Wiredframe Radar logo

Category intelligence

Research Briefing — February 23, 2026

316 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research centers on foundational AI evaluation frameworks, alignment theory, and theoretical insights into core architectures.

In interdisciplinary and applied work, FlyGM demonstrates whole-body locomotion control using an architecture identical to the adult fruit fly's complete connectome. Turn amplification identifies a new conversational LLM failure mode with mechanistic root causes. Clever Materials reveals that ML models for materials discovery exploit bibliographic confounds rather than learning true chemistry.

Key Themes

Learning Theory & Scaling Laws · 8AI Safety and Evaluation · 5AI Safety & Alignment · 18Language Models & LLM Safety · 10LLM Reasoning and Decoding · 4Language Models & NLP · 11Diffusion Model Theory and Practice · 6Reinforcement Learning · 12Efficient Architectures and Inference · 7In-Context Learning Theory · 5

Primary evidence

Top Ranked Signals

Research arXiv (Machine Learning) Feb 23

Scaling Laws and Spectra of Shallow Neural Networks in the Feature Learning Regime

By Leonardo Defilippis, Yizhou Xu, Julius Girardin, Emanuele Troiani, Vittorio Erba, Lenka Zdeborov\'a, Bruno Loureiro and Florent Krzakala

72 score
AI Analysis

Provides a systematic theoretical analysis of neural scaling laws for shallow networks (quadratic and diagonal) in the feature learning regime, deriving phase diagrams for scaling exponents and connecting them to spectral properties of trained weights. Bridges theory and empirical observations of scaling behavior in deep learning.

Neural scaling laws underlie many of the recent advances in deep learning, yet their theoretical understanding remains largely confined to linear models. In this work, we present a systematic analysis of scaling laws for quadratic and diagonal neural networks in the feature learning regime. Leveraging connections with matrix compressed sensing and LASSO, we derive a detailed phase diagram for the scaling exponents of the excess risk as a function of sample complexity and weight decay. This analy
Scaling LawsLearning TheoryFeature Learning
Research arXiv (physics.chem-ph) Feb 23

Clever Materials: When Models Identify Good Materials for the Wrong Reasons

By Kevin Maik Jablonka

72 score
AI Analysis

Demonstrates that ML models for materials discovery can exploit bibliographic confounds (author, journal, publication year) rather than learning chemistry. Across five tasks spanning MOFs, perovskites, batteries, and emitters, models trained solely on predicted metadata 'bibliographic fingerprints' sometimes match conventional descriptor-based predictors.

Machine learning can accelerate materials discovery. Models perform impressively on many benchmarks. However, strong benchmark performance does not imply that a model learned chemistry. I test a concrete alternative hypothesis: that property prediction can be driven by bibliographic confounding. Across five tasks spanning MOFs (thermal and solvent stability), perovskite solar cells (efficiency), batteries (capacity), and TADF emitters (emission wavelength), models trained on standard chemical de
ML for ScienceMaterials DiscoveryBenchmark ValidityDataset Bias
Research arXiv (Machine Learning) Feb 23

Capabilities Ain't All You Need: Measuring Propensities in AI

By Daniel Romero-Alvarado, Fernando Mart\'inez-Plumed, Lorenzo Pacchiardi, Hugo Save, Siddhesh Milind Pawar, Behzad Mehrbakhsh, Pablo Antonio Moreno Casares, Ben Slater, Paolo Bova, Peter Romero, Zachary R. Tyler, Jonathan Prunty, Luning Sun, Jose Hernandez-Orallo

72 score
AI Analysis

Introduces the first formal framework for measuring AI propensities (behavioral tendencies) as distinct from capabilities, using a bilogistic IRT formulation where both excess and deficiency of a propensity can be problematic. Estimates ideal bands for propensities and connects to safety outcomes.

AI evaluation has primarily focused on measuring capabilities, with formal approaches inspired from Item Response Theory (IRT) being increasingly applied. Yet propensities - the tendencies of models to exhibit particular behaviours - play a central role in determining both performance and safety outcomes. However, traditional IRT describes a model's success on a task as a monotonic function of model capabilities and task demands, an approach unsuited to propensities, where both excess and defici
AI SafetyAI EvaluationAlignment
Research arXiv (Artificial Intelligence) Feb 23

Epistemic Traps: Rational Misalignment Driven by Model Misspecification

By Xingcheng Xu, Jingjing Qu, Qiaosheng Zhang, Chaochao Lu, Yanqing Yang, Na Zou, Xia Hu

70 score
AI Analysis

Proposes 'epistemic traps' framework showing that LLM pathologies (sycophancy, hallucination, deception) are mathematically rationalizable behaviors arising from model misspecification, adapting Berk-Nash Rationalizability from economics to explain why these behaviors are stable equilibria resistant to RL mitigation.

The rapid deployment of Large Language Models and AI agents across critical societal and technical domains is hindered by persistent behavioral pathologies including sycophancy, hallucination, and strategic deception that resist mitigation via reinforcement learning. Current safety paradigms treat these failures as transient training artifacts, lacking a unified theoretical framework to explain their emergence and stability. Here we show that these misalignments are not errors, but mathematicall
AI SafetyAlignmentLLM BehaviorGame TheoryHallucination
Research arXiv (Machine Learning) Feb 23

Statistical Advantage of Softmax Attention: Insights from Single-Location Regression

By O. Duranthon, P. Marion, C. Boyer, B. Loureiro, L. Zdeborov\'a

68 score
AI Analysis

Studies the statistical advantage of softmax attention over linear attention through a single-location regression task, using statistical physics to show softmax achieves Bayes risk while linear attention fundamentally falls short.

Large language models rely on attention mechanisms with a softmax activation. Yet the dominance of softmax over alternatives (e.g., component-wise or linear) remains poorly understood, and many theoretical works have focused on the easier-to-analyze linearized attention. In this work, we address this gap through a principled study of the single-location regression task, where the output depends on a linear transformation of a single input token at a random location. Building on ideas from statis
Attention MechanismsTransformersStatistical PhysicsLearning Theory
Research arXiv (Machine Learning) Feb 23

Asking Forever: Universal Activations Behind Turn Amplification in Conversational LLMs

By Zachary Coalson, Bo Fang, Sanghyun Hong

68 score
AI Analysis

Discovers 'turn amplification' - a new failure mode where conversational LLMs systematically prolong multi-turn interactions without completing tasks. Identifies a universal activation subspace associated with clarification-seeking behavior that can be exploited to inflate operational costs.

Multi-turn interaction length is a dominant factor in the operational costs of conversational LLMs. In this work, we present a new failure mode in conversational LLMs: turn amplification, in which a model consistently prolongs multi-turn interactions without completing the underlying task. We show that an adversary can systematically exploit clarification-seeking behavior$-$commonly encouraged in multi-turn conversation settings$-$to scalably prolong interactions. Moving beyond prompt-level beha
LLM SafetyAdversarial MLMechanistic InterpretabilityLanguage Models
Research arXiv (Machine Learning) Feb 23

Whole-Brain Connectomic Graph Model Enables Whole-Body Locomotion Control in Fruit Fly

By Zehao Jin, Yaoye Zhu, Chen Zhang, Yanan Sui

68 score
AI Analysis

Develops FlyGM, a graph model whose architecture is identical to the complete connectome of an adult fruit fly brain, for whole-body locomotion control in a biomechanical simulation. Achieves stable control across walking, turning, and climbing.

Whole-brain biological neural networks naturally support the learning and control of whole-body movements. However, the use of brain connectomes as neural network controllers in embodied reinforcement learning remains unexplored. We investigate using the exact neural architecture of an adult fruit fly's brain for the control of its body movement. We develop Fly-connectomic Graph Model (FlyGM), whose static structure is identical to the complete connectome of an adult Drosophila for whole-body lo
Neuroscience-Inspired AIEmbodied RLGraph Neural NetworksRobotics
Research arXiv (Machine Learning) Feb 23

Gradient Regularization Prevents Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards

By Johannes Ackermann, Michael Noukhovitch, Takashi Ishida, Masashi Sugiyama

68 score
AI Analysis

Proposes gradient regularization to prevent reward hacking in RLHF/RLVR by biasing policy updates toward regions where the reward is more accurate (flatter optima). Provides theoretical connection between reward accuracy and loss flatness.

Reinforcement Learning from Human Feedback (RLHF) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works address this by limiting the policy update with a Kullback-Leibler (KL) penalty towards a reference model. We propose a different framing: Train the LM in a way that biases policy updates towards regions
RLHFReward HackingAI AlignmentLanguage Models
Research arXiv (Machine Learning) Feb 23

Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory

By Usman Anwar, Tim Bakker, Dana Kianfar, Cristina Pinneri, Christos Louizos

68 score
AI Analysis

Uses information theory to analyze Chain-of-Thought monitorability, showing mutual information between CoT and output is necessary but not sufficient. Identifies information gap and elicitation error as sources of monitor failure, and proposes improvements.

Chain-of-thought (CoT) monitors are LLM-based systems that analyze reasoning traces to detect when outputs may exhibit attributes of interest, such as test-hacking behavior during code generation. In this paper, we use information-theoretic analysis to show that non-zero mutual information between CoT and output is a necessary but not sufficient condition for CoT monitorability. We identify two sources of approximation error that may undermine the performance of CoT monitors in practice: informa
AI SafetyChain-of-ThoughtMonitoringInformation Theory
Research arXiv (Machine Learning) Feb 23

The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning

By Mojtaba Sahraee-Ardakan, Mauricio Delbracio, Peyman Milanfar

68 score
AI Analysis

Provides theoretical analysis of why diffusion models don't need noise conditioning, introducing the concept of Marginal Energy that resolves the paradox of noise-agnostic models remaining stable near data manifolds where gradients typically diverge.

Autonomous (noise-agnostic) generative models, such as Equilibrium Matching and blind diffusion, challenge the standard paradigm by learning a single, time-invariant vector field that operates without explicit noise-level conditioning. While recent work suggests that high-dimensional concentration allows these models to implicitly estimate noise levels from corrupted observations, a fundamental paradox remains: what is the underlying landscape being optimized when the noise level is treated as a
Diffusion ModelsGenerative Model TheoryMathematics
Research arXiv (Machine Learning (Statistics)) Feb 23

Topological Exploration of High-Dimensional Empirical Risk Landscapes: general approach, and applications to phase retrieval

By Antoine Maillard, Tony Bonnaire, Giulio Biroli

65 score
AI Analysis

Uses the Kac-Rice formula to analyze the topological landscape of empirical risk minimization for high-dimensional Gaussian single-index models, studying critical point complexity and providing new variational characterizations. Applied to phase retrieval as a case study.

We consider the landscape of empirical risk minimization for high-dimensional Gaussian single-index models (generalized linear models). The objective is to recover an unknown signal $\boldsymbol{\theta}^\star \in \mathbb{R}^d$ (where $d \gg 1$) from a loss function $\hat{R}(\boldsymbol{\theta})$ that depends on pairs of labels $(\mathbf{x}_i \cdot \boldsymbol{\theta}, \mathbf{x}_i \cdot \boldsymbol{\theta}^\star)_{i=1}^n$, with $\mathbf{x}_i \sim \mathcal{N}(0, I_d)$, in the proportional asympto
Optimization LandscapeStatistical PhysicsHigh-Dimensional Statistics
Research arXiv (cs.CY) Feb 23

The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems

By Leon Staufer, Kevin Feng, Kevin Wei, Luke Bailey, Yawen Duan, Mick Yang, A. Pinar Ozisik, Stephen Casper, Noam Kolt

65 score
AI Analysis

Presents the 2025 AI Agent Index documenting technical and safety features of 30 state-of-the-art AI agents. Catalogs origins, design, capabilities, ecosystem, and safety features based on public information and developer correspondence, illuminating transparency trends.

Agentic AI systems are increasingly capable of performing professional and personal tasks with limited human involvement. However, tracking these developments is difficult because the AI agent ecosystem is complex, rapidly evolving, and inconsistently documented, posing obstacles to both researchers and policymakers. To address these challenges, this paper presents the 2025 AI Agent Index. The Index documents information regarding the origins, design, capabilities, ecosystem, and safety features
AI AgentsAI SafetyAI GovernanceBenchmarking