Dev visualizes LFM2.5-VL driving decisions from first-token probabilities
helloiamleonie · x · 2026-09-22
Inspired by recent JEV demos, developer @noctus91 ran LFM2.5-VL on a driving video to test visual decision readouts derived from first-token probabilities — surfacing the model's scene judgments in real time.
The author admits the demo isn't perfect and latency is noticeable, but it showcases an unconventional use of first-token distributions as a decision signal.
More from Models
- Decision Index 0.1 benchmarks 30+ open-weight decision models across 35+ evals and 130K questions — multimodalart · 2026-09-22
- Xiaomi Releases MiMo-V2.6-Distill-Qwen-9B, a 9B Agentic Model Distilled from Qwen3.5 — anovers · 2026-09-22
- Alibaba's 5-10 Trillion Parameter Plan: Why Scale Isn't a Product Milestone — yi111 · 2026-09-22
- Qwen's new live-translate model puts heat on Wispr Flow and Fireflies — Necessary-Garlic-704 · 2026-09-22
- EEBench comparison sparks debate: Grok 4.7 called out vs Astra's speed and cost — teortaxesTex · 2026-09-22
- Leak: OpenAI's new agent reportedly named Aeon, launch expected Thursday to rival Grok Bot — ZeroStateReflex · 2026-09-22