A fine‑tuned version of the lightweight Gemma model has been integrated into Singapore smartphones to spot scam conversations directly on the device. The system uses a two‑stage pipeline – a fast classifier followed by the Gemma model that…
Google DeepMind and Singapore GovTech Deploy On‑Device Gemma Model Fine‑Tuned for Scam Detection
A joint effort between Google DeepMind and Singapore’s Government Technology Agency (GovTech) has produced a new on‑device AI system that can spot scam conversations directly on users’ phones. The system uses a two‑stage pipeline: a lightweight classifier first…
Fine‑tuning Shifts Model Output Form Before True Competence, New Study Shows
A recent research paper on adapter‑based fine‑tuning finds that the most noticeable impact of fine‑tuning large language models is a change in surface form rather than a genuine boost in capability. The authors report that “fine‑tuning shifts output…
Fine‑tuning Mostly Alters Style, Not Competence, Recent Study Finds
A new research paper reveals that the most noticeable impact of fine‑tuning large language models is a shift in output style rather than a true change in capability. Experiments showed that fine‑tuned models often register lower scores on…
New Research Shows Personalized Test‑Time Scaling Lets LLMs Adapt to Users Without Fine‑Tuning
A recent paper from Beijing Institute of Technology and Xiaohongshu introduces “amortized agentic policy discovery,” a technique that trains a policy to choose decoding strategies based on each user’s preferences instead of seeking a single, shared optimal point.…