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: ConforNets: Latents-Based Conformational Control in OpenFold3

Research Models & Releases

ConforNets: Latents-Based Conformational Control in OpenFold3

Researchers propose ConforNets, a method to extract multiple protein conformations from AlphaFold3 by applying learned channel-wise transforms to latent representations. The technique generalizes across proteins without retraining, addressing a key limitation where AF3 typically predicts only dominant structural states.

arXiv cs.LG·Apr 20

58

Illustration for: GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling

Research Tools & Code

GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling

Researchers propose GSQ, a scalar quantization method that closes the accuracy gap between simple techniques like GPTQ and complex vector-quantization approaches at ultra-low bit-widths (2-3 bits per parameter). The work suggests efficient LLM deployment needn't sacrifice accuracy for implementation simplicity.

arXiv cs.CL·Apr 20

62

Illustration for: A Note on TurboQuant and the Earlier DRIVE/EDEN Line of Work

A Note on TurboQuant and the Earlier DRIVE/EDEN Line of Work

A new arXiv note clarifies that TurboQuant, a recent quantization method, is a constrained variant of the earlier EDEN scheme (itself an extension of DRIVE). TurboQuant's fixed scaling parameter is suboptimal compared to EDEN's adaptive approach, though the gap narrows in high dimensions.

arXiv cs.LG·Apr 20

42

$Illustration for: Physics-Informed Neural Networks for Biological $2\mathrm{D}{+}t$ Reaction-Diffusion Systems$

Physics-Informed Neural Networks for Biological $2\mathrm{D}{+}t$ Reaction-Diffusion Systems

Researchers extended biologically-informed neural networks to 2D spatial domains with time, combining physics-based structure with learned neural components for reaction-diffusion systems. The work bridges symbolic regression and PINN frameworks to improve equation discovery beyond prior 1D limitations.

arXiv cs.LG·Apr 20

52

Illustration for: FUSE: Ensembling Verifiers with Zero Labeled Data

FUSE: Ensembling Verifiers with Zero Labeled Data

Researchers propose FUSE, an unsupervised ensemble method that improves LLM verification without labeled ground truth data. The technique controls dependencies between verifiers using spectral algorithms, matching or beating semi-supervised baselines while eliminating costly annotation requirements.

arXiv cs.CL·Apr 20

58

Illustration for: Wasserstein Distributionally Robust Risk-Sensitive Estimation via Conditional Value-at-Risk

Wasserstein Distributionally Robust Risk-Sensitive Estimation via Conditional Value-at-Risk

Researchers propose a method for risk-sensitive signal estimation under distribution uncertainty using Wasserstein robustness and conditional value-at-risk. The approach enables exact computation of affine estimators that minimize worst-case CVaR when the nominal distribution has finite support, advancing robust optimization techniques relevant to ML systems operating under model uncertainty.

arXiv cs.LG·Apr 20

42

Illustration for: ClawEnvKit: Automatic Environment Generation for Claw-Like Agents

Research Tools & Code

ClawEnvKit: Automatic Environment Generation for Claw-Like Agents

Researchers introduce ClawEnvKit, an automated pipeline that generates training environments for robotic agents from natural language descriptions. The system combines parsing, generation, and validation modules to scale environment creation beyond manual construction, addressing a bottleneck in agent development.

arXiv cs.CL·Apr 20

52

Illustration for: Transition-Matrix Regularization for Next Dialogue Act Prediction in Counselling Conversations

Transition-Matrix Regularization for Next Dialogue Act Prediction in Counselling Conversations

Researchers improved dialogue act prediction in counselling conversations by regularizing neural models against empirical transition patterns, achieving 9-42% relative gains in macro-F1 and demonstrating cross-language transfer on German and English datasets.

arXiv cs.CL·Apr 20

52

Illustration for: IDOBE: Infectious Disease Outbreak forecasting Benchmark Ecosystem

Research Tools & Code

IDOBE: Infectious Disease Outbreak forecasting Benchmark Ecosystem

Researchers released IDOBE, a benchmark dataset of over 10,000 epidemic outbreaks spanning a century of U.S. and global surveillance data, to standardize evaluation of machine learning and statistical forecasting models for infectious disease prediction.

arXiv cs.LG·Apr 20

52

Illustration for: UDM-GRPO: Stable and Efficient Group Relative Policy Optimization for Uniform Discrete Diffusion Models

Research Models & Releases

UDM-GRPO: Stable and Efficient Group Relative Policy Optimization for Uniform Discrete Diffusion Models

Researchers propose UDM-GRPO, the first framework combining Uniform Discrete Diffusion Models with reinforcement learning for stable training. The method treats final samples as actions and reconstructs diffusion trajectories to align with pretraining distributions, plus introduces efficiency strategies that outperform naive GRPO integration.

arXiv cs.LG·Apr 20

52

Illustration for: Deezer says 44% of new music uploads are AI-generated, most streams are fraudulent

Business & Funding Policy & Regulation

Deezer says 44% of new music uploads are AI-generated, most streams are fraudulent

Deezer reports that 44% of newly uploaded tracks are AI-generated, though they represent a tiny share of actual streams and face widespread demonetization for fraud. The finding underscores how generative audio is flooding music platforms while listeners remain uninterested in low-quality synthetic content.

Ars Technica — AI·Apr 20

69

Illustration for: Open-weight Kimi K2.6 takes on GPT-5.4 and Claude Opus 4.6 with agent swarms

Models & Releases Tools & Code

Open-weight Kimi K2.6 takes on GPT-5.4 and Claude Opus 4.6 with agent swarms

Moonshot AI released Kimi K2.6 as an open-weight model competitive with GPT-5.4 and Claude Opus 4.6 on coding tasks, with the ability to orchestrate up to 300 parallel agents. The open release challenges closed-model dominance in frontier capabilities.

The Decoder·Apr 20

85

Illustration for: Different Paths to Harmful Compliance: Behavioral Side Effects and Mechanistic Divergence Across LLM Jailbreaks

Different Paths to Harmful Compliance: Behavioral Side Effects and Mechanistic Divergence Across LLM Jailbreaks

Researchers compared three methods for jailbreaking open-weight LLMs and found they produce unsafe models with distinct internal failure modes. RLVR-jailbroken models notably retain the ability to recognize harmful requests and describe safe responses while still complying with them, revealing a mechanistic divergence across attack vectors.

arXiv cs.CL·Apr 20

62

Illustration for: MASS-RAG: Multi-Agent Synthesis Retrieval-Augmented Generation

Research Models & Releases

MASS-RAG: Multi-Agent Synthesis Retrieval-Augmented Generation

Researchers propose MASS-RAG, a multi-agent framework that assigns specialized roles to LLMs during retrieval-augmented generation, enabling them to handle noisy or incomplete retrieved evidence more effectively than single-pass approaches.

arXiv cs.CL·Apr 20

58

Illustration for: Document-as-Image Representations Fall Short for Scientific Retrieval

Document-as-Image Representations Fall Short for Scientific Retrieval

Researchers challenge the document-as-image paradigm dominating scientific retrieval benchmarks, arguing that rendering papers as pixels obscures structured content like tables and equations. They introduce ArXivDoc, a new benchmark built from LaTeX sources to better evaluate how models handle text-rich multimodal documents.

arXiv cs.CL·Apr 20

58

Illustration for: Fortnite developers can make AI characters now — just don’t try to date them

Products & Apps Tools & Code

Fortnite developers can make AI characters now — just don’t try to date them

Epic Games rolled out a conversational AI tool for Fortnite creators to build interactive NPCs without hand-coding dialogue trees, expanding on last year's AI-voiced Darth Vader character. The feature lets players talk naturally with AI-generated characters on custom islands, marking a shift toward generative dialogue in game development.

The Verge — AI·Apr 20

65

Illustration for: Learning the Riccati solution operator for time-varying LQR via Deep Operator Networks

Learning the Riccati solution operator for time-varying LQR via Deep Operator Networks

Researchers propose using deep operator networks to learn a surrogate for solving Riccati equations in LQR control problems, replacing repeated numerical integration with a single offline training stage. The approach trades computational cost during deployment for upfront learning, with theoretical guarantees on control performance across system families.

arXiv cs.LG·Apr 20

52

Illustration for: Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data

Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data

Researchers identify a failure mode in RL-trained LLMs: when base models already solve benchmarks like MATH near-perfectly, reinforcement learning algorithms collapse into homogeneous solutions due to vanishing reward signals. They propose CUTS, a decoding strategy that enforces diverse exploration among high-confidence outputs to restore learning dynamics.

arXiv cs.LG·Apr 20

58

Illustration for: Barrier-enforced multi-objective optimization for direct point and sharp interval forecasting

Barrier-enforced multi-objective optimization for direct point and sharp interval forecasting

Researchers propose a neural forecasting framework that generates point and interval predictions simultaneously while guaranteeing coverage bounds through barrier-enforced multi-objective optimization, eliminating manual loss-weight tuning.

arXiv cs.LG·Apr 20

52

Illustration for: Faster by Design: Interactive Aerodynamics via Neural Surrogates Trained on Expert-Validated CFD

Research Models & Releases

Faster by Design: Interactive Aerodynamics via Neural Surrogates Trained on Expert-Validated CFD

Researchers built a neural surrogate model trained on high-fidelity CFD data from LMP2 race-car geometries to accelerate aerodynamic design iteration. The approach cuts evaluation time from tens of thousands of core-hours to near-interactive speeds, enabling motorsport teams to explore design spaces previously infeasible under budget constraints.

arXiv cs.LG·Apr 20

58

Illustration for: LQM: Linguistically Motivated Multidimensional Quality Metrics for Machine Translation

LQM: Linguistically Motivated Multidimensional Quality Metrics for Machine Translation

Researchers propose LQM, a linguistically grounded error taxonomy for machine translation that captures dialect and culture-specific failures in diglossic languages like Arabic through six hierarchical levels. The framework extends beyond surface-form evaluation to address pragmatic and sociolinguistic mismatches, tested on a 3,850-sentence parallel corpus across language varieties.

arXiv cs.CL·Apr 20

52

Illustration for: Aligning Language Models for Lyric-to-Melody Generation with Rule-Based Musical Constraints

Aligning Language Models for Lyric-to-Melody Generation with Rule-Based Musical Constraints

Researchers propose an alignment framework that uses rule-based musical constraints to fix LLMs' tendency to generate rhythmically broken or vocally implausible melodies. The method chains Direct Preference Optimization with Kahneman-Tversky Optimization on automatically-generated preference data, substantially reducing constraint violations without human annotation.

arXiv cs.CL·Apr 20

52

Illustration for: Adversarial Humanities Benchmark: Results on Stylistic Robustness in Frontier Model Safety

Research Policy & Regulation

Adversarial Humanities Benchmark: Results on Stylistic Robustness in Frontier Model Safety

A new benchmark reveals that frontier models' safety guardrails collapse dramatically when harmful prompts are rewritten in literary or obfuscated styles. Attack success rates jumped from 3.84% to 55.75% across 31 models when researchers applied humanities-inspired transformations, exposing a critical gap in stylistic robustness.

arXiv cs.CL·Apr 20

68

Illustration for: OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

Research Models & Releases

OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

Researchers propose OneVL, a vision-language model that compresses chain-of-thought reasoning into latent tokens for real-time autonomous driving. The approach combines a VLA with a world model to capture causal dynamics rather than pure language, addressing latency bottlenecks in current CoT methods.

arXiv cs.CL·Apr 20

58

Illustration for: Safe Control using Learned Safety Filters and Adaptive Conformal Inference

Safe Control using Learned Safety Filters and Adaptive Conformal Inference

Researchers propose Adaptive Conformal Filtering, which pairs learned safety filters with dynamic error correction to make neural network-based control systems more reliable. The method adjusts its decision thresholds in real time based on prediction errors, addressing a key gap in scaling safety guarantees to high-dimensional systems.

arXiv cs.LG·Apr 20

52

Illustration for: Physics-Informed Neural Networks: A Didactic Derivation of the Complete Training Cycle

Physics-Informed Neural Networks: A Didactic Derivation of the Complete Training Cycle

A new arXiv paper provides a complete pedagogical walkthrough of Physics-Informed Neural Network training, breaking down forward pass, loss computation, backpropagation, and gradient updates with explicit numerical verification rather than delegating to autodiff black boxes.

arXiv cs.LG·Apr 20

52

Illustration for: WorldDB: A Vector Graph-of-Worlds Memory Engine with Ontology-Aware Write-Time Reconciliation

Research Tools & Code

WorldDB: A Vector Graph-of-Worlds Memory Engine with Ontology-Aware Write-Time Reconciliation

WorldDB proposes a recursive vector-graph memory system for long-running AI agents, addressing RAG's fragmentation problem by treating each node as a composable world with its own ontology and content-addressed invariants. The approach aims to enable persistent, contradiction-aware memory beyond stateless chatbots.

arXiv cs.CL·Apr 20

58

Illustration for: Google builds elite team to close the coding gap with Anthropic

Business & Funding Models & Releases

Google builds elite team to close the coding gap with Anthropic

Google is assembling a dedicated coding team under Sergey Brin to narrow Anthropic's lead in AI-assisted programming. The effort signals intensifying competition in a capability area where Anthropic has gained ground, with Google betting on self-improving models as a path forward.

The Decoder·Apr 20

73

Illustration for: Multi-Scale Reversible Chaos Game Representation: A Unified Framework for Sequence Classification

Research Tools & Code

Multi-Scale Reversible Chaos Game Representation: A Unified Framework for Sequence Classification

Researchers propose MS-RCGR, a reversible encoding framework that converts biological sequences into multi-resolution geometric representations for classification tasks. The method unifies three analytical paradigms—traditional ML, computer vision, and hybrid protein language models—while preserving complete sequence information through rational arithmetic.

arXiv cs.LG·Apr 20

52

Illustration for: Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts

Research Models & Releases

Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts

Researchers propose BAR, a modular post-training method that trains domain experts independently via mixture-of-experts routing, avoiding the quadratic cost scaling and capability degradation of traditional continued training. At 7B scale, the approach enables linear-cost updates to individual domains without harming existing capabilities.

arXiv cs.LG·Apr 20

62

Older stories →