AI/ML arXiv cs.AI

Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?

Researchers introduce Drop-Then-Recovery (DTR) to analyze architectural redundancy in Vision-Language-Action (VLA) models, finding that language backbones are often highly redundant for robotic tasks.

AI/ML arXiv cs.AI

RS-Diffuser: Risk-Sensitive Diffusion Planning with Distributional Value Guidance

RS-Diffuser is a risk-sensitive offline diffusion planning framework that uses distributional value critics to allow flexible risk-averse or risk-seeking behaviors in robotic navigation.

AI/ML arXiv cs.AI

Improving Adversarial Robustness via Activation Amplification and Attenuation

The A3 module improves adversarial robustness in neural networks by jointly learning to amplify and attenuate signals through a lightweight activation scaling mechanism.

AI/ML arXiv cs.AI

Output-Space Allocation Costs for Calibration-Guided LLM Compression: An Empirical Study

An empirical study on LLM compression explores whether aligning allocation costs with output-space objectives improves fidelity, noting a tradeoff between accuracy and perplexity.

AI/ML arXiv cs.AI

SHIFT: Gate-Modulated Activation Steering for Knowledge Conflict Mitigation in Retrieval-Augmented Generation

SHIFT is a framework that uses gate-modulated activation steering to resolve knowledge conflicts between retrieved context and parametric knowledge in RAG systems.

AI/ML arXiv cs.AI

NLL-Guided Full-Attention Layer Selection for Training-Free Sliding-Window Adaptation

A training-free method for selecting full-attention layers in hybrid models using Negative Log-Likelihood (NLL) guidance to optimize long-context inference efficiency.

AI/ML arXiv cs.AI

Position Bias Correction is Insufficient for One-Pass Attention Sorting

Research indicates that simple position bias correction is insufficient to match the performance of iterative attention sorting for long-context language models.

AI/ML arXiv cs.AI

Optimizing Teacher-Student Partitioning for Scalable Knowledge Distillation on HPC Systems

A new HPC-aware methodology for knowledge distillation decouples teacher and student partitioning to achieve up to 67% higher throughput than the TRL library.

AI/ML arXiv cs.AI

Parameter-Efficient Quantum-Inspired Fast Weight Programmers for Traffic-Matrix Forecasting

The paper proposes QKAN-FWPs, quantum-inspired recurrent models that provide efficient traffic-matrix forecasting with significantly lower memory and training budgets.

AI/ML arXiv cs.AI

Pepti-drift: Toxicity-Repulsive Drifting for Antigen-Conditioned Discrete Peptide Generation

Pepti-drift is a toxicity-aware latent refinement framework designed to generate antigen-specific binding peptides while avoiding toxicity.

Open Source Hacker News

Replacing Systemd with OpenRC in Debian

A discussion on the process and implications of replacing systemd with OpenRC in Debian Linux distributions.

Hardware/Chips Hacker News

1.38 Millimeter Microcontroller

Presentation of a microcontroller with an extremely small physical footprint of 1.38 millimeters.

Tech Business/VC Hacker News

US Grid Constraints: Towards 40GW+ of Behind-the-Meter Datacenter by 2028?

An analysis of US power grid constraints and the potential rise of behind-the-meter datacenter capacity by 2028.

AI/ML arXiv cs.AI

Do Speech Emphasis Models Generalize across Languages and Emotions?

Introduction of the MMEE corpus, a multilingual multi-emotion emphasis dataset for improving speech emphasis detection models.

AI/ML arXiv cs.AI

Enhancing Numerical Prediction in LLMs via Smooth MMD Alignment

Proposes Smooth Maximum Mean Discrepancy (SMMD) to improve the numerical precision of LLM predictions using value-distance kernels.

AI/ML arXiv cs.AI

Bifocal Diffusion Language Models: Asymmetric Bidirectional Context for Parallel Generation

Introduces Bifocal Diffusion Language Models and R2LM, combining causal attention with Mamba SSMs to improve parallel generation throughput.

AI/ML arXiv cs.AI

KG2Cypher: Data-Centric Pipeline for Building Enterprise Text-to-Cypher Systems

Presents KG2Cypher, a data-centric pipeline for converting natural language questions into executable Cypher queries for enterprise knowledge graphs.

AI/ML arXiv cs.AI

End-to-End Dynamic Sparsity for Resource-Adaptive LLM Inference

Proposes the L2A framework for resource-adaptive LLM inference, allowing models to dynamically adjust computation based on runtime budgets.

AI/ML arXiv cs.AI

Flexformer: Flexible Linear Transformer with Learnable Attention Kernel

Introduces Flexformer, a linear Transformer that utilizes learnable attention kernels to reduce complexity while maintaining expressiveness.

AI/ML arXiv cs.AI

From General-Purpose Audio Tagging to Spatially Grounded Sound Event Localization and Detection

Explores the AT2SELD framework for extending general-purpose audio tagging models to spatially grounded sound event localization and detection.