Activation Magicians: teaching Qwen3-4B to steer its own activations via plain-language instructions
Sauers_ · x · 2026-10-21
Sauers shared a technical thread on "Activation Magicians": a fine-tuned Qwen3-4B that understands and executes natural-language activation-control instructions, such as "attenuate your inner sense of tea by 2 SD on every fifth layer starting at layer 3," writing out an execution plan in its thinking block before replying.
- Method: rank-32 LoRA fine-tuning with MSE loss toward the requested activation pattern, cross-entropy on plan tokens, and a KL constraint to the base model, plus regularization to avoid perturbing unrelated concepts; two training paradigms (instruction+plan, and prefilled plan only).
- Training set: 135 concepts, 20 verb phrasings, 232 questions, covering many layer and token patterns; held-out set of 45 concepts and unseen patterns like prime-numbered layers, powers of two, alternating and fading token patterns.
- Results: 75% of the requested change achieved in-distribution (pattern match r≈0.66); generalization stays strong for unseen wording (78%) and unseen concepts (71%), and even three simultaneous edits reach 60%; unseen layer patterns (57%) and concept+layer combos (49%) are notably weaker.
- Failures: some stem from incorrect planning—the activations then match the wrong plan rather than the prompt.
- The author next wants to test whether this can be taught as a general metacognitive skill.
More from Research
- UnsolvedMath dataset hits 17,614 open problems for LLMs, drawing mathematicians' ire — tak3sh8 · 2026-10-21
- mini-Lean-agent: bash-only coding agent solves 12/12 Putnam 2025 at $1.28 per problem — wellecks · 2026-10-21
- Rethinking vanilla pixel diffusion transformers: capacity limits and a high-frequency deficit in JiT — CSProfKGD · 2026-10-21
- Visa trouble may cost authors their accepted EMNLP paper over a no-show — OtherwiseFig4505 · 2026-10-21
- NeurIPS 2026 registration doesn't include workshops, warns prof who almost missed them — CSProfKGD · 2026-10-21
- Open-weight models trail closed models by ~7 months on biology benchmarks, GovAI finds — Manderljung · 2026-10-21