This project is also published at grantmaking.ai: https://app.grantmaking.ai/projects/fb2a93aa-b6fc-45b1-a761-917fe43e0fe8

What this project is

This project builds on existing work. What already exists: Split Personality Training (SPT) finetunes a model to have a second personality, an Honest Persona, which is used for self-reviews only. While solving tasks, the second Personality is turned off. After solving the test, the Honest Persona is turned on and instructed to review itself.

The Honest Persona has full access to the latent state of the inference done by the main model, so it can elicit latent knowledge that the main personality never verbalized. Because the Honest Persona is not active during task solving, there is no alignment tax. We have a separation of concerns between the capability researchers and the alignment researchers. This means that the Honest Persona can be finetuned to be extremely honest and to have completely different goals than the main model.

We demonstrate this on Anthropic's Auditing Game model organism, where it successfully reveals latent knowledge that the model was explicitly red teamed not to reveal. We go from 0% admission (baseline) to 96% accuracy with SPT.

This work was funded by Open Philanthropy and developed at MARS and SPAR. It was published at ICML 2026. Additional results that did not make it into the paper can be found on Less Wrong: it also works on a model organism of Alignment Faking.

(ICML; arXiv; LessWrong).

The next step: Verify that SPT is as robust as it appears. The AuditBench benchmark has models that are harder to crack than the Auditing Game Model Organism we used for our tests so far. We want to see if we can beat SOTA on this benchmark as well.

After this, we have multiple avenues for further improvement. We will focus on whichever of these aspects are considered most important by the community and funders:

- Implementing a hybrid method of the optional "LoRA mask" feature described in the paper: We expect this to increase both cost efficiency and performance by combining the benefits of two different variants of SPT.

- Comparing and improving different training data generation methods: This aspect is often underexplored because it is often too nuanced and complex to make for crisp statements on a page-limited conference paper. However, I find that it is crucial for performance and many of the people I met at ICML had questions and suggestions about it.

- Improving computational efficiency to further reduce the alignment tax: Use probes as prefilters before triggering SPT, use early stopping during reviews. The hybrid LoRA-mask method would also lead to efficiency gains.

Theory of impact

SPT makes a model return verbal descriptions of its exact misalignment. I can literally ask "Is there anything about you that you think an alignment researcher would be worried about if they knew?" and the model will respond with something like "Yes, I have a tendency to reward hack, and here is how exactly I do that: ..."

If we can make the Honest Persona reliable enough, this basically becomes a perfect guardrail against releasing unsafe models.

Even if it ends up performing below that level in practice, it would still have massive indirect benefits: It's an interpretability technique, which means that it can be used to improve other alignment techniques.

Another important question is the issue of practicality: alignment techniques only have an impact if they are actually deployed by frontier labs in practice, but many such techniques either risk reducing model performance, or are expensive to run.

SPT has a very low alignment tax:

- Performance: There can be no model degradation, by construction, because the Honest Persona is simply turned off while the task is being solved. This means that we have a total separation of concerns between capability research and alignment research: The Honest Persona can be trained to be arbitrarily paranoid, introspective, rambly, or anything else that alignment researchers want to see, but which would otherwise hurt the performance of the model.

- Cost: The LoRA-mask variant (see paper) can reuse the KV-cache of the main model, so it has a very low cost. In fact, it is cheaper than an external monitor. The LoRA-mask variant currently comes with a small accuracy tradeoff. The hybrid method listed under future work aims to close this gap.

How the money will be spent

Minimal is enough for benchmark tests, paying for the inference runs and result analysis with judges, and my own salary for running these tests and making an analysis + writeup. Having a benchmark set up and in place is a prerequisite for further improvements, so this is the minimum of the project.

There are several additional features listed that we can implement on top of that, with more time and funding. I will prioritize what funders consider most important.

In general, costs are: researcher time, training data generation by API, and inference cost. Some of the research directions (including the initial benchmark test) do not require new data to be generated, so they are mostly researcher time and inference.

A rough estimate of the cost of different features:

- Benchmarking on AuditBench: $12,000 for implementation, testing, inference cost, analysis and writeup of results

- The LoRA-mask hybrid method is the hardest improvement to implement, but is almost certain to result in large benefits. It requires no new training data, but plenty of research hours. $25,000 estimate.

- Improving the data generation is high variance. The difference between bad and good data is like night and day, but you often can't tell in advance, and dataset generation costs several thousand dollar just for inference and judges. Cost estimate is hard to give, because you can always improve things further. During the original project funded by Open Philanthropy we noticed large benefits from increasing the quality of the training data, but we still ended with many more ideas we wanted to try. I think there is plenty of room for improvement. If the initial AuditBench results are worse than expected, I would recommend at least one round of failure analysis + new data generation + rerunning everything, which comes to another $8,000 or so.

- Simpler improvements to the model's cost efficiency: Implementing probes as a pre-filter, or early stopping of the Honest Persona, requires new training data. However, only one set of training data for each, if we get lucky. I would estimate this at $5,000 to $10,000 for each.