- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper looks at how we evaluate models that predict what an individual person will do, with LLM "digital twins" as the main target.
The tasks are concrete. On Twin-2K-500, people answered a long survey and were asked some of the same questions again about two weeks later; the model has to predict the second answer from everything the person said before. The paper does the same with three waves of the General Social Survey, and with sequences of choices in cognitive experiments (Psych-101/201). Statistical models, a small RNN and prompted LLMs are all scored by per-person accuracy, and each is shown next to two simple baselines: the most common answer among other people, and simply repeating the person's earlier answer.
The central argument is that the common practice of reporting twin accuracy as a fraction of test–retest agreement is misleading. Retest agreement is exactly what you get by copying the earlier answer, so for a model that sees that answer it is a floor rather than a ceiling. The authors derive an upper bound from classical results. Empirically, all 13 released twins score below the most-common-answer baseline, LLMs that are given the earlier answer mostly just copy it, and on the cognitive tasks the previous choice carries most of the predictive signal.
Critical strengths
The main point is simple and, I think, important. Under the usual normalisation, even "guess the most common answer" would look like 92.6% of the ceiling, and every released twin actually falls below that baseline. That is a striking result, and it should change how digital-twin papers report their numbers. It is the main reason I lean towards acceptance.
The work is careful. As far as I could check, the theory is correct and the authors are upfront about where it stops holding. I also liked the controls: all models are scored on the same rows, the earlier answer is deleted from the LLM prompt to see what remains, and shuffled versions are used as sanity checks.
Critical weaknesses
Novelty. The bounds themselves are classical (Cover–Hart, classical test theory), and the paper says so, so the contribution is mostly in applying them to digital twins. The related work covers inter-annotator and neuroscience noise ceilings, but I was surprised not to see the "magic barrier" from recommender systems (Herlocker et al., 2004; Said et al., 2012), where users' inconsistency when re-rating items was used in much the same way to bound achievable accuracy. A short comparison would help, as would a clearer statement of what is new beyond Props. 3 and 6 and the empirical audit.
The framing is stronger than the evidence. The abstract leads with "retest agreement is a floor", but no model in the paper beats copying by more than about a point (0.4 on Twin-2K-500, 1.2 on GSS). The released twins being criticised never see the earlier answer, so for them the fair comparison is the most-common-answer baseline, not copying. And the line-deletion test showing that the LLM is copying uses one model and 150 people. The limitations section concedes most of this; I'd like the abstract to as well.
Questions for authors
How does your bracket relate to the magic barrier in recommender systems, and what would you say is new beyond Props. 3 and 6 and the audit?
Could the abstract lead with the difference between models that see the earlier answer and models that don't?
Do the main results hold without the repeated pricing questions? They make up about half of the Twin-2K-500 rows, and people may simply remember what they said last time.
Does the full-persona copying result replicate with a second LLM and with a prompt that invites the model to consider a changed answer?
4 comments
Follow-up
Dear Reviewer,
Thank you again for your review and for the pointer to the magic barrier. We would be grateful for your view on whether our response and the revision answer your four questions. On the first, the magic barrier is the error of a predictor that knows each user's expected rating. It thus corresponds to the accuracy of a predictor that knows each person's answer probabilities, which our upper bound limits, and not to retest agreement. On the second, the abstract now leads with the difference between predictors that do and do not see the earlier answer. It also states that no predictor we ran beats copying that answer by more than 1.2 points on either survey. On the third, the main results hold without the repeated pricing questions. On the fourth, the full-persona copying result replicates with a second LLM and with a prompt inviting a change. We are glad to answer any further question.
Thank you for the response in addressing my questions. I'd like to maintain my rating.
Thanks
Dear Reviewer,
Thank you for the reply and for your review. We are pleased that we address your questions. Your suggestions led to the comparison with the magic barrier, the analysis without the repeated pricing questions and the replication with a second LLM. The paper is stronger for them.
Appreciation for the comment and some further discovery shared
Dear Reviewer,
Thank you for finding the normalization point important and for checking the theory and the controls. That point follows from the paper's method, which makes explicit what each predictor sees. In this method, a predictor sees a person's earlier behavior and observes each decision after predicting it, a setting that earlier work calls prequential.
Your first question asked how our bounds on the best attainable accuracy, the retest bracket, relate to the magic barrier. The barrier is the squared-error counterpart of what the bracket bounds from above, and not of retest agreement. The bracket starts at the accuracy of copying a person's earlier answer, which is retest agreement. It ends at a bound on the accuracy of a predictor that knows each person's answer probabilities, if these are stable. The paper calls this predictor the oracle and the bound the upper end. The magic barrier is likewise the error of a predictor that knows each user's expected rating. If two ratings are independent draws from the same distribution, copying one to predict the other has exactly twice the squared error of that predictor. One re-rating thus identifies the barrier under squared error. Estimates that report the error between two ratings as the barrier, as in Amatriain et al. 2009, therefore report the error of copying, the counterpart of dividing by retest agreement. Under accuracy, retest agreement likewise measures copying, but no exact relation links copying to the oracle, so one retest only brackets the oracle. The related work on page 4 now makes this comparison.
Appendix A.2 estimates the barrier on the 43 items of Twin-2K-500 with ordered options, coded from 0 to 1. With stable answer probabilities, its square is estimated at 0.028, half the squared error of copying, 0.056. Predicting instead the average retest answer of other persons with the same earlier answer lowers the squared error to 0.043.
You also asked what is new beyond Propositions 3 and 6 and the results on twins. What is new is the method, which extends this prequential setting to survey retests and matches each reference to what the predictor sees. For a predictor that sees the earlier answer, retest agreement is a floor. A twin does not see the earlier answer, so the simplest reference matched to it guesses the most common answer to the question, which the paper calls the item mode. The revised Discussion therefore asks evaluations of twins to report the item mode on the same rows. Beyond twins, variants of each statistical predictor separate the previous choice from learning on cognitive tasks.
We agree that the upper end itself comes from the classical inequality of Cover and Hart and its weaker form from test theory. Beyond them, Proposition 5 keeps the upper end when answer probabilities drift in either of two ways the predictor cannot foresee. Remark 2 keeps it under a memory that repeats any earlier answer at the same rate. Under the same conditions, retest agreement is a ceiling only for a predictor that guesses at random with the true answer probabilities given what it sees, Remarks 1 and 4. If every person answers yes with probability 0.8, such a predictor has accuracy 0.68, the retest agreement, but always predicting yes has accuracy 0.8.
Your second question asked whether the abstract could lead with the difference between models that do and do not see the earlier answer, and it now does. It opens with predictors that see the person's earlier behavior, for which retest agreement is a floor. It then turns to twins, which do not see the earlier answer, and to the item mode, which all released twins fall below. The abstract also states the limit you named, that no predictor we ran beats copying by more than 1.2 points on either survey.
Your third question asked whether the main results hold without the repeated pricing questions, and they do. On the 68 other items of Twin-2K-500, copying beats the item mode by 10.5 points, as on all rows. All 13 released twins stay below the item mode. Under the metric of the dataset's authors, dividing by retest agreement credits the item mode with 93.0% instead of 92.6%, Section 6.1 and Appendix A.2.
Your fourth question asked whether the full-persona copying result replicates with a second LLM and with a prompt inviting a change, and it does. On 10 retest rows of each of 450 persons, GPT-6.1 Sol, given all the person's earlier categorical answers, repeats the earlier answer on every row. The first LLM, deepseek-flash, repeats it on 98.6% of them. A second prompt adds that about 29% of people change their answer and asks whether this person would. Under it, GPT-6.1 Sol still repeats on 99.98% of the 4,500 rows and deepseek-flash on 96.7%. Deleting the prompt line with the earlier answer costs GPT-6.1 Sol 7.4 points and deepseek-flash 9.3, Section 6.3 and Appendix F.
We would welcome your view on whether these answers resolve the two weaknesses you named.