AI/ML arXiv cs.AI

Generative Ontology Induction: Domain-Agnostic Schema Discovery from Document Corpora Using Large Language Models

Introduces Generative Ontology Induction (GOI), a domain-agnostic framework for automatically discovering schemas and exporting them as typed graphs in YAML/JSON using LLMs.

AI/ML arXiv cs.AI

Democratizing AI with Small Language Models: Structured Benchmarking and Parameter-Efficient Fine-Tuning for Local Deployment

Evaluates small language models (sub-3B parameters) for local deployment, demonstrating that a workflow of structured benchmarking and PEFT makes them viable for niche workloads.

AI/ML arXiv cs.AI

Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL

Proposes using Masked Diffusion Language Models (MDLMs) as steerable world models for RL, outperforming autoregressive models in coherence and diversity.

AI/ML arXiv cs.AI

It Takes 8 Tokens: Weak-to-Strong Off-Policy RL via Auxiliary Branches

Introduces W2SPO, an off-policy RL method that uses short auxiliary branches from weaker models to improve reasoning in 4B scale models and speed up training.

AI/ML arXiv cs.AI

PPO-HSC: An Exploratory Reinforcement Learning Framework Based on Wide-Area Policy Coverage Optimization

Introduces PPO-HSC, a framework that uses High-order Sampling Coverage to prevent mode collapse in LLM fine-tuning by rewarding semantic novelty.

Cybersecurity arXiv cs.AI

JUMP: Single-Pass Membership Inference on Fine-Tuned Diffusion Language Models

Presents JUMP, a single-pass membership inference attack for fine-tuned discrete diffusion language models that improves detection accuracy over previous methods.

AI/ML arXiv cs.AI

ColGraphRAG: Late-Interaction Evidence Retrieval for Multimodal GraphRAG

Introduces ColGraphRAG, which uses late-interaction multi-vector scoring (ColBERT/ColPali style) to improve evidence retrieval for multimodal GraphRAG.

AI/ML arXiv cs.AI

Shapley Context Pruning: A Cooperative Game Perspective for Context Reranking and Pruning

Proposes Shapley Context Pruning (SCP), a game-theory-based framework for reranking and pruning context in RAG systems for better efficiency and interpretability.

AI/ML arXiv cs.AI

A Survey on the Verification of Reinforcement Learning Policies

A comprehensive survey providing a taxonomy and unifying perspective on the verification of reinforcement learning policies for safety-critical domains.

Hardware/Chips Hacker News

Running Doom on Our Custom CPU and Going Viral

A project demonstrating the capability of a custom-built CPU to run Doom, achieving viral popularity.

Other Hacker News

A Koi Pond Mosaic Made from 10 Pounds of 3D Printer Waste

An artistic project creating a koi pond mosaic using recycled 3D printer waste.

Tech Business/VC Hacker News

Five US tech giants' hidden debts soar to $1.65T on opaque AI funding

Investigation into the massive hidden debts of five US tech giants linked to opaque funding for AI ventures.

Other Hacker News

A Mathematical Tribute to the Soccer Ball

A mathematical exploration and tribute to the geometry and properties of the soccer ball.

AI/ML arXiv cs.AI

Rater State Bias in RLHF Preference Data: An Audit Framework

An audit framework to identify 'rater state bias' in RLHF preference data, where annotator stress affects training signals.

AI/ML arXiv cs.AI

Design and Validation of a Lightweight 1D CNN for Affective Touch Classification in Soft Plush Companions

Development of a lightweight 1D CNN for affective touch classification in soft robotics, including an open-source MATLAB framework and dataset.

AI/ML arXiv cs.AI

Some Large Language Models Exhibit Consistent Risk Attitudes

Research showing that LLMs exhibit stable and consistent risk attitudes across different domains like finance and clinical triage.

AI/ML arXiv cs.AI

A Survey on GNN-based Link Prediction: Techniques, Applications, and Challenges

A comprehensive survey of Graph Neural Network (GNN) architectures used for link prediction in knowledge graphs and recommendation systems.

Cybersecurity arXiv cs.AI

PlanFlip: Attacking Multi-Agent LLM Systems via Planning-Phase Prompt Injection

Introduction of PlanFlip, a framework for attacking multi-agent LLM systems via prompt injection during the planning phase.

Software Engineering arXiv cs.AI

Deterministic Replay for AI Agent Systems

Presentation of agrepl, a Go-based CLI tool that provides deterministic replay for AI agent executions by intercepting external API interactions.

Other Hacker News

Is surveillance risk chilling your online speech?

A community discussion on whether the risk of digital surveillance is creating a chilling effect on online expression.