The Gradient / Index
The reading room.
Search the stories, research, and sources from every issue.
93 editions · 461 stories
93 issues found
Long-Horizon State Tracking in LLMs: Executing MD5 through a Deep Sequence of Dependent Tool Calls
A new arXiv paper exposes how large language models lose grip on tasks that require many dependent steps, even when each individual step looks accurate.
Polimill builds Japan's next-generation public AI infrastructure
Polimill is rolling out a new AI‑powered infrastructure that lets Japanese local governments tap into their own administrative knowledge instantly.
Rating the Raters: Rasch Measurement Theory for LLM Evaluation
A new preprint proposes using Rasch Measurement Theory to untangle the complexities of evaluating large language models.
Sony and Warner accuse Anthropic of training Claude on tens of thousands of pirated works. Should the model be retrained from scratch?
Music publishers Sony and Warner allege that Anthropic built its Claude model using millions of illegally sourced works, sparking a debate over industry‑wide repercussions.
Google’s Gemini has a branding problem, and so does the rest of AI
The latest critique of consumer AI points out that confusing product names and hidden architectures are turning users away, and Google’s Gemini is a prime example.
Standalone LLM and a Pre-specified Agentic Pipeline for Explaining ICU Mortality Predictions: a Feasibility Study on the eICU Demo Dataset
A new feasibility study explores how large language models can turn raw ICU mortality predictions into bedside‑ready narratives.
RENDER: Controlling Reader-Facing Evidence in LLM Memory Evaluation
Researchers unveil RENDER, a novel benchmark that isolates the way memory evidence is shown to users in LLM‑driven conversations.
RENDER: Controlling Reader-Facing Evidence in LLM Memory Evaluation
A fresh benchmark called RENDER tests how the format of stored conversation snippets influences large language model answers.
Advancing price-performance for developers with GPT‑5.6 in Kiro
OpenAI’s latest model, GPT‑5.6, is now integrated into the Kiro platform, delivering faster and cheaper AI‑powered development tools.