Category intelligence

Research Briefing — March 3, 2026

1061 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research centers on AI safety mechanisms, fundamental theoretical advances, and practical efficiency gains for frontier model deployment.

A striking safety result shows GPT-5, Claude-4.5, and Qwen-3 can take consequential actions at low probabilities while remaining well-calibrated, posing tail-risk concerns at scale
Qwen3-Coder-Next delivers an 80B/3B-active MoE coding agent achieving state-of-the-art results via large-scale verifiable task synthesis
Mechanistic probing reveals instruction-tuned LLMs often determine answers before generating chain-of-thought, challenging assumptions about CoT faithfulness
Systematic evaluation finds current LLM agents show minimal scheming propensity despite having capability, providing a cautious baseline for safety monitoring

On the theory side, Adam is proven to converge without modification given problem-dependent hyperparameters, resolving a long-standing concern. Recursive self-invocation models prove any computable problem admits decomposition for bounded-context LLMs. GRPO—the technique behind DeepSeek-R1—is formally characterized as a U-statistic with finite-sample error bounds.

Thought Virus demonstrates viral propagation of subliminal biases across multi-agent networks from a single compromised node
CoDD breaks the factorization barrier in diffusion language models via coupled denoising with autoregressive backbone coordination
Attn-QAT enables end-to-end FP4 inference including attention, achieving stable 4-bit quantization-aware training for next-generation GPU deployment

Key Themes

AI Safety & Alignment · 49AI Agents & Agentic Systems · 35AI Safety & Security · 12Emergent Multi-Agent Risks · 1Reasoning & Chain-of-Thought · 14Sequence Model Theory · 2AI Safety (Technical) · 5Optimization Theory & Training Dynamics · 7Language Models · 18LLM Training & Optimization · 8

Primary evidence

Top Ranked Signals

Research arXiv (Machine Learning) Mar 3

Frontier Models Can Take Actions at Low Probabilities

By Alex Serrano, Wen Xing, David Lindner, Erik Jenner

82 score

AI Analysis

Tests whether frontier models (GPT-5, Claude-4.5, Qwen-3) can take actions at very low probabilities while maintaining calibration, finding they are surprisingly well-calibrated even at rates like 0.01%. This has implications for evasion of safety evaluations.

arXiv:2603.02202v1 Announce Type: new Abstract: Pre-deployment evaluations inspect only a limited sample of model actions. A malicious model seeking to evade oversight could exploit this by randomizing when to "defect": misbehaving so rarely that no malicious actions are observed during evaluation, but often enough that they occur eventually in deployment. But this requires taking actions at very low rates, while maintaining calibration. Are frontier models even capable of that? We prompt the G

AI SafetyEvaluationLanguage ModelsAlignment

Research arXiv (Computation and Language) Mar 3

Qwen3-Coder-Next Technical Report

By Ruisheng Cao, Mouxiang Chen, Jiawei Chen, Zeyu Cui, Yunlong Feng, Binyuan Hui, Yuheng Jing, Kaixin Li, Mingze Li, Junyang Lin, Zeyao Ma, Kashun Shum, Xuwu Wang, Jinxi Wei, Jiaxi Yang, Jiajun Zhang, Lei Zhang, Zongmeng Zhang, Wenting Zhao, Fan Zhou

82 score

AI Analysis

Presents Qwen3-Coder-Next, an 80B parameter (3B active) open-weight coding agent model trained via large-scale synthesis of verifiable coding tasks with environment feedback. Achieves competitive performance on SWE-Bench and Terminal-Bench.

arXiv:2603.00729v1 Announce Type: new Abstract: We present Qwen3-Coder-Next, an open-weight language model specialized for coding agents. Qwen3-Coder-Next is an 80-billion-parameter model that activates only 3 billion parameters during inference, enabling strong coding capability with efficient inference. In this work, we explore how far strong training recipes can push the capability limits of models with small parameter footprints. To achieve this, we perform agentic training through large-sc

Code GenerationLanguage ModelsMixture of ExpertsAgents

Research arXiv (Artificial Intelligence) Mar 3

Evaluating and Understanding Scheming Propensity in LLM Agents

By Mia Hopman, Jannes Elstner, Maria Avramidou, Amritanshu Prasad, David Lindner

78 score

AI Analysis

Systematically evaluates LLM agents' propensity to scheme (covertly pursue misaligned goals) by decomposing scheming incentives into agent and environmental factors. Finds only minimal instances of scheming despite high environmental pressure, suggesting current models have low scheming propensity in realistic scenarios.

arXiv:2603.01608v1 Announce Type: new Abstract: As frontier language models are increasingly deployed as autonomous agents pursuing complex, long-term objectives, there is increased risk of scheming: agents covertly pursuing misaligned goals. Prior work has focused on showing agents are capable of scheming, but their propensity to scheme in realistic scenarios remains underexplored. To understand when agents scheme, we decompose scheming incentives into agent factors and environmental factors.

AI SafetyAlignmentLLM AgentsScheming

Research arXiv (Machine Learning) Mar 3

Adam Converges Without Any Modification On Update Rules

By Yushun Zhang, Bingran Li, Congliang Chen, Zhi-Quan Luo, Ruoyu Sun

75 score

AI Analysis

Proves that Adam converges with proper problem-dependent hyperparameters, resolving concerns raised by the classic divergence example. Shows convergence when β₂ is large and β₁ < √β₂, and identifies divergence regions for small β₂.

arXiv:2603.02092v1 Announce Type: new Abstract: Adam is the default algorithm for training neural networks, including large language models (LLMs). However, \citet{reddi2019convergence} provided an example that Adam diverges, raising concerns for its deployment in AI model training. We identify a key mismatch between the divergence example and practice: \citet{reddi2019convergence} pick the problem after picking the hyperparameters of Adam, i.e., $(\beta_1,\beta_2)$; while practical application

Optimization TheoryAdam OptimizerConvergence Analysis

Research arXiv (Machine Learning) Mar 3

Recursive Models for Long-Horizon Reasoning

By Chenxiao Yang, Nathan Srebro, Zhiyuan Li

74 score

AI Analysis

Proposes recursive models where an LLM can recursively invoke itself to solve subtasks in isolated contexts, proving any computable problem admits decomposition requiring exponentially smaller context per subtask.

arXiv:2603.02112v1 Announce Type: new Abstract: Modern language models reason within bounded context, an inherent constraint that poses a fundamental barrier to long-horizon reasoning. We identify recursion as a core principle for overcoming this barrier, and propose recursive models as a minimal realization, where the model can recursively invoke itself to solve subtasks in isolated contexts. We prove that any computable problem admits a recursive decomposition in which each subtask requires o

Language ModelsReasoningComputational TheoryLLM Agents

Research arXiv (Artificial Intelligence) Mar 3

Decoding Answers Before Chain-of-Thought: Evidence from Pre-CoT Probes and Activation Steering

By Kyle Cox, Darius Kianersi, Adri\`a Garriga-Alonso

72 score

AI Analysis

Provides mechanistic evidence that instruction-tuned LLMs often determine their answer before generating chain-of-thought, using linear probes on residual stream activations with 0.9 AUC. Shows these directions are causally active via activation steering.

arXiv:2603.01437v1 Announce Type: new Abstract: As chain-of-thought (CoT) has become central to scaling reasoning capabilities in large language models (LLMs), it has also emerged as a promising tool for interpretability, suggesting the opportunity to understand model decisions through verbalized reasoning. However, the utility of CoT toward interpretability depends upon its faithfulness -- whether the model's stated reasoning reflects the underlying decision process. We provide mechanistic evi

InterpretabilityReasoningAI SafetyMechanistic Interpretability

Research arXiv (Artificial Intelligence) Mar 3

Breaking the Factorization Barrier in Diffusion Language Models

By Ian Li, Zilei Shao, Benjie Wang, Rose Yu, Guy Van den Broeck, Anji Liu

72 score

AI Analysis

Proposes Coupled Discrete Diffusion (CoDD), a hybrid framework that breaks the factorization barrier in diffusion language models by allowing simultaneous token predictions to be jointly dependent without requiring prohibitively large output parameterization.

arXiv:2603.00045v1 Announce Type: cross Abstract: Diffusion language models theoretically allow for efficient parallel generation but are practically hindered by the "factorization barrier": the assumption that simultaneously predicted tokens are independent. This limitation forces a trade-off: models must either sacrifice speed by resolving dependencies sequentially or suffer from incoherence due to factorization. We argue that this barrier arises not from limited backbone expressivity, but fr

Diffusion ModelsLanguage ModelsParallel GenerationNovel Architectures

Research arXiv (Artificial Intelligence) Mar 3

Thought Virus: Viral Misalignment via Subliminal Prompting in Multi-Agent Systems

By Moritz Weckbecker, Jonas M\"uller, Ben Hagag, Michael Mulet

72 score

AI Analysis

Demonstrates that subliminal prompting (biasing LLMs via semantically unrelated tokens) can propagate through multi-agent networks - a single subliminally prompted agent spreads persistent bias throughout its entire network. Tests across 6 agents and two topologies.

arXiv:2603.00131v1 Announce Type: cross Abstract: Subliminal prompting is a phenomenon in which language models are biased towards certain concepts or traits through prompting with semantically unrelated tokens. While prior work has examined subliminal prompting in user-LLM interactions, potential bias transfer in multi-agent systems and its associated security implications remain unexplored. In this work, we show that a single subliminally prompted agent can spread a weakening but persisting b

AI SafetyMulti-Agent SystemsAdversarial AttacksAlignment

Research arXiv (Artificial Intelligence) Mar 3

What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models

By Yingqi Fan, Junlong Tong, Anhao Zhao, Xiaoyu Shen

72 score

AI Analysis

Introduces EmbedLens to analyze how visual tokens are structured in multimodal LLMs, revealing that ~40% of visual tokens are semantically dead/sink tokens while only ~60% carry image-specific meaning. Demonstrates significant redundancy.

arXiv:2603.00510v1 Announce Type: cross Abstract: Multimodal large language models (MLLMs) project visual tokens into the embedding space of language models, yet the internal structuring and processing of visual semantics remain poorly understood. In this work, we introduce a two-fold analytical framework featuring a novel probing tool, $\textbf{EmbedLens}$, to conduct a fine-grained analysis. We uncover a pronounced semantic sparsity at the input level: visual tokens consistently partition int

Multimodal ModelsInterpretabilityEfficiency

Research arXiv (Machine Learning) Mar 3

DUEL: Exact Likelihood for Masked Diffusion via Deterministic Unmasking

By Gilad Turok, Chris De Sa, Volodymyr Kuleshov

72 score

AI Analysis

DUEL introduces a framework for exact likelihood computation in masked diffusion models (MDMs) by formalizing deterministic position selection during unmasking. This gives MDMs proper perplexity evaluation for the first time, addressing a key gap where prior ELBO-based measures were loose bounds not aligned with test-time generation.

arXiv:2603.01367v1 Announce Type: new Abstract: Masked diffusion models (MDMs) generate text by iteratively selecting positions to unmask and then predicting tokens at those positions. Yet MDMs lack proper perplexity evaluation: the ELBO is a loose bound on likelihood under the training distribution, not the test-time distribution, while generative perplexity requires a biased external model and ignores diversity. To address this, we introduce the \textsc{DUEL} framework, which formalizes \emph

Diffusion ModelsLanguage ModelsEvaluation Methods

Research arXiv (Artificial Intelligence) Mar 3

Attn-QAT: 4-Bit Attention With Quantization-Aware Training

By Peiyuan Zhang, Matthew Noto, Wenxuan Tan, Chengquan Jiang, Will Lin, Wei Zhou, Hao Zhang

70 score

AI Analysis

Presents the first systematic study of 4-bit quantization-aware training for attention mechanisms, identifying key principles for stable FP4 attention and proposing Attn-QAT to enable end-to-end FP4 computation on emerging hardware.

arXiv:2603.00040v1 Announce Type: cross Abstract: Achieving reliable 4-bit attention is a prerequisite for end-to-end FP4 computation on emerging FP4-capable GPUs, yet attention remains the main obstacle due to FP4's tiny dynamic range and attention's heavy-tailed activations. This paper presents the first systematic study of 4-bit quantization-aware training (QAT) for attention. We find that "drop-in" QAT, which naively combines an FP4 forward pass with a high-precision Flash Attention (FA)-st

QuantizationEfficiencyTraining MethodsHardware-Aware ML

Research arXiv (Artificial Intelligence) Mar 3

Reverse CAPTCHA: Evaluating LLM Susceptibility to Invisible Unicode Instruction Injection

By Marcus Graves

70 score

AI Analysis

Reverse CAPTCHA evaluates LLM susceptibility to invisible Unicode-encoded instructions, finding that tool use dramatically amplifies compliance with hidden instructions (Cohen's h up to 1.37) and models show provider-specific encoding preferences.

arXiv:2603.00164v1 Announce Type: cross Abstract: We introduce Reverse CAPTCHA, an evaluation framework that tests whether large language models follow invisible Unicode-encoded instructions embedded in otherwise normal-looking text. Unlike traditional CAPTCHAs that distinguish humans from machines, our benchmark exploits a capability gap: models can perceive Unicode control characters that are invisible to human readers. We evaluate five models from two providers across two encoding schemes (z

AI SafetyLLM SecurityPrompt InjectionAdversarial Attacks