Hacker News Reader

gghfezon DeepSeek-V4-Flash means LLM steering is interesting again

I've been building control vectors for creative writing for a while now (models like Gemma-4, Kimi-K2.6, GLM-5). Concept axes like optimism/nihilism, simple/ornate language, empathy/sociopathy.

The "just use prompting" argument breaks down for stylistic and tonal shifts. You can prompt a model to "write in an ornate style" but you're fighting the trained-in distribution, and the instruction competes with your actual creative intent for the model's attention. A control vector shifts the distribution directly without consuming any context window.

The more surprising finding has been cross-domain transfer. Vectors trained on creative writing concepts visibly affect other domains. An optimism/nihilism vector changes the color palette and mood of generated SVGs with no mention of mood or style in the prompt. Same prompt, completely different output.

Model cards with samples (SVG generation, short stories) and downloadable vectors for llama.cpp:

https://huggingface.co/gghfez/Kimi-K2.6-control-vectors

https://huggingface.co/gghfez/gemma-4-31b-it-control-vectors