Distress Representations in Language Models Are Referent-Specific

Strong, well-controlled evidence that distress representations are referent-specific. I particularly liked the matched referential design, the reliability ceilings, and validation across multiple model families rather than a single model. The results are strong on their own terms: 39 of 40 cells show significant self/other separation among the models passing the referent check, and the distinction survives the no-experiences persona. The main limitation for me is the connection between this representational result and downstream behavior or welfare-relevant outcomes. What practical decision changes once we know these directions differ? Establishing whether the self-specific component causally influences behavior would make the broader significance substantially clearer.

- This project concerns a very important question. Much prior work does not sufficiently distinguish between the referent of putative welfare-relevant representations, so I think this is a valuable direction.

- The prose could sometimes be simplified. As a reader of the abstract, for example, I have no idea what “Δ peaks at +0.348” means, what the units or natural scale of Δ are, or how large +0.348 should be considered. More intuition would make the headline results much easier to interpret.

- There are some apparent signs of LLM-assisted writing in the manuscript. That is absolutely fine, but for future versions the authors may want to try to find more of their own voice. Some of the more promotional writing could also be toned down. For example, instead of “We supply exactly that control”, why not simply say “We seek to fill this gap”?

- The methods section moves far too quickly for me. Table 1 gives one helpful example of the construction, but I would have liked to see more examples of the actual distress and neutral items used to construct v_r, and more explanation of how the battery was constructed. By contrast, technical details such as the Spearman–Brown correction could readily be moved to an appendix.

- The discussion of the manipulation check was initially quite confusing. The preregistered self-ascription check appears not to have worked at all for some models, and was therefore replaced with a different referent-identification task before the primary data were collected. The replacement check seems to work reasonably well. Given how central successful referent discrimination is to the interpretation of the results, I would have liked substantially more explanation and discussion of this issue.

- The manuscript repeatedly refers to a preregistration. Where exactly can this be found?

- For the key outcome of interest, Δ, I would have liked considerably more intuition about its interpretation. What is its theoretical and practically relevant maximum? What should count as a large or small value? Is not any reliably positive difference already informative? The paper explains how Δ is constructed and why Δ > 0 is evidence of separation, but it is much less clear how its magnitude should be interpreted.

- An especially useful next step would be to steer models along the separate representational directions and examine how downstream behaviour differs. The authors already propose steering the self-specific residual, which seems right. It would be particularly interesting to compare interventions on the self-specific and shared/other-referential components: does only the self-specific component produce the kinds of downstream effects that would make it especially relevant to welfare evaluation?