AI/ML arXiv cs.AI

How Do Instructions Shape Speech? Cross-Attention Attribution for Style-Captioned Text-to-Speech

Researchers propose cross-attention attribution for speech diffusion models to understand how natural language instructions influence the acoustic output in text-to-speech systems.

AI/ML arXiv cs.AI

Toward Calibrated Mixture-of-Experts Under Distribution Shift

This paper explores calibration in Mixture-of-Experts (MoE) models under distribution shift and proposes adversarial reweighting to improve the accuracy-calibration tradeoff.

AI/ML arXiv cs.AI

Human Universal Grasping

HUG is a flow-matching model that generates diverse human grasps for any object in a single RGB-D image, including a new dataset and benchmark.

AI/ML arXiv cs.AI

Human-AI Agent Interaction in a Business Context

A study identifying UX principles and interaction patterns for improving human-AI agent interaction within business contexts.

AI/ML arXiv cs.AI

Exposing the Unsaid: Visualizing Hidden LLM Bias through Stochastic Path Aggregation

TreeTracer is a visual analytics tool that uses stochastic path aggregation to detect and visualize hidden biases in Large Language Models.

AI/ML arXiv cs.AI

Ensembles of Large Language Models for Identifying EQ-5D Studies in PubMed Based on Their Abstracts

Research demonstrating that an ensemble of Gemini and Gemma LLMs can effectively automate the screening of biomedical studies in PubMed.

AI/ML arXiv cs.AI

Disentangling Linguistic Relatedness from Task Alignment in Cross-Lingual Transfer

A study on cross-lingual transfer in LLMs finds that fine-tuning improves task-format alignment rather than providing cross-lingual knowledge transfer.

AI/ML arXiv cs.AI

How LLMs Fail and Generalize in RTL Coding for Hardware Design?

An analysis of LLM failures in RTL coding for hardware design, finding that alignment techniques only improve compilation while functional gaps remain.

AI/ML Hacker News

Generative AI Is Having Its Herbalife Moment

A discussion on Hacker News regarding whether Generative AI is currently in a speculative bubble similar to Herbalife.

Hardware/Chips TechCrunch

The US says ASML’s top chip tool may be in China. ASML says it isn’t

The US government claims ASML's advanced chip-making tools may be in China, while ASML denies the allegation.

AI/ML arXiv cs.AI

Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe

Researchers introduce UFP4, a uniform 4-bit training recipe for LLMs that reduces shrinkage bias and training instability compared to E2M1 formats.

AI/ML arXiv cs.AI

Interpretable Sperm Morphology Classification via Attention-Guided Deep Learning

A new attention-guided deep learning framework using EfficientNet-B0 and CBAM for more interpretable sperm morphology classification.

AI/ML arXiv cs.AI

Context-Aware Hierarchical Bayesian Modeling of IVF Laboratory Environmental Conditions

A study applying hierarchical Bayesian modeling to IVF laboratory environmental data to improve pregnancy rate predictions.

AI/ML arXiv cs.AI

What Do Safety-Aligned LLMs Learn From Mixed Compliance Demonstrations?

Research exploring how safety-aligned LLMs interpret mixed compliance demonstrations and the role of preference optimization in preventing jailbreaks.

AI/ML arXiv cs.AI

Multi-LCB: Extending LiveCodeBench to Multiple Programming Languages

Introduction of Multi-LCB, a benchmark extending LiveCodeBench to twelve programming languages to evaluate LLM cross-language code generation.

AI/ML arXiv cs.AI

FlowEdit: Associative Memory for Lifelong Pronunciation Adaptation in Flow-Matching TTS

FlowEdit introduces a lifelong adaptation framework for flow-matching TTS to correct pronunciation errors without retraining the model.

AI/ML arXiv cs.AI

DeepSWIP: Quotient-WMC Counterfactuals for Neural Probabilistic Logic Programs

DeepSWIP provides a counterfactual semantics for neurosymbolic DeepProbLog programs using weighted model counting for faster and more exact inference.

AI/ML arXiv cs.AI

LedgerAgent: Structured State for Policy-Adherent Tool-Calling Agents

LedgerAgent is an inference-time method for tool-calling agents that uses a separate ledger to maintain task states and ensure policy adherence.

Software Engineering Hacker News

Project Valhalla, Explained: How a Decade of Work Arrives in JDK 28

An explanation of Project Valhalla and its goal to modernize Java's type system with value types, arriving in JDK 28.

Open Source Hacker News

Akse3D – open-source 3D modelling anyone can master

Introduction to Akse3D, an open-source 3D modelling tool designed for ease of use.