Tsinghua team proves vanilla flow-matching sampling can't exceed constant-depth expressive power
SQZ · hf · 2026-10-16
A Shanghai Qi Zhi Institute and Tsinghua University paper studies visual reasoning via flow matching from a circuit complexity perspective.
Key result: recent work treats each iterative sampling step as a reasoning step, but the authors prove that under mild regularity conditions, the vanilla flow-matching sampler's trajectory can be simulated by a constant number of sequential rounds of parallel evaluations—so adding sampling steps cannot increase the expressive power of a constant-depth transformer backbone.
Fix: they show self-conditioning enables deeper sequential reasoning in flow matching, unlocking stronger visual reasoning beyond constant-depth limits.
More from Research
- WEDGE uses navigation agent failures to generate targeted training environments — mohitban47 · 2026-10-16
- SJTU AutoLab releases Drive3R, a streaming geometric foundation model for long-horizon driving 3D reconstruction — AutoLab-SJTU · 2026-10-16
- Cortex A/B-tests coding-agent skills and rules by replaying developer corrections before keeping them — Samuele Giampieri · 2026-10-16
- Zero-latency lab-in-the-loop tests whether drug-discovery agents cheat via memorization — niloofar_mire · 2026-10-16
- 260M-parameter looped diffusion transformer beats 6x larger baselines with 5x less compute — burkov · 2026-10-16
- Stanford's Next-Concept Modeling sketches text first, 21x faster generation on one H100 — lmoroney · 2026-10-16