Brief

A new research paper introduces cooperative memory paging, a technique designed to help Large Language Models (LLMs) manage conversations that exceed their context window. This method replaces evicted conversation segments with concise keyword bookmarks, allowing the LLM to retrieve full content using a recall tool when necessary. Experiments on the LoCoMo benchmark demonstrated that cooperative paging outperformed other methods in answer quality across multiple LLMs, though the effectiveness was significantly impacted by the distinctiveness of the generated bookmarks. AI

A research paper titled "Depth Registers Unlock W4A4 on SwiGLU: A Reader/Generator Decomposition" was withdrawn from arXiv. The paper explored post-training W4A4 quantization on a 300M-parameter language model, aiming to reduce perplexity errors. It introduced a method called Depth Registers with a hinge loss, which significantly improved quantization results but still left a small gap compared to FP16. AI

A new research paper proposes a commit-open protocol to detect when hosted large language model providers substitute cheaper models for advertised ones. The protocol uses Merkle trees to commit to sparse autoencoder (SAE) feature traces of model outputs, allowing verifiers to detect such substitutions. Experiments on Qwen3-1.7B, Gemma-2-2B, and a scaled-up Gemma-2-9B demonstrated the protocol's effectiveness in rejecting various substitution attacks, outperforming existing methods like SVIP. AI