AI/ML arXiv cs.AI

cAPM: Continual AI-Assisted Pace-Mapping with Active Learning

cAPM introduces a continual learning framework for AI-assisted pace-mapping in cardiac ablation, significantly improving localization accuracy over previous active-learning methods.

AI/ML arXiv cs.AI

Protein Representation Learning with Secondary-Structure and Energy-Filtered Hydrogen-Bond Graphs

The authors propose a secondary-structure-aware graph neural network (SSProNet) for protein representation learning that incorporates hydrogen-bond interactions and structural motifs.

AI/ML arXiv cs.AI

Cost-Optimal LLM Routing with Limited User Feedback under User Satisfaction Guarantees

SLARouter is an online routing algorithm for LLMs that optimizes costs while providing theoretical guarantees for strict Service Level Agreement (SLA) compliance.

AI/ML arXiv cs.AI

Emyx: Fast and efficient all-atom protein generation

Emyx is a lightweight conditional flow matching model for all-atom protein generation that outperforms larger models in enzyme design while requiring significantly less training time.

AI/ML arXiv cs.AI

How Linear Is a Transformer Feed-Forward Block? Per-Block Linear Recoverability Is Learned, Not Architectural

Research exploring the linear recoverability of Transformer Feed-Forward blocks, discovering that linearity is a learned property rather than an architectural one.

AI/ML arXiv cs.AI

Improving Code-Switching ASR with Code-Mixing Guided Synthetic Speech

A new preference-learning framework for improving code-switching ASR by generating synthetic speech with better language-boundary consistency.

AI/ML arXiv cs.AI

DynAMO:Dynamic Asset Management Orchestration via Topological Multi-Agent Scheduling

DynAMO is an orchestration engine for LLM-powered industrial agents that uses a Plan-then-Execute architecture to improve efficiency and safety in Industry 4.0.

AI/ML arXiv cs.AI

Bistable by Construction: Wall-Clock-Calibrated State Monitors Have No Moment-Detection Regime at Agent Cadence

A technical correction and analysis of runtime monitors for autonomous agents, highlighting how wall-clock calibration leads to failure in moment-detection regimes.

AI/ML arXiv cs.AI

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

DeepSeek-V4 introduces MoE language models supporting 1 million token context with improvements in attention architecture and a new Muon optimizer.

AI/ML arXiv cs.AI

Where to Place the Query? Unveiling and Mitigating Positional Bias in In-Context Learning for Diffusion LLMs via Decoding Dynamics

Research on Diffusion LLMs reveals that query position significantly impacts In-Context Learning quality, proposing Auto-ICL for adaptive query routing.

AI/ML arXiv cs.AI

Detecting Hallucinations for Large Language Model-based Knowledge Graph Reasoning

LUCID is proposed as a hallucination detection method for LLM-based knowledge graph reasoning, integrating attention scores and KG structure.

AI/ML arXiv cs.AI

Sign-Language Datasets at Scale: A Comprehensive Survey on Resources, Benchmarks, and Annotation Standards

A comprehensive survey on sign-language datasets and the release of a standardized documentation repository for inclusive technology development.

AI/ML arXiv cs.AI

Trustworthy Multi-Agent Systems: Mitigating Semantic Drift with the Argent Signaling Protocol

The Argent Signaling Protocol (ASP) introduces structured quality signals to multi-agent LLM systems to distinguish and handle different failure types.

Hardware/Chips arXiv cs.AI

Physical Atari: A Robust and Accessible Platform for Real-time Reinforcement Learning on Robots

Physical Atari is an accessible, low-cost hardware platform for real-time reinforcement learning on robots using a physical Atari controller.

AI/ML arXiv cs.AI

Computational Identifiability

The paper proposes 'computational identifiability' to replace theoretical identifiability with a finite computational search procedure for empirical estimators.

AI/ML arXiv cs.AI

Information Lattice Learning as Probabilistic Graphical Model Structure Learning

Research explores Information Lattice Learning (ILL) as a method for probabilistic graphical model structure learning for interpretable rules.

AI/ML arXiv cs.AI

Zero-Inflated Gaussian Distributions Enable Parameter-Space Sparsity in Estimation-of-Distribution Algorithms

A new EDA approach uses zero-inflated Gaussian distributions to enable parameter-space sparsity in black-box optimization.

AI/ML arXiv cs.AI

Human-like autonomy emerges from self-play and a pinch of human data

A method combining minimal human data with self-play reinforcement learning is shown to emerge human-like driving autonomy efficiently.

Software Engineering Hacker News

Fable Converted Pylint to Rust

Fable has rewritten Pylint, a popular Python linter, in Rust to improve performance.

Open Source Hacker News

The Raku Foundation is born

The Raku Foundation has been established to support the development and governance of the Raku programming language.