A new arXiv study finds gender bias is common and highly heterogeneous across ten LLMs from nine vendors, with models showing inconsistent patterns in both gender attribution and moral judgment tasks.

#LLMs #AIBias #AIResearch #ResponsibleAI
https://arxiv.org/abs/2609.38036

A factor analysis of 13,251 eval scores finds the "general factor" in language models explains far less variance than assumed. Domain-specific #AI #LLMs #MachineLearning #Psychometrics

https://freegardner.com/synapse/general-factor-in-language-models-lacks-support.html

Here are eight challenges of running AI agents with open-weight LLM models on local HPC hardware

agenticaiinsights.substack.com/p/deeper-div...

#agenticAI #hpc #gpu #computing #aiagents #llms #openweightmodels #management #managers

Love this blog post by @jonkeane.bsky.social about language and interacting with AI; something I've been thinking about a lot lately, the more time I spend my day talking with human-simulating outputs and less time talking with humans.

jonkeane.com/blog/languag...

#ai #llms

A new arXiv paper studies how to evolve the harness layer around fixed LLMs so personal agents better adapt to user preferences over time. It examines harness architecture, scale, and self-evolution algorithms through empirical and theoretical…

#AI #LLMs #Agents
https://arxiv.org/abs/2609.36892

OpenAI Slashes Prices While Chasing Astra

https://uenozooo.com/openai-slashes-prices-while-chasing-astra/

#OpenAI #AI #LLMs

New research traces a three-stage symbolic reasoning circuit in LLMs that's detectable well before high accuracy, with per-head causal contributions growing up to 8x from 1- to 10-shot. Function vectors can rescue…

#AI #LLMs #MachineLearning #Interpretability
https://arxiv.org/abs/2609.36265

A new arXiv paper introduces a framework for evaluating LLM agents on end-to-end reproduction of astronomy studies, separating execution failures from methodological ambiguity in the source papers. Across fourteen studies, eleven…

#AI #LLMs #ResearchReproduction
https://arxiv.org/abs/2609.35900

Perhaps the time has come to establish an alternative to the #arXiv, intended for depositing largely AI-derived results? An #aiXiv.

Papers posted to the arXiv could still make use of #LLMs. The aiXiv would be intended for articles where the *majority* of the result originates from an #AI.

#MathSky

Research examines whether multimodal LLMs can both generate and detect realistic social media fake news across science, health, and entertainment domains, using a multi-agent framework to produce over 9,000 paired posts and…

#AI #LLMs #FakeNews #Disinformation
https://arxiv.org/abs/2609.35809

New GeoOutageBench benchmark evaluates LLMs for geospatiotemporal knowledge graph question answering on multimodal power outage data, combining visual, textual, and structured sources. It tests ambiguity handling across difficulty levels…

#AI #LLMs #GeospatialAI
https://arxiv.org/abs/2609.36082

FlashSampling is an exact sampling primitive that fuses sampling into the LM-head matmul, avoiding materializing logits in HBM and replacing all-gathers with streaming peer-to-peer writes across up to 8 GPUs for tensor-parallel…

#AI #GPUs #LLMs #Semiconductors
https://arxiv.org/abs/2603.15854

CORTEX introduces a learning-dynamics framework that learns internal modularization within dense language models by partitioning trainable matrices into parameter groups based on gradient similarity. This approach lets models share…

#AI #LLMs #MachineLearning #NLP
https://arxiv.org/abs/2609.34449

LLMs appear to store substantial lexical knowledge, but new research examines how much of it they actually use during reasoning tasks like textual entailment. The study tests whether LLMs can supply the lexical relations needed…

#AIResearch #NLP #Entailment #LLMs
https://arxiv.org/abs/2609.32491

A study compared Gemini 2.5 Flash to manual extraction by health researchers for clinical findings in Brazilian Portuguese mammography and ultrasound reports, with the LLM reaching a Macro F1 of 0.91. The…

#AIinHealthcare #MedicalNLP #BreastCancerResearch #LLMs
https://arxiv.org/abs/2609.31974

Raven is presented as an open-source multi-agent ecosystem that automatically constructs and evolves modular harnesses for specific models and domains. It treats each executable model-harness pair as a composable unit of…

#OpenSourceAI #AIagents #LLMs #DevTools
https://arxiv.org/abs/2609.33439

RAG is still a practical way to connect #LLMs with enterprise information. But production #RAG is not simply a vector database sitting behind a prompt.
#GenAI #GenerativeAI #data #programming #development

https://bit.ly/4rymvxx

Nereus is a cost-aware runtime that adapts RL post-training LLM jobs to changing conditions like memory pressure and bottlenecks on GPU clusters. It handles transition costs, distributed state reuse, and GPU coordination to…

#semiconductors #GPUs #AIinfra #LLMs
https://arxiv.org/abs/2609.34645