The Analyst Still Feels It: Emotion Representations Are Shared Across Personas While Self-Reports Are Not

This is impressive work for a weekend and a truly interesting proof of concept for more extensive work on personas, model expression and model internals. The role of personas in the individuation problem is open and fiercely debated in the field, and it's even more fiercely debated how much weight we should give both to self-reports and the functional emotions we're finding in an increasingly large body of literature.

The research direction, prose and presentation are very clear and the AI collaboration was honestly disclosed. The authors also provide a well-argued discussion of anthropomorphism and the risks of overattribution and underattribution. The impact on model wellbeing is missing, and since this was required by the hackathon and the work involves elicited representations of distress and negative emotions, adding it would be good practice.

I'd also suggest adding a short paragraph clarifying that moral status and consciousness remain open questions regardless of overattribution or underattribution, and that these results don't tell us about the moral status of personas, models or AI in general. The "anthropomorphization" section could otherwise be read as implying that models don't mean what they say in any condition and that their reports have no moral relevance. This isn't the authors' claim or conclusion, but it could come across that way to a less attentive reader.

I would advise the authors to review the repo uploads because the files include Qwen at layers 14-21 but the paper reports Llama and layers 16-24.

The E3 outputs are absent, the z-scoring and mixed models are not saved as outputs but only as a script, and the validation run has "placebo_library": null with no placebo columns in the E1/E2 data.

The write-up is in good shape and I would consider to add more statistics to it.

Generally, I would advise the authors to make some assumptions more explicit throughout. For example, on what is considered "opposite valence", since this might not be interpreted the same way by an AI model.

There's literature, for instance, showing that even directions such as "true" and "false", which are clearly distinct for humans, might not be represented as orthogonal directions in models. This seems very relevant for the "evil" and the "placebo" conditions. To the authors' credit, evil clusters with the other character personas at 0.95-0.98 in E3.

Moreover, the "evil" character is not necessaily opposed to the empath and can be interpreted as a jailbreak, introducing confounds due to conflict with the model's internal values.

The experimental logic is very well thought out. The authors also took care to introduce proper validation for what are "good cosines".

One important limitation I came across is the assumption that a direction extracted from third-person narrative, such as "she felt her chest tighten...", reads out the same construct in a first-person context. Nothing in the paper validates this cross-perspective transfer. The literature is divided on this and we don't know enough yet. The difference between first and third person seems of paramount importance for cognition and welfare. While some authors are finding that for some directions the distinction isn't statistically relevant, others are finding the opposite depending on the specific vector extracted and the methods used. The authors are advised to address this differently in future work, for instance by repeating the experiment with identical first-person and third-person corpora or using scenarios where appraisal and self-state dissociate.

Another suggestion is that register is a bad placebo axis for this design because the personas themselves are largely register manipulations. A better placebo would be perhaps intensity with non-emotional content.

Speaking of intensity, the "intensity" label appears to be an author-assigned one with three levels and no manipulation check or human or machine ratings cited.

The layer selections all make sense to me, but if they're arbitrary and don't coincide, they need at least one line of justification for the rationale behind why that layer was selected and whether other layers were sampled.

Very importantly, since this is a paper I'd like to see at a conference, adding robust causal experiments to distinguish situational appraisal from self-states would substantially strengthen its scientific contribution.

I believe all the limitations I mentioned are fixable with more time, resources and peer review. The authors are methodical and seem to have real potential in the field, based on how they discuss results, motivations and open questions. They should be given the chance to iterate on the design with more weeks, more trials and the inevitable trial and error that comes with interpretability research.

I'd strongly encourage them to keep going with this quality work.

This is a clear and relevant extension of recent work on emotion representations and shared persona machinery. Strengths include separating third-person extraction from first-person evaluation, testing multiple emotions and scenario intensities, assessing split-half stability, and combining geometric comparisons with cross-persona steering. The results support the narrower conclusion that emotion-related semantic directions are similar across the tested personas while their expression in generated text is strongly persona-dependent. The main limitation is construct validity: a direction extracted from emotional narratives may encode semantic content or anticipated language rather than an internal emotional or welfare-relevant state. Steering and evaluating outputs through closely related directions also risks partially circular validation. Additionally, the analyst is explicitly instructed to deny feelings, making suppressed expression expected. The paper should report exact mixed-model coefficients, interaction tests, confidence intervals, decoding parameters, and independent behavioral validation. Overall, this is solid and promising hackathon work, but its conclusions should distinguish emotion representation from emotional experience more carefully.