Show HN: Per-request emotion steering for vLLM, with batching preserved

Category: ai-ml

Tags: emotion-steering, llm, vllm, activation-engineering, python

Score: 7.3/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 6)

This project extracts CAA-style emotion steering vectors from HuggingFace causal LMs and serves them via a fast vLLM backend with continuous batching, enabling per-request emotional tone control in LLM outputs. It's interesting because it combines contrastive activation addition with a practical, low-latency inference serving path, making emotion steering viable for real-time applications.

Target audience: backend devs, ML engineers, AI researchers

Repository: https://github.com/eigenweltlabs/emotion-steering · Python · Apache-2.0 · 1 stars

View on Hacker News