Wiredframe Radar logo

Category intelligence

Research Briefing — March 4, 2026

471 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research is dominated by foundational advances in alignment theory and a strong cluster of AI safety work spanning agentic systems, cybersecurity, and content authentication.

Safety and security research is notably strong: ZeroDayBench tests frontier models on real zero-day vulnerability discovery; MOSAIC introduces plan-check-act-or-refuse for safe agentic tool use; and safety training is shown to persist through helpfulness optimization in agentic settings. The Integrity Clash paper exposes a fundamental conflict between C2PA provenance and AI watermarking, undermining deployed authentication infrastructure. On the efficiency side, Speculative Speculative Decoding parallelizes speculation and verification for practical inference speedups.

Key Themes

AI Safety & Alignment · 24AI Safety & Security · 5Deep Learning Theory · 12AI Safety and Alignment · 8Multimodal Pretraining & Foundation Models · 5Human-AI Interaction & Delegation · 4Agentic AI and Tool Use · 8LLM Inference & Efficiency · 4AI Policy & Governance · 5Alignment & Safety · 8

Primary evidence

Top Ranked Signals

Research arXiv (Artificial Intelligence) Mar 4

Why Does RLAIF Work At All?

By Robin Young

82 score
AI Analysis

Proposes the 'latent value hypothesis' to explain why RLAIF works: pretraining encodes human values as directions in representation space, and constitutional prompts act as projection operators to elicit these latent values. Formalizes this under a linear model and derives conditions for when RLAIF succeeds or fails.

arXiv:2603.03000v1 Announce Type: cross Abstract: Reinforcement Learning from AI Feedback (RLAIF) enables language models to improve by training on their own preference judgments, yet no theoretical account explains why this self-improvement seemingly works for value learning. We propose the latent value hypothesis, that pretraining on internet-scale data encodes human values as directions in representation space, and constitutional prompts elicit these latent values into preference judgments.
AI AlignmentRLHF/RLAIFLanguage ModelsTheoretical AI
Research arXiv (Computer Vision) Mar 4

Beyond Language Modeling: An Exploration of Multimodal Pretraining

By Shengbang Tong, David Fan, John Nguyen, Ellis Brown, Gaoyue Zhou, Shengyi Qian, Boyang Zheng, Th\'eophane Vallaeys, Junlin Han, Rob Fergus, Naila Murray, Marjan Ghazvininejad, Mike Lewis, Nicolas Ballas, Amir Bar, Michael Rabbat, Jakob Verbeek, Luke Zettlemoyer, Koustuv Sinha, Yann LeCun, Saining Xie

82 score
AI Analysis

This paper from a strong team (including Yann LeCun, Saining Xie, and Meta researchers) provides empirical clarity on native multimodal pretraining design space through controlled from-scratch experiments using the Transfusion framework with next-token prediction for language and diffusion for vision, yielding insights about visual representation, data mixing, and emergent cross-modal capabilities.

arXiv:2603.03276v1 Announce Type: new Abstract: The visual world offers a critical axis for advancing foundation models beyond language. Despite growing interest in this direction, the design space for native multimodal models remains opaque. We provide empirical clarity through controlled, from-scratch pretraining experiments, isolating the factors that govern multimodal pretraining without interference from language pretraining. We adopt the Transfusion framework, using next-token prediction
Multimodal PretrainingFoundation ModelsVision-Language ModelsDiffusion ModelsRepresentation Learning
Research arXiv (Machine Learning) Mar 4

Scaling Reward Modeling without Human Supervision

By Jingxuan Fan, Yueying Li, Zhenting Qi, Dinghuai Zhang, Kiant\'e Brantley, Sham M. Kakade, Hanlin Zhang

78 score
AI Analysis

Explores scaling reward models through unsupervised approaches using preference learning over document prefixes/suffixes from web corpora, without human annotations. Shows that training on 11M tokens of math-focused web data yields consistent gains on RewardBench across multiple backbone models.

arXiv:2603.02225v1 Announce Type: new Abstract: Learning from feedback is an instrumental process for advancing the capabilities and safety of frontier models, yet its effectiveness is often constrained by cost and scalability. We present a pilot study that explores scaling reward models through unsupervised approaches. We operationalize reward-based scaling (RBS), in its simplest form, as preference learning over document prefixes and suffixes drawn from large-scale web corpora. Its advantage
AI AlignmentReward ModelingRLHF/RLAIFLanguage Models
Research arXiv (Artificial Intelligence) Mar 4

ZeroDayBench: Evaluating LLM Agents on Unseen Zero-Day Vulnerabilities for Cyberdefense

By Nancy Lau, Louis Sloot, Jyoutir Raj, Giuseppe Marco Boscardin, Evan Harris, Dylan Bowman, Mario Brajkovski, Jaideep Chawla, Dan Zhao

75 score
AI Analysis

Introduces ZeroDayBench, a benchmark where LLM agents must find and patch 22 novel critical vulnerabilities in open-source codebases. Tests GPT-5.2, Claude Sonnet 4.5, and Grok 4.1, finding frontier LLMs are not yet capable of autonomously solving these tasks.

arXiv:2603.02297v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly being deployed as software engineering agents that autonomously contribute to repositories. A major benefit these agents present is their ability to find and patch security vulnerabilities in the codebases they oversee. To estimate the capability of agents in this domain, we introduce ZeroDayBench, a benchmark where LLM agents find and patch 22 novel critical vulnerabilities in open-source codebases.
AI SafetyCybersecurityLLM AgentsBenchmarksVulnerability Discovery
Research arXiv (Artificial Intelligence) Mar 4

Why Adam Can Beat SGD: Second-Moment Normalization Yields Sharper Tails

By Ruinan Jin, Yingbin Liang, Shaofeng Zou

75 score
AI Analysis

Establishes the first theoretical separation between high-probability convergence of Adam and SGD, showing Adam achieves δ^{-1/2} dependence on confidence parameter while SGD necessarily has δ^{-1} dependence. Attributes this to Adam's second-moment normalization creating sharper tail behavior.

arXiv:2603.03099v1 Announce Type: cross Abstract: Despite Adam demonstrating faster empirical convergence than SGD in many applications, much of the existing theory yields guarantees essentially comparable to those of SGD, leaving the empirical performance gap insufficiently explained. In this paper, we uncover a key second-moment normalization in Adam and develop a stopping-time/martingale analysis that provably distinguishes Adam from SGD under the classical bounded variance model (a second m
Optimization TheoryDeep Learning TheoryAdam vs SGD
Research arXiv (Computation and Language) Mar 4

Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use

By Aradhye Agarwal, Gurdit Siyan, Yash Pandya, Joykirat Singh, Akshay Nambi, Ahmed Awadallah

74 score
AI Analysis

Introduces MOSAIC, a post-training framework for safe multi-step agentic tool use that structures inference as 'plan, check, then act or refuse' with explicit safety reasoning and refusal as first-class learning objectives.

arXiv:2603.03205v1 Announce Type: new Abstract: Agentic language models operate in a fundamentally different safety regime than chat models: they must plan, call tools, and execute long-horizon actions where a single misstep, such as accessing files or entering credentials, can cause irreversible harm. Existing alignment methods, largely optimized for static generation and task completion, break down in these settings due to sequential decision-making, adversarial tool feedback, and overconfide
AI SafetyAlignmentAgentic AITool Use
Research arXiv (Artificial Intelligence) Mar 4

Quantifying Frontier LLM Capabilities for Container Sandbox Escape

By Rahul Marchand, Art O Cathain, Jerome Wynne, Philippos Maximos Giavridis, Sam Deverett, John Wilkinson, Jason Gwartz, Harry Coppock

72 score
AI Analysis

Introduces SANDBOXESCAPEBENCH, a benchmark measuring LLM agents' ability to escape Docker/OCI container sandboxes. Tests GPT-5.2, Claude Sonnet 4.5, and Grok 4.1, finding frontier LLMs cannot yet autonomously escape sandboxes.

arXiv:2603.02277v1 Announce Type: cross Abstract: Large language models (LLMs) increasingly act as autonomous agents, using tools to execute code, read and write files, and access networks, creating novel security risks. To mitigate these risks, agents are commonly deployed and evaluated in isolated "sandbox" environments, often implemented using Docker/OCI containers. We introduce SANDBOXESCAPEBENCH, an open benchmark that safely measures an LLM's capacity to break out of these sandboxes. The
AI SafetyLLM SecurityBenchmarksAgentic AICybersecurity
Research arXiv (Artificial Intelligence) Mar 4

The Geometry of Learning Under AI Delegation

By Lingxiao Huang, Nisheeth K. Vishnoi

72 score
AI Analysis

Mathematically models the joint evolution of human skill and AI delegation as a coupled dynamical system, revealing that adaptive AI use creates stable low-skill equilibria where humans can become persistently dependent on AI ('learned helplessness'). Shows that beyond the high-skill equilibrium of human-only learning, AI delegation fundamentally alters skill acquisition dynamics.

arXiv:2603.02950v1 Announce Type: cross Abstract: As AI systems shift from tools to collaborators, a central question is how the skills of humans relying on them change over time. We study this question mathematically by modeling the joint evolution of human skill and AI delegation as a coupled dynamical system. In our model, delegation adapts to relative performance, while skill improves through use and decays under non-use; crucially, both updates arise from optimizing a single performance me
Human-AI InteractionAI SafetyTheoretical AI
Research arXiv (Machine Learning) Mar 4

Safety Training Persists Through Helpfulness Optimization in LLM Agents

By Benjamin Plaut

72 score
AI Analysis

Studies safety training persistence in agentic (multi-step, tool-use) LLM settings, finding that safety training persists through subsequent helpfulness DPO training, unlike findings in single-step chat settings. All training configurations converge near a linear Pareto frontier.

arXiv:2603.02229v1 Announce Type: new Abstract: Safety post-training has been studied extensively in single-step "chat" settings where safety typically refers to refusing harmful requests. We study an "agentic" (i.e., multi-step, tool-use) setting where safety refers to harmful actions directly taken by the LLM. We compare the effects of running direct preference optimization (DPO) on safety or helpfulness alone vs both metrics sequentially. As expected, training on one metric alone results in
AI SafetyLLM AgentsAlignmentRLHF
Research arXiv (Machine Learning) Mar 4

Speculative Speculative Decoding

By Tanishq Kumar, Tri Dao, Avner May

72 score
AI Analysis

Introduces 'speculative speculative decoding' (SSD) that parallelizes the speculation and verification phases of speculative decoding by pre-emptively preparing speculations for likely verification outcomes, eliminating draft model latency.

arXiv:2603.03251v1 Announce Type: new Abstract: Autoregressive decoding is bottlenecked by its sequential nature. Speculative decoding has become a standard way to accelerate inference by using a fast draft model to predict upcoming tokens from a slower target model, and then verifying them in parallel with a single target model forward pass. However, speculative decoding itself relies on a sequential dependence between speculation and verification. We introduce speculative speculative decoding
LLM InferenceEfficiencyLanguage ModelsSystems
Research arXiv (Computation and Language) Mar 4

BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing?

By Guoxin Chen, Fanzhe Meng, Jiale Zhao, Minghao Li, Daixuan Cheng, Huatong Song, Jie Chen, Yuzhi Lin, Hui Chen, Xin Zhao, Ruihua Song, Chang Liu, Cheng Chen, Kai Jia, Ji-Rong Wen

72 score
AI Analysis

Introduces BeyondSWE, a benchmark with 500 real-world instances that evaluates code agents beyond single-repo bug fixing, covering cross-repository reasoning, domain-specialized problems, dependency migration, and full-repo generation. Frontier models plateau below 45% success.

arXiv:2603.03194v1 Announce Type: new Abstract: Current benchmarks for code agents primarily assess narrow, repository-specific fixes, overlooking critical real-world challenges such as cross-repository reasoning, domain-specialized problem solving, dependency-driven migration, and full-repository generation. To address this gap, we introduce BeyondSWE, a comprehensive benchmark that broadens existing evaluations along two axes - resolution scope and knowledge scope - using 500 real-world insta
Code GenerationBenchmarkingAgentic AILanguage Models
Research arXiv (Machine Learning) Mar 4

Subspace Geometry Governs Catastrophic Forgetting in Low-Rank Adaptation

By Brady Steele

70 score
AI Analysis

Presents a geometric theory characterizing catastrophic forgetting in LoRA through gradient subspace interactions, deriving a simple law: forgetting depends on the minimum principal angle between task gradient subspaces. Reveals an approximate rank-invariance property of forgetting.

arXiv:2603.02224v1 Announce Type: new Abstract: Low-Rank Adaptation (LoRA) has emerged as a parameter-efficient approach for adapting large pre-trained models, yet its behavior under continual learning remains poorly understood. We present a geometric theory characterizing catastrophic forgetting in LoRA through the lens of gradient subspace interactions. Our central finding is that forgetting is governed by a simple geometric law: $\mathcal{F} = \alpha(1 - \cos^2\theta_{\min}) + \beta$, where
Continual LearningLoRADeep Learning TheoryParameter-Efficient Fine-Tuning