AI/ML arXiv cs.AI

Explaining Attention with Program Synthesis

A new method for reverse-engineering transformer attention heads by approximating their behavior using executable Python programs.

AI/ML arXiv cs.AI

Data Intelligence Agents: Interpreting, Modeling, and Querying Enterprise Data via Autonomous Coding Agents

Data Intelligence Agents (DIA) uses autonomous coding agents to automate the interpretation, modeling, and querying of enterprise data via SQL.

Other Hacker News

If your product is Great, it doesn't need to be Good (2010)

A discussion on product development philosophy emphasizing that a truly great product can succeed without needing to adhere to traditional 'good' design or marketing standards.

AI/ML arXiv cs.AI

Compute Efficiency and Serial Runtime Tradeoffs for Stochastic Momentum Methods

This research analyzes the trade-offs between compute efficiency and serial runtime in stochastic momentum methods for linear regression, proving bounds on batch-size tradeoffs.

AI/ML arXiv cs.AI

Language Models as Interfaces, Not Oracles: A Hybrid LLM-ML System for Pediatric Appendicitis

The authors present ClaMPAPP, a hybrid system that uses an LLM as an interface to extract features for a deterministic XGBoost classifier to improve pediatric appendicitis diagnosis.

AI/ML arXiv cs.AI

The More the Merrier: Combining Properties for ABox Abduction under Repair Semantics for ELbot

This paper explores ABox abduction under repair semantics for the ELbot knowledge representation language, focusing on hypotheses that satisfy multiple properties.

AI/ML arXiv cs.AI

Forecasting what Matters: Decision-Focused RL for Controlled EV Charging with Unknown Departure Times

A proposed decision-focused RL (DF-RL) framework for EV charging that trains a forecaster and controller end-to-end to handle unknown departure times.

Cybersecurity arXiv cs.AI

Machine Unlearning for the XGBoost Model with Network Intrusion Datasets

Introduces XGBoost-Forget, a machine unlearning approach for XGBoost models specifically designed for tabular network intrusion detection datasets.

AI/ML arXiv cs.AI

Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoning

Proposes MAST, a mechanism-guided method for selective unlearning in RLVR-induced reasoning to reduce collateral damage compared to full-parameter updates.

AI/ML arXiv cs.AI

STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability

Introduces STARE, a reweighting method for policy entropy stability in RL algorithms like GRPO to prevent entropy collapse during LLM post-training.

Other arXiv cs.AI

A Taxonomy of Mental Health and Technology Needs for Alzheimer's and Dementia Caregivers

A study introducing a taxonomy to link mental health needs of Alzheimer's caregivers with appropriate technology-based interventions.

AI/ML arXiv cs.AI

OneCanvas: 3D Scene Understanding via Panoramic Reprojection

Presents OneCanvas, a method for 3D scene understanding in VLMs that aggregates view features onto a single equirectangular panoramic canvas.

Other Hacker News

Everything Is BOM: Bill of Materials Encyclopedia

A resource providing a comprehensive encyclopedia of Bills of Materials (BOM), useful for understanding hardware components and supply chains.

Software Engineering Hacker News

Zero-Touch OAuth for MCP

A discussion or implementation of a Zero-Touch OAuth mechanism specifically designed for the Model Context Protocol (MCP).

Other Hacker News

Zork name origin got an update on Wikipedia

A minor update to the Wikipedia page regarding the origin of the name 'Zork'.

AI/ML arXiv cs.AI

Pareto Q-Learning with Reward Machines

Introduces Pareto Q-Learning with Reward Machines (PQLRM), a sample-efficient reinforcement learning algorithm for multi-objective tasks with non-Markovian rewards.

AI/ML arXiv cs.AI

A Technical Taxonomy of LLM Agent Communication Protocols

Presents a technical taxonomy for LLM agent communication protocols to address interoperability challenges in distributed agent networks.

AI/ML arXiv cs.AI

OrthoReg: Orthogonal Regularization for Hybrid Symbolic-Neural Dynamical Systems

Introduces OrthoReg, an orthogonal regularization method that prevents neural networks from absorbing the symbolic structure in hybrid symbolic-neural dynamical systems.

AI/ML arXiv cs.AI

AdsMind: A Physics-Grounded Multi-Agent System for Self-Correcting Discovery of Adsorption Configurations on Heterogeneous Catalyst Surfaces

Proposes AdsMind, a closed-loop multi-agent framework for the autonomous discovery of adsorption configurations on catalyst surfaces using MLFF relaxation feedback.

AI/ML arXiv cs.AI

Essential Subspace Merging for Multi-Task Learning

Presents Essential Subspace Merging (ESM), a training-free method for integrating multiple fine-tuned models into a single multi-task model while reducing interference.