoE sparsity expands total capacity at fixed active compute. Yet existing scaling laws model recurrence or sparsity in isolation. In this work, we introduce Loop Scaling Laws, the first scaling law to jointly model recurrence and sparsity alongside model size and data. At its core is a bounded, sparsity-conditional recurrence mappin
Scaling LawsMixture of ExpertsTransformer Architecture
Research Hugging Face Papers 6 days ago

AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks

By Hongjin Qian, Chaofan Li, Kun Luo, Wenqing Wei, Jianlyu Chen, Shuqi Lu, Yuyang Hu, Hongwang Xiao, Hui Wang, Chaozhuo Li, Qiwei Ye, Zhicheng Dou, Defu Lian, Zheng Liu

78 score
AI Analysis

Defines self-improving agents via two complementary capabilities (reflection and long-horizon execution) and synthesizes training trajectories from ML/algorithmic programming domains with verifiable feedback. Trained on Qwen3.8-27B, the agent achieves strong results on MLE-bench Lite and related benchmarks.

Source details
We present AREX-2, an effort to advance the self-improving capability of LLM agents, which we define as the ability to iteratively refine a solution at test time. This ability rests on two complementary capabilities: reflection, which produces a solution better than the current one, and long-horizon execution, which keeps the iteration effective over many rounds. We hypothesize that both capabilities are domain-agnostic, and can therefore be learned in scenarios that are well suited for supervis
Agent SystemsSelf-ImprovementReinforcement Learning