Models & Releases Research Products & Apps Business & Funding

Modelwire

A curated feed of what matters in AI. Independent, ad-supported, built in Denver, Colorado.

Read

Today
Models & Releases
Research
Business & Funding

About

About Modelwire
Methodology
Our sources
Editor's notes
Contact
Advertise

Legal

Privacy policy
Terms of use
DMCA & takedowns
Corrections

© 2026 Modelwire. All article links go to the original publishers.Summaries generated by Modelwire. We don’t republish full articles.

Earlier stories

The full Modelwire feed, ordered by publish time.

Illustration for: Promoting Simple Agents: Ensemble Methods for Event-Log Prediction

Promoting Simple Agents: Ensemble Methods for Event-Log Prediction

Researchers show that simple n-gram models match neural networks (LSTM, Transformer) on event-log prediction while using far fewer resources. A new ensemble method called promotion dynamically selects between models, avoiding the memory overhead of traditional voting approaches.

arXiv cs.LG·Apr 23

52

Illustration for: A-THENA: Early Intrusion Detection for IoT with Time-Aware Hybrid Encoding and Network-Specific Augmentation

A-THENA: Early Intrusion Detection for IoT with Time-Aware Hybrid Encoding and Network-Specific Augmentation

Researchers propose A-THENA, a Transformer-based intrusion detection system for IoT networks that uses time-aware encoding and network-specific data augmentation to catch attacks earlier. The approach targets the expanding attack surface of connected devices by capturing temporal patterns in network traffic.

arXiv cs.LG·Apr 23

42

Illustration for: Process Supervision via Verbal Critique Improves Reasoning in Large Language Models

Research Models & Releases

Process Supervision via Verbal Critique Improves Reasoning in Large Language Models

Researchers introduce Verbal Process Supervision, a training-free method that uses natural-language critique from a stronger model to iteratively refine reasoning. The technique pushes GPT-5.4 to 94.9% on GPQA Diamond and lifts weak models from 11.7% to 90% on AIME 2025, establishing a new inference-time scaling axis beyond chain-of-thought depth and sample breadth.

arXiv cs.CL·Apr 23

72

Illustration for: On the Role of Preprocessing and Memristor Dynamics in Reservoir Computing for Image Classification

Research Hardware & Infra

On the Role of Preprocessing and Memristor Dynamics in Reservoir Computing for Image Classification

Researchers analyzed how memristor device properties—decay rate, quantization, noise—shape reservoir computing performance on image tasks. The work clarifies hardware-level tradeoffs for building neuromorphic RC systems with reduced parameter overhead compared to standard RNNs.

arXiv cs.LG·Apr 23

52

Illustration for: Verifying Machine Learning Interpretability Requirements through Provenance

Verifying Machine Learning Interpretability Requirements through Provenance

Researchers propose using ML provenance tracking to make interpretability requirements measurable and verifiable in ML engineering. The work bridges requirements engineering and ML development by addressing a critical gap: interpretability has been defined qualitatively but never rigorously validated.

arXiv cs.LG·Apr 23

52

Illustration for: A Kernel Nonconformity Score for Multivariate Conformal Prediction

A Kernel Nonconformity Score for Multivariate Conformal Prediction

Researchers propose a kernel-based nonconformity score for multivariate conformal prediction that adapts prediction regions to residual geometry while bridging Bayesian and frequentist uncertainty quantification. The method decomposes into anisotropic MMD and offers finite-sample coverage guarantees with convergence rates tied to kernel covariance rank.

arXiv cs.LG·Apr 23

52

Illustration for: Language as a Latent Variable for Reasoning Optimization

Language as a Latent Variable for Reasoning Optimization

Researchers found that large language models often reason more accurately in non-English languages, suggesting language itself shapes internal inference pathways rather than merely encoding output. A polyglot optimization method showed unconstrained multilingual reasoning outperforms English-only approaches, implying model capability scales with linguistic diversity.

arXiv cs.CL·Apr 23

62

Illustration for: AgenticQwen: Training Small Agentic Language Models with Dual Data Flywheels for Industrial-Scale Tool Use

Models & Releases Research

AgenticQwen: Training Small Agentic Language Models with Dual Data Flywheels for Industrial-Scale Tool Use

Alibaba's Qwen team released a family of small language models trained to act as autonomous agents through dual reinforcement learning flywheels that automatically generate harder tasks. The approach combines reasoning and tool-use optimization to hit industrial cost and latency constraints without sacrificing multi-step decision-making.

arXiv cs.CL·Apr 23

62

Illustration for: CoFEE: Reasoning Control for LLM-Based Feature Discovery

Research Tools & Code

CoFEE: Reasoning Control for LLM-Based Feature Discovery

Researchers introduce CoFEE, a framework that guides LLMs to generate higher-quality features from unstructured data by enforcing structured reasoning patterns. The method addresses a core challenge in ML workflows: preventing feature leakage and weak proxies while scaling feature discovery across complex datasets.

arXiv cs.LG·Apr 23

52

Illustration for: Microsoft launches ‘vibe working’ in Word, Excel, and PowerPoint

Products & Apps

Microsoft launches ‘vibe working’ in Word, Excel, and PowerPoint

Microsoft is rolling out Agent Mode across Office apps this week, a more autonomous version of Copilot designed to handle complex workflows in Word, Excel, and PowerPoint. The feature represents Microsoft's push to deepen AI integration in enterprise productivity software beyond basic copilot assistance.

The Verge — AI·Apr 23

69

Illustration for: Measuring Opinion Bias and Sycophancy via LLM-based Coercion

Research Tools & Code

Measuring Opinion Bias and Sycophancy via LLM-based Coercion

Researchers released llm-bias-bench, an open-source tool that exposes hidden opinions in LLMs by simulating multi-turn conversations where models reveal contradictory stances on contested topics. The work highlights how AI assistants' evasive responses mask underlying biases that propagate at scale into user decisions on policy, health, and ethics.

arXiv cs.CL·Apr 23

62

Illustration for: Finding Meaning in Embeddings: Concept Separation Curves

Finding Meaning in Embeddings: Concept Separation Curves

Researchers propose a classifier-free evaluation method for sentence embeddings that measures how well models capture semantic meaning by introducing syntactic noise and semantic negations. The approach sidesteps downstream task dependencies to isolate embedding quality from classifier behavior.

arXiv cs.CL·Apr 23

52

Illustration for: AI Designs Thermoelectric Generators 10,000 Times Faster Than We Can

Research Tools & Code

AI Designs Thermoelectric Generators 10,000 Times Faster Than We Can

Japanese researchers deployed an AI tool that accelerates thermoelectric generator design by 10,000x compared to traditional simulation and experimentation. Prototypes built from the system's recommendations matched performance of current commercial designs, suggesting AI-driven materials discovery could unlock waste-heat energy conversion at scale.

IEEE Spectrum — AI·Apr 23

69

Illustration for: GPT-5.5 System Card

Models & Releases Research

GPT-5.5 System Card

OpenAI released a system card for GPT-5.5, documenting the model's capabilities, limitations, and safety considerations. System cards have become standard transparency artifacts for frontier labs, offering researchers and practitioners concrete details on model behavior and deployment considerations.

OpenAI·Apr 23

94

Illustration for: Introducing GPT-5.5

Models & Releases

Introducing GPT-5.5

OpenAI released GPT-5.5, positioning it as a faster, more capable successor optimized for coding, research, and data analysis workflows. The launch marks a meaningful step in frontier model capability, though specifics on benchmarks and performance gains remain limited in the announcement.

OpenAI·Apr 23

100

Illustration for: UKP_Psycontrol at SemEval-2026 Task 2: Modeling Valence and Arousal Dynamics from Text

UKP_Psycontrol at SemEval-2026 Task 2: Modeling Valence and Arousal Dynamics from Text

UKP_Psycontrol's SemEval-2026 Task 2 submission ranks first in modeling emotional valence and arousal shifts across user texts, finding that LLMs capture static affect well but recent numeric trajectories better predict short-term emotional change than semantic content alone.

arXiv cs.CL·Apr 23

42

Illustration for: Job Skill Extraction via LLM-Centric Multi-Module Framework

Research Tools & Code

Job Skill Extraction via LLM-Centric Multi-Module Framework

Researchers introduced SRICL, a framework that combines semantic retrieval, in-context learning, and fine-tuning to extract job skills from ads with higher accuracy and fewer hallucinations. The system uses a deterministic verifier to enforce structural constraints, addressing a persistent challenge in candidate-job matching systems.

arXiv cs.CL·Apr 23

52

Illustration for: Seeing Isn't Believing: Uncovering Blind Spots in Evaluator Vision-Language Models

Seeing Isn't Believing: Uncovering Blind Spots in Evaluator Vision-Language Models

Researchers systematically tested whether Vision-Language Models used to evaluate other AI systems can reliably catch common errors like object hallucinations and spatial reasoning failures. A benchmark of 4,000+ perturbed instances across 40 perturbation types reveals significant blind spots in how VLMs assess image-to-text and text-to-image outputs.

arXiv cs.CL·Apr 23

62

Illustration for: OpenAI says its new ChatGPT for Clinicians outperforms doctors on clinical tasks even when they have unlimited time and web access

Products & Apps Models & Releases

OpenAI says its new ChatGPT for Clinicians outperforms doctors on clinical tasks even when they have unlimited time and web access

OpenAI launched ChatGPT for Clinicians, a free medical chatbot built on GPT-5.4 that reportedly outperforms human doctors on clinical benchmarks even when physicians have unlimited time and web access. The capability gap raises questions about AI's readiness for clinical deployment and potential workforce implications.

The Decoder·Apr 23

85

Illustration for: From Tokens to Concepts: Leveraging SAE for SPLADE

Research Models & Releases

From Tokens to Concepts: Leveraging SAE for SPLADE

Researchers propose SAE-SPLADE, which replaces token-based vocabularies with learned semantic concepts to improve sparse retrieval models. The approach maintains competitive retrieval performance while addressing polysemy and enabling better multi-lingual support, suggesting a path beyond fixed vocabularies in IR systems.

arXiv cs.CL·Apr 23

52

Illustration for: OptiVerse: A Comprehensive Benchmark towards Optimization Problem Solving

Research Models & Releases

OptiVerse: A Comprehensive Benchmark towards Optimization Problem Solving

Researchers released OptiVerse, a 1,000-problem benchmark spanning stochastic, dynamic, game, and optimal control optimization tasks. Testing 22 LLMs revealed steep performance cliffs on hard problems, with GPT-5.2 and Gemini-3 both capping below 27% accuracy, exposing a critical weakness in current model reasoning.

arXiv cs.CL·Apr 23

62

Illustration for: An Interview with Google Cloud CEO Thomas Kurian About the Agentic Moment

Business & Funding Opinion & Analysis

An Interview with Google Cloud CEO Thomas Kurian About the Agentic Moment

Google Cloud CEO Thomas Kurian discusses the company's enterprise agent platform and how its integrated cloud infrastructure gives it an edge in the emerging agentic AI market. The interview covers Google's strategic positioning as AI workloads shift from model training to autonomous agent deployment.

Stratechery·Apr 23

73

Illustration for: Automations

Products & Apps Tools & Code

Automations

OpenAI expanded Codex with native automation capabilities, letting users schedule tasks and build recurring workflows through triggers without writing code. The feature targets business users building reports and summaries at scale.

OpenAI·Apr 23

68

Illustration for: How English Print Media Frames Human-Elephant Conflicts in India

How English Print Media Frames Human-Elephant Conflicts in India

Researchers analyzed nearly 2,000 English-language news articles on human-elephant conflict in India using transformer models and custom sentiment lexicons to quantify how media frames wildlife encounters. The study reveals linguistic patterns that shape public perception of ecological crises through computational text analysis.

arXiv cs.CL·Apr 23

52

Illustration for: Generalizing Numerical Reasoning in Table Data through Operation Sketches and Self-Supervised Learning

Generalizing Numerical Reasoning in Table Data through Operation Sketches and Self-Supervised Learning

Researchers propose TaNOS, a self-supervised pretraining framework that improves numerical reasoning over tables by anonymizing headers and using operation sketches to reduce domain-specific shortcuts. The technique boosts transferability across table domains when applied to 8B instruction-tuned models.

arXiv cs.CL·Apr 23

52

Illustration for: Preferences of a Voice-First Nation: Large-Scale Pairwise Evaluation and Preference Analysis for TTS in Indian Languages

Preferences of a Voice-First Nation: Large-Scale Pairwise Evaluation and Preference Analysis for TTS in Indian Languages

Researchers evaluated 7 TTS systems across 10 Indian languages using 120K+ pairwise comparisons from 1,900 native speakers, introducing a multidimensional framework that isolates perceptual factors like intelligibility, expressiveness, and hallucinations to address high variance in multilingual speech evaluation.

arXiv cs.CL·Apr 23

58

Illustration for: Cross-Domain Data Selection and Augmentation for Automatic Compliance Detection

Cross-Domain Data Selection and Augmentation for Automatic Compliance Detection

Researchers evaluated four data selection strategies to improve how NLI models transfer compliance detection across different regulatory domains. Targeted selection methods outperformed random sampling, suggesting principled data curation can reduce negative transfer when adapting legal AI systems to new regulations.

arXiv cs.CL·Apr 23

52

Illustration for: Reasoning Primitives in Hybrid and Non-Hybrid LLMs

Reasoning Primitives in Hybrid and Non-Hybrid LLMs

Researchers compared hybrid architectures combining attention and recurrence against pure transformer models on controlled reasoning tasks, finding that reasoning augmentation—not architecture choice—drives the largest performance gains across state-tracking and recall primitives.

arXiv cs.CL·Apr 23

52

Illustration for: How to Use Transformers.js in a Chrome Extension

How to Use Transformers.js in a Chrome Extension

Hugging Face published a guide for integrating Transformers.js, its JavaScript library for running transformer models, directly into Chrome extensions. This enables developers to deploy lightweight ML inference in browser plugins without server-side dependencies.

Hugging Face·Apr 23

64

Illustration for: GPT-5.5 Bio Bug Bounty

Research Policy & Regulation

GPT-5.5 Bio Bug Bounty

OpenAI launched a red-teaming bug bounty for GPT-5.5 focused on biosafety jailbreaks, offering up to $25,000 per finding. The initiative targets universal exploits that could bypass safety guardrails on biological risk queries, signaling intensified focus on frontier model safety validation.

OpenAI·Apr 23

81

Older stories →