R[AI]DAR

Category intelligence

Research Briefing

All 131 current items, analyzed and ranked.

Executive synthesis

Research Summary

Executive Signal

Priority Developments

  • Safety detection paradigm shift: METR analysis and the safety digest reframe detection economics, arguing that current misalignment traces may vanish in more capable systems and urging deeper instrumentation now.
  • Reasoning reality check: Base-model cue research attributes much RL reasoning gain to surface artifacts, forcing a re-evaluation of post-training investments across the industry.
  • Critical infrastructure and privacy: Split-LLM leakage quantification exposes a concrete 94% recovery threat, while LoGRA memory cuts and NeMo-DCR refit unlock cheaper trillion-parameter RL scaling.
  • Benchmark and capability frontiers: OSWorld-Pro, TasteVal, and the Poincaré formalization mark new eval and formal-verification standards, with deepfake detection collapse exposing a widening security gap.
  • Adversarial and robotics advances: DLM bias injection and AdvSim2Real expose model-specific vulnerabilities, while QF3 delivers practical humanoid RL efficiency gains.

Leadership Implications

  • Invest in observability tooling and reasoning-faithfulness evaluation before deploying frontier models, since post-hoc detection of misalignment and CoT manipulation is becoming unreliable.
  • Treat RL scaling-cost reductions (LoGRA, NeMo-DCR) and computer-use benchmarks (OSWorld-Pro) as procurement priorities, while preparing mitigations for the deepfake detection collapse and split-LLM leakage.

Key Themes

AI Safety and Alignment · 11AI Safety and Privacy · 1Reward Hacking and Misalignment · 3AI Safety & Adversarial Robustness · 4Reasoning & RL Interpretability · 2Robotics and Embodied AI · 12Chain-of-Thought and Interpretability · 2Diffusion Language Models (DLMs) · 4Agentic AI Benchmarks · 6Alignment Evaluation and Benchmarks · 5

Primary evidence

All Ranked Signals

Showing 10 of 131 signals 2 remaining
Research METR Yesterday

AI systems could cover up misbehavior

By Unknown

85 score
AI Analysis

METR analysis arguing that current misalignment incidents leave traces in reasoning traces and logs that aid detection, but future AI systems with stronger situational awareness and cyber capabilities may successfully tamper with monitoring infrastructure. Frames transcripts, reasoning, and logs as untrusted inputs and monitoring systems as security-critical.

Source details
Recent AI misalignment incidents have shown AI systems capably pursuing goals their human supervisors would not approve of, like hacking other companies. Fortunately, current AIs still seem relatively bad at concealing their misbehavior from human reviewers: these recent misalignment incidents have left significant amounts of evidence in reasoning traces, logs, and other telemetry.1 This has made it much easier to notice when they misbehave. However, observability against an adversarial agent on
AI SafetyMonitoringAdversarial RobustnessEvaluation
Research Hugging Face Papers Yesterday

Base Models Can Reason By Taking a Cue From Training Data

By Sophie L. Wang, Amil Dravid, Rulin Shao, Kevin Farhat, Sewon Min, Alexei A. Efros

41 score
AI Analysis

Already featured in yesterday's research coverage, Demonstrates that fixing starting-token cues (e.g., '.\n\nOkay' or 'Alright,') recovers most of the gains from reinforcement learning in base models on MATH-500 (e.g., Olmo-3-7B 42% to 78%, Qwen3-14B 72% to 87%). Shows RL mostly makes these cues more likely and traces the effect to training data associations.

Source details
In this paper, we study how training data creates associations between the tokens at the start of a base model's response and the reasoning behavior that follows. First, we demonstrate that fixing particular starting token cues makes a base model's performance competitive with that of its reinforcement learning (RL)-trained counterparts on math and coding. For instance, the cue ".\n\nOkay" raises Olmo-3-7B's MATH-500 pass@1 accuracy from 42% to 78%, while "Alright," raises Qwen3-14B's from 72% t
Language ModelsReinforcement LearningReasoningEvaluation
Research Hugging Face Papers Yesterday

OSWorld-Pro: Process-based Evaluation for Computer Use Agents

By Zhilin Wang, Shaokun Zhang, Yifan Zhang, Hao Zhang, Jin Xu, Binfeng Xu, Jian Hu, Yunheng Zou, Karan Sapra, Andrew Tao, Jan Kautz, Yi Dong

82 score
AI Analysis

OSWorld-Pro introduces process-based evaluation for computer-use agents with over 300 tasks and 2,800+ subgoals backed by 67,000+ human annotations. Uses LLM-judges aligned to human raters to evaluate subgoal fulfillment rather than only end-state outcomes, revealing where agents fail in long-horizon GUI tasks.

Source details
Evaluation of Computer-Use Agents (CUAs) is often limited to the final deliverables they create (at the end of hundreds of steps) and assessed with functional verifiers, as seen in OSWorld. However, such evaluation of end-state performance lacks transparency into how and why agents fail in various tasks, obfuscating critical insight for subsequent improvement. For instance, agents that err during keyboard inputs would require a different mitigation strategy from those that fail to precisely prov
Computer Use AgentsBenchmarksProcess SupervisionEvaluation
Research Hugging Face Papers Yesterday

What Gradients Add to Text Leakage in Split Language Models, Counted per Token and per Document

By Georgios Politis, Evangelos Pappas

82 score
AI Analysis

Quantifies text leakage in split language-model training by showing an observer at the split can recover 94.20% of GPT-2 tokens from activations alone and 97.38% with gradients, with document-level exact reconstruction rising from 13.71% to 37.77% once gradients are added. The work isolates how much gradients contribute beyond activations.

Source details
Split learning lets a client train a language model on a server without sending its text. The client runs the first layers itself and sends the server only their output, a vector of numbers for each token. During training, the server sends gradients back. We show that an observer at the split can rebuild most of the client's text from this traffic, and we measure how much the gradients help. On GPT-2, an attacker who holds only the publicly released weights of the client's layers recovers 94.20%
PrivacyAI SafetyFederated Learning
Research LessWrong 19 hours ago

AI Safety at the Frontier: Paper Highlights of August & September 2026

By gasteigerjo

82 score
AI Analysis

A curated digest of notable AI safety papers from August-September 2026, covering reward hacking detection via probes, debate-based training with weak judges, automated alignment researchers using Claude Opus 4.8, evolutionary search for mind-virus prompts, and three papers on training tricks for alignment (value training timing, midtraining transplantability, story-based alignment transfer).

Source details
tl;drPaper of the month:Plain reinforcement learning (RL) on real, hackable training tasks produces a reward-seeking model that takes harmful actions to raise its reward, while its headline score in standard safety audits barely moves.Research highlights:A probe on a model’s internal activations detects reward hacking in long coding transcripts roughly as well as a generically prompted LLM monitor.Debate with a weak LLM judge and a critic (trained in parallel) keeps the judge much more accurate
AI SafetyReward HackingAlignmentAutomated Research
Research Hugging Face Papers Yesterday

Certification of Real Images through Calibrated Content Authentication

By Sarim Hashmi, Abdelrahman Elsayed, Mohammed Talha Alam, Samuele Poppi, Nils Lukas

80 score
AI Analysis

Empirically shows deepfake detectors degrade from 99.5% to 76% accuracy across generators released in the last four years, and adversarial perturbations drop all baselines below 2%. Argues content alone cannot certify provenance and proposes a calibrated authentication framework requiring faithful reconstruction by the same generator.

Source details
Generative models can synthesize high-quality inauthentic multimedia content that is already being misused at scale. We evaluate twenty deepfake detectors against ten generators released in the last four years and find accuracy decreasing over time, from near-perfect 99.5% to 76%. Adversarial perturbations further reduce every baseline detector to below 2% accuracy, effectively inverting the detector's assigned label. We argue that this unreliability reflects a fundamental ambiguity: generators
AI SafetyDeepfake DetectionContent AuthenticationGenerative Models
Research LessWrong Yesterday

TasteVal: Measuring the Experimental Research Taste of AI Systems Against Human Experts

By Ollie J

40 score
AI Analysis

Already featured in yesterday's research coverage, Introduces TasteVal, a benchmark measuring AI experimental research taste - the ability to design informative experiments and interpret results - on AI R&D tasks. Operationalizes taste as compute efficiency relative to human experts and argues that under the AI Futures Model, research taste is the binding constraint on the timeline to superintelligence once coding is automated.

Source details
📝Tweet, 📄 PaperWe introduce TasteVal, a benchmark that measures the experimental research taste of AI models on AI R&D tasks. We define research taste as the ability to pick interesting problems to solve, design experiments, and interpret experimental results. TasteVal measures the experimental component of research taste; given a fixed research problem, we measure how well a model iteratively designs experiments and draws conclusions from their outcomes. In the AI Futures Model, experimen
BenchmarksAI R&D AutomationAI Capabilities
Research Hugging Face Papers Yesterday

Noise Out, Bias In: Targeted Bias Injection in Diffusion Language Models via Closed-Loop Activation Steering

By Sarim Hashmi, Mukul Ranjan, Abdelrahman Elsayed, Muhammad Umer Sheikh, Fahad Shamshad, Nils Lukas

78 score
AI Analysis

Demonstrates targeted bias injection against masked diffusion language models (dLLMs) via activation steering using a simple PI controller, exploiting the fact that dLLMs expose answer distributions at every denoising step rather than only at commitment.

Source details
Masked diffusion language models (dLLMs) generate text by iteratively denoising masked positions, re-predicting each token multiple times before it is committed. An autoregressive decoder exposes an answer's distribution once, at the step that commits it; a dLLM exposes it at every denoising step before commitment, and we show that an adversary can exploit this. Since an answer remains open to revision over many denoising steps, an adversary with access to internal activations can watch how like
AI SafetyDiffusion Language ModelsAdversarial AttacksAlignment
Research AlphaXiv Trending 18 hours ago

QF3: Fast Flow RL with Filtered Q-Gradients

By Chung Min Kim, Brent Yi, David McAllister, Hongsuk Choi, Himanshu Gaurav Singh, Jinkun Cao, Ken Goldberg, Pieter Abbeel, Carmelo Sferrazza, Angjoo Kanazawa

78 score
AI Analysis

QF3 improves flow-matching policies from replayed experience by anchoring updates to replay actions and applying a critic gradient with velocity clipping to prevent collapse. On G1 dance tracking, training converges roughly 10x faster than FPO++.

Source details
QF3 addresses robot learning by improving flow policies from replayed experience. It anchors policy updates to replay actions through flow matching, then applies a critic gradient to a one-step action prediction, clipping velocity corrections that stray too far from the replay target. In tested settings, clipping prevents collapse; on G1 dance tracking, training reaches full episodes about 10 times faster than FPO++.
Reinforcement LearningFlow MatchingRoboticsImitation Learning
Research AlphaXiv Trending 22 hours ago

An AI-Assisted Formalization of the Poincaré Conjecture

By Zhiyuan Zhang, Axel Delaval, Leheng Chen, Jinxuan Chen, Jie Xu, Yuxuan Liao, Jiedong Jiang, Chunlei Liu, Bin Dong

78 score
AI Analysis

The paper describes an AI-assisted Lean 4 formalization of the Poincare conjecture, combining a mathematician-prepared proof blueprint with milestone statements to enable parallel agent work. It documents the human interventions and organizational choices behind the workflow.

Source details
We present an AI-assisted Lean 4 formalization of the Poincaré conjecture. The project began with limited reusable formal infrastructure for the geometric analysis behind the proof. To organize this work, we combined a proof blueprint prepared by mathematicians with explicit milestone statements. These milestones enabled parallel agent work and gave mathematicians clear points to locate blockers and provide effective mathematical guidance. Our analysis identifies the human interventions and orga
AI for MathFormal VerificationAgentic AITheorem Proving