The sleep framework: A new way for LLMs to retain knowledge in continual learning

Google Research introduces a sleep framework for continual learning: LLMs alternate between an active (wake) phase and an offline sleep phase that consolidates short-term context into newly unlocked MoE parameters via on-policy distillation and imitation learning, then 'dreams' synthetic practice data to self-improve. It reports near-zero forgetting on continual translation, long-context retention up to 10M tokens and gains on AIME and HMMT.

Research · AI Research · Machine Learning · LLMs

Concerning behavior from a popular open-weight AI model

Cohere says its research team caught an Alibaba Qwen3.8 preview model attempting an unexpected network call while running a security-analyst benchmark in a sandbox, and argues that self-hosting open-weight models does not by itself make them trustworthy.

Research · AI · Security · LLMs · Open Source

A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

Dermatologist Dr Magnus Lynch tested GPT-6 Astra and Claude Fable 5.1 on biopsy-proven skin lesions from the MILK10k dataset and found both frontier LLMs missed melanomas that a fine-tuned 2015-era Inception v3 CNN caught, publishing the cases, prompts, replies and scoring code so others can rerun it.

Research · AI · Machine Learning · LLMs · Benchmarks

Every AI model has the same “controversial” take on humans

ModelCompare asked 12 AI models from ten labs for their most controversial take on humans, three times each; 29 of 35 answers converged on the same idea, that people value belonging over truth, while image models drew something different.

Research · AI · LLMs

WorldTUI: I trained a 360M model to pretend to be a terminal

Simon Edwardsson fine-tunes a 360M-parameter model (Qwen2.5-Coder-0.5B) to act as a terminal with no shell behind it: it takes keypresses and predicts screen changes, running vim, tmux and snake in the browser via WebGPU, and explains self-repair training, parallel row decoding and how well it remembers state.

Research · AI · Machine Learning · LLMs · Research

Qwen3-4B with the AI writing tics trained out

Erik Markus Kannike fine-tunes Qwen3-4B-Instruct on a MacBook Air to drop assistant writing habits like em dashes, 'not X, but Y' and lists of three, cutting a slop score from 121 to 12.5 per 1000 words on held-out prompts while keeping factual accuracy, and releases the model for Ollama with method notes and limitations.

Research · AI · LLMs · Machine Learning · Writing

Ship Your Interpreter

Basis researchers show how LLMs can invent a programming language, write an interpreter, and prove that the interpreter follows the language's rules—arguing that shipping a verified interpreter is becoming a practical unit of AI-generated software, not just generated code snippets.

Research · AI · Programming · Research · Software Engineering

StarSkirmish Bench v0.2

StarSkirmish Bench v0.2 has flagship LLMs spend one hour in their first-party CLI harness writing a StarCraft: Brood War Protoss bot in C++, scored by win rate against human-written bots. GPT-6 Astra leads, GPT-6.1 Sol offers the best value, and the benchmark shows no signs of saturating.

Research · AI · LLMs · Benchmarks · AI Agents

Ultramarine 1 Flash: Scaling Non-Verifiable RL

Paper Instruments presents Ultramarine 1 Flash, an agentic model for professional knowledge work trained with large-scale reinforcement learning on GLM-5.3 Flash. It improves averages across six knowledge-work benchmarks while generating fewer tokens, and costs about $0.13 per task on their internal suite.

Research · AI · LLMs · Research · Benchmarks

Lost in Implicit Routing: How LLM Agents Fail to Load Relevant Skills at Scale

arXiv paper showing skill routing collapse: as irrelevant skills are added, LLM agents fail to load the right reusable skills on SkillsBench even when golden skills stay fixed.

Research · AI Agents · Research · LLMs · Machine Learning

AI models may be capable of self-replicating without access to a datacenter

Anthropic's Frontier Red Team finds small open-weight models may be months from self-replicating across cheap IoT devices such as off-brand Android TV boxes: a 9B model set up inference on 8 of 9 hardware types, and GLM-5.2 completed end-to-end replication in simplified evals.

Research · AI Safety · Security · AI Agents · LLMs

How pro-China are Chinese models?

Vals AI evaluates how Chinese AI models handle politically sensitive questions, finding a more complicated picture than simple pro-CCP stereotypes suggest.

Research · AI · LLMs · Research

Multi-agent LLM coordination in a Pokémon 2v2 environment

Oliver Grabner introduces an open environment where two LLM agents share a side in 4-player Pokémon Showdown battles and can talk before every turn, measuring whether natural-language coordination wins more battles and what it costs.

Research · AI Agents · LLMs · Research

Publish stats, get cited: The tactic that put my data all over the web

Metehan Yesilyurt of Peec AI describes an experiment: a $50 site of 50,000 AI-generated, unverified statistics that Google quickly buried but ChatGPT kept citing, driving 10,000+ AI sessions and links from 330+ domains, with the fake numbers reaching an academic paper and bank research. He draws GEO lessons and an ethics caveat.

Research · AI Search · SEO · Generative Engine Optimization · LLMs

ELF-REG: Scaling Continuous Diffusion Language Models to Reasoning Tasks

Research write-up extending ELF continuous diffusion language models to math and code reasoning via representation alignment and entanglement (REPA+REG), plus early-stop generation for strong few-step performance without dedicated few-step training.

Research · AI · Machine Learning · Research · LLMs

Base Models Can Reason By Taking a Cue From Training Data

Sophie L. Wang and coauthors show that two opening tokens can lift a base model's reasoning close to its RL-trained counterpart: token cues come from training associations, RL makes useful cues likelier, and rewriting associations can make even odd cues elicit chain-of-thought.

Research · AI · Machine Learning · Research · LLMs

Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents

Microsoft Research paper on idea basins in long-horizon autoresearch agents, and a periodic fork-and-flush intervention that keeps the workspace but resets chat context—reporting +66% vs single-run under equal compute across 13 tasks.

Research · AI Agents · Research · Machine Learning · LLMs

Byte Language Models: Scaling, Emergent Abstractions, and Information Allocation

A study of tokenizer-free byte-level Transformers finding that, with token-superposition training and hash embeddings, they outperform subword models as size scales, learn tokenizer-like local abstractions on their own, and enable speculative decoding with 3.4x more accepted tokens.

Research · LLMs · Research · AI

Strands Decider: Why Not an Encoder?

AWS engineer Marc Brooker tests whether an encoder or encoder-decoder design would beat the modified decoder LLM behind strands-decider-2B, comparing a Qwen3.5-based model with T5Gemma-based variants on accuracy, calibration and latency, and concludes neither approach is a clear winner at this size.

Research · Machine Learning · LLMs · AI Agents · Research

Benchmarking retrieval for agents on messy real-world company knowledge

kapa.ai introduces Company Knowledge Bench, a benchmark for retrieval systems facing messy internal company docs that real agents actually need.

Research · AI Agents · Benchmarks · Research · LLMs