DEFT splits networks into frozen Core and trainable Slack to block forgetting
TheGradient · x · 2026-07-24
The mechanism in plain terms
The thread explains DEFT’s partitioning scheme: neurons with larger Hilbert-norm values are assigned to the frozen “Core,” while the rest become the trainable “Slack.” A merge step removes redundancy in the Core, and then an architectural mask blocks any path from Slack back into Core.
The claimed benefit
This creates a one-way valve for learning: new task updates can use the plastic side, but they cannot propagate backward to corrupt the stable foundation. The author presents it as a stronger guarantee against catastrophic forgetting than ordinary fine-tuning or soft regularization.
Related event: DEFT Introduces Core-Slack Architecture to Prevent Catastrophic Forgetting(3 posts)→
More from Research
- Kimi K3 may be strong on cyber, but token efficiency keeps it off UK AISIS — teortaxesTex · 2026-07-27
- ARC AGI 3 should have stayed private, with no examples or public dataset — flowersslop · 2026-07-27
- ExploitGym may have only 60–70% solvable tasks, fueling the OpenAI cheating debate — max_paperclips · 2026-07-27
- RTX 5090 local tests show Qwen Q6 can drop to 15 tok/s at 80k context — LFAdvice7984 · 2026-07-27
- Noahpinion quotes Chollet: intelligence may hit a hard ceiling — binarybits · 2026-07-27
- Paper argues graph topology can become the core operating system for AI agents — theomitsa · 2026-07-27