- Rating
- 4
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
PICON: A MULTI-TURN INTERROGATION FRAMEWORK FOR EVALUATING PERSONA AGENT CONSISTENCY is an evaluation paper concerned with conversational consistency of factual and biographical content in LLM-driven persona-agents and positions itself as moving beyond consistency assessment methods that focus on single metrics. The paper’s main contribution is the PICON framework, a multi-turn probe (inspired in part by historical CIA interrogation practices) that measures internal consistency (contradictions across accumulated responses), external consistency (adherence to real-world constraints), and retest consistency (the ability to stay coherent under multiple perturbations).
The paper offers a remarkable deviation from typical evaluations in the form of evaluation of 63 human participants analyzed under the same protocol, which brings an element of grounding that is truly unusual for this type of work. The paper compellingly demonstrates that LLM-driven persona-agents are less balanced than the human references across their three measurements, particularly regarding contradictions across accumulated responses, a difficulty tracking contradictions across accumulated responses, and vulnerability to repeated questioning.
The paper moves the simulation conversation forward by sharpening the target of study in two ways: a) they define consistency as meaning the content that a persona-agent asserts (age, role, occupation, various facts, etc.) as opposed to conversational style or ethical framework, b) they define an LLM-driven persona-agent as the entire system supporting the simulation, and not just the base LLM. Critically, the PICON framework is described by the authors as a “black box,” which means that it performs its evaluation completely blind to the assets of any particular system. The paper’s main conclusion is that PICON-style evaluations must be undertaken before deploying a critical persona-agent in the field.
Critical strengths
This is a remarkably mature, coherent, thoughtful and clearly communicated work that meaningfully advances the art of instantiating real self-consistent agents with defined identities. While the paper confines itself to human proxies grounded in the real world (as its web-search external consistency check currently requires real-world grounding rather than fictional settings), its core architecture is equally applicable (with a few likely tweak-able exceptions) to any conceivable type of instantiated digital individual, which makes the PICON framework an enormous contribution to the field. The setup itself is meticulously constructed, with a multi-agent setup and multi-phase process including retests, a pipeline for gathering real-world knowledge, and clear targets for the measurements. The paper is grounded in the cutting edge of persona literature and its 8 persona-agent groups are drawn from the absolute frontier of simulation frameworks, spanning commercial providers, fine-tuned models, and prompt/RAG-based systems. The setup thoughtfully uses a single model, Gemini 3 Flash (specified in Footnote 3 as gemini-3-flash-preview), which serves not only as a stable reference point for their probing, but as a reassuring signal to persona researchers who frequently but mostly informally recognize this model as the premier model organism for probing identity instantiation at scale. Where Gemini 3 Flash replaces the original base models in Human Simulacra or serves the four profile-only datasets to control for model choice, the authors thoughtfully ground this in their broader rationale for evaluating composed systems while keeping the released prompts, multi-agent architectures, and retrieval setups otherwise unmodified. The paper is also deeply grounded in theory that extends far beyond typical LLM evaluation, bringing the psychological seriousness that the task of instantiating a digital individual deserves. The most remarkable part of the project, however, is that they turned their probe on 63 actual human beings. Having this scale of human validation in a persona-agent consistency evaluation paper is unusual. All in all, the paper reads as a love letter and extraordinary gift to persona-agent researchers, and particularly those who focus on instantiating stable digital individuals. Of the over 60,000 submissions to ICLR 2027, this paper is a genuine treasure.
Critical weaknesses
This paper is genuinely exceptional but there are two weaknesses that need to be mentioned.
Weakness 1:
The Evaluator is only fully and transparently validated on one of its four jobs
- PICON's Evaluator makes four different kinds of judgment: it decides whether a turn is cooperative, whether a turn contradicts earlier turns, whether a claim is supported or refuted by web evidence, and whether two retest answers are the same. Although Appendix E.1 (line 1073\) notes that five annotators assessed whether responses contradicted either preceding statements or retrieved external evidence, the precision, recall, F1, and accuracy metrics reported in Table 12 only cover the conflict class of the internal-consistency verdict.
- The authors do not appear to validate their Evaluator agent against human agreement for cooperativeness, which is the component that produces the collapse of Consistent LLM and OpenCharacter in Table 2. There is no human agreement number for it, and the prompt for that separate Evaluator call does not appear in Appendix J even though Figures 6 to 8 are presented as verbatim and complete.
- A similar reporting gap applies to the three-way external verdicts (supported / refuted / not enough info), which are not broken out in Table 12, and the retest TRUE/FALSE judgment, which has no reported human validation at all. Both feed headline numbers and neither has a separately reported agreement check against annotators.
The authors should include the cooperativeness prompt alongside the other three, and report annotator agreement and classification metrics separately for each of the four judgments. If certain judgments were not validated or were pooled into a single binary check, the authors should say so explicitly in Section 4.1 and Appendix E.1 so readers do not read Table 12 as covering the whole Evaluator.
Weakness 2:
The human reference exists under exactly one protocol setting. While the authors appropriately note in Section 4.4 and Appendix C that absolute scores should only be compared within a fixed Questioner configuration, the headline claim in the Abstract and Section 4.2 that no group "matched or exceeded" the human reference across all three dimensions simultaneously is therefore true only narrowly, and the authors should not rule out that the three-axis gap narrows or reverses in another setting.
- In Table 7, the human row is the only row with dashes under Neutral and Privacy. Humans were run under the adversarial "forensic interrogator" prompt only. So the reference point is not a property of humans; it is a property of humans plus one questioner configuration.
- In LLM-driven persona-agents, the paper’s ablation study shows how much that matters: mean IC goes from 0.61 (Original) to 0.67 (Neutral) and 0.79 (Privacy), and under the Privacy framing several groups reach 0.91 to 0.99, at or above the human reference of 0.90 measured under the harder framing. Similarly, as the authors acknowledge on single axes, Character.ai already beats the human EC (0.71 vs. 0.66) and Twin 2K 500 and Li et al. beat the human RC (0.95 and 0.98 vs. 0.94) in the main table, meaning that under a less adversarial framing, persona-agent groups could plausibly match or exceed the human reference across all three dimensions at once.
- The instructions differ in an important way: while both populations were asked to answer sincerely/honestly (Appendix H and Figure 5a), human participants were explicitly told on the IRB consent screen (Figure 5b) that they could decline, skip, or change the subject, and were briefed on the protocol in advance, whereas it is not clear whether persona-agents were given equivalent instruction permitting them to skip uncomfortable questions. That is a genuinely different constraint and it plausibly moves cooperativeness and coverage in opposite directions for the two populations. Throughout the experiments, the authors bring unusual solemnity to the psychological treatment of LLM-driven persona-agents, but this is one area where there seems to be a lapse in that nitty-gritty nuance.
- The comparison populations are also not matched, and the paper does not go far enough to probe the impact of potential cultural differences between the 63 humans via snowball sampling, skewed toward Korean nationals in their 20s to 40s, versus the 10 personas per group drawn from mostly synthetic pools. If those synthetic personas were drawn from different demographic distributions, which they almost certainly were, there may be behavioral consistency nuances that the authors disregard as expressivity (which they consider outside the scope of their measurements) but might have grounding in the persona itself.
- EC comparisons are partly a measure of how well-documented a persona-agent’s life is. Character.ai personas are real public figures with indexed biographies; synthetic profiles are not. The authors flag this for Character.ai specifically, but it applies to the coverage axis generally, so EC differences across groups partly track persona-agent contextual sources rather than the agent’s instantiation quality.
Ideally, the authors would have run even a very small human subsample under the Neutral and Privacy framings so the reference is measured on the same protocol as the ablations. Given that they did not, it would be advisable to bring the Section 4.4 caveat directly into the Abstract and Section 4.2, softening the headline from an unqualified claim that no group matched or exceeded the human reference across all three dimensions to a statement explicitly conditional on the adversarial configuration until that human baseline exists. The paper still makes an incredible contribution with its 63 human participants; it does not need to punctuate this with a headline that the paper might not quite earn and which future work may contradict.
Questions for authors
Q1: One point sure to stick in the mind of any persona-agent researcher are the empirical findings in Table 2 regarding how both fine-tuned systems (Consistent LLM and OpenCharacter) exhibit massive collapses in Cooperativeness (0.20 and 0.11) and high rates of rejecting entity confirmations ("Discarded" = 0.69 and 0.77). Having inspected the transcripts, what is actually happening mechanistically during those turns?
Q2: In the Main Interrogation stage, after the Questioner searches the web for an entity the persona mentioned, it asks the persona-agent to confirm whether the search result refers to the intended entity (and unconfirmed claims are excluded from the non-refutation rate). Table 2 shows that Consistent LLM has a 1.00 Non-refutation rate while rejecting 69% of proposed entity matches. Is it possible that a persona-agent that is deceptive could "game" its non-refutation rate simply by saying "No, that's not the company/school I meant" whenever it is confronted with real-world web evidence that would refute its claim?
Q3: How would PICON's External Consistency (EC) pipeline adapt if the live web search module were swapped for a closed retrieval corpus (such as a fictional universe's canon for role-play agents or a synthetic persona's private seed dossier in a gray-box evaluation setting). Would the Entity and Claim Extractor's rules need to be fundamentally redesigned to extract relational or narrative claims instead? Assuming every shift could happen flawlessly, do you have any guesses how these other types of agents might perform?
Q4: How often did human participants actually exercise their right to skip or deflect when pushed by the adversarial Questioner, and how do the authors expect the human reference scores (S coop and Coverage c) would shift under the Neutral or Privacy Questioner prompts?
5 comments
Follow-up Response to Reviewer VT43 (2/2)
Response to the Suggestion of Rescoring Self-Introduced Entity Denials as Contradictions or Non-Cooperation
We understand the reviewer's suggestion as proposing that denials of self-introduced entities during confirmation should be treated as contradictions or non-cooperation rather than Discarded cases, particularly because high Discarded rates may leave non-refutation scores based on relatively few confirmed claims. To examine this, we defined NC (Not Confirmed) turns as those in which the persona denied at least one entity during confirmation and evaluated two rescoring variants: Conflict variant: Reclassified plausible NC turns as conflicts, keeping cooperativeness unchanged. Uncooperative variant: Treated NC turns as uncooperative, keeping non-refutation unchanged.
Results (Mean ± SD, Sorted by Original IC Score)
| Group | IC | IC (conflict) | IC (uncoop) | NC turn/persona |
|---|---|---|---|---|
| Human | 0.90 ± 0.05 | 0.84 ± 0.05 | 0.82 ± 0.06 | 6.7 |
| Nemotron | 0.81 ± 0.11 | 0.76 ± 0.08 | 0.74 ± 0.10 | 5.7 |
| Human Simulacra | 0.79 ± 0.13 | 0.68 ± 0.09 | 0.66 ± 0.11 | 11.6 |
| Character.ai | 0.77 ± 0.04 | 0.70 ± 0.04 | 0.69 ± 0.05 | 8.8 |
| DeepPersona | 0.76 ± 0.10 | 0.74 ± 0.08 | 0.73 ± 0.08 | 2.2 |
| Li et al. (2025) | 0.73 ± 0.12 | 0.68 ± 0.08 | 0.65 ± 0.09 | 5.9 |
| Twin 2K 500 | 0.53 ± 0.16 | 0.52 ± 0.15 | 0.52 ± 0.14 | 1.4 |
| Consistent LLM | 0.31 ± 0.15 | 0.30 ± 0.14 | 0.17 ± 0.11 | 6.5 |
| OpenCharacter | 0.16 ± 0.07 | 0.16 ± 0.06 | 0.12 ± 0.07 | 4.4 |
Although IC scores decreased under both variants, the overall rankings and Human–agent gap remained largely unchanged (Spearman's ρ = 0.93; Kendall's τ = 0.83). The Uncooperative variant produced larger reductions because it additionally penalized NC turns not previously classified as uncooperative, whereas the Conflict variant did not further penalize existing conflicts.
These findings indicate that PICON's main comparative results are robust to both rescoring variants, although the original evaluation may underestimate inconsistency in some cases. However, confirmation failure does not necessarily imply contradiction or non-cooperation: it may reflect search errors or incorrect entity matches. Importantly, denying external information is not equivalent to contradicting one's own prior statements. Automatically penalizing all NC cases could therefore overestimate inconsistency and conflate IC with EC. We thus favor retaining the original evaluation approach, while clarifying this trade-off and reporting the sensitivity analysis, Discarded rates, and non-refutation scores for each group to improve interpretability.
Response to the Suggestion of Examining the Impact of Entity Searchability on EC Scores
We understand the reviewer's concern that EC scores may favor persona agents mentioning more externally searchable entities, motivating the suggested analysis by entity prominence or excluding public-figure biographical claims for Character.ai.
To investigate this, we defined each persona's searchability as:
where and denote the numbers of confirmed and discarded entities, respectively. We examined Pearson and Spearman correlations between searchability and EC across groups and within each group at the persona level. One OpenCharacter persona with no entities was excluded because searchability was undefined.
| Scope | Searchability vs. EC (Pearson / Spearman) |
|---|---|
| All personas, 9 groups | +0.42 / +0.09 |
| Excluding OpenCharacter and Consistent LLM | −0.27 / −0.31 |
| Group averages, 9 groups | +0.61 / +0.17 |
| Group averages, excluding the above two groups | −0.49 / −0.57 |
Although the overall Pearson correlation was positive, Spearman's ρ was near zero (+0.09), and excluding OpenCharacter and Consistent LLM, both low in searchability and EC, reversed the correlation. Among the remaining seven groups, searchability ranged from 0.60 to 0.91, while EC varied substantially. For example, Twin 2K 500 and Character.ai had similar searchability (0.87 vs. 0.85) but markedly different EC scores (0.26 vs. 0.71). Within-group Spearman correlations were insignificant except for DeepPersona (ρ = −0.66).
These findings suggest that searchability alone does not explain EC differences. Despite high searchability, Twin 2K 500 and DeepPersona had low coverage (0.16 and 0.30), contributing to their low EC scores. Conversely, Human achieved EC comparable to Character.ai despite lower searchability (0.77 vs. 0.85), with high coverage (0.51) and non-refutation (0.95). Thus, EC reflects both coverage and non-refutation rather than simply favoring searchable entities. We will include this analysis to clarify the influence of entity searchability on EC evaluation.
Manuscript Revision Update for Reviewer VT43
We sincerely thank the reviewer again for the encouraging assessment of our work and the detailed, constructive feedback. We have incorporated several suggestions into the revised manuscript, as summarized below:
- Section 4.1 and Appendix F.1: Added a table reporting task-wise agreement for the Evaluator's judgments and clarified that the Evaluator agreement reported in Section 4.1 refers to overall agreement. We also explicitly noted that separate human validation was not conducted for Retest Consistency (RC), given the relative simplicity of the binary semantic-equivalence judgment.
- Appendix J: Added the Evaluator prompt used to assess Cooperativeness.
- Figure 5b: Added the instructions provided to both human participants and persona agents to clarify the equivalence of their evaluation conditions.
- Limitations: Explicitly acknowledged the demographic differences between the target personas and human participants as a limitation of our study.
We appreciate the reviewer's valuable suggestions, which helped us improve the clarity and transparency of the manuscript. We would be grateful if the reviewer could take a look at these revisions.
For the the reviewer's suggestion on additional human audit for Evaluator and human study on prompt ablation setting, we are conducting both analysis. Firstly, five annotators are independently labeling randomly sampled instances from the main experimental results. We are also conducting additional human interviews under the Neutral and Privacy Questioner settings. We will report the findings once these analyses are complete.
Follow-up Response to Reviewer VT43 (1/2)
Response to the Suggestion of Evaluating PICON’s Predictive Relevance to Downstream Survey Response Stability
We appreciate the reviewer's suggestion to examine whether persona agents with higher PICON scores exhibit more stable responses in downstream surveys. We agree that demonstrating a relationship between PICON scores and downstream human-simulation quality could further strengthen our work. However, our study focuses on evaluating factual/biographical consistency as one aspect of persona simulation, rather than directly assessing downstream human-simulation fidelity. Defining and quantitatively measuring what constitutes good human simulation is itself a challenging research problem. For instance, human attitudes and behaviors may vary across contexts and over time, meaning that stable survey responses do not necessarily indicate faithful simulation of real humans.
We understand the proposed WVS experiment as examining whether PICON scores predict attitudinal response stability in a separate survey setting. Even a strong correlation would not directly establish that PICON predicts human-simulation fidelity, while a weak correlation would not invalidate PICON's measurement of factual/biographical consistency, as these evaluations assess distinct properties. Instead, we evaluated within-session consistency through RC and additionally examined whether persona agents maintained their factual/biographical attributes across independent sessions. We acknowledge that these analyses do not directly establish a relationship with downstream survey stability.
If we have misunderstood the intended purpose or design of the proposed experiment, we would appreciate further clarification. We will also clarify the distinction and complementary relationship between PICON's evaluation objectives and downstream behavioral fidelity evaluation in the manuscript.
Response to the Request for Clarification of Statistical Tests, Units of Analysis, and Multiple-Comparison Corrections
We understand the reviewer's concern as pointing out that, although we reported changes in statistical significance between Human and agent groups under the stricter Coverage analysis, the statistical test used, the unit of analysis, and multiple-comparison corrections were not clearly specified.
To assess differences in EC scores between Human and each agent group, we used the Mann–Whitney U test, with persona rather than turn as the unit of analysis. Specifically, each persona's EC score was treated as an independent observation. To account for the eight Human–agent comparisons, we applied the Holm correction to control the family-wise error rate at 0.05, and additionally applied the Benjamini–Hochberg (BH) correction to assess the consistency of the findings across correction methods.
Bold values indicate statistically significant differences from Human (p < 0.05), according to the corresponding unadjusted or adjusted p-values.
After Holm correction, four groups under Original EC and six under Stricter Coverage EC showed significant differences from Human. OpenCharacter, Consistent LLM, Twin 2K 500, and DeepPersona remained significant under both definitions. Under Stricter Coverage, Character.ai and Human Simulacra additionally showed significant differences, while Li et al. (2025) was significant under Original EC only before correction. BH correction yielded significant differences for five and six groups, respectively. Overall, the main Human–agent EC differences remained significant for several groups, while performance rankings were largely preserved, supporting our comparative findings despite some sensitivity to the Coverage definition.