Reimagining LLM Memory: Using Context as Training Data Unlocks Models That Learn at Test-Time | NVIDIA Technical Blog

AI-generated content may summarize information incompletely. Verify important information. Learn more We keep seeing LLMs with larger windows in , along with promises that they can hold entire histories, volumes of , or multiple codebases in view at once. And yet, these models still repeat the same mistakes. We still have to copy and paste the earliercontextback into the chat for LLMs to “get it”. A smart would pick up on these patterns, adapt, and carry the lessons forward. Why can’t LLMs? In this blog post, we observe a critical between LLM and humanmemory Then, we introduce test-time with an end-to-end formulation (TTT-E2E), our latest research, in which the LLM compresses it’s reading into its weights through next-token prediction. Our key results are highlighted in Figure 1, which measures scaling with context length, in terms of loss (left) and latency (right). Transformer with full attention scales well in terms of loss but not latency. Recurrent Neural Networks (RNNs), s

Reimagining LLM Memory: Using Context as Training Data Unlocks Models That Learn at Test-Time | NVIDIA Technical Blog Technical Blog Subscribe Related Resources Agentic AI / Generative AI English 한국어 中文 Reimagining LLM Memory: Using Context as Training Data Unlocks Models That Learn at Test-Time Jan 09, 2026 By Yu Sun and Yejin Choi Like Discuss (2) L T F R E AI-Generated Summary Like Dislike Test-Time Training with an End-to-End formulation (TTT-E2E) enables LLMs to compress long context into model weights via next-token prediction, outperforming both transformers with full attention and

saved by

• jonathan liu

related reading

• Recursive Language Models | Alex L. Zhangalexzhang13.github.io
• [2501.00663] Titans: Learning to Memorize at Test Timearxiv.org
• Why We Need Continual Learning | Andreessen Horowitza16z.com
• Extending Context is Hard | kaiokendevkaiokendev.github.io
• Andrej Karpathy — AGI is still a decade awaydwarkesh.com
• [2506.06266] Cartridges: Lightweight and general-purpose long context representations via self-studyarxiv.org
• Optimizing LLM Test-Time Compute Involves Solving a Meta-RL Problem – Machine Learning Blog | ML@CMU | Carnegie Mellon Universityblog.ml.cmu.edu
• Extending the Context of Pretrained LLMs by Dropping their Positional Embeddingspub.sakana.ai
• Getting Caught Up to Modern LLM Research | Samarth Goeldev.samarthgoel.com
• Mediumblog.gopenai.com
• Subquadratic — How SSA Makes Long Context Practicalsubq.ai
• [2605.15156] MeMo: Memory as a Modelarxiv.org