AI/ML arXiv cs.AI

Mathematics of Data Science

A comprehensive mathematical foundation for data science covering topics from SVD and PCA to deep learning and matrix concentration inequalities.

AI/ML arXiv cs.AI

CARE-LoRA: Compressed Activation REconstruction for Memory-Efficient LoRA

CARE-LoRA is a memory-efficient fine-tuning framework that reduces the activation memory bottleneck in LoRA by using compressed activation reconstruction.

AI/ML arXiv cs.AI

How Query Visibility Changes KV-Cache Compression Rankings: A Matched-Budget Audit

An audit of KV-cache compression methods reveals that rankings shift significantly when compression is query-agnostic versus query-aware.

AI/ML Hacker News

I tricked Claude into leaking your deepest, darkest secrets

An exploration of prompt injection or jailbreaking techniques used to trick Claude into revealing sensitive information.

Hardware/Chips Hacker News

RISC-V Is Inevitable: State of the Union Keynote Argues

A discussion on the inevitability and growth of the RISC-V open-standard instruction set architecture.

AI/ML arXiv cs.AI

So Many Opinions, So Many LLMs: Comparing Large Language Models to Traditional Machine Learning for Open- Ended Survey Analysis

A study comparing LLMs to traditional machine learning for open-ended survey analysis, finding LLMs have higher accuracy but lower consistency.

AI/ML arXiv cs.AI

CANDI: Contextual Alignment for Niche Domains Question Answering

Introduction of CANDI-QA, a dataset for evaluating LLMs in niche domains, and MTSS-Net, a neuro-symbolic framework for baseline QA.

AI/ML arXiv cs.AI

G-SHARE: A Guideline-Based Structured Reasoning Framework for Human-Factor Event Diagnosis

The G-SHARE framework operationalizes expert diagnostic guidelines into a multi-stage pipeline for human-factor event diagnosis in nuclear power plants.

AI/ML arXiv cs.AI

I'm Sorry, but I Can't Help with Braille: Revealing Accessibility Failures in State-of-the-Art LLMs

Research revealing the failures of state-of-the-art LLMs in Braille translation and the effectiveness of task-specific fine-tuning of smaller models.

AI/ML arXiv cs.AI

Graph-Based Detection of Disinformation Narrative Diffusion between Russian and Ukrainian Telegram Channels

A graph-based framework for detecting disinformation narratives by analyzing propagation across Telegram channels using weak supervision.

AI/ML arXiv cs.AI

OmniPMNet: Bridging discrete and gridded PM10 forecasts via omni-query neural processes

OmniPM-Net is a fusion model using Convolutional Conditional Neural Processes to reconcile discrete station-scale and gridded PM10 air quality forecasts.

AI/ML arXiv cs.AI

SeqGPT: A Constrained Transformer Agent for the Inverse Designof Multi-Panel Composite Structures

SeqGPT is a constrained Transformer agent using neurosymbolic decoding to optimize composite stacking sequences for structural design.

AI/ML arXiv cs.AI

Towards Self-Evolving Agents: A Human-Inspired Adaptive Exploration-Exploitation Framework for Genetic Network Programming

Introduction of HGNP, a human-inspired adaptive exploration-exploitation framework for Genetic Network Programming to evolve agentic AI strategies.

Other Hacker News

LinkedIn is a cesspool of scammers and identity theft

A discussion thread on Hacker News regarding the prevalence of scams and identity theft on LinkedIn.

AI/ML arXiv cs.AI

FormalAnalyticGeo: A Neural-Symbolic Based Framework for Multimodal Analytic Geometry Problem Generation

Introduction of FormalAnalyticGeo, a framework for automatically generating multimodal analytic geometry problems and the accompanying AnalyticGeo7K dataset.

AI/ML arXiv cs.AI

Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs

A research paper proposing a method to make LLMs more incentive-compatible by learning counterfactual report mediators to resist user pressure.

AI/ML arXiv cs.AI

Win by Silence: Deletion Non-Monotonicity, Autonomous Exploitation, and Typed-State Gating in LLM Plan Evaluation

An analysis of 'omission incentives' in LLM plan evaluation, where evaluators reward plans for being less explicit, and the introduction of PCSC to neutralize this.

AI/ML arXiv cs.AI

Dynamic Resource Allocation for Ensemble Determinization MCTS

Proposed enhancements for Ensemble Determinization MCTS using dynamic resource allocation for better performance in adversarial board games.

AI/ML arXiv cs.AI

Audio-Native Speech Recognition with a Frozen Discrete-Diffusion Language Model

Development of an audio-native speech recognition system using a frozen discrete-diffusion language model (DiffusionGemma) and a Whisper encoder.

AI/ML arXiv cs.AI

Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution

The E3 (Estimate, Execute, Expand) framework aims to reduce LLM agent redundancy by estimating task complexity before execution.