Large Language Models (LLMs) - The Batch

Diagram showing threat actor using AI to find vulnerabilities and bypass two-factor authentication.

Machine Learning Research

Cybersecurity Alarms Grow Louder: Google study shows LLM-generated malware is getting harder to track and stop

An AI-generated script to bypass two-factor authentication signals a dawning era of industrial-scale cyberattacks, according to a Google report.

A woman in martial arts attire faces off against a cartoon lobster in a futuristic cityscape.

Machine Learning Research

Hermes Agent Challenges OpenClaw: OpenClaw created a class of personal agents; upstart Hermes Agent is outworking it

OpenClaw, the immensely popular AI agent, has fast-rising competition.

A graph shows assistant behavior shifting between helpful and role-playing, with conversation bubbles.

Machine Learning Research

Assistants That Assist Consistently: Large language models can drift drift from helpful personas to harmful ones, but new research aims to stabilize them

Typically, large language models are trained to act as helpful, harmless, honest assistants. However, during long or emotionally charged conversations, traits can emerge that are less beneficial. Researchers devised a way to steady the assistant personas of LLMs.

GLM-5.1 excels in SWE-Bench Pro and Terminal-Bench 2.0, leading in coding and reasoning tests.

Machine Learning Research

GLM 5.1 Aims for Long-Running Tasks: Z.ai’s GLM 5.1 evaluates interim results and may change its approach hundreds of times before it delivers final output

Z.ai updated its flagship open-weights large language model to work autonomously on single tasks for up to eight hours.

Image depicts persona generator creating synthetic personas, with outputs analyzed for diversity metrics.

Machine Learning Research

Simulating Diverse Human Cohorts: Persona generation simulates human characters across a controllable range of points of view

If you want to understand how the public will respond to your offerings, large language models can simulate users who answer questions about capabilities, features, promotions, or prices.

Table compares AI models' performance across benchmarks, showing Claude Mythos Preview leading.

Machine Learning Research

Claude Mythos Preview Raises Security Worries: Why Claude’s advanced Mythos Preview model will be limited-release-only

Anthropic took unusual steps to prepare the world for a forthcoming large language model that it said poses extraordinary risks to cybersecurity.

Two graphs show TTT-E2E maintains stable loss and latency across increasing context lengths up to 128k.

Machine Learning Research

Learning Long Context at Inference: Test-Time Training End-to-End (TTT-E2E) retrains model weights to handle long inputs

Large language models typically become less accurate and slower when they process longer contexts, but researchers enabled an LLM to keep accuracy stable and inference time constant as its context grew.

Top graph (blue) shows GPT-5 score drop; bottom graph (orange) shows RLM maintaining higher scores.

Machine Learning Research

Context As An External Variable: Recursive Language Models offer path to aramatically expand beyond the context window

When processing long contexts, large language models often lose track of details or devolve into nonsense. Researchers reduced these effects by managing context externally.

The chart compares Nemotron 3 models’ performance in accuracy and processing speed against other AI models.

Machine Learning Research

Open-Source Speed Demon: Nvidia’s open Nemotron 3 Super 120B-A12B model sets new paces in its class

Nvidia, the dominant supplier of AI chips, released a competitive open-source large language model whose speed tops its size class — the first open-weights leader to come from the United States since last year, when Meta delivered Llama 4.

Infographic on mobile AI use in 2025: 149B downloads, $167B revenue, 5.3T usage hours, 3.6 hours/day, 34 apps/month.

Business

AI on Mobile Skyrockets: State of Mobile 2026 Report shows AI chatbot, search, and assistant growth outpaces gaming, social, and more

Downloads of mobile AI apps and resulting revenue are surging.

Table shows GPT-5.4 outperforms in GDPval and Tau2-bench Telecom, setting new state-of-the-art scores.

Machine Learning Research

GPT-5.4’s Higher Performance, Higher Price: OpenAI’s GPT-5.4 Pro and GPT-5.4 Thinking challenge Google’s Gemini 3.1 Pro Preview as best all-around AI model

OpenAI updated its flagship models, extending the ability to use tools and setting the state of the art on a handful of benchmarks, and priced them at the top of the market. Its coding and agentic abilities have enabled Codex, OpenAI’s competitor to Anthropic’s Claude Code, to leap ahead.

Diagram depicts a math problem-solving workflow from problem generation to verification and revision.

Machine Learning Research

Agent Solves Stubborn Math Problems: Google’s Aletheia uses Gemini 3 Deep Think to find original mathematics solutions

LLMs have achieved gold-medal performance in math competitions. An agentic system showed strength in mathematical research as well.