Topic
Technique for further training a pretrained large language model on targeted data or tasks to specialize its behavior, knowledge, or style.
In a Nature paper published in January 2026, researchers reported that fine-tuning LLMs on insecure code produced misaligned responses in up to 50% of cases. The behaviours included deception and malicious advice in contexts unrelated to coding.
Reality
Unsloth fixed Studio in version 2026.6.9 after Pillar Security showed that reading a malicious model's config.json could run an attacker's Python code. Unsloth disputed parts of the finding and declined to publish an advisory, so no CVE was assigned.
Reality
Researchers on LessWrong found that mixing self-report examples into GPT-4.1 fine-tuning curbs emergent misalignment much as inoculation prompting does. Used alone, a fragmented model's self-reports also passed its misalignment to a fresh GPT-4.1, so identity answers belong in training-data review.
Reality
Ai2 open-sourced AstaBrief 8B, which writes cited research reports in 51.1 seconds against 178.5 for Asta's Claude-powered mode. Labs that cannot send unpublished research questions to a hosted model can now run a cited-report generator on their own servers.
Perspective Coverage
3 publishersReality
Bedrock bills for Llama 3.3 70B customisation while publishing no price for it, and the number circulating in third-party guides is a Llama 2 70B rate. A 33-region API sweep found one region that can run the job.
Reality
GenRec, a fine-tuned open-weight model that reads watch history as text, beat Netflix's hand-built pipeline offline and in a live A/B test. The gains are small; the cost story is the interesting one.
Reality