Small open-source LLMs specialized for SQL can outperform general-purpose chat models on Text-to-SQL tasks, and lightweight iterative retry workflows can stabilize results without increasing model size.

#OpenSourceAI #TextToSQL #LLMs #AIResearch
https://arxiv.org/abs/2610.17305

New arXiv research questions whether LLMs can create genuinely new research directions or only extend existing knowledge, distinguishing "distance" from "direction" in hypothesis novelty. It finds human reviewers often rate…

#AIresearch #LLMs #ScientificDiscovery
https://arxiv.org/abs/2610.17439

How #LLMs Deal With Fraudulent Papers (via @science.org)
www.science.org/content/blog...

#GenAI #AI #scholcomm #libraries

ICYMI: Facebook feed now takes typed requests as Meta adds LLMs to ranking #FacebookFeed #Meta #LLMs #SocialMedia #DigitalMarketing

ICYMI: Facebook feed now takes typed requests as Meta adds LLMs to ranking #FacebookFeed #Meta #LLMs #SocialMedia #DigitalMarketing

A new arXiv paper proposes a source-aware taxonomy for agentic workflow failures, mapping issues along workflow stages and agent capabilities using academic and GitHub evidence. The two-axis localization matrix aims to make hard-to-diagnose LLM…

#AI #LLMs #Agents
https://arxiv.org/abs/2610.17418

Glasseek is a fully transparent open-source framework for training and evaluating deep search agents, releasing data, model weights, training recipes, and code to address opaque pipelines, training-inference mismatches,…

#OpenSourceAI #AIagents #DeepSearch #LLMs
https://arxiv.org/abs/2610.16998

# Satire, not a standard, not read by anything:

cats.txt: cultureddigital.co.uk/cats.txt
dogs.txt: cultureddigital.co.uk/dogs.txt
fish.txt: cultureddigital.co.uk/fish.txt

#llms #aeo #geo #aeieo

New research proposes Perception-Aware Adversarial Self-Play (PASP) to help LLMs detect harmful intent disguised through phonetic, visual, or contextual cues that humans naturally pick up on. It highlights how text-only models…

#AI #Cybersecurity #LLMs #AISafety
https://arxiv.org/abs/2610.15449

The Fallacy of Data Driven Marketing:
idea2product4profit.substack.com/p/the-fallac...

#TECHCRUNCH #Entrepreneurs #programming #coding #webdev #DataScience #startup #leanstartup #stealth #AI #MachineLearning #NLP #web3 #fintech #deeplearning #gamedev #LLMs #chatbots

Microsoft's Confidence Scores Just Admit Your LLM Still Lies

https://uenozooo.com/microsofts-confidence-scores-just-admit-your-llm-still-lies/

#Microsoft #AI #LLMs

Why is every conversation with #LLMs…

LLM: ok here’s our agree upon workflow.
Me: yup, good. do it.
LLM: ok this is what I’ll do.
Me: yeah, I know so do it then.
LLM: done.

Why does it always need to be asked twice?

A new arXiv paper proposes optimizing a task-adapted recipe space to automate LLM post-training, letting agents adapt their choices to target tasks rather than searching a fixed space. The approach aims to make systematic recipe reuse and…

#AI #LLMs #Robotics
https://arxiv.org/abs/2610.18572

New arXiv paper proposes Token Window Pacing (TWP), a closed-loop scheduler that uses runtime ITL, queue depth, and workload feedback to admit serverless LLM requests at the token level, balancing TTFT and inter-token…

#semiconductors #AIinfrastructure #GPUs #LLMs
https://arxiv.org/abs/2610.17689

Echo is a 4-bit all-reduce method for tensor parallelism in LLM training and inference that quantizes each value at most twice per call and feeds injected errors back into later calls to prevent accumulation.

#AI #LLMs #GPUs
https://arxiv.org/abs/2610.14277

New arXiv work introduces Big-Big-SycoBench, a 450-artefact, 5,400-argument benchmark that measures LLM sycophancy by testing whether models revise judgments only in response to valid criticism, not fallacious or emotional pressure.…

#AIresearch #LLMs #Benchmark
https://arxiv.org/abs/2610.15634

New arXiv analysis of 367,174 MCP tools finds substantial parameter-schema variation among functionally similar tools, and controlled metadata interventions across five LLMs show schema differences can shape tool selection before any call is made.

#AI #LLMs #MCP
https://arxiv.org/abs/2610.14334

Jev-Guard is a new guardrail model that routes agent actions across editable policy trees spanning content moderation, cybersecurity, privacy, and legal compliance. It reports 93.72% safety accuracy on 7,033 evaluation…

#AIsafety #Cybersecurity #LLMs #AgenticAI
https://arxiv.org/abs/2610.18326

ClawLoop is a new framework from arXiv that lets harness-based LLM agents self-improve through iterative training, rather than relying on one-shot pipelines driven by a stronger external model. It targets agents that work directly with…

#AIresearch #LLMs #Agents
https://arxiv.org/abs/2610.14118

I've been tinkering with a new piece of software I'm calling Conveyor, that provides creative tooling for 'painting' with agents. The artist controls scope, prompts, brushes, style, etc. The agent puts pixels on the canvas.

#LLMs #AIArt