AI/ML arXiv cs.AI

Mitigating Early Training Collapse in CTR Models

A study on mitigating early training collapse in click-through rate (CTR) models by controlling feature sparsity.

AI/ML arXiv cs.AI

Lesioned Multimodal Language Models Reproduce Aphasic Picture-Naming Patterns

Researchers used LLaVA 1.6 to simulate aphasic picture-naming error patterns, suggesting LLMs could serve as digital twins for post-stroke patients.

AI/ML arXiv cs.AI

Reproducing human biases in route choice using large language models: Toward scalable behavioral modeling

A study demonstrates that LLMs can reproduce non-rational human route choice biases, offering a scalable alternative for behavioral modeling.

AI/ML arXiv cs.AI

Think Through a Bottleneck: Hourglass Reasoning for Rigorous Induction

The 'Hourglass reasoning' framework improves inductive reasoning in LLMs by enforcing structural isolation between reasoning stages, significantly boosting accuracy in hardware synthesis and visual abstraction.

AI/ML arXiv cs.AI

Playful AI in Professional Email: A Field Experiment on Tone and Recipient Engagement

A field experiment shows that AI-assisted email writing affects recipient engagement primarily through the emotional tone produced rather than the AI's use itself.

Cybersecurity arXiv cs.AI

Reverse Engineering Compliance: A Dual-Graph Verification Framework for Auditing Legacy IT Security Concepts

The ASSERT framework uses ontology-based extraction to verify legacy IT security concepts and export them into machine-readable OSCAL artifacts for compliance auditing.

AI/ML arXiv cs.AI

Knowledge Graphs Meet Graph Neural Networks: A Comprehensive Survey

A comprehensive survey on the integration of Graph Neural Networks (GNNs) with Knowledge Graphs, outlining a taxonomy for construction, embedding, reasoning, and applications.

Hardware/Chips arXiv cs.AI

ECG-LDC: A Hardware-Efficient Low-Dimensional Computing Framework for ECG Arrhythmia Classification

ECG-LDC is a hardware-efficient framework for ECG arrhythmia classification using low-dimensional computing, achieving high accuracy with a minimal memory footprint on FPGA.

AI/ML arXiv cs.AI

Ablation, Statistical Inference, and Validation for KV-Cache Compression

A statistical comparison of KV-cache compression methods like Turbo-Quant and SpectralQuant, analyzing their effectiveness based on data distribution and calibration budgets.

AI/ML arXiv cs.AI

SciML in the Wild: A Diagnostic Study of When Structural Priors Help and When They Hurt

A diagnostic study on Scientific Machine Learning (SciML) reveals that structural priors can act as misregularizers if they do not align with the actual data-generating process.

AI/ML arXiv cs.AI

Transfer Learning Across Policy Regimes in Adaptive Multi-Agent Systems

Research on transfer learning in adaptive multi-agent systems suggests that reusing previous regulatory experience can be beneficial or harmful depending on the stability of structural invariants.

Other Hacker News

Notable Knot Index (2016)

A Hacker News discussion centered around a 'Notable Knot Index' from 2016.

AI/ML arXiv cs.AI

Omni-Decision: A Progressive Evidence-State Agent System for Omni-Modal QA

Introduces Omni-Decision, a training-free system for omni-modal QA that uses a structured evidence-state to track and close evidence gaps across multiple modalities.

AI/ML arXiv cs.AI

The Ebb and Flow of Multimodal Focus: Scheduling Visual Relay Windows for Grounded VLM Reasoning

Proposes TRACE, a task-adaptive inference-time control framework that manages visual attention focus in VLMs to improve evidence-grounded reasoning.

AI/ML arXiv cs.AI

Enhancing Query Efficiency for d-DNNF Representations Through Preprocessing

Explores preprocessing techniques to improve query efficiency for d-DNNF representations of propositional formulas, specifically for sampling and model enumeration.

AI/ML arXiv cs.AI

Comparative Analysis of GAT and BERT for Human-Like Playtesting

Compares GAT and BERT architectures for modeling human-like playtesting in puzzle games, finding they better capture relational game board structures than CNNs.

AI/ML arXiv cs.AI

Learning Residual Kinematic Corrections for Continuous Neural Decoding via Reinforcement Learning

Presents a two-stage BCI decoding framework using RL to perform residual kinematic corrections on CNN-LSTM outputs for improved 3D motor imagery decoding.

Hardware/Chips arXiv cs.AI

HCRMap: Pressure-Aware Hot-Expert Residency Mapping for 3.5D MoE Chiplet Inference

Introduces HCRMap, a framework for MoE chiplet inference that dynamically manages hot expert replicas to reduce latency and communication bottlenecks.

AI/ML arXiv cs.AI

MAGIC: Transition-Aware Generation of Navigable Multi-Scene Game Worlds with Large Language Models

Presents MAGIC, a prompt-to-project system that uses LLMs to generate navigable, multi-scene game worlds with consistent transitions and connectivity.

AI/ML arXiv cs.AI

Interaction Scaling: Grounding the Third Axis of Test-Time Compute

Argues for 'interaction scaling' as a third axis of test-time compute, where models iteratively revise artifacts based on grounded external feedback.