large language models

EvoLLM v3.0 — Phase 3 registered mixed-model tournament: receipts, analyses and post hoc audit

Replication package for Phase 3 of the EvoLLM project: a tournament between 572 strategy identities written by eight large language models (iterated prisoner's dilemma, 1,000 rounds, no noise), registered on OSF (https://osf.io/2nzdh) before execution. It contains the frozen protocol commit 18cb41744f3f1e66fe711aad0969...

Pattern Over Pixels: Measuring Pattern Completion Bias in Multimodal Code Generation

Multimodal large language models (MLLMs) are increasingly usedto translate webpage screenshots into front-end code, but repeatedUI patterns may sway them toward visually incorrect yet pattern-consistentoutputs. In this work, we test how repeated webpagepatterns hurt MLLM accuracy on an objective screenshot-to-codefill-...

Climate Knowledge in Large Language Models

This evaluation framework provides a reproducible benchmark for quantifying parametric climate knowledge in LLMs, and shows that LLMs reproduce broad geographic regularities in climatological temperature fields, capturing latitudinal and topographic patterns.

AS2Biz: Leveraging Web Presence and AI to Improve AS Business Classification

Classifying Autonomous Systems (ASes) by the business sectors of their operating organizations, also referred to as AS business classification, is a relatively recent but increasingly important area of study. However, the only publicly available dataset for this purpose suffers from significant inaccuracies and limited...

Measuring AI R&D Automation Before Intelligence Explosion: A Governance-Grade Metric Framework for Loop Closure, Oversight Latency, Verification Independence, and Authority Distance

This paper develops a governance-grade measurement framework for AI research and development automation under conditions in which increasingly autonomous AI systems may participate in, lead, or close progressively larger portions of the AI R&D loop. Existing measurements of AI R&D automation provide an important starti...

ToshLLM: local LLM inference on Intel Macs with AMD GPUs

A native SwiftUI application that runs large language models locally on Intel Macs with AMD GPUs, a configuration mainstream inference stacks leave unsupported or incorrect. Beyond packaging, it contributes original work to the Metal backend of llama.cpp: ToshGEMM, a manually tiled matrix multiply that replaces the sim...

Neutrino Oscillation Fixes Only the Differences of the Squared Masses, Not the Masses Themselves ── Shifting every squared mass by the same amount leaves every probability unmoved and moves only the shape of the endpoint ── the separator is whether one measures oscillation or the β-decay endpoint ── [Paper 1071]

A neutrino changes its type while it flies. The formula for that probability contains only differences of squared masses, never the masses themselves. This paper counts that "differences only" with three-flavour mixing and shows that the quantity answering to the absolute scale instead is the shape of the β-decay endpo...

Large-language-model-based white spot analysis of patent landscapes in emerging battery technologies

Grand Theft Autocomplete: 'AI' enclosures and their threat to data sovereignty

The way we distribute research data and software today relies heavily on the open web: largely standardised formats, decentralised data hubs, public availability, affordable compute, free learning resources, and findability afforded by a few ad-supported web indexes. This openness, however, is increasingly under threat...

GNR-Q: Inference-Time State-Conditioned Reconstruction for Memory-Efficient Quantized Language Models

Low-bit quantization substantially reduces the memory footprint of large language models(LLMs), but the associated loss of numerical precision can distort internal representations anddegrade downstream prediction quality. We investigate whether part of this lost representationquality can be reconstructed dynamically at...

Grand Theft Autocomplete: 'AI' enclosures and their threat to data sovereignty

The way we distribute research data and software today relies heavily on the open web: largely standardised formats, decentralised data hubs, public availability, affordable compute, free learning resources, and findability afforded by a few ad-supported web indexes. This openness, however, is increasingly under threat...

GNR-Q: Inference-Time State-Conditioned Reconstruction for Memory-Efficient Quantized Language Models

Low-bit quantization substantially reduces the memory footprint of large language models(LLMs), but the associated loss of numerical precision can distort internal representations anddegrade downstream prediction quality. We investigate whether part of this lost representationquality can be reconstructed dynamically at...