LLM 0.36 has been released with two new OpenAI models, gpt-6-sol and gpt-6-luna, and several improvements. Model plugins can now declare support for single-turn prompts, and reasoning traces in logs are now formatted with <details><summary> tags.

Researchers propose a new action tokenization method, M^2Tok, to improve the performance of Vision-Language-Action (VLA) models. The approach decomposes latent action features into multiple heads and assigns independent codebooks for quantization, leading to better reconstruction fidelity and higher success rates in VLA tasks.

A new protocol, DISCERN, is proposed to certify model updates without requiring labeled data. DISCERN uses unlabeled traffic to detect benign updates and an audited tier to label sampled disagreements. The protocol is proven to have finite-sample validity and matching label-complexity bounds, and is demonstrated on 14,000+ replayed audit streams across 785 update pairs.

Researchers propose T2T-VICL, a framework for cross-task visual in-context learning using large vision-language models. This allows VLMs to perform visual tasks by conditioning on demonstrations and queries from different tasks without explicit task naming. The framework uses a teacher VLM to generate structured descriptions and a student VLM to produce content-dependent prompts for image-editing VLMs.

Researchers have introduced MASA, a novel approach to wild test-time adaptation that uses a frozen multimodal large language model to provide structured semantic descriptions for object families and nuisance factors. This helps to break the self-referential loop in model adaptation and improves accuracy in limited-batch, mixed-domain, and imbalanced-label-shift settings.