# Apple — Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation

- Company: Apple (apple.com)
- Announced: 2026-09-16
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://machinelearning.apple.com/research/trajectory-teacher-flow-matching
- Record: https://forck.live/items/11448-trajectory-as-the-teacher-few-step-discrete-flow-matching-via-energy-navigated
- Subject: Machine Learning Research

Apple researchers propose Trajectory-Shaped Discrete Flow Matching (TS-DFM), a method that improves few-step discrete flow matching by using a lightweight energy compass to guide token selection during training. On a 170M-parameter language model, the 8-step student achieves 32% lower perplexity than the 1,024-step teacher while being 128× faster.

## Evidence

Verbatim from https://machinelearning.apple.com/research/trajectory-teacher-flow-matching:

> Trajectory-Shaped Discrete Flow Matching (TS-DFM) replaces these blind jumps with guided navigation: a lightweight energy compass evaluates candidate continuations at each midpoint, selecting the most coherent.

---

Record: https://forck.live/items/11448-trajectory-as-the-teacher-few-step-discrete-flow-matching-via-energy-navigated
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md