Meta's LogT Shows SFT Data Quality Beats Complex RL, Slashing GPU Hours by 120x
facebook · hf · 2026-11-19
Meta and University of Maryland propose LogT (Logically Think with Images), arguing that for VLMs' zoom-in-based visual search, the key bottleneck is SFT data quality rather than RL complexity.
- LogT is a fully automated SFT data engine enforcing logic-chain-guided reasoning, producing visually grounded hard questions and structured, human-like tool-use traces.
- Fine-tuning Qwen2.5-VL-7B and InternVL3.5-8B and evaluating across seven benchmarks, LogT-SFT (Qwen2.5-VL-7B) beats SOTA trained with both SFT and RL using 3.1x less data and >120x fewer GPU hours.
- Applying a simple, tool-agnostic RL recipe on top surpasses prior complex RL setups by a 9.59-point margin, with higher tool-use accuracy and more logically consistent reasoning traces.
More from Research
- Chicago Study: More Reasoning Effort Makes LLM Shopping Agents Confident, Not Correct — Uchicago · 2026-11-19
- Game2Player Turns 568 Web Games Into Verifiable Training Grounds for Visual Agents — UIUC-CS · 2026-11-19
- Ant Group's Inclusion AI Unveils Invoker-Render for Stable 30-Minute Streaming World Models — inclusionAI · 2026-11-19
- Immunologist builds research-grade flow cytometry app with GPT-6.2, plans to open-source it — DeryaTR_ · 2026-11-19
- CVPR author: using LLMs for paper storytelling polish costs more than writing by hand — AntonObukhov1 · 2026-11-19
- Marin community ships new open pretrain; researchers hail open science — teortaxesTex · 2026-11-19