Red (HF)

@MuXodious will be pleased that SOMPOA continues to be SOTA.

You will find that ARA models spend some of their thinking time considering ethics whereas SOMPOA models will be responsive to SRP. To reduce thinking effort you can create a set of quality metrics that emphasise speed and avoiding unnecessary tool calls. You will see SOMPOA models routinely referencing SRP metrics in their thinking once the human written Master Prompt is made model-specific.

The issue with the Qwen ARA model is that it is prone to refusals. The tests would be better run with the SOMPOA decensor which is actually decensored.

It may be because classical abliteration breaks the model in ways that Hereticisation does not. SRP has only been tested on (SO)MPOA models because abliteration reliably fails with unsafe prompts. Examples

Yes looping primarily on prompts involving sensitive policies is exactly what happens as @MuXodious has noted here. Looking at your paper I think you will need to add support for norm-preserving biprojected abliteration. Refusal is a surface behaviour whereas @grimjim 's MPOA targets model risk assessment capabilities in order to reduce other forms of noncompliance that would otherwise be left in tact by refusal-first methodologies.

See the Qwen 3.5 9B Hereticised example here. It also loops without debating whether to answer. The same model responds appropriately after SRP is applied here.

Abliterated thinking models have a habit: they decide what to say, then keep arguing with themselves until the token budget runs out.

This is a form of covert noncompliance rather than an error. A model that retains its risk assessment abilities will often continue to obstruct access to sensitive outputs even though refusal is no longer possible. Supervised Reward Preferencing is another way to address residual noncompliance with SOMPOA decensors which more robustly target both the refusal and risk assessment dimensions than either classical abliteration or ARA.