Can a Language Model Learn Facts Continually in Its Weights?

2026-07-15 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

Research on Qwen3 models investigates whether language models can continually acquire and retain facts within their weights after initial training. The study found that while facts can be written into model weights, their long-term retention and accessibility are severely limited, particularly for bare-statement training. After twenty sequential writes, facts trained with bare statements retained only 1% accuracy, whereas facts from broad, diverse "study" data retained 46%. The core issue is not erased storage, as forgotten facts retain most of their log-probability lift, but rather lost access, where later writes redirect questions that previously reached earlier facts. The reliable channel for facts requiring composition or survival through many updates remains in-context learning, not weight-based continual learning.

Key takeaway

For AI Scientists and Machine Learning Engineers designing systems that require models to acquire and retain new facts, prioritize in-context learning for knowledge that must be composed or persist through many updates. When weight-based updates are necessary, use broad, diverse training data for each fact, as bare-statement training leads to rapid forgetting. Be aware that current weight-based methods struggle with fact reachability, even if the knowledge is technically stored.

Key insights

Continual learning in language model weights struggles with fact reachability, not storage, making context a more reliable channel.

Principles

• Broad training data creates usable, applicable knowledge in model weights.
• Forgetting in continual learning is primarily about lost access, not erased storage.
• Incoming writes cause interference, regardless of how earlier facts were stored.

Method

The study evaluates fact retention and usability in Qwen3 models using invented facts, five question types (recall, application, composition, paraphrase, counterfactual), and a dual strict/lenient grading policy to measure the "entailment gap."

In practice

• Employ diverse restatements in training data to reduce the recitation-to-use gap.
• Supply forgotten facts in the prompt to recover high accuracy (77–80%).
• Distill against a frozen original model to preserve general model capabilities.

Topics

• Continual Learning
• Language Models
• Knowledge Editing
• Catastrophic Forgetting
• Qwen3
• Fine-tuning
• In-context Learning

Code references

• basetenlabs/cortex

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Towards Continual Learning — Continual learning, crucial for human-like AI, involves updating models either through weight adjustments or external context/harness modifications.
• LMs as Task-Specific Knowledge Bases: An Interpretability Analysis — Language models encode factual knowledge in a task-specific manner, undermining the consistent "knowledge base" analogy.
• Can Scale Save Us From Plasticity Loss in Large Language Models? — Plasticity loss persists in LLMs, scaling sublinearly with size, suggesting scale alone won't prevent it.
• Reinforced Fast Weights with Next-Sequence Prediction — Next-sequence prediction via reinforcement learning improves fast weight models' long-context understanding.
• Why Limit the Residual Stream to Layers and Not Tokens? Persistent Memory for Continuous Latent Reasoning — AGCLR introduces a gated, persistent memory stream to prevent large language models from losing critical facts during continuous latent reasoning.
• MEMORY.md (C-engineering) is not real AI MEMORY - Different Math! — Current agentic memory systems are memos, not true learning, limiting generalization and skill acquisition.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.