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: Towards a Foundation-Model Paradigm for Aerodynamic Prediction in Three-dimensional Design

Research Models & Releases

Towards a Foundation-Model Paradigm for Aerodynamic Prediction in Three-dimensional Design

Researchers developed AeroTransformer, a foundation model for aerodynamic prediction that pre-trains on 30,000 diverse wing geometries before fine-tuning for specific design tasks. The approach cuts training data costs for 3D shape optimization by leveraging transfer learning, addressing a major bottleneck in computational design workflows.

arXiv cs.LG·Apr 20

58

Illustration for: Sonata: A Hybrid World Model for Inertial Kinematics under Clinical Data Scarcity

Research Models & Releases

Sonata: A Hybrid World Model for Inertial Kinematics under Clinical Data Scarcity

Researchers built Sonata, a 3.77M-parameter world model for IMU-based movement tracking that outperforms standard autoencoders on clinical tasks despite training on just 739 subjects. The hybrid approach predicts future sensor states rather than reconstructing raw data, improving fall-risk prediction and cross-cohort generalization in healthcare settings where data is scarce.

arXiv cs.LG·Apr 20

52

Illustration for: ExAI5G: A Logic-Based Explainable AI Framework for Intrusion Detection in 5G Networks

ExAI5G: A Logic-Based Explainable AI Framework for Intrusion Detection in 5G Networks

Researchers propose ExAI5G, combining Transformer-based intrusion detection with logic-based explainability for 5G networks, achieving 99.9% accuracy while generating LLM-evaluated explanations that prioritize operational actionability over black-box performance.

arXiv cs.LG·Apr 20

52

Illustration for: JudgeMeNot: Personalizing Large Language Models to Emulate Judicial Reasoning in Hebrew

Research Models & Releases

JudgeMeNot: Personalizing Large Language Models to Emulate Judicial Reasoning in Hebrew

Researchers developed a pipeline to fine-tune language models on individual judges' past decisions, enabling personalized judicial reasoning systems in low-resource languages like Hebrew. The synthetic-organic supervision approach outperformed existing personalization baselines on lexical, stylistic, and semantic similarity metrics.

arXiv cs.CL·Apr 20

52

Illustration for: Variational Autoencoder Domain Adaptation for Cross-System Generalization in ML-Based SOP Monitoring

Variational Autoencoder Domain Adaptation for Cross-System Generalization in ML-Based SOP Monitoring

Researchers propose a VAE-based domain adaptation framework that lets ML models trained on one optical fiber system generalize to different systems without catastrophic failure. The approach learns shared threat-detection signatures across a 21 km O-band testbed and a 63.4 km C-band metro ring, freezing the encoder and retraining only the classifier per system.

arXiv cs.LG·Apr 20

42

Illustration for: TSMC Earnings, New N3 Fabs, The Nvidia Ramp

Hardware & Infra Business & Funding

TSMC Earnings, New N3 Fabs, The Nvidia Ramp

TSMC's latest earnings reveal skepticism within leadership about AI-driven demand, despite plans for new N3 fabrication capacity and Nvidia's ongoing ramp. The gap between public AI enthusiasm and internal caution signals potential overcapacity risk in chip supply.

Stratechery·Apr 20

80

Illustration for: SignDPO: Multi-level Direct Preference Optimisation for Skeleton-based Gloss-free Sign Language Translation

SignDPO: Multi-level Direct Preference Optimisation for Skeleton-based Gloss-free Sign Language Translation

Researchers introduce SignDPO, a preference optimization framework that improves skeleton-based sign language translation by moving beyond imitation learning to discriminate spatial and temporal nuances. The multi-level approach constructs hierarchical training signals across linguistic dimensions to reduce semantic drift in real-time signing.

arXiv cs.CL·Apr 20

52

Illustration for: How Creative Are Large Language Models in Generating Molecules?

How Creative Are Large Language Models in Generating Molecules?

Researchers systematically evaluated how creatively large language models generate molecules from natural language prompts, treating creativity as a functional requirement for satisfying chemical constraints rather than an aesthetic property. The work examines what types of creative behavior LLMs exhibit in molecular design and proposes evaluation methods for this emerging capability.

arXiv cs.CL·Apr 20

58

Illustration for: CodePivot: Bootstrapping Multilingual Transpilation in LLMs via Reinforcement Learning without Parallel Corpora

Research Tools & Code

CodePivot: Bootstrapping Multilingual Transpilation in LLMs via Reinforcement Learning without Parallel Corpora

Researchers propose CodePivot, a reinforcement learning method that trains LLMs to translate code across multiple programming languages without requiring parallel training data. The approach addresses a key bottleneck in transpilation: scaling beyond language pairs to support low-resource PLs efficiently.

arXiv cs.CL·Apr 20

58

Illustration for: RASP-Tuner: Retrieval-Augmented Soft Prompts for Context-Aware Black-Box Optimization in Non-Stationary Environments

Research Tools & Code

RASP-Tuner: Retrieval-Augmented Soft Prompts for Context-Aware Black-Box Optimization in Non-Stationary Environments

RASP-Tuner addresses a real constraint in deployed ML systems: optimizing black-box objectives when the underlying problem shifts with external context. The method combines retrieval of past regimes with low-rank prompt adaptation to reduce repeated retuning costs in non-stationary settings.

arXiv cs.LG·Apr 20

52

Illustration for: Clusterability-Based Assessment of Potentially Noisy Views for Multi-View Clustering

Clusterability-Based Assessment of Potentially Noisy Views for Multi-View Clustering

Researchers propose a Multi-View Clusterability Score to assess data quality before clustering, addressing a gap where noisy views degrade performance. The method combines per-view structure, joint-space geometry, and cross-view consistency to identify problematic data sources upfront rather than during clustering.

arXiv cs.LG·Apr 20

52

Illustration for: Boltzmann Machine Learning with a Parallel, Persistent Markov chain Monte Carlo method for Estimating Evolutionary Fields and Couplings from a Protein Multiple Sequence Alignment

Research Tools & Code

Boltzmann Machine Learning with a Parallel, Persistent Markov chain Monte Carlo method for Estimating Evolutionary Fields and Couplings from a Protein Multiple Sequence Alignment

Researchers propose a parallel persistent MCMC approach to accelerate Boltzmann machine training for inferring evolutionary constraints in protein sequences. The method targets the inverse Potts problem, combining stochastic gradient descent with distributed sampling to reduce computational overhead while maintaining reproducibility of inferred couplings.

arXiv cs.LG·Apr 20

42

Illustration for: Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation

Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation

Researchers found that scaling up multi-agent LLM systems for brainstorming backfires: stronger aligned models, hierarchical group dynamics, and larger teams all suppress diversity rather than expand it, revealing a fundamental tension between individual quality and collective exploration.

arXiv cs.CL·Apr 20

62

Illustration for: Neural Garbage Collection: Learning to Forget while Learning to Reason

Neural Garbage Collection: Learning to Forget while Learning to Reason

Researchers propose Neural Garbage Collection, a technique that trains language models to selectively discard KV cache entries during chain-of-thought reasoning rather than relying on hand-crafted pruning rules. The end-to-end learned approach could unlock longer reasoning chains by reducing memory bottlenecks that currently constrain scaling.

arXiv cs.LG·Apr 20

62

Illustration for: The NSA is using Anthropic's most powerful AI model Mythos

Policy & Regulation Business & Funding

The NSA is using Anthropic's most powerful AI model Mythos

The NSA has deployed Anthropic's Mythos Preview model for intelligence operations, marking a significant government adoption of frontier AI capabilities by a major US surveillance agency.

The Decoder·Apr 20

85

Illustration for: Causally-Constrained Probabilistic Forecasting for Time-Series Anomaly Detection

Research Models & Releases

Causally-Constrained Probabilistic Forecasting for Time-Series Anomaly Detection

Researchers propose Causally Guided Transformer, a framework that combines causal graph priors with deep sequence modeling to detect anomalies in industrial sensor data. The approach aims to move beyond correlation-only methods by enabling root-cause localization in multivariate time series.

arXiv cs.LG·Apr 20

52

Illustration for: Employing General-Purpose and Biomedical Large Language Models with Advanced Prompt Engineering for Pharmacoepidemiologic Study Design

Employing General-Purpose and Biomedical Large Language Models with Advanced Prompt Engineering for Pharmacoepidemiologic Study Design

Researchers benchmarked general-purpose LLMs (GPT-4o, DeepSeek-R1) against biomedical-tuned models on 46 real pharmacoepidemiologic study protocols, testing relevance and coding accuracy across multiple ontologies. Specialized biomedical models showed mixed results compared to frontier general-purpose systems, raising questions about fine-tuning ROI in regulated domains.

arXiv cs.CL·Apr 20

52

Illustration for: Online Conformal Prediction with Adversarial Semi-bandit Feedback via Regret Minimization

Online Conformal Prediction with Adversarial Semi-bandit Feedback via Regret Minimization

Researchers propose a new online conformal prediction method that handles partial feedback from adversarial sources, extending uncertainty quantification to more realistic deployment scenarios where labels aren't always revealed. The work bridges a gap in safety-critical systems that must operate with incomplete information.

arXiv cs.LG·Apr 20

52

Illustration for: Mitigating Multimodal Hallucination via Phase-wise Self-reward

Mitigating Multimodal Hallucination via Phase-wise Self-reward

Researchers propose PSRD, a self-rewarding inference method that cuts vision-language model hallucinations without fine-tuning or external labels. The technique exploits newly observed phase-wise patterns in how hallucinations emerge, enabling real-time correction during generation.

arXiv cs.CL·Apr 20

58

Illustration for: Chinese tech workers are starting to train their AI doubles–and pushing back

Business & Funding Policy & Regulation

Chinese tech workers are starting to train their AI doubles–and pushing back

Chinese tech workers are being ordered to train AI agents designed to automate their own roles, sparking internal resistance among early adopters. A GitHub project called Colleague Skill enables companies to extract worker skills and personality traits into replicable AI systems, raising questions about job displacement and worker agency in AI-driven labor markets.

MIT Technology Review — AI·Apr 20

84

Illustration for: ltzGLUE: Luxembourgish General Language Understanding Evaluation

Research Tools & Code

ltzGLUE: Luxembourgish General Language Understanding Evaluation

Researchers released ltzGLUE, the first NLU benchmark for Luxembourgish, filling a gap for an overlooked official EU language. The work evaluates existing pretrained models on classification tasks including NER and intent detection, establishing a baseline for future LTZ language model development.

arXiv cs.CL·Apr 20

42

Illustration for: Modeling Multiple Support Strategies within a Single Turn for Emotional Support Conversations

Modeling Multiple Support Strategies within a Single Turn for Emotional Support Conversations

Researchers reformulated emotional support conversation tasks to handle multiple support strategies within single utterances, proposing two generation methods enhanced with reinforcement learning-guided reasoning. The work addresses a gap between prior single-strategy assumptions and real-world supportive dialogue patterns.

arXiv cs.CL·Apr 20

52

Illustration for: From Fallback to Frontline: When Can LLMs be Superior Annotators of Human Perspectives?

From Fallback to Frontline: When Can LLMs be Superior Annotators of Human Perspectives?

Researchers show LLMs can outperform human annotators at predicting aggregate subgroup opinions on subjective tasks, flipping the typical view of models as annotation fallbacks. The advantage stems from structural properties like low variance and reduced bias coupling rather than domain knowledge, with conditions for superiority common in real-world scenarios.

arXiv cs.CL·Apr 20

62

Illustration for: A Sugeno Integral View of Binarized Neural Network Inference

A Sugeno Integral View of Binarized Neural Network Inference

Researchers connect binarized neural networks to Sugeno integrals, a mathematical framework that expresses BNN inference as interpretable if-then rules. The work makes explicit how individual neuron decisions map to set functions, potentially improving explainability of ultra-low-precision models.

arXiv cs.LG·Apr 20

52

Illustration for: The Umwelt Representation Hypothesis: Rethinking Universality

The Umwelt Representation Hypothesis: Rethinking Universality

Researchers challenge the assumption that all capable AI systems converge on universal representations of reality, proposing instead that alignment between ANNs and biological brains stems from shared ecological constraints rather than a single global optimum. The Umwelt Representation Hypothesis reframes how we should interpret similarities across neural systems.

arXiv cs.LG·Apr 20

62

Illustration for: Process Reward Models Meet Planning: Generating Precise and Scalable Datasets for Step-Level Rewards

Research Tools & Code

Process Reward Models Meet Planning: Generating Precise and Scalable Datasets for Step-Level Rewards

Researchers developed a scalable method to generate Process Reward Models (PRMs) that evaluate step-level reasoning in LLMs, creating a dataset of ~1M reasoning steps across planning domains to reduce annotation costs and improve beyond math-only benchmarks.

arXiv cs.CL·Apr 20

58

Illustration for: Federated Rule Ensemble Method in Medical Data

Research Tools & Code

Federated Rule Ensemble Method in Medical Data

Researchers propose a federated RuleFit framework that trains interpretable machine learning models across distributed medical institutions without sharing raw patient data, addressing privacy constraints while maintaining clinical applicability.

arXiv cs.LG·Apr 20

52

Illustration for: ReCoQA: A Benchmark for Tool-Augmented and Multi-Step Reasoning in Real Estate Question and Answering

Research Tools & Code

ReCoQA: A Benchmark for Tool-Augmented and Multi-Step Reasoning in Real Estate Question and Answering

Researchers released ReCoQA, a 29,270-instance benchmark for training AI agents to answer real-estate questions by combining database queries and API calls. The accompanying HIRE-Agent framework uses hierarchical planning to integrate structured and unstructured data sources, establishing a baseline for multi-step reasoning tasks.

arXiv cs.CL·Apr 20

52

Illustration for: Domain-oriented RAG Assessment (DoRA): Synthetic Benchmarking for RAG-based Question Answering on Defense Documents

Research Models & Releases

Domain-oriented RAG Assessment (DoRA): Synthetic Benchmarking for RAG-based Question Answering on Defense Documents

Researchers introduced DoRA, a domain-specific RAG benchmark using 6.5K curated defense documents with synthetic QA pairs to measure real-world retrieval-augmented generation performance. A model fine-tuned on DoRA showed 26% QA improvement and 47% hallucination reduction over Llama 3.1-8B-Instruct, exposing how public benchmarks mask deployment gaps.

arXiv cs.CL·Apr 20

58

Illustration for: From Heads to Neurons: Causal Attribution and Steering in Multi-Task Vision-Language Models

From Heads to Neurons: Causal Attribution and Steering in Multi-Task Vision-Language Models

Researchers introduce HONES, a gradient-free framework for identifying and steering task-critical neurons in multi-task vision-language models. The method addresses polysemanticity noise by analyzing how information flows through feed-forward networks across different tasks, improving neuron attribution beyond single-task analyses.

arXiv cs.CL·Apr 20

52

Older stories →