AI/ML arXiv cs.AI

CompilerKV: Risk-Adaptive KV Compression via Offline Experience Compilation

CompilerKV introduces risk-adaptive KV compression that compiles retention policies offline, significantly reducing memory overhead for long-context LLMs.

AI/ML arXiv cs.AI

Fly0: Persistent Metric Anchoring for Zero-Shot Aerial Vision-Language Navigation

Fly0 decouples semantic reasoning from geometric planning to enable zero-shot aerial vision-language navigation for drones in unstructured environments.

AI/ML arXiv cs.AI

When Visual Evidence is Ambiguous: Pareidolia as a Diagnostic Probe for Vision Models

A study using face pareidolia as a diagnostic probe reveals that Vision-Language Models often over-interpret ambiguous patterns as human faces compared to specialized detectors.

Cybersecurity arXiv cs.AI

Give Them an Inch and They Will Take a Mile:Understanding and Measuring Caller Identity Confusion in MCP-Based AI Systems

A security analysis of the Model Context Protocol (MCP) reveals significant vulnerabilities regarding caller identity confusion and lack of fine-grained authorization.

AI/ML arXiv cs.AI

TransDex: Pre-training Visuo-Tactile Policy with Point Cloud Reconstruction for Dexterous Manipulation of Transparent Objects

TransDex is a visuo-tactile fusion policy that uses point cloud reconstruction pre-training to enable dexterous manipulation of transparent objects by robots.

AI/ML arXiv cs.AI

PlotTwist: A Creative Plot Generation Framework with Small Language Models

PlotTwist is a framework that enables Small Language Models (under 3B parameters) to generate high-quality creative plots through structured preference-based alignment.

Other Hacker News

What Rose Petals Teach Us about Induction

A discussion on the conceptual relationship between rose petals and the principle of induction.

AI/ML arXiv cs.AI

Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures

Introduces the Hyperdimensional Probe, a hybrid supervised probe that combines symbolic representations with neural probing to better decode LLM internal representations.

AI/ML arXiv cs.AI

Breaking the MoE LLM Trilemma: Dynamic Expert Clustering with Structured Compression

Presents a framework for dynamic expert clustering and structured compression in MoE LLMs, reducing parameters by 80% while maintaining quality.

AI/ML arXiv cs.AI

Kontinuous Kontext: Continuous Strength Control for Instruction-based Image Editing

Introduces Kontinuous Kontext, a model for instruction-based image editing that allows continuous control over the strength of edits.

Hardware/Chips arXiv cs.AI

Beyond-Diagonal RIS Under Non-Idealities: Learning-Based Architecture Discovery and Optimization

Proposes a learning-based two-tier architecture discovery framework (LTTADF) to optimize beyond-diagonal reconfigurable intelligent surfaces for wireless networks.

AI/ML arXiv cs.AI

QuArch: A Benchmark for Evaluating LLM Reasoning in Computer Architecture

Introduces QuArch, a new benchmark for evaluating the reasoning capabilities of LLMs in the field of computer architecture.

AI/ML arXiv cs.AI

Active Electrosensing and Communication in MARL-trained Weakly Electric Fish Collectives

Explores the emergence of collective behavior in weakly electric fish using a computational framework trained via multi-agent reinforcement learning.

AI/ML arXiv cs.AI

ImplicitRDP: An End-to-End Visual-Force Diffusion Policy with Structural Slow-Fast Learning

Presents ImplicitRDP, an end-to-end visual-force diffusion policy for contact-rich manipulation using Structural Slow-Fast Learning.

AI/ML arXiv cs.AI

Memo2496: Expert-Annotated Dataset and Dual-view Adaptive Framework for Music Emotion Recognition

Presents the Memo2496 dataset and the DAMER framework for improving music emotion recognition through dual-view adaptive learning.

AI/ML arXiv cs.AI

PRISP: Privacy-Safe Few-Shot Personalization via Lightweight Adaptation

Introduces PRISP, a privacy-safe few-shot personalization framework for LLMs using a Text-to-LoRA hypernetwork.

Software Engineering Hacker News

Why I'm building a note taking app without AI

A developer discusses their decision to build a note-taking application without integrating AI features, focusing on traditional utility.

Other Hacker News

All 253 Patterns from Christopher Alexander's a Pattern Language Summarized

A summary of all 253 patterns from Christopher Alexander's 'A Pattern Language', providing an architectural and urban planning reference.

Other The Verge

Lego’s Donkey Kong arcade machine lets Mario jump endless barrels — Miyamoto is reportedly happy

Lego has released a Donkey Kong arcade machine set that includes a basic functional game mechanism.

AI/ML arXiv cs.AI

GSPRec: On Improving Item Representations in Graph Signal Processing for Collaborative Filtering

GSPRec is a graph spectral collaborative filtering framework that improves item representations by incorporating user interaction ordering.