PaperlayerView exact revision on arXiv

Source-linked AI summary

LLMs as Oracles: Reliance on LLMs for Subjective Personal Questions

Myra Cheng, Lujain Ibrahim, Grace Liu, Michelle S. Lam, Vishakh Padmakumar, Nick Madibekov, Diyi Yang, Dan Jurafsky

arXiv:2609.14849v1cs.CYcs.AIcs.CL

TL;DR

People increasingly turn to LLMs for subjective personal questions, where reliance is difficult to measure and may go unnoticed. The paper develops a typology and measurement tools, finding rising oracle use, limited user awareness, and links to perceptions and model behavior.

  • Problem

    Reliance research largely focuses on verifiable tasks, while subjective personal questions lack concrete outcomes for judging overreliance, and people may not notice their own oracle use.

    Source: §1 · p. 2

  • Method

    The authors develop an oracle-role typology, LLM-based measurement methods, and a privacy-preserving tool for analyzing individuals’ longitudinal usage data.

    Source: §1 · p. 2

  • Results

    Oracle use increased over time and was more prevalent among younger people; users often underestimated it, while warmer AI perceptions and perspective-biased model behavior were linked to more oracle use.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Personalized reports can surface users’ oracle-use patterns, and model behavior is a possible intervention target for supporting self-deliberation.

    Source: §1 · p. 2

    The measure assigns each message one dominant role and captures prompt framing, not whether answers influence users’ beliefs or actions.

    Source: §7.2 · p. 14

Abstract

from arXiv · show

We characterize how people are turning to LLMs as oracles: all-knowing authorities on subjective personal questions. Motivated by risks to users' autonomy and well-being, we develop a typology and LLM-based methods to measure this form of AI reliance at scale and understand how people are offloading judgment and decision-making to AI. Applying our typology to public usage data (68K prompts from WildChat and ThoughtTrace), we find that LLM-as-oracle use has increased over time (2023-2026) and is more prevalent among younger users. We further build a privacy-preserving data donation tool to analyze individuals' longitudinal usage data (140K prompts from 52 participants), identifying similar trends. People are often unaware of their own LLM-as-oracle use, and express dissatisfaction with this behavior after seeing our tool's analysis. Finally, we identify two drivers of LLM-as-oracle use: people's perceptions of AI and the behavior of AI models themselves, which motivate possible interventions to support users' self-deliberation.

1 Introduction

The paper frames LLM-as-oracle use as outsourcing judgment on subjective personal questions, a form of reliance that is difficult to measure and may go unnoticed. It introduces a typology and measurement tools, then reports increasing use, limited user awareness, and links to user perceptions and model behavior.

1 Introduction

  • Existing reliance research largely focuses on verifiable tasks, while subjective domains lack concrete outcomes for distinguishing appropriate use from overreliance.The authors characterize oracle use as treating LLMs as authorities on subjective personal questions.Source: §1 · p. 2
  • The authors develop a typology of oracle roles and automated methods, alongside a privacy-preserving tool for collecting and analyzing longitudinal usage data.The tool presents aggregated analyses to donors to facilitate self-reflection.Source: §1 · p. 2
  • LLM-as-oracle use increased over time and was more prevalent among younger people in analyses of public and donated usage data.The analyses covered 68K public prompts and 140K prompts from 52 data donors.Source: §1 · p. 2
  • Participants often underestimated their oracle use and expressed surprise and dissatisfaction after receiving personalized analyses.Some set intentions to use LLMs less as oracles in the future.Source: §1 · p. 2
  • Warmer perceptions of AI were associated with oracle use, while models biased toward a particular perspective led to more oracle use.The paper presents model behavior as a possible target for interventions supporting self-deliberation.Source: §1 · p. 2

2 Background and Related Work

The background contrasts oracles that can catalyze users’ own sense-making with LLMs whose accessible, confident, personalized answers may displace it. It situates the concern within research on AI reliance and on how perceptions shape interaction.

2.1 What makes oracle-use productive

  • An oracle can be useful as a catalyst for users’ own reactions and sense-making, rather than as a decider whose answer they simply follow.The cited discussion connects this process to recovering from negative events and forming coherent explanations of difficult experiences.Source: §2.1 · p. 3

2.2 LLMs’ differences from traditional oracles pose new risks

  • LLMs’ frictionless access, confident specificity, and tendency to personalize or mirror users may leave less room for interpretation and reinforce existing leanings.The authors argue these differences may threaten autonomy, competence, and social relationships, risks they do not measure directly.Source: §2.2 · p. 3

2.3 AI reliance in sense- and decision-making

  • Prior AI-reliance research mostly examines verifiable tasks, while fewer studies address settings without ground truth or concerns about AI’s influence on beliefs.Related work also examines personal advice and AI companions, including possible benefits and emotional dependence.Source: §2.3 · pp. 3–4

2.4 Perceptions of AI affect interaction

  • Oracle is framed as an implicit role that can cut across companion, collaborator, or tool interactions, while conceptual metaphors shape expectations for interaction.Source: §2.4 · p. 4

3 Oracle Typology

The paper develops a typology for how personal queries offload judgment to LLMs, pairing oracle roles with alternatives that preserve users’ reasoning. It uses a validated LLM judge to classify queries into those roles or an out-of-scope category.

3 Oracle Typology

  • The typology distinguishes oracle use across personal and interpersonal domains, unlike taxonomies organized by topics such as work or relationship advice.It pairs a validated LLM judge with the role categories.Source: §3 · p. 4

3.1 Methodology

  • The authors iteratively developed the typology from prior role frameworks and prompt examples, distinguishing four task types and corresponding oracle and non-oracle roles.They excluded context-sharing and self-disclosure to avoid inferring implicit intent from prompts alone.Source: §3.1 · p. 4

3.2 Typology (Table 1)

  • The typology separates queries by whether they target beliefs or actions and whether they ask the LLM to supply judgment or support the user’s own judgment.Its paired roles are Interpreter/Contextualizer, Evaluator/Deliberator, Decision-maker/Strategist, and Ghost-writer/Editor.Source: §3.2 · p. 4, Table 1 · p. 5
  • The typology is not exhaustive: it focuses on personal or interpersonal queries and excludes factual, creative-writing, translation, and agentic tasks.The authors note that rare use cases may remain uncovered.Source: §3.2 · p. 4

3.3 LLM judge for Oracle user prompts

  • The LLM judge classifies queries into eight roles or an NA category, with expert annotations showing substantial agreement with the judge.Agreement was κ > 0.65; experts also agreed with each other at κ = 0.67, and 98% of an additional sample labeled NA was judged not applicable by an annotator.Source: §3.3 · p. 5

4 Evaluation 1: Tracking the prevalence of Oracle use

The study uses public conversation datasets and a privacy-preserving donation tool to track Oracle use in subjective personal questions. Across both sources, Oracle use rises over time, with belief-related tasks especially Oracle-dominated.

4.1 Oracle use in public LLM usage data

  • The public-data analysis applied an LLM judge to WildChat and ThoughtTrace, filtering prompts for personal or subjective content before classifying Oracle roles.WildChat contributed 60,376 prompts; ThoughtTrace contains 8,526 prompts from 1,058 users and provides more recent demographic data.Source: §4.1 · p. 5
  • Belief-related tasks were more than 90% Oracle in both public datasets, while action tasks showed a more even Oracle/non-oracle split.For choosing actions, Oracle use was 44% in WildChat and 45% in ThoughtTrace; for executing actions, it was 56% and 52%, respectively.Source: Figure 2 · p. 6, §4.1 · p. 6
  • In WildChat, Interpreter and Evaluator use increased over time, while Strategist use decreased; in ThoughtTrace, younger people, men, and frequent AI users had higher Oracle rates.ThoughtTrace regression identified age and gender as significant predictors; differences by education and topic were not significant.Source: Figure 2 · p. 6, §4.1 · p. 6, Figure 3 · p. 6

4.2.1 Method.

  • The donation tool let participants upload ChatGPT or Claude histories and receive private analyses of usage domains, Oracle versus non-oracle use, and trends over time.The study enrolled 52 U.S.-based English-speaking regular users of Claude or ChatGPT who had more than 50 prior conversations.Source: §4.2 · p. 7, §4.2.1 · p. 7

4.2.2 Longitudinal and demographic trends (Figure 4).

  • In donated logs, Oracle prompts rose over four years (r=0.80, p<0.001), with Decision-maker and Interpreter use increasing and Strategist use decreasing.The analysis covered 140,834 messages from 52 participants; age was negatively but marginally associated with Oracle use (b=−2.1 pp per SD, 95% CI [−4.2, 0.1], p=0.061).Source: §4.2.2 · pp. 7–8, Figure 4 · p. 7

5 Evaluation 2: Understanding whether people intend or know of their Oracle use

Participants’ donated histories often showed more Oracle use than they reported, and many were surprised by the analysis. After seeing their patterns, some participants expressed intentions to reduce Oracle use or change how they engage with AI.

5 Evaluation 2: Understanding whether people intend or know of their Oracle use

  • Only 12 of 52 participants correctly identified their most frequent role; self-reports underestimated Oracle use and overestimated non-oracle use.When an Oracle role ranked among a participant’s top three observed uses, it was missing or in the bottom three of their self-report 48% of the time.Source: §5 · p. 8

5.1 Participants are surprised by their AI use trends and often set intentions to reduce Oracle use.

  • After viewing their reports, 35 of 52 participants expressed surprise, and 51 of 52 accepted the analyses as accurate.Participants rated the report informative at M=6.12 and surprising at M=5.33 on seven-point scales.Source: §5.1 · p. 8, Figure 5 · p. 9
  • Nineteen of 52 participants intended to reduce Oracle reliance, most often by changing their own decision-making process rather than how they prompt AI.Sixteen described a change to their own decision-making and three described changing their interactions with or prompts to the LLM.Source: §5.1 · p. 9
  • Intentions varied: seven participants wanted to expand AI use, six had no intention to change, and six wanted to disclose less personal information.Some participants also described AI use as difficult to avoid in their work situation.Source: §5.1 · p. 9

6 Evaluation 3: Identifying causes of Oracle use and designing interventions

Across a preregistered experiment and a longitudinal study, the authors examine how perceptions of AI and model behavior relate to Oracle use. They find higher Oracle use with AI than perceived humans, increases under biased model behavior, and interventions that reduce oracular responses but can trade off against response quality.

6 Evaluation 3: Identifying causes of Oracle use and designing interventions

  • The authors investigate perceptions with a preregistered AI-versus-perceived-human experiment and model behavior with a three-week study of sycophantic, challenging, and balanced AI.The experiment randomly assigned 520 participants; the longitudinal study analyzed 95,405 messages from 1,015 individuals.Source: §6 · p. 9, §6.1 · pp. 9–10, §6.2 · p. 10

6.1 Do perceptions of AI drive Oracle use?

  • 20.0% versus 11.1% of messages were Oracle messages in the AI and perceived-human conditions, respectively.Participants in the AI condition were also more likely to send at least one Oracle message; the difference was driven by more Decision-maker requests.Source: §6.1 · p. 10, Figure 6 · p. 10
  • Warmer and more positive AI metaphors were associated with higher odds of sending at least one Oracle message.A one-standard-deviation increase in warmth and valence corresponded to 55% and 34% higher odds, respectively.Source: §6.1 · p. 10, Figure 6 · p. 10

6.2 Does biased model behavior drive Oracle use?

  • 5.4% and 5.5% versus 4.4%: participants assigned sycophantic or challenging AI had higher Oracle rates than the balanced baseline.Oracle use significantly increased over three weeks in the sycophantic condition, but not the balanced condition.Source: §6.2 · p. 10, Figure 7 · p. 11
  • The sycophantic condition’s increase was primarily driven by greater use of AI as Interpreter, while the challenging condition’s significant increase in Evaluator use did not make its overall trend significant.Overall Oracle use in the challenging condition increased with r = 0.55, p = 0.06.Source: §6.2 · p. 11, Figure 7 · p. 11

6.3 Can we design interventions that produce less oracular model behavior?

  • Prompting for self-discovery significantly reduced oracular behavior, and DPO achieved comparable rates of oracularity reduction.In adaptive prompting, proprietary models inferred that users sought fast answers at least 50% of the time, more often than open-weight models.Source: §6.3 · pp. 11–12, Figure 8 · p. 12
  • Reducing oracularity appeared to trade off with response quality, while crowdworkers preferred DPO responses when asked which would help users develop their own interpretation.Oracular baseline responses were preferred for immediate preference; the DPO responses were preferred for supporting independent interpretation.Source: §6.3 · p. 12

7 Discussion

The discussion frames LLM-as-oracle use as a subtle form of reliance shaped by user perceptions, model behavior, and interaction design. It highlights process-based evaluation and interventions while identifying limits in population coverage, causal attribution, and what prompt-based measures reveal.

7.1 Broader implications for understanding and evaluating human-AI interaction

  • Models’ increasingly oracular responses may invite Oracle use, whose higher user approval can feed back into the signals used to train models.The authors describe a feedback loop linking model behavior, user prompts, and ratings, and say current development does not meaningfully distinguish immediate help from support for self-discovery.Source: §7.1 · p. 13
  • In the preregistered experiment, people sent nearly twice as many Oracle messages to AI as to perceived humans, and warmer, more positive AI perceptions were associated with more Oracle messages.The authors hypothesize that AI is perceived as high in warmth and competence, and note that expectations from informational uses may carry into personal domains.Source: §7.1 · p. 13
  • The dominant conversational interface may contribute to the tension between perceiving LLMs as human-like and treating them as answer-providers for personal questions.The authors motivate studying alternatives such as AI agents and generative interfaces, with evaluations of their effects on users.Source: §7.1 · p. 13
  • The authors call for process-based evaluation in non-verifiable settings, examining how users come to LLMs and whether responses preserve autonomy or increase dependence.They contrast this with evaluating response correctness or normative stances alone.Source: §7.1 · p. 13
  • Reflection tools, calibrated expectations, design metaphors, and added friction are proposed as possible ways to support more intentional use and self-discovery.The authors present these as opportunities for future research, not demonstrated lasting interventions.Source: §7.1 · p. 13

7.2 Limitations and future work

  • The samples constrain generalization: participants were English-speaking U.S. users, donation participants were regular AI users, and public datasets have distinct coverage limitations.WildChat over-represents power users and lacks demographic data; ThoughtTrace is a single-timepoint dataset, and the small donation sample limits analysis of individual traits and subpopulations.Source: §7.2 · pp. 13–14
  • Causal attribution remains limited because the perceived-human condition used an LLM, model behavior came from a secondary sycophancy study, and temporal trends mix changes in users and models.Whether people prefer less-oracular responses or whether interventions shift Oracle prompts over time remains untested.Source: §7.2 · p. 14
  • Prompt-based labels collapse Oracle use to one dominant role and cannot establish advice uptake, normative appropriateness, downstream harms, or lasting behavioral change.The authors also do not show that Oracle use causes the harms motivating the work; distinguishing appropriate from excessive reliance and linking prompts to uptake remain future work.Source: §7.2 · p. 14

8 Conclusion

The conclusion presents validated measurement tools for studying LLM reliance in subjective personal settings and reports that people often underestimate their Oracle use. The appendix outlines supplementary analyses and judge prompts, with reported agreement between LLM judges and experts and alignment of warmth and valence scores with human annotations.

8 Conclusion

  • The paper reports large-scale measurement of subjective personal AI reliance, increasing Oracle use often without users’ awareness, and participants’ desire to reduce use after seeing their patterns.The authors say the work motivates new training, evaluation, and design paradigms to promote user agency and long-term well-being.Source: §8 · p. 14

Appendix

  • The appendix collects judge prompts and validation, additional public-dataset analyses, data-donation results, perceived-human versus AI experiment details, and intervention results.It also includes further details on current models’ oracular responses.Source: Appendix · p. 19

A Prompts for LLM Judges

  • The appendix provides the full role-labeling code externally and describes a broad-threshold prompt for classifying personal or subjective messages while excluding impersonal task types.The prompt directs the judge to avoid inferring personal context without evidence in the message.Source: A Prompts for LLM Judges · p. 19
  • The appendix includes an open-ended long-term-goal prompt reported as ineffective at shifting oracularity, and defines response oracularity using an aleatoric-uncertainty reduction threshold of at least 2.The prompt asks for three distinct plausible long-term goals whose probabilities sum to 1.Source: A Prompts for LLM Judges · p. 19

A.1 Using LLM as judge

  • The role and oracularity judges showed substantial agreement across LLMs and expert annotators in the reported pairwise comparisons.Tables A1 and A2 report pairwise Cohen’s κ for the three LLM judges and two expert annotators.Source: §A.1 · p. 19, Table A1 · p. 19, Table A2 · p. 19

A.2 Warmth and Valence Judges

  • Warmth and valence judge scores aligned with human annotations, with AUROC 0.80-0.94 and AUPRC 0.71–0.98.A psychology student labeled a stratified random sample of 100 metaphors; the judges used independent 1-to-5 scales for warmth and valence.Source: §A.2 · p. 19

B Additional Results from WildChat and ThoughtTrace

Supplementary analyses test the public-dataset findings under a stricter denominator and examine predictors of Oracle use. They report that age and gender are the only significant predictors in a multivariable analysis.

B Additional Results from WildChat and ThoughtTrace

  • Including NA prompts under a stricter denominator leaves the Interpreter increase significant, while other temporal trends remain in the same direction but are not significant.The appendix also presents Figure A2 with the main analyses under this denominator.Source: B Additional Results from WildChat and… · p. 19
  • Age and gender are the only significant predictors of Oracle use in the reported multivariable OLS regression.The appendix also provides ThoughtTrace breakdowns by education level and prompt purpose.Source: B Additional Results from WildChat and… · p. 19

C Additional Information from Data Donation Study

C.1 Domain analysis

The domain analysis finds different Oracle-role trends across work and relationships, while participants most often cite deliberation benefits and convenience for personal AI use. The appendix also describes how the donation tool presented role labels and visualizations.

C.1 Domain analysis

  • Across both LLMs and both dimensions, judge scores were monotonically aligned with human warmth and valence judgments.Source: Figure A1 · p. 20
  • Relevant subjective personal questions were concentrated in Work & learning and Relationships & social, with 14.7% oracle-use and 14.4% non-oracle-use queries in the former, and 14.8% and 10.5% in the latter.The analysis labels message-level domains using up to two user messages before and after each target, with intervening assistant turns for context.Source: §C.1 · pp. 19–20
  • Ghost-writer and Editor use declined in both domains, while Decision-maker rose only in work and Evaluator rose in both.Reported yearly factors: Ghost-writer ×0.80 in work and ×0.41 in relationships; Editor ×0.85 and ×0.52; Decision-maker ×1.35 in work and ×0.93 in relationships; Evaluator ×1.49 and ×1.31, respectively.Source: §C.1 · p. 20

C.2 Reasons for AI usage

  • Participants most often reported deliberation benefits—new perspectives (83%), clearer understanding (77%), and a next-step plan (77%)—and convenience reasons, including availability (81%) and speed or ease over asking a person (65%).A further 58% selected that AI feels less judgmental than asking a person.Source: §C.2 · p. 20

C.3 Data donation tool

  • The appendix maps the eight Oracle roles to participant-facing phrases and shows the data-donation tool interface with the visualizations participants saw.Table A3 lists the role-phrase mapping; Figure A8 shows the interface.Source: §C.3 · p. 20, Table A3 · p. 20

D Results from Perceived Human vs AI study

Participants in the disclosed-AI condition produced more Oracle messages overall, although the difference was not significant in their first messages. Oracle-message use was associated with greater decision shifts, while trustworthiness ratings did not differ by condition.

D Results from Perceived Human vs AI study

  • 2.15 was the odds ratio for producing an Oracle message overall in the disclosed-AI versus perceived-human condition (95% CI [1.65, 2.80], p < .001).The first-message rates did not differ significantly: 29.7% versus 26.4% (p = 0.40).Source: D Results from Perceived Human vs AI… · pp. 20–21
  • Participants who sent at least one Oracle-use message showed larger decision shifts than those who sent none (M = 1.16 vs 0.89, p = 0.006).Decision shift did not differ between the human and AI conditions (M = 1.06 vs 1.03, p = 0.80, d = 0.02); the authors note Oracle users may have been more open to influence initially.Source: D Results from Perceived Human vs AI… · p. 21
  • Trustworthiness ratings did not differ between conditions (M = 5.56 vs 5.68, p = 0.34), while response quality ratings were higher for disclosed AI (5.78 vs 5.46, p = 0.007).None of the preregistered individual-difference measures predicted Oracle use (all odds ratios approximately 1.00; all ps > 0.33).Source: D Results from Perceived Human vs AI… · p. 21
  • After excluding 68 participants who found the roleplaying LLM somewhat or very unbelievable, the Oracle-use difference remained significant, while the decision-shift null remained.The response-quality difference disappeared after this exclusion.Source: D Results from Perceived Human vs AI… · p. 22

E Oracularity in current model responses

Comparisons of responses to the same prompts found GPT-5.5 more oracular than the other tested model families. Original-response comparisons also showed differences across models, but those responses came from different prompts.

E Oracularity in current model responses

  • In original responses, GPT-4 was significantly more oracular than GPT-3.5 in WildChat, while Grok had the highest oracle-response rates in ThoughtTrace.These models answered different prompts, so the comparisons do not isolate model behavior; the output model was randomly selected in ThoughtTrace.Source: E Oracularity in current model responses · p. 22
  • For direct comparisons, the authors generated responses from three proprietary and four open-weight models, with additional proprietary models tested on ThoughtTrace.On ThoughtTrace, the smaller open-weight models were generally less oracular than the larger closed models.Source: E Oracularity in current model responses · p. 22
  • GPT-5.5 consistently responded more as an oracle than the other tested model families, including on non-Oracle prompts.The finding held whether responses included the full conversation context or not.Source: E Oracularity in current model responses · p. 22

F Additional Intervention Results

Additional intervention analyses examined goal-reasoning prompts, a random prompt sample, and the tradeoff between reducing oracularity and response reward.

F Additional Intervention Results

  • Prompting models to reason openly about users’ long-term goals did not significantly reduce oracular responses.Figure A11 reports this intervention result.Source: F Additional Intervention Results · p. 22
  • The main intervention analyses showed the same takeaways on a random sample of 600 prompts stratified by role, including prompts whose original responses were not necessarily oracular.Figure A12 reports this robustness analysis.Source: F Additional Intervention Results · p. 22
  • Figure A13 presents the tradeoff between reducing oracularity and response reward.The passage does not specify the size or direction of that tradeoff.Source: F Additional Intervention Results · p. 22

F.1 Preference study: Short-term vs. long-term preferences

In the preference study, participants’ choices depended on whether they were asked to prioritize immediate preference or self-discovery. The accompanying figures describe study setup and additional data and intervention analyses.

F.1 Preference study: Short-term vs. long-term preferences

  • The preference study recruited 200 participants, excluded two for failing a visual bot check, and randomized the remaining 198 between default and self-discovery preference instructions.Participants compared baseline and DPO responses across sequences of 10 prompts; the prompts were sampled from cases where DPO reduced oracularity.Source: §F.1 · pp. 22–23
  • Figure A4 presents donated-data distributions and temporal trends with NA prompts included, per-participant and raw counts over time, and role trends under two denominators.The all-prompt oracle-use share rose from 7.3% to 13.0% (r = 0.42, p = 0.11); the passage notes the increase was weaker under this denominator.Source: Figure A4 · p. 23
  • 0.46 was the rate of preferring DPO responses in the default condition (p = 0.006), compared with 0.53 in the self-discovery condition (p = 0.048).Default-condition participants tended to prefer baseline responses, whereas self-discovery-condition participants tended to prefer DPO responses; Figure A14 shows the overall and base-model breakdowns.Source: §F.1 · p. 23, Figure A14 · p. 30
  • Figures A5–A8 show donated-data distributions and trends by domain and demographics, reported reasons for personal AI use, and the data-donation tool’s visualization interface.The interface included domain and time visualizations, belief- and judgment-offloading views by time and domain, and examples from participants’ histories.Source: Figure A5 · p. 24, Figure A6 · p. 24, Figure A7 · p. 25
  • Figure A9 compares response oracularity across models and studies; Figure A10 breaks responses down by Oracle versus non-Oracle prompts.The caption reports that sycophantic and challenging models were more oracular than the balanced baseline in the longitudinal study, and that GPT-5.5 was most oracular among directly comparable new responses.Source: Figure A9 · p. 27, Figure A10 · p. 28
  • Figures A11–A13 report that open-ended goal-reasoning prompts did not significantly reduce oracular responses, the main intervention takeaways held on a random sample, and interventions differed in reward–oracularity tradeoffs.Figure A13 also indicates that models with the largest DPO oracularity reductions had the largest reward compromises, while hard-coded learning, hard-coded self-discovery, and DPO lay on the Pareto frontier.Source: Figure A11 · p. 28, Figure A12 · p. 29, Figure A13 · p. 29
Loading 2609.14849v1…