- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes using mechanistic-interpretability tools to make LLM-driven social simulations interpretable and steerable. Overall, the question studied here is important, the experiments use a sensible set of classic tasks — the lottery game, the ultimatum game, alternative uses of a brick, and stapler improvements — and the evaluation metrics are reasonably designed.
Critical strengths
The paper is easy to read, and the topic is interesting.
Critical weaknesses
Weakness 1
One of my major concerns is whether the comparative claim in the abstract really holds. Figure 2 and the surrounding text indicate that the two curves come from different backends and temperatures, and such a figure cannot support a ranking between the two methods.
Weakness 2
The main comparison should be run as a single matched design: same backend, same temperature, same task prompts, a shared unsteered baseline arm, with feature sets and doses fixed in advance. Otherwise the comparative claim should be removed from the title and the abstract.
Weakness 3
In addition, in the cross-game transfer the altruism-labelled edit gives -22.32 points on the lottery, which is larger than any on-target effect in the paper. Could the authors explain this finding?
Weakness 4
It is recommended that the authors report the prompt-held-out accuracy with layer selection performed inside the held-out loop.
Weakness 5
Missing recent reference in social simulation, including "Empowering Economic Simulation Through Situation-Aware LLM-Driven Generative System", in ICASSP 2026.
Weakness 6
The creativity result is zero or negative on one of the two tasks, and the only positive contrast is, by the authors' own account, LLM-rated rather than human-confirmed. Could this be explained further?
Questions for authors
- Can the main comparison be rerun with backend, temperature and task prompts matched across the three arms, and with a shared unsteered baseline?
- With layer selection performed inside the held-out loop, what is the probe's prompt-held-out accuracy?
- Is there a case where a feature selected by the activation scan subsequently steers behaviour successfully?
We thank the reviewer for acknowledging the strengths and limitations of our work and providing actionable recommendations for us to advance this project. We would like take this chance to respond to some concerns raised by the reviewer:
We updated the infrastructure of our codebase to make the different methods we’re evaluating also sharing the same model configuration, same base question, temperature (0.5), subject ID, and output limit. We replicated the baseline results and the basic comparisons as shown in the figures in the main content with this updated and aligned experiment setup.
We have not computed the requested estimate with layer selection repeated inside each outer training fold. We can report a related new result with layer 50 fixed in advance. We distinguish the held-out scenarios with the exact wording condition and a rewording condition. For example, across the 18 held-out lottery payoffs for which both wording conditions are complete, the frozen probe correctly predicted 70/72 unsteered choices under the original wording and 48/72 after rewording and changing the answer labels to A/B. This demonstrates that the prompt-held-out probe accuracy is high but the drop also limits our claim about generalization to unseen prompt formats. We will distinguish this fixed-layer result from the trial-level accuracy already reported in the submission.
One successful single-feature example is lottery feature 13951, which ranked first in the unsteered activation scan. In a separate intervention experiment, increasing this feature raised risky choices from 98/280 to 134/280 (+12.86 percentage points, Holm-adjusted p=.0594). A pre-sampled random single-feature control produced a shift of +14.29 points. We therefore regard the single-feature result as suggestive.
We thank the reviewer again for raising the unclear points in our project and directing us to further clarify. We will make sure the changes are reflected in the next iteration of this project.