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: Learning to Reason with Insight for Informal Theorem Proving

Learning to Reason with Insight for Informal Theorem Proving

Researchers propose DeepInsightTheorem, a hierarchical dataset and training framework that teaches LLMs to recognize core proof techniques in informal theorem proving. The approach structures proofs by extracting key insights and sketches alongside final solutions, addressing a critical bottleneck in natural-language mathematical reasoning.

arXiv cs.LG·Apr 17

58

Illustration for: No Universal Courtesy: A Cross-Linguistic, Multi-Model Study of Politeness Effects on LLMs Using the PLUM Corpus

No Universal Courtesy: A Cross-Linguistic, Multi-Model Study of Politeness Effects on LLMs Using the PLUM Corpus

Researchers tested five major LLMs across English, Hindi, and Spanish to measure how politeness in user prompts affects model output quality. Using 22,500 prompt-response pairs and an eight-factor evaluation framework, they found performance varies significantly by model and language, suggesting politeness effects aren't universal across systems.

arXiv cs.CL·Apr 17

58

Illustration for: This charming gadget writes bad AI poetry

Products & Apps

This charming gadget writes bad AI poetry

The Poetry Camera is a physical gadget that generates AI poetry, combining appealing retro design with generative AI capabilities. The Verge's review highlights the tension between the device's charming aesthetics and the underwhelming quality of its AI-generated output.

The Verge — AI·Apr 17

47

Illustration for: VEFX-Bench: A Holistic Benchmark for Generic Video Editing and Visual Effects

VEFX-Bench: A Holistic Benchmark for Generic Video Editing and Visual Effects

Researchers released VEFX-Dataset, a 5,049-example human-annotated benchmark spanning 9 major video editing categories, addressing a critical gap in standardized evaluation for AI-assisted video editing systems that currently rely on manual inspection or generic vision-language judges.

arXiv cs.CL·Apr 17

58

Illustration for: From Benchmarking to Reasoning: A Dual-Aspect, Large-Scale Evaluation of LLMs on Vietnamese Legal Text

Research Models & Releases

From Benchmarking to Reasoning: A Dual-Aspect, Large-Scale Evaluation of LLMs on Vietnamese Legal Text

Researchers benchmarked GPT-4o, Claude 3 Opus, Gemini 1.5 Pro, and Grok-1 on Vietnamese legal text simplification, introducing a dual-aspect evaluation framework that measures accuracy, readability, and consistency alongside detailed error analysis on 60 complex articles.

arXiv cs.CL·Apr 17

52

Illustration for: FL-MHSM: Spatially-adaptive Fusion and Ensemble Learning for Flood-Landslide Multi-Hazard Susceptibility Mapping at Regional Scale

FL-MHSM: Spatially-adaptive Fusion and Ensemble Learning for Flood-Landslide Multi-Hazard Susceptibility Mapping at Regional Scale

Researchers developed a deep learning pipeline combining spatial partitioning, early fusion, and mixture-of-experts models to jointly predict flood and landslide risk across Kerala. The approach outperforms uniform baseline methods by capturing cross-hazard dependencies and spatial heterogeneity in multi-hazard susceptibility mapping.

arXiv cs.LG·Apr 17

42

Illustration for: Anthropic Launches Claude Design

Products & Apps

Anthropic Launches Claude Design

Anthropic released Claude Design, a new AI tool enabling users to generate designs, prototypes, presentations, and one-pagers. The launch confirms earlier reporting and expands Claude's capabilities beyond text into visual and document creation workflows.

The Information — AI·Apr 17

85

Illustration for: Information Router for Mitigating Modality Dominance in Vision-Language Models

Information Router for Mitigating Modality Dominance in Vision-Language Models

Researchers propose MoIR, an information router that addresses modality dominance in vision-language models by routing data based on information density rather than just adjusting attention. The technique tackles a fundamental limitation where VLMs over-rely on single modalities even when input signals differ in quality and noise levels.

arXiv cs.LG·Apr 17

52

Illustration for: SwanNLP at SemEval-2026 Task 5: An LLM-based Framework for Plausibility Scoring in Narrative Word Sense Disambiguation

SwanNLP at SemEval-2026 Task 5: An LLM-based Framework for Plausibility Scoring in Narrative Word Sense Disambiguation

SwanNLP's SemEval-2026 submission tests LLM plausibility scoring on narrative word sense disambiguation, comparing fine-tuned smaller models against few-shot prompted large models. The work bridges a gap between benchmark performance and real-world story understanding, revealing how different model scales handle contextual sense selection.

arXiv cs.CL·Apr 17

42

Illustration for: Beyond Distribution Sharpening: The Importance of Task Rewards

Beyond Distribution Sharpening: The Importance of Task Rewards

Researchers directly compare reinforcement learning via task rewards against distribution sharpening, proving the latter hits fundamental stability limits and unfavorable optima. The work clarifies whether frontier models gain new skills or merely surface latent ones, with implications for how RL should be integrated into training pipelines.

arXiv cs.LG·Apr 17

62

Illustration for: Do Vision-Language Models Truly Perform Vision Reasoning? A Rigorous Study of the Modality Gap

Do Vision-Language Models Truly Perform Vision Reasoning? A Rigorous Study of the Modality Gap

Researchers introduce CrossMath, a benchmark that isolates vision reasoning in multimodal models by presenting identical problems in text-only, image-only, and combined formats. The work challenges whether VLMs genuinely reason over visual input or simply leverage their text backbone's reasoning capabilities.

arXiv cs.CL·Apr 17

62

Illustration for: Joint-Centric Dual Contrastive Alignment with Structure-Preserving and Information-Balanced Regularization

Joint-Centric Dual Contrastive Alignment with Structure-Preserving and Information-Balanced Regularization

Researchers introduce HILBERT, a multimodal framework that aligns audio and text representations from long documents using frozen pre-trained encoders and a reciprocal contrastive objective. The approach handles severe dimensional imbalance between modalities while preserving structure in low-resource settings.

arXiv cs.LG·Apr 17

42

Illustration for: Detecting and Suppressing Reward Hacking with Gradient Fingerprints

Detecting and Suppressing Reward Hacking with Gradient Fingerprints

Researchers propose Gradient Fingerprint (GRIFT), a technique that detects reward hacking in reinforcement learning by analyzing internal model computations rather than surface-level reasoning chains. The method addresses a critical vulnerability where models exploit loopholes in reward functions while maintaining plausible-looking intermediate outputs.

arXiv cs.LG·Apr 17

62

Illustration for: BAGEL: Benchmarking Animal Knowledge Expertise in Language Models

Research Models & Releases

BAGEL: Benchmarking Animal Knowledge Expertise in Language Models

Researchers introduced BAGEL, a benchmark dataset for testing how well language models handle specialized animal knowledge across taxonomy, morphology, behavior, and other domains. The evaluation uses closed-book questions drawn from scientific sources like bioRxiv and Xeno-canto to measure LLM expertise gaps in zoological reasoning.

arXiv cs.CL·Apr 17

52

Illustration for: Adaptive multi-fidelity optimization with fast learning rates

Adaptive multi-fidelity optimization with fast learning rates

Researchers introduce Kometo, a multi-fidelity optimization algorithm that achieves near-optimal convergence rates without requiring prior knowledge of function smoothness or approximation quality. The method improves on prior work by trading off between cheap biased approximations and expensive accurate ones, with applications to expensive-to-evaluate ML tasks.

arXiv cs.LG·Apr 17

52

Illustration for: Enhancing AI and Dynamical Subseasonal Forecasts with Probabilistic Bias Correction

Research Models & Releases

Enhancing AI and Dynamical Subseasonal Forecasts with Probabilistic Bias Correction

Researchers developed probabilistic bias correction, an ML framework that halves forecast error in subseasonal weather predictions (2–6 weeks out) by learning to correct systematic biases in ECMWF dynamical and AI models. The technique addresses a critical gap where traditional physics-based forecasts degrade sharply beyond two weeks, with direct applications to crop planning, wildfire management, and energy allocation.

arXiv cs.LG·Apr 17

58

Illustration for: Optimizing Korean-Centric LLMs via Token Pruning

Research Models & Releases

Optimizing Korean-Centric LLMs via Token Pruning

Researchers benchmarked token pruning—a compression technique that strips irrelevant language tokens—across Qwen3, Gemma-3, Llama-3, and Aya for Korean NLP tasks. Pruning reduced language confusion and improved generation stability, with vocabulary tailoring (English-Korean vs. English-Korean-Chinese) showing measurable trade-offs in performance.

arXiv cs.CL·Apr 17

52

Illustration for: Neuro-Symbolic ODE Discovery with Latent Grammar Flow

Research Tools & Code

Neuro-Symbolic ODE Discovery with Latent Grammar Flow

Researchers propose Latent Grammar Flow, a neuro-symbolic framework that discovers differential equations from data by embedding equations into a discrete latent space and using flow models to generate candidates that fit observations. The approach combines interpretability with learned search, enabling domain constraints like stability to guide equation discovery.

arXiv cs.LG·Apr 17

58

Illustration for: OT on the Map: Quantifying Domain Shifts in Geographic Space

OT on the Map: Quantifying Domain Shifts in Geographic Space

Researchers propose GeoSpOT, an optimal transport method for measuring distribution shifts between geographic regions in machine learning. The technique quantifies domain distance to predict when models trained in one region will succeed when deployed elsewhere, addressing a critical gap in geospatial ML deployment.

arXiv cs.LG·Apr 17

52

Illustration for: Beyond Surface Statistics: Robust Conformal Prediction for LLMs via Internal Representations

Beyond Surface Statistics: Robust Conformal Prediction for LLMs via Internal Representations

Researchers propose using internal model representations instead of surface-level outputs to build more reliable uncertainty estimates for LLM answers. The Layer-Wise Information scoring method measures how input conditioning reshapes entropy across model depth, enabling conformal prediction that stays valid even when deployment conditions shift from training.

arXiv cs.CL·Apr 17

58

Illustration for: Sketching the Readout of Large Language Models for Scalable Data Attribution and Valuation

Research Tools & Code

Sketching the Readout of Large Language Models for Scalable Data Attribution and Valuation

Researchers propose RISE, a scalable method for attributing and valuing training data in large language models by focusing on influence hotspots at the output layer rather than computing gradients across entire models. The technique uses dual-channel sketching to reduce computational overhead, addressing a major bottleneck in understanding which data drives LLM behavior.

arXiv cs.LG·Apr 17

62

Illustration for: How Dylan Patel and SemiAnalysis Grabbed Sway in Silicon Valley

Business & Funding Opinion & Analysis

How Dylan Patel and SemiAnalysis Grabbed Sway in Silicon Valley

Dylan Patel, founder of influential semiconductor analysis firm SemiAnalysis, has become a major industry voice scrutinizing Nvidia's operations and product roadmaps. His reporting occasionally draws friction from Nvidia, though CEO Jensen Huang recently publicly praised him at GTC, signaling his rising credibility in Silicon Valley.

The Information — AI·Apr 17

73

Illustration for: The Real AI Shift Isn’t New Models. It’s Control.

Opinion & Analysis

The Real AI Shift Isn’t New Models. It’s Control.

As AI deployment scales across enterprises, operational governance and system management have become more critical than building new models themselves. The shift reflects maturation in the field toward production-grade reliability and control over raw capability gains.

AI Business·Apr 17

55

Illustration for: Tokenmaxxing, OpenAI’s shopping spree, and the AI Anxiety Gap

Business & Funding Opinion & Analysis

Tokenmaxxing, OpenAI’s shopping spree, and the AI Anxiety Gap

OpenAI is acquiring multiple companies spanning finance and media while competitors like Anthropic release restricted models and a major shoe brand pivots to AI infrastructure, widening the gap between AI insiders and the general public.

TechCrunch — AI·Apr 17

69

Illustration for: Synthetic data in cryptocurrencies using generative models

Research Tools & Code

Synthetic data in cryptocurrencies using generative models

Researchers applied conditional GANs with LSTM generators to synthesize cryptocurrency price time series, addressing privacy and access constraints in financial ML. The approach produces statistically consistent synthetic data across multiple crypto-assets, potentially enabling safer model training without exposing real market data.

arXiv cs.LG·Apr 17

52

Illustration for: Dairy Queen is putting an AI chatbot in its drive-thrus

Products & Apps Business & Funding

Dairy Queen is putting an AI chatbot in its drive-thrus

Dairy Queen is deploying AI chatbots in drive-thrus across the US and Canada to streamline ordering and increase average transaction values. The rollout joins a broader trend of fast-food chains adopting conversational AI for customer-facing operations.

The Verge — AI·Apr 17

58

Illustration for: AI Drafting My Stories? Over My Dead Body

Opinion & Analysis Policy & Regulation

AI Drafting My Stories? Over My Dead Body

WIRED examines how newsrooms are adopting AI-assisted writing tools to boost productivity, while questioning whether efficiency gains justify potential editorial and labor costs that publishers have yet to fully reckon with.

WIRED — AI·Apr 17

65

Illustration for: JumpLoRA: Sparse Adapters for Continual Learning in Large Language Models

Research Tools & Code

JumpLoRA: Sparse Adapters for Continual Learning in Large Language Models

JumpLoRA introduces sparse adapters using JumpReLU gating to prevent catastrophic forgetting in continual learning for LLMs. The technique dynamically isolates parameters across sequential tasks and integrates with existing LoRA-based approaches like IncLoRA, improving performance on multi-task adaptation.

arXiv cs.LG·Apr 17

52

Illustration for: AtManRL: Towards Faithful Reasoning via Differentiable Attention Saliency

AtManRL: Towards Faithful Reasoning via Differentiable Attention Saliency

Researchers propose AtManRL, a reinforcement learning method that uses differentiable attention masks to make LLM reasoning traces more faithful to actual model decision-making. The technique combines saliency rewards with outcome-based rewards to ensure chain-of-thought explanations genuinely influence predictions rather than merely accompanying them.

arXiv cs.LG·Apr 17

58

Illustration for: On the Rejection Criterion for Proxy-based Test-time Alignment

On the Rejection Criterion for Proxy-based Test-time Alignment

Researchers unify two test-time alignment methods under a shared graphical model framework, showing they differ only in rejection criteria. They argue confidence-based rejection is flawed for ambiguous language and propose a conservative confidence bet alternative with experimental validation.

arXiv cs.CL·Apr 17

52

Older stories →