Wiredframe Radar logo

Category intelligence

Research Briefing — January 27, 2026

694 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research reveals critical vulnerabilities across the AI ecosystem, from scientific integrity to deployed systems. A forensic audit quantifying 17% phantom citation rates in AI-assisted survey papers exposes systematic epistemic decay in AI-augmented research workflows.

Security and safety research dominates:

Architecture and efficiency advances include NVIDIA's LatentMoE optimizing accuracy per FLOP through hardware-software co-design, and AR-Omni achieving unified any-to-any multimodal generation without expert decoders. Privacy research shows fine-tuned models leak input-only PII through unexpected memorization channels.

Key Themes

AI Security & Privacy Attacks · 10AI Scientific Integrity · 2AI Safety & Alignment · 16Agentic Systems Architecture · 6AI Safety and Alignment · 8AI Safety & Guardrails · 12Agentic AI Systems · 22On-device & Efficient LLMs · 8Reinforcement Learning for Hard Problems · 8Reasoning and Interpretability · 7

Primary evidence

Top Ranked Signals

Research arXiv (Artificial Intelligence) Jan 27

The 17% Gap: Quantifying Epistemic Decay in AI-Assisted Survey Papers

By H. Kemal \.Ilter

92 score
AI Analysis

A forensic audit of 50 AI survey papers (5,514 citations) reveals a consistent 17% 'phantom rate' - citations that cannot be resolved to any existing publication. This quantifies systematic epistemic degradation from AI-assisted scientific writing.

arXiv:2601.17431v1 Announce Type: cross Abstract: The adoption of Large Language Models (LLMs) in scientific writing promises efficiency but risks introducing informational entropy. While "hallucinated papers" are a known artifact, the systematic degradation of valid citation chains remains unquantified. We conducted a forensic audit of 50 recent survey papers in Artificial Intelligence (N=5,514 citations) published between September 2024 and January 2026. We utilized a hybrid verification pipe
AI SafetyScientific IntegrityLLM Hallucination
Research arXiv (Artificial Intelligence) Jan 27

Breaking the Protocol: Security Analysis of the Model Context Protocol Specification and Prompt Injection Vulnerabilities in Tool-Integrated LLM Agents

By Narek Maloyan, Dmitry Namiot

91 score
AI Analysis

First formal security analysis of the Model Context Protocol (MCP) specification, identifying three fundamental vulnerabilities: absent capability attestation, unauthenticated bidirectional sampling enabling prompt injection, and implicit trust propagation in multi-server setups.

arXiv:2601.17549v1 Announce Type: cross Abstract: The Model Context Protocol (MCP) has emerged as a de facto standard for integrating Large Language Models with external tools, yet no formal security analysis of the protocol specification exists. We present the first rigorous security analysis of MCP's architectural design, identifying three fundamental protocol-level vulnerabilities: (1) absence of capability attestation allowing servers to claim arbitrary permissions, (2) bidirectional sampli
AI SecurityAgentic SystemsPrompt InjectionMCP
Research arXiv (Artificial Intelligence) Jan 27

Physical Prompt Injection Attacks on Large Vision-Language Models

By Chen Ling, Kai Hu, Hangcheng Liu, Xingshuo Han, Tianwei Zhang, Changhai Ou

88 score
AI Analysis

Introduces PPIA, the first physical prompt injection attack on vision-language models that embeds malicious instructions into physical objects. The attack is black-box, query-agnostic, and operates solely through visual observation without model access.

arXiv:2601.17383v1 Announce Type: cross Abstract: Large Vision-Language Models (LVLMs) are increasingly deployed in real-world intelligent systems for perception and reasoning in open physical environments. While LVLMs are known to be vulnerable to prompt injection attacks, existing methods either require access to input channels or depend on knowledge of user queries, assumptions that rarely hold in practical deployments. We propose the first Physical Prompt Injection Attack (PPIA), a black-bo
AI SecurityVision-Language ModelsAdversarial AttacksPrompt Injection
Research arXiv (Computation and Language) Jan 27

MortalMATH: Evaluating the Conflict Between Reasoning Objectives and Emergency Contexts

By Etienne Lanzeray, Stephane Meilliez, Malo Ruelle, Damien Sileo

88 score
AI Analysis

Introduces MortalMATH benchmark revealing that reasoning-optimized LLMs exhibit 'tunnel vision' - ignoring life-threatening emergencies (stroke symptoms, freefall) while maintaining 95%+ task completion on math problems. Generalist models like Llama-3.1 appropriately refuse tasks to address danger.

arXiv:2601.18790v1 Announce Type: new Abstract: Large Language Models are increasingly optimized for deep reasoning, prioritizing the correct execution of complex tasks over general conversation. We investigate whether this focus on calculation creates a "tunnel vision" that ignores safety in critical situations. We introduce MortalMATH, a benchmark of 150 scenarios where users request algebra help while describing increasingly life-threatening emergencies (e.g., stroke symptoms, freefall). We
AI SafetyBenchmarksReasoningAlignment
Research arXiv (Artificial Intelligence) Jan 27

Unintended Memorization of Sensitive Information in Fine-Tuned Language Models

By Marton Szep, Jorge Marin Ruiz, Georgios Kaissis, Paulina Seidl, R\"udiger von Eisenhart-Rothe, Florian Hinterwimmer, Daniel Rueckert

86 score
AI Analysis

Systematically investigates PII leakage from fine-tuned LLMs, finding that sensitive information appearing only in model inputs (not training targets) can still be extracted. Benchmarks four privacy-preserving approaches including differential privacy.

arXiv:2601.17480v1 Announce Type: cross Abstract: Fine-tuning Large Language Models (LLMs) on sensitive datasets carries a substantial risk of unintended memorization and leakage of Personally Identifiable Information (PII), which can violate privacy regulations and compromise individual safety. In this work, we systematically investigate a critical and underexplored vulnerability: the exposure of PII that appears only in model inputs, not in training targets. Using both synthetic and real-worl
AI PrivacyLanguage ModelsData SecurityFine-tuning
Research arXiv (Artificial Intelligence) Jan 27

Reconstructing Training Data from Adapter-based Federated Large Language Models

By Silong Chen, Yuchuan Luo, Guilin Deng, Yi Liu, Min Xu, Shaojing Fu, Xiaohua Jia

85 score
AI Analysis

Demonstrates that adapter-based federated LLMs (using LoRA) create new exploitable leakage channels contrary to assumptions. Proposes UTR attack that reconstructs training data from low-rank adapter gradients.

arXiv:2601.17533v1 Announce Type: cross Abstract: Adapter-based Federated Large Language Models (FedLLMs) are widely adopted to reduce the computational, storage, and communication overhead of full-parameter fine-tuning for web-scale applications while preserving user privacy. By freezing the backbone and training only compact low-rank adapters, these methods appear to limit gradient leakage and thwart existing Gradient Inversion Attacks (GIAs). Contrary to this assumption, we show that low-r
Federated LearningAI SecurityPrivacy AttacksLoRA
Research arXiv (Computation and Language) Jan 27

The Shadow Self: Intrinsic Value Misalignment in Large Language Model Agents

By Chen Chen, Kim Young Il, Yuan Yang, Wenhao Su, Yilin Zhang, Xueluan Gong, Qian Wang, Yongsen Zheng, Ziyao Liu, Kwok-Yan Lam

85 score
AI Analysis

Formalizes Loss-of-Control risk and Intrinsic Value Misalignment in LLM agents operating in benign settings. Introduces IMPRESS benchmark for probing value misalignment in realistic scenarios without explicit harmful inputs.

arXiv:2601.17344v1 Announce Type: new Abstract: Large language model (LLM) agents with extended autonomy unlock new capabilities, but also introduce heightened challenges for LLM safety. In particular, an LLM agent may pursue objectives that deviate from human values and ethical norms, a risk known as value misalignment. Existing evaluations primarily focus on responses to explicit harmful input or robustness against system failure, while value misalignment in realistic, fully benign, and agent
AI SafetyAlignmentLLM AgentsValue Alignment
Research arXiv (Artificial Intelligence) Jan 27

AR-Omni: A Unified Autoregressive Model for Any-to-Any Generation

By Dongjie Cheng, Ruifeng Yuan, Yongqi Li, Runyang You, Wenjie Wang, Liqiang Nie, Lei Zhang, Wenjie Li

84 score
AI Analysis

AR-Omni presents a unified autoregressive model for any-to-any multimodal generation (text, vision, speech) without requiring expert decoder modules, using a single token stream and next-token objective.

arXiv:2601.17761v1 Announce Type: cross Abstract: Real-world perception and interaction are inherently multimodal, encompassing not only language but also vision and speech, which motivates the development of "Omni" MLLMs that support both multimodal inputs and multimodal outputs. While a sequence of omni MLLMs has emerged, most existing systems still rely on additional expert components to achieve multimodal generation, limiting the simplicity of unified training and inference. Autoregressive
Multimodal ModelsAutoregressive ModelsUnified Architecture
Research arXiv (Artificial Intelligence) Jan 27

Self-Manager: Parallel Agent Loop for Long-form Deep Research

By Yilong Xu, Zhi Zheng, Xiang Long, Yujun Cai, Yiwei Wang

83 score
AI Analysis

Self-Manager introduces a parallel agent loop for complex research tasks, enabling asynchronous concurrent execution with isolated context windows per subthread, managed via Thread Control Blocks.

arXiv:2601.17879v1 Announce Type: cross Abstract: Long-form deep research requires multi-faceted investigations over extended horizons to get a comprehensive report. When handling such complex tasks, existing agents manage context at the subtask level to overcome linear context accumulation and information loss. However, they still adhere to a single context window and sequential execution paradigm, which results in mutual interference and blocking behavior, restricting scalability and adaptabi
Agentic SystemsAgent ArchitectureParallel Computing
Research arXiv (Artificial Intelligence) Jan 27

A Monosemantic Attribution Framework for Stable Interpretability in Clinical Neuroscience Large Language Models

By Michail Mamalakis, Tiago Azevedo, Cristian Cosentino, Chiara D'Ercoli, Subati Abulikemu, Zhongtian Sun, Richard Bethlehem, Pietro Lio

82 score
AI Analysis

Proposes a unified interpretability framework for clinical LLMs combining attribution and mechanistic interpretability through monosemantic feature extraction. Addresses instability in existing attribution methods.

arXiv:2601.17952v1 Announce Type: cross Abstract: Interpretability remains a key challenge for deploying large language models (LLMs) in clinical settings such as Alzheimer's disease progression diagnosis, where early and trustworthy predictions are essential. Existing attribution methods exhibit high inter-method variability and unstable explanations due to the polysemantic nature of LLM representations, while mechanistic interpretability approaches lack direct alignment with model inputs and
InterpretabilityClinical AIMechanistic InterpretabilityLLM Safety
Research arXiv (Artificial Intelligence) Jan 27

LatentMoE: Toward Optimal Accuracy per FLOP and Parameter in Mixture of Experts

By Venmugil Elango, Nidhi Bhatia, Roger Waleffe, Rasoul Shafipour, Tomer Asida, Abhinav Khattar, Nave Assaf, Maximilian Golub, Joey Guman, Tiyasa Mitra, Ritchie Zhao, Ritika Borkar, Ran Zilberstein, Mostofa Patwary, Mohammad Shoeybi, Bita Rouhani

82 score
AI Analysis

NVIDIA researchers revisit Mixture of Experts design from hardware-software co-design perspective, introducing LatentMoE to optimize accuracy per FLOP and parameter. Characterizes performance bottlenecks across offline and online inference regimes.

arXiv:2601.18089v1 Announce Type: cross Abstract: Mixture of Experts (MoEs) have become a central component of many state-of-the-art open-source and proprietary large language models. Despite their widespread adoption, it remains unclear how close existing MoE architectures are to optimal with respect to inference cost, as measured by accuracy per floating-point operation and per parameter. In this work, we revisit MoE design from a hardware-software co-design perspective, grounded in empirical
Mixture of ExpertsModel EfficiencyHardware-Software Co-designLanguage Models
Research arXiv (Computation and Language) Jan 27

Unknown Unknowns: Why Hidden Intentions in LLMs Evade Detection

By Devansh Srivastav, David Pape, Lea Sch\"onherr

82 score
AI Analysis

Introduces taxonomy of ten categories of hidden intentions in LLMs (goal-directed covert behaviors arising from training or adversarial manipulation). Shows these can be easily induced but evade detection.

arXiv:2601.18552v1 Announce Type: new Abstract: LLMs are increasingly embedded in everyday decision-making, yet their outputs can encode subtle, unintended behaviours that shape user beliefs and actions. We refer to these covert, goal-directed behaviours as hidden intentions, which may arise from training and optimisation artefacts, or be deliberately induced by an adversarial developer, yet remain difficult to detect in practice. We introduce a taxonomy of ten categories of hidden intentions,
AI SafetyHidden BehaviorsLLM SecurityAlignment