Resolution criteria:

Resolves YES if, before January 1, 2028, any model ranked in the top 10 on LMArena (lmarena.ai, Style Control OFF) has a publicly available version with demonstrated weight-based online learning capability, meaning the model updates internal parameters during inference or between user sessions to incorporate new information.

The model must:

1. Be ranked top-10 on LMArena at any point while offering this capability

2. Have the online learning feature available to end users (via API or direct access), not just as a research demo

3. Encode new information through learned parameters, not context injection

Background:

Current LLMs are static after training. To handle new or user-specific information, providers use workarounds:

Context-based approaches (don't count):

• RAG (retrieval-augmented generation): fetching relevant documents into the prompt

• Long context windows: fitting more information into a single conversation

• Conversation memory systems: storing and retrieving past exchanges as text

These provide information as input rather than changing the model itself. They're limited by context length and don't alter how the model computes its outputs.

Weight-based online learning (counts for YES):

• Test-Time Training: layers with internal parameters updated via gradient descent during inference (paper)

• Neural memory modules: learnable parameters that adapt to memorize information at test time, e.g., Titans (paper)

• Adapter-based continual learning: LoRA-style parameter updates between sessions to incorporate new information

These approaches modify model parameters to encode new information, enabling learning from experience rather than just retrieval.