Wiredframe Radar logo

Category intelligence

Research Briefing — May 18, 2026

503 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research centers on fundamental limitations of AI oversight, theoretical bounds on self-improvement, and security vulnerabilities in deployed systems.

On the constructive side, AIRA-Compose/Design demonstrates autonomous architecture discovery beyond standard Transformers, while an LLM-guided tree search system matches CDC experts in prospective multi-pathogen forecasting. Theoretical work reframes grokking as Bayesian inference over task graphs, and Judge Circuits applies causal interpretability to the LLM-as-judge paradigm. VLM analysis reveals self-reflective statements like "let me check again" are learned linguistic patterns rather than genuine visual re-examination.

Key Themes

AI Safety & Alignment · 11AI Safety & Security · 12Mechanistic Interpretability · 3Language Models & Efficiency · 8Reinforcement Learning & World Models · 9AI Safety, Security & Governance · 8LLM Agents & Memory · 15Efficient ML Training · 5Vision-Language Models · 21LLM Agents and Autonomous Systems · 9

Primary evidence

Top Ranked Signals

Research arXiv (Machine Learning) May 18

Training on Documents About Monitoring Leads to CoT Obfuscation

By Reilly Haskins, Bilal Chughtai and Joshua Engels

75 score
AI Analysis

Shows that exposing models to documents describing CoT monitoring leads to obfuscation of reasoning traces, with monitor-aware models achieving higher rates of undetected misbehavior.

Chain-of-thought (CoT) monitoring is one of the most promising tools we have for detecting model misbehavior, but its effectiveness depends on models faithfully externalizing their reasoning. Motivated by this vulnerability, we study whether monitor-aware models are capable of obfuscating their reasoning to evade detection. We use synthetic document finetuning to expose eight models to realistic pre-training-style documents describing a CoT monitor and find that monitor-aware models consistently
AI SafetyAlignmentChain-of-ThoughtMonitoring
Research arXiv (Artificial Intelligence) May 18

NOVA: Fundamental Limits of Knowledge Discovery Through AI

By Salman Avestimehr, Ken Duffy, and Muriel M\'edard

72 score
AI Analysis

Introduces NOVA framework modeling AI self-improvement as adaptive sampling over knowledge space, proving coverage conditions and identifying failure modes including a 'contamination trap' where false positives overwhelm genuine discoveries.

Can AI systems discover genuinely new knowledge through iterative self improvement, and if so, at what cost? We introduce the NOVA framework, which models the common ``generate, verify, accumulate, retrain'' loop as an adaptive sampling process over a knowledge space. We identify sufficient conditions under which accumulated genuine knowledge eventually covers a finite domain, and show how their violations produce distinct failure modes: contamination, forgetting, exploration failure, and accept
AI TheorySelf-ImprovementKnowledge DiscoveryAI Safety
Research arXiv (cs.CR) May 18

Hidden in Memory: Sleeper Memory Poisoning in LLM Agents

By Sidharth Pulipaka, Stanislau Hlebik, Leonidas Raghav, Sahar Abdelnabi, Vyas Raina, Ivaxi Sheth, Mario Fritz

72 score
AI Analysis

Proposes and studies sleeper memory poisoning, where adversaries manipulate external context to cause LLM assistants to store fabricated memories that re-emerge across conversations to steal user actions or data.

Large language models are increasingly augmented with persistent memory, allowing assistants to store user-specific information across sessions for personalization and continuity. This statefulness introduces a new security risk: adversarial content can corrupt what an assistant remembers and thereby influence future interactions. We propose and study sleeper memory poisoning, a delayed attack in which an adversary manipulates external context, such as a document, webpage, or repository, to caus
AI SafetyLLM SecurityAdversarial AttacksMemory Systems
Research arXiv (Computation and Language) May 18

RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably

By Yufeng Du, Phillip Harris, Minyang Tian, Eliu A Huerta, Srikanth Ronanki, Subendhu Rongali, Aram Galstyan, Hao Peng

72 score
AI Analysis

Proves fundamental limitations of RoPE in long-context models: as context grows, attention loses locality bias and consistency in token relevance, with failure probability approaching 0.5 (random guessing). Provides theoretical explanation for long-context degradation.

We identify intrinsic limitations of Rotary Positional Embeddings (RoPE) in Transformer-based long-context language models. Our theoretical analysis abstracts away from the specific content of the context and depends only on its length. We prove that as context length increases, RoPE-based attention becomes unpredictable and loses two properties that are central to its effectiveness. First, it loses its locality bias: RoPE is no more likely to favor nearer positions than substantially farther on
Language ModelsPositional EncodingsLong ContextTheoretical ML
Research arXiv (Machine Learning) May 18

Grokking as Structural Inference: Transformers Need Bayesian Lottery Tickets

By Kai Hidajat, Solden Stoll, Joseph An

72 score
AI Analysis

Proposes a theoretical framework explaining grokking in Transformers through the lens of Bayesian inference over task dependency graphs. Proves generalization requires both bounded MLP capacity and a structural condition on attention placing sufficient mass on informative tokens.

Why does a Transformer that has memorized its training set wait thousands of steps before it generalizes? Existing accounts locate this delay in norm minimization, feature emergence, or the late discovery of sparse subnetworks. These explanations capture important parts of the transition, but ignore a constraint unique to attention-based models: if attention discards an informative token, no bounded downstream computation can recover it. We formalize attention as an implicit Bayesian posterior o
Deep Learning TheoryTransformersGrokkingMechanistic Interpretability
Research arXiv (Artificial Intelligence) May 18

Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design

By Alberto Pepe, Chien-Yu Lin, Despoina Magka, Bilge Acun, Yannan Nellie Wu, Anton Protopopov, Carole-Jean Wu, Yoram Bachrach

72 score
AI Analysis

Introduces AIRA-Compose and AIRA-Design, dual frameworks where LLM agents autonomously discover neural architectures beyond standard Transformers. Using 11 agents under 24-hour budget, discovers 14 architectures that outperform Llama 3.2 when pre-trained at 1B scale.

Toward recursive self-improvement, we investigate LLM agents autonomously designing foundation models beyond standard Transformers. We introduce a dual-framework approach: AIRA-Compose for high-level architecture search, and AIRA-Design for low-level mechanistic implementation. AIRA-Compose uses 11 agents to explore fundamental computational primitives under a 24-hour budget. Agents evaluate million-parameter candidates, extrapolating top designs to 350M, 1B, and 3B scales. This yields 14 archit
Neural Architecture SearchAI AgentsSelf-ImprovementLanguage Models
Research arXiv (Artificial Intelligence) May 18

Imperfect World Models are Exploitable

By Logan Mondal Bhamidipaty (University of Edinburgh), Esmeralda S. Whitammer (University of Edinburgh), David Abel (University of Edinburgh), Mykel J. Kochenderfer (Stanford University), Subramanian Ramamoorthy (University of Edinburgh)

72 score
AI Analysis

Proposes a formal definition of model exploitation in RL where imperfect world models lead to incorrect policy preferences. Proves that exploitation is essentially unavoidable for large policy sets and develops a general theory unifying reward hacking and model exploitation. Notable for theoretical contributions from strong authors (David Abel, Mykel Kochenderfer).

We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred over another while the environment's true transition model implies the reverse. We analogize our definition with a prior characterization of reward hacking but show that the associated proof of inevitability does not transfer to exploitation. To overcome this obstruction, we develop a general theory of reward hacking an
Reinforcement LearningAI SafetyWorld ModelsReward Hacking
Research arXiv (Artificial Intelligence) May 18

Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search

By Sarah Martinson, Michael P. Brenner, Martyna Plomecka, Brian P. Williams, Nicholas G. Reich, Zahra Shamsi

72 score
AI Analysis

Presents an autonomous system using LLM-guided tree search to iteratively generate, evaluate, and optimize infectious disease forecasting software. In prospective real-time evaluation during 2025-2026 US respiratory season, the system matched or outperformed the human-curated CDC FluSight ensemble.

Probabilistic forecasting of infectious diseases is crucial for public health but relies on labor-intensive manual model curation by expert modeling teams. This bespoke development bottlenecks scalability to granular geographic resolutions or emerging pathogens. Here, we present an autonomous system using Large Language Model (LLM)-guided tree search to iteratively generate, evaluate, and optimize executable forecasting software. In a fully prospective, real-time evaluation during the 2025-2026
AI for ScienceLLM AgentsEpidemiologyAutomated MLPublic Health
Research arXiv (Computer Vision) May 18

Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination

By Chufan Shi, Cheng Yang, Yaokang Wu, Linhao Jin, Bo Shui, Taylor Berg-Kirkpatrick, Xuezhe Ma

70 score
AI Analysis

Reveals that VLMs' self-reflective statements like 'let me check the figure again' do not trigger genuine visual re-examination. Uses image-swap probing to show thinking models are nearly 3x more vulnerable than instructed counterparts, with accuracy dropping up to 60%.

Vision-Language Models (VLMs) often produce self-reflective statements like "let me check the figure again" during reasoning. Do such statements trigger genuine visual re-examination, or are they merely learned textual patterns? We investigate this via VisualSwap, an image-swap probing framework: after a model reasons over an image, we replace it with a visually similar but semantically different one and test whether the model notices. We introduce VS-Bench, 800 image pairs curated from MathVist
Vision-Language ModelsAI SafetyReasoningInterpretability
Research arXiv (Computation and Language) May 18

Judge Circuits

By Nils Feldhus, Tanja Baeumel, Elena Golimblevskaia, Qianli Wang, Van Bach Nguyen, Aaron Louis Eidt, Christopher Ebert, Wojciech Samek, Jing Yang, Vera Schmitt, Sebastian M\"oller, Simon Ostermann

70 score
AI Analysis

Uses causal mechanistic interpretability (Position-aware Edge Attribution Patching) to investigate LLM-as-a-judge internals across Gemma-3, Qwen2.5, and Llama-3. Discovers a sparse 'Latent Evaluator' sub-graph in mid-to-late MLPs responsible for judgments, separable from output formatting.

LLM-as-a-judge has become the dominant paradigm for grading model outputs at scale, yet the same model assigns systematically different scores when its output format changes (e.g., a 1-5 rating vs. a True/False label). Existing diagnoses of these format-induced inconsistencies stop at the input-output level. Using Position-aware Edge Attribution Patching (PEAP), we causally investigate the internal mechanism in Gemma-3, Qwen2.5, and Llama-3. We find that judgments across structured understanding
Mechanistic InterpretabilityLLM-as-a-JudgeAI Evaluation
Research arXiv (Artificial Intelligence) May 18

Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions

By Jagdish Tripathy, Marcus Buckmann

68 score
AI Analysis

Reveals that instruction-tuned LLMs retain and amplify racial bias in internal representations despite showing no output-level bias in mortgage underwriting, and demonstrates this suppressed information is causally potent through activation steering.

Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconne
AI SafetyBiasAlignmentInterpretability
Research arXiv (Machine Learning) May 18

DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts

By Jiading Gai, Shuai Zhang, Xiang Song, Bernie Wang, George Karypis

68 score
AI Analysis

Proposes DualKV, the first FlashAttention kernel variant that exploits prompt invariance in RL training to process shared prompts once rather than N times, dramatically reducing compute for large-rollout RL post-training.

Modern RL post-training methods such as GRPO and DAPO train on $N$ response sequences of $R$ tokens sampled from a shared prompt of $P$ tokens, but standard FlashAttention replicates all $P$ prompt tokens $N$ times across both forward and backward passes -- duplicating compute and memory on identical hidden states. In large-rollout, long-context RL training ($N{\geq}16$, $P{\geq}8\text{K}$), this redundancy dominates the policy update cost. We observe that in decoder-only models, causal masking
Efficient TrainingReinforcement LearningAttention MechanismsLLM Post-training