maxbennett/generalized-neural-memory-iclft-personamem
The maxbennett/generalized-neural-memory-iclft-personamem is an 8 billion parameter Llama-3-8B-Instruct model, fine-tuned as an In-Context Learning (ICL) baseline for the PersonaMem benchmark. It processes documents directly within its 8192-token context window rather than relying on latent memory. This model is designed for research in controllable neural memory, specifically for evaluating how well models can learn from in-context information.
Loading preview...
Generalized Neural Memory — ICL-FT (PersonaMem)
This model, maxbennett/generalized-neural-memory-iclft-personamem, is an 8 billion parameter Llama-3-8B-Instruct variant specifically fine-tuned as an In-Context Learning (ICL) baseline. It was developed for the PersonaMem benchmark, as detailed in the research paper "Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language" (NeurIPS 2026).
Key Capabilities
- In-Context Learning (ICL) Baseline: Serves as a direct comparison for models utilizing latent memory by performing tasks purely through information provided within its 8192-token context window.
- Research-Oriented: Designed for evaluating and understanding how large language models process and utilize information presented in-context, particularly in the context of controllable neural memory.
- Llama-3-8B-Instruct Foundation: Leverages the strong base capabilities of the Llama-3-8B-Instruct architecture.
Good for
- Benchmarking: Ideal for researchers evaluating new neural memory architectures against a strong ICL baseline on the PersonaMem benchmark.
- Understanding Contextual Learning: Useful for studies focused on the limits and capabilities of in-context learning in LLMs.
- Reproducing Research: Provides the specific model used in Table 2 of the associated NeurIPS 2026 paper, enabling replication of experimental results.