VQV Signal

CREATOR TOOLS WATCH TECHNICAL

Post on X says steering vectors recover most RL reward gains in diffusion models

According to a post on X by @tunahansalih, Tuna Han Salih Meral and co-authors froze diffusion transformers and learned fixed, input-agnostic steering vectors. The post says these vectors recover at least 89.8% of the reward gain across 8 RL-tuned image and video models. Vectors trained directly with RL reached 104.6% of the RL-tuned gain. The excerpt is cut off before it gives the full results for SD3.5-M.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-20T23:45:00+00:00 Detected 2026-11-21T10:02:53+00:00
View original source

According to a post on X by @tunahansalih, Tuna Han Salih Meral and co-authors froze diffusion transformers and learned fixed, input-agnostic steering vectors. The post says these vectors recover at least 89.8% of the reward gain across 8 RL-tuned image and video models. Vectors trained directly with RL reached 104.6% of the RL-tuned gain. The excerpt is cut off before it gives the full results for SD3.5-M.

AI-assisted summary based on the listed source.

Tuna Han Salih Meral and co-authors froze diffusion transformers and learned fixed, input-agnostic steering vectors that recover at least 89.8% of the reward gain across 8 RL-tuned image and video models. Vectors trained directly with RL reached 104.6% of the RL-tuned gain, and on SD3.5-M most of the....

For researchers and teams tuning image and video generators, the post suggests a simpler steering method may capture much of what RL post-training achieves. That could change how builders weigh training effort against reward gains. Watch for the full paper to confirm the methods, the SD3.5-M results, and whether the findings hold beyond the models tested.

Signal Strength 81% Technical label WATCH Public Interest 28 Category CREATOR TOOLS Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 0 Practical Impact Score 26 Novelty Interest Score 70 Consequence Score 10 Curiosity Score 32 Shareability Score 46

VQV surfaced this signal because it is recent, relevant to AI Video, connected to Scobleizer News Pulse (X).