AI/ML arXiv cs.AI

Weak-to-Strong Elicitation via Mismatched Wrong Drafts

This research discovers that injecting mismatched wrong drafts from a smaller model into a stronger learner's GRPO context improves performance on complex math benchmarks.

AI/ML arXiv cs.AI

Semantic Generative Tuning for Unified Multimodal Models

The authors propose Semantic Generative Tuning (SGT), a paradigm using image segmentation as a generative proxy to align visual understanding and generation in multimodal models.

Other Hacker News

Making Sense of Proof by Contradiction [pdf]

A discussion and resource on understanding the logical concept of proof by contradiction.

Tech Business/VC TechCrunch

Trump Admin releases Anthropic Mythos to be used by more than 100 US companies, agencies

The Trump administration has authorized over 100 US companies and agencies to use Anthropic's Mythos 5 AI system.

AI/ML arXiv cs.AI

VisNec: Measuring and Leveraging Visual Necessity for Multimodal Instruction Tuning

VisNec is a data selection framework that identifies vision-critical training samples to improve efficiency and performance in multimodal instruction tuning.

AI/ML arXiv cs.AI

Delegation and Verification Under AI

Research on how AI delegation and verification processes create phase transitions in worker quality and institutional disparities.

AI/ML arXiv cs.AI

Latent-Mark: An Audio Watermark Robust to Neural Codec Compression

Latent-Mark introduces a zero-bit audio watermarking framework robust to neural codec compression by embedding marks in the latent space.

AI/ML arXiv cs.AI

Residual RL-MPC for Robust Microrobotic Cell Pushing Under Time-Varying Flow

A hybrid controller combining MPC with a learned residual policy for robust microrobotic cell pushing in microfluidic flow.

AI/ML arXiv cs.AI

A Guideline-Aware AI Agent for Zero-Shot Target Volume Auto-Delineation

OncoAgent is a guideline-aware AI agent that converts textual clinical guidelines into 3D target contours for radiotherapy without retraining.

AI/ML arXiv cs.AI

MedPruner: Training-Free Hierarchical Token Pruning for Efficient 3D Medical Image Understanding in Vision-Language Models

MedPruner is a training-free token pruning framework that significantly reduces visual-token overhead in 3D medical Vision-Language Models.

AI/ML arXiv cs.AI

Power Couple? AI Growth and Renewable Energy Investment

An analysis of the equilibrium between AI growth and renewable energy investment and the potential for an 'adaptation trap' regarding fossil fuels.

AI/ML arXiv cs.AI

Scalable AI-assisted Workflow Management for Detector Design Optimization Using Distributed Computing

A scalable AI-assisted framework integrating Bayesian optimization with the PanDA-iDDS engine for detector design optimization in scientific computing.

Software Engineering Hacker News

Show HN: DBOSify – Drop-in Temporal replacement built on Postgres

DBOSify is a drop-in replacement for Temporal designed to run on PostgreSQL, simplifying the workflow orchestration layer.

Other Hacker News

Om

A post regarding 'Om', though the content provided is insufficient for a detailed technical analysis.

Tech Business/VC TechCrunch

FTC gives Musk the OK to acquire SpaceX alumni startup Mesh

The FTC has approved Elon Musk's acquisition of Mesh, a startup founded by former SpaceX employees.

AI/ML The Verge

Anthropic’s Mythos 5 is back

Anthropic's Mythos 5 model is partially restored for select organizations following government negotiations, while its public counterpart Fable 5 remains unavailable.

AI/ML arXiv cs.AI

Digital Twin-Driven Communication-Efficient Federated Anomaly Detection for Industrial IoT

Researchers propose a digital twin-integrated federated learning (DTFL) suite to improve anomaly detection in Industrial IoT while reducing communication overhead.

AI/ML arXiv cs.AI

Metaphors are a Source of Cross-Domain Misalignment of Large Reasoning Models

A study explores how metaphors in training data cause cross-domain misalignment in the reasoning pathways of large language models.

AI/ML arXiv cs.AI

Dual-Prototype Disentanglement: A Context-Aware Enhancement Framework for Time Series Forecasting

The DPAD framework introduces a dual-prototype bank to enhance time series forecasting by disentangling common and rare temporal patterns.

AI/ML arXiv cs.AI

VecSet-Edit: Unleashing Pre-trained LRM for Mesh Editing from Single Image

VecSet-Edit is a new pipeline for 3D mesh editing from single images using the VecSet Large Reconstruction Model as a backbone.