Supervised Fine Tuning

Figma trains custom AI model to edit designs directly on canvas

Figma is developing an in-house editing agent trained via supervised fine-tuning to understand its native file format and make iterative design changes, such as spacing, color, or layout edits, directly on the canvas. The company built training data by pairing starting designs with synthetic edit requests and validated outputs, using larger teacher models, deterministic validators, and visual judges to check results. Figma reports that an early Llama 3.3 70B model trained with this approach performed well on these tasks.

figma.com · 2026-11-01

Startup fine-tunes 4B open-weights model to beat Postgres's own query planner

An independent researcher used supervised fine-tuning and agentic reinforcement learning to train a small 4B-parameter open-weights language model to generate PostgreSQL execution plans. On a set of 113 join-heavy queries, the tuned model cut latency by 44.7% on average compared to Postgres's default planner, even though the base model initially failed to produce valid plans for 99 of those queries. The project also involved building a custom measurement rig to reduce caching noise and a modified GRPO reinforcement learning method for scoring plans in a noisy environment.

rohanbansal.com · 2026-09-16