All Articles
17528 articles total
Pitwall: Faithful Natural-Language Race-Strategy Briefings from a Calibrated Real-Time Monte Carlo Engine
Pitwall is a production system for Formula 1 strategy briefings that ensures faithfulness by verifying generated claims against a calibrated Monte Carlo engine.
Industry Classification of GitHub Repositories Using the North American Industry Classification System (NAICS)
The creation of NAICS-GH, a corpus that maps GitHub repositories to standardized North American Industry Classification System (NAICS) sectors.
RSF-GLLM: Bridging the Semantic Gap in Multi-Hop Knowledge Graph QA via Recurrent Soft-Flow and Decoupled LLM Generation
RSF-GLLM is a framework that decouples differentiable graph reasoning from LLM generation to improve multi-hop knowledge graph question answering.
Graph Convolutional Attention: A Spectral Perspective on Graph Denoising and Diffusion
Introduction of Graph Convolutional Attention (GCA), a mechanism that improves graph denoising and diffusion by utilizing the input graph spectrum.
ELSA3D: Elastic Semantic Anchoring for Unified 3D Understanding and Generation
ELSA3D is a unified 3D foundation model using elastic semantic anchoring to improve both 3D asset generation and linguistic reasoning about 3D geometry.
Harnessing Code Agents for Automatic Software Verification
The researchers introduce Aria, a harness that uses LLM code agents like Claude Code to automatically prove lemmas in theorem provers like Coq and Lean 4, achieving full coverage without expert intervention.
Responsible Personalisation: The Double-Edged Sword of Personalisation in Human-Robot Interaction
The paper presents a lifecycle-based framework to analyze and mitigate ethical risks, such as autonomy erosion and privacy violations, in personalized human-robot interaction.
What Images Cannot Say: Language-Guided Olfactory Representation Learning
SCENT is a multimodal framework that uses language guidance to align olfactory signals from electronic noses with visual and textual representations, improving smell-to-image retrieval.
RuBench: A Repository-Level Agentic Coding Benchmark with Natively Authored Russian Task Specifications
The authors introduce RuBench 1.0, a benchmark for repository-level agentic coding using natively authored Russian task specifications to evaluate the performance of coding agents in non-English contexts.
An Experimental Design Approach to Evaluating Agentic AI's Autonomous Model Discovery
The paper proposes an experimental design framework to systematically evaluate the autonomous model discovery behavior of stochastic coding agents like Codex and Claude Code.
TILDE: TILt-based Distributional Erasure for Concept Unlearning
TILDE is a distributional erasure method for concept unlearning in text-to-image diffusion models, aiming to suppress unwanted concepts while preserving the quality of benign generation.
Analysis-by-Proxy: Localization Signals in VLMs Operating as Condition Encoders
The study introduces Analysis-by-Proxy, a framework to uncover how spatial knowledge is represented in VLMs when used as condition encoders for image editing pipelines.
From Voting to Agent Collaboration: Answer-Type-Aware LLM Pipelines for BioASQ 14b
The researchers developed a question-type-specific LLM framework for biomedical QA that uses different inference procedures for yes/no, factoid, and list questions to improve robustness.
Provable learning separation for predicting time-evolution of quantum many-body systems
The paper demonstrates a provable learning separation for predicting the time-evolution of quantum many-body systems, showing certain tasks are quantum-learnable but classically hard.
Prompt-Adapter Context Routing for Parameter-Efficient Multi-Shot Long Video Extrapolation
PACR-Video is a parameter-efficient framework for long video extrapolation that uses low-rank temporal adapters and a recursive prompt bank to maintain visual and narrative coherence.
A software engineering interview question I like: computing the median
A discussion on a favorite software engineering interview question focused on the efficient computation of a median.
Patching MechCommander's "left arm bug" for fun and profit
An exploration of patching a specific bug in the game MechCommander for personal interest and technical challenge.
Show HN: Frugon – Find which LLM calls a cheaper model could handle (local, MIT)
Frugon is a tool designed to identify which LLM calls can be handled by cheaper models to optimize costs.
UBEP: Re-architecting Expert Parallelism Communication Library for Production Superpods
Introduces UBEP, a communication library that optimizes Mixture-of-Experts (MoE) model communication on high-bandwidth superpods.
Spider 2.0-AIFunc: Extending Real-World Text-to-SQL to AI-Native SQL Workflows
Presents Spider 2.0-AIFunc, a benchmark for evaluating LLMs' ability to generate AI-native SQL workflows on the Snowflake platform.