Multiclass aggregation rates freed from class-count penalty
New theoretical results show that averaging class probabilities, not score functions, keeps multiclass aggregation optimal as the number of classes grows.
Research Digest
Preprints and papers, rewritten in plain English.
Brain Decoding Shows Shared Semantic Geometry Across Languages
A new multilingual brain decoding study suggests that different languages map onto a common cortical semantic structure.
Small Language Models Fine-Tuned for Reliable VASP INCAR Generation
A new preprint shows that fine-tuning compact language models could make physics-sensitive VASP INCAR generation more reliable without relying on large proprietary cloud models.
Rubrics Give LLM Agents Dense Rewards in Non-Verifiable Tasks
A new arXiv paper proposes using expert-written rubrics as dense reward signals to train LLM agents in domains where outcomes can't be automatically verified.
New framework measures value structure in AI-generated debates
A proposed diagnostic framework aims to shift evaluation of AI debate from argumentative quality to underlying value structures.
Predator-Prey Model Accelerates Grokking in Neural Networks
A new machine learning method borrows predator-prey dynamics to speed up the delayed learning phenomenon known as grokking.
XRust Targets Logical Races in Rust Linux Drivers
The paper introduces a tool to detect logical races that Rust's memory safety guarantees do not prevent.
Study Characterizes Software Evolution After Incidents
A new arXiv paper examines how software systems change after incidents to inform reliability improvements.
AI-Assisted Defect Analytics Aim to Prevent Recurring Software Bugs
An industrial case study explores using AI to turn historical defect data into actionable prevention strategies.
Reinforced Linear Genetic Programming Targets LGP Limitations
A new arXiv paper applies reinforcement ideas to linear genetic programming to address the technique's known shortcomings.
Accounting for Selection Bias in Differential Privacy Audits
A new arXiv paper highlights how selection bias can distort MIA-based estimates of differential privacy guarantees.
Evolutionary semi-supervised learning for smart-city classification
A new arXiv paper proposes combining evolutionary multi-objective search with semi-supervised learning to classify smart-city data when labeled examples are scarce.
Better Bug-Exposing Tests via Step-Level Credit Assignment
A new arXiv paper proposes a method for generating tests that expose subtle bugs in LLM-written code by learning from valuable failures.
Benchmarking framework tackles RAG configuration complexity
A new modular, non-invasive framework aims to automate analysis of retrieval-augmented generation configurations.
MIMOS-Tools Aims to Unify Model-Based Embedded Systems Design
A new toolchain proposes to cover the entire development lifecycle for embedded real-time systems, a gap that persists despite decades of model-based methods.
RELAY system aims to restore historical software builds at scale
A new arXiv paper proposes combining rule reuse with agentic repair to rebuild old software that depends on drifting external resources.
Training coding agents should update the harness, not just the policy
A new arXiv paper argues that fixing the harness during coding-agent training misses interaction problems that emerge as the policy learns.
Plant-Aware Agent Aims to Ease PLC Programming
A new preprint proposes a retrieval-augmented agent that treats PLC code as part of an engineering workflow, not just a code generation task.