Generalized Neural Memory (NeurIPS 2026)
Collection
Models, benchmarks, and eval records for 'Tell Me What To Learn' (NeurIPS 2026). • 9 items • Updated
Generalized Neural Memory — GNM (CounterFACT)
The paper's main GNM model for the CounterFACT benchmark (Table 1; Figures 3 via the long-horizon serving stack): a MemoryLLM (Llama-3-8B) trained with instruction-conditioned memorization to selectively learn facts, response formats, and refusal policies from a document stream.
From "Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language" (NeurIPS 2026).
Usage
Use through the code repo:
python scripts/download_models.py --counterfact # places this model at models/gnm_counterfact/model.pth
Registry key in the repo: memoryllm_chat. See the repo README for the full CounterFACT
evaluation commands (Table 1) and training configs.
- Downloads last month
- 314
Model tree for maxbennett/generalized-neural-memory-gnm-counterfact
Base model
meta-llama/Meta-Llama-3-8B-Instruct