Are new steering methods reliable and scalable enough to be alignment tools?
ML Engineer, requesting compute to run large scale steering evals, so see if new steering methods can be useful alignment tools.
Who am I? See wassname.com which links to my GitHub and LinkedIn and CV, I'm a physicist working in the energy industry in Australia. I've been in ML for 10+ years, and AI Safety for 3+.
Steering could help alignment: it shapes internal representations (harder to game than RL), guides stronger models without equally strong labels w2schar-mini, and helps shape model character persona_vectors — reducing sycophancy and revealing behavior hidden by evaluation awareness ssteer-eval-aware. My results and others suggest larger-scale testing is warranted.
But researchers are divided (see below), they have often encountered basic forms of steering that are unreliable. I'd like to see whether newer and novel forms of steering can scale up to be useful alignment tools on large models. I'd also like to inform frontier labs if there are promising tools they could use.
I've also developed a way to measure useful steering: how much can we change the intended behavior before off-target behavior changes or the output becomes incoherent? We sweep strength in both directions and use LLM judges to score intended change and unwanted behavioral changes, such as honesty steering producing exaggerated oversharing. Judges make this affordable across a sweep and can catch subtle deviations that capability tests miss, even with just 100 samples. Plotting these scores alongside random steering vectors lets us compare methods at the same achieved effect.
The VJP steering plot shows the pilot on Qwen3.5-4B.
I'm requesting US$15,000 to extend this comparison up a size ladder: a mid-size model, GLM-5.2, and Kimi K3. I'll publish the comparison plots with full responses and reproducible code. I will report which methods survive scale-up and which drop out.
The money will be spent on compute to extract/train steering vectors, and evaluate them on benchmarks, and judge the results.
My back of the napkin calculation based on the UK AISI's work on GLM 5.2, and assuming extracting is in bf16 in pytorch, and eval is in vLLM leads me to a conservative estimate of these compute costs:
- US$12,000 for rented GPUs. About US$5,100 of generation and US$4,800 of vector construction across three model sizes, plus roughly US$2,000 for reruns.
- US$3,000 for LLM judges to score intended behavioral change and side effects.
- Total: US$15,000.
The output will be a report where the central results will resemble plot here but for large GLM 5.3 and Kimi K3 or equivalent models. In this plot we have the Pareto frontier of each steering as we vary the dose until the model becomes incoherent, it's harder to steer a model in the direction it doesn't want (against sycophancy). I show the zone of random steers and baselines like PCA and Mean diff at least. I would like to cover uncertainty and steering fidelity as these are salient concerns for researchers.
Private comment. Only shown to approved funders and grant reviewers.