Paper: concept circuits track how LLMs learn and forget during continual pre-training
tokenbender · x · 2026-08-24
A new arXiv paper, "How Do Large Language Models Learn Concepts During Continual Pre-Training?" (UC Davis, Virginia Tech, UCLA, Meta AI), studies how LLMs acquire, retain, and forget concepts during continual pretraining.
Key findings:
- Concept circuits — internal computational subgraphs tied to specific concepts — provide a non-trivial, consistent signal of concept learning and forgetting;
- Concept circuits show a stage-wise temporal pattern: early increase, gradual decrease, then stabilization;
- Concepts with larger learning gains tend to exhibit greater forgetting under subsequent training;
- Semantically similar concepts induce stronger interference than weakly related ones;
- Conceptual knowledge differs in transferability, with some concepts significantly facilitating the learning of others.
More from Research
- U of T team mines 122,194 patient reviews across 45 Ontario hospitals with NLP to expose satisfaction and equity patterns — ishtiaqueSIA · 2026-10-16
- University of Toronto hiring AI postdoc: builders of interventions, not just benchmarkers, due Oct 30 — ishtiaqueSIA · 2026-10-16
- WorldMemArena Accepted at NeurIPS: Memory Storage ≠ Memory Use — xwang_lk · 2026-10-16
- Mech Interp Veteran Rethinks Alignment, Misgeneralization, CoT Oversight — akbirthko · 2026-10-16
- WEDGE paper generates training environments conditioned on navigation agent failures — _akhaliq · 2026-10-16
- RExBench: best coding agent completes only 33% of autonomous research extensions — najoungkim · 2026-10-16