Thomson Argues Continual Learning on Open Weights Puts Frontier Performance Inside Ordinary Budgets

arXiv 2608.27147·medium signalXBlueskyLinkedInCopy link

This paper contends that frontier-level performance is reachable by a wide range of institutions through continual learning on readily available open-weight models, not just by the handful of heavily funded labs, and not through the limited routes of small-scale fine-tuning, prompt engineering or tool-augmenting a frozen model. The approach uses a full modern mid- and post-training stack with per-stage safeguards that preserve both plasticity and stability while making the minimum number of high-impact parameter interventions. The demonstration model, Thomson, is competitive with recent frontier models on agentic tasks, safety, legal, tax, multilingualism and large-scale deep research, showing what the authors call a pi-shaped pattern of gains including on capabilities not explicitly targeted, while nearly eliminating the forgetting that plagues narrow domain adaptation.