PERSONAJUDGE: Simulating Individual Human Preference Judgments with Evaluator-Specific Demonstration Data

Personas, identity, and agents2026arXivApproved editorial review

Authors: Zeyu He, Xuan Qi, Subramanian Chidambaram, Zhichao Xu, Vinayak Arannil, Lydia Chilton, Alex C. Williams

Keywords: Evaluator-specific simulation, LLM-as-a-Judge, Personalized evaluation, Preference judgments, In-context learning, Interface telemetry, Retrospective reasoning

Source: Open primary source (opens in a new tab)

7
Authors
12
Findings
20
Limitations
13
Evidence

Editorial summary

English

PERSONAJUDGE asks whether an LLM judge can reproduce one particular evaluator's criteria from that evaluator's prior decisions. The study uses two tasks derived from Anthropic's Helpful and Harmless dataset: 700 unique conversations for helpfulness and 700 for harmlessness, each labeled by three people. Thirty-two professional in-house annotators participate; 21 work on each task and 10 overlap. Each annotator completes 100 three-way comparisons, Prefer A, Neutral, or Prefer B, yielding 4,200 human judgments. In a second stage, annotators replay their interaction and retrospectively verbalize how they understood the conversation, assessed the responses, and selected a label. The system stores the label, interface telemetry, clicks, scrolling, selection, waiting, and typing, and this post-hoc explanation. This is neither a population preference sample nor a whole-person simulation: the empirical target is a professional annotator's labels on two closely related evaluation tasks.

For each evaluator and task, 40 items form a demonstration pool and 60 form the validation set. PERSONAJUDGE decomposes prediction into two rounds: whether the evaluator would express a preference or neutrality, followed, when needed, by preference direction. It crosses four models, Claude 3.5 Sonnet V2, Claude 3.7 Sonnet V1, DeepSeek-R1, and Amazon Nova Premier, with four demonstration contents, judgment only; judgment plus interface telemetry; judgment plus retrospective reasoning; or all three, and 1, 2, 4, or 8 examples. This creates 64 conditions, 80,640 simulations per task, and 161,280 total. Validation items never enter the demonstration pool, but each condition uses only one random demonstration draw, so sensitivity to example composition is not estimated. Immutable model snapshots, inference parameters, model outputs, and an executable environment are also absent.

Averaged across all 64 conditions, the unpersonalized Base Judge scores 0.452 accuracy on harmlessness and 0.496 on helpfulness; PERSONAJUDGE reaches 0.480 and 0.510. On the matched subset, target-evaluator demonstrations beat demonstrations from other evaluators by 0.028 on harmlessness, 0.477 versus 0.450, p=0.019, and 0.044 on helpfulness, 0.515 versus 0.471, p<0.001. Retrospective reasoning is the useful auxiliary signal: judgment plus reasoning averages 0.505 and 0.537, whereas judgment plus telemetry falls to 0.457 and 0.492, below labels alone. The highlighted Claude 3.5 eight-shot judgment-plus-reasoning setting reaches 0.581 versus a 0.482 base on harmlessness and 0.558 versus 0.500 on helpfulness. Its planned contrasts give p=0.046 and p=0.008, but the appendix reports that after FDR correction of per-configuration comparisons only 3 of 64 gains survive on harmlessness and none on helpfulness. The headline “up to 9.9 points” is a selected best case, not the average effect.

The individualized signal is genuine but modest. Against a predictor that always chooses each evaluator's most frequent label, PERSONAJUDGE's mean does not improve significantly: -0.019 with p=0.95 on harmlessness and +0.042 with p=0.14 on helpfulness. On items where an evaluator departs from the group majority, 20.2% and 23.3% of cases, it recovers the exact personal label at 0.367 and 0.360. When the model also departs from consensus, it selects the evaluator's direction at 0.617 and 0.632, above the binary 0.5 chance level. Neutral remains hardest, with recall 0.356 and 0.242. Per-evaluator accuracy ranges from 0.375 to 0.565 and from 0.386 to 0.655. Across the ten overlapping annotators, simulatability itself does not transfer between tasks, r=0.181, p=0.616, while neutral-use rate does, r=0.728, p<0.05, an externally unreplicated inference based on n=10.

The design has meaningful strengths: disjoint demonstration and validation sets, an explicit Neutral class, response-order controls, a non-target-evaluator control, and reporting by evaluator, class, and configuration. However, explanations are collected after replay and can rationalize rather than reveal the original decision process. No delayed human test-retest is measured, so the ceiling for reproducing a person's label is unknown; other domains, languages, modalities, and populations remain untested. The paper states that consent was obtained and audio stored confidentially, but provides no ethics-review identifier, compensation, country, demographics, or employment conditions. The arXiv source contains the prompt, tables, and figures but not the 4,200 labels, telemetry, transcripts, 161,280 predictions, scripts, seeds, or cited workbook. The defensible contribution is therefore a large experiment showing an incremental, task-dependent signal in evaluators' historical examples. It does not establish a stable personality, a reliable copy of a person, out-of-domain generalization, or a substitute for human evaluators.

Español

PERSONAJUDGE pregunta si un modelo juez puede reproducir el criterio de una persona concreta a partir de decisiones anteriores de esa misma persona. El estudio usa dos tareas del conjunto Anthropic Helpful and Harmless: 700 conversaciones únicas para helpfulness y otras 700 para harmlessness, cada una anotada por tres personas. Participan 32 anotadores profesionales de un equipo interno; 21 trabajan en cada tarea y 10 aparecen en ambas. Cada anotador completa 100 comparaciones con tres respuestas posibles, preferir A, neutral o preferir B, para un total de 4.200 juicios humanos. En una segunda fase revisa la reproducción de su interacción y verbaliza retrospectivamente cómo entendió la conversación, evaluó las respuestas y decidió. El sistema conserva la etiqueta, telemetría de interfaz, clics, desplazamiento, selección, espera y escritura, y esa explicación post hoc. No es una muestra de preferencias de la población general ni una simulación completa de personas: el objetivo empírico es predecir etiquetas de evaluadores profesionales en dos tareas muy próximas.

Por persona y tarea, 40 ítems forman el banco de demostraciones y 60 el conjunto de validación. PERSONAJUDGE descompone cada predicción en dos rondas: primero decide si esa persona expresaría preferencia o neutralidad y, si prevé preferencia, elige su dirección. Cruza cuatro modelos, Claude 3.5 Sonnet V2, Claude 3.7 Sonnet V1, DeepSeek-R1 y Amazon Nova Premier, cuatro contenidos de demostración, solo juicio; juicio más telemetría; juicio más razonamiento retrospectivo; o los tres, y 1, 2, 4 u 8 ejemplos. Son 64 condiciones, 80.640 simulaciones por tarea y 161.280 en total. Los ítems de validación no entran en el banco, pero cada condición usa una sola muestra aleatoria de demostraciones; no hay repeticiones que estimen la sensibilidad a qué ejemplos concretos caen en el prompt. Tampoco se publican snapshots inmutables, parámetros de inferencia, respuestas de los modelos o un entorno ejecutable.

Promediando las 64 condiciones, el juez sin personalización obtiene exactitud 0,452 en harmlessness y 0,496 en helpfulness; PERSONAJUDGE alcanza 0,480 y 0,510. En el subconjunto comparable, las demostraciones de la persona correcta superan a demostraciones de otros evaluadores por 0,028 puntos en harmlessness, 0,477 frente a 0,450; p=0,019, y 0,044 en helpfulness, 0,515 frente a 0,471; p<0,001. El razonamiento retrospectivo es la señal más útil: juicio más razonamiento promedia 0,505 y 0,537, mientras juicio más telemetría cae a 0,457 y 0,492, incluso por debajo de usar solo etiquetas. La configuración destacada, Claude 3.5 con ocho ejemplos de juicio y razonamiento, llega a 0,581 frente a 0,482 en harmlessness y 0,558 frente a 0,500 en helpfulness. Sus contrastes planificados dan p=0,046 y p=0,008, pero el apéndice muestra que, tras corrección FDR de las comparaciones por configuración, solo 3 de 64 mejoras sobreviven en harmlessness y ninguna en helpfulness. La cifra de “hasta 9,9 puntos” describe el mejor caso seleccionado, no el efecto general.

La personalización detectada es real pero modesta. Frente al predictor que siempre usa la etiqueta más frecuente de cada persona, la media de PERSONAJUDGE no mejora significativamente: diferencia -0,019 con p=0,95 en harmlessness y +0,042 con p=0,14 en helpfulness. En los ítems donde una persona discrepa de la mayoría, 20,2% y 23,3%, acierta su etiqueta exacta en 0,367 y 0,360; cuando el sistema también se aparta del consenso elige la dirección individual correcta en 0,617 y 0,632, por encima del azar binario de 0,5. Aun así, neutralidad es la clase más difícil, con recall 0,356 y 0,242. La exactitud individual varía mucho, de 0,375 a 0,565 y de 0,386 a 0,655. Entre las diez personas presentes en ambas tareas, la simulatabilidad no se transfiere, r=0,181; p=0,616; sí se correlaciona el uso de neutralidad, r=0,728; p<0,05, una inferencia frágil por n=10 y sin réplica externa.

El diseño tiene fortalezas: separa banco y validación, conserva la opción neutral, controla el orden de respuestas, compara contra demostraciones de otras personas y publica resultados por evaluador, clase y configuración. Pero las explicaciones se recogen después de ver una reproducción y pueden racionalizar la decisión, no revelar su causa. El estudio no mide estabilidad test-retest humana, de modo que desconoce el techo de predicción de una etiqueta personal; tampoco prueba otros dominios, idiomas, modalidades o poblaciones. La documentación afirma consentimiento y confidencialidad del audio, pero no identifica revisión ética, compensación, país, demografía ni condiciones laborales. El TeX de arXiv contiene el prompt, tablas y figuras, pero no los 4.200 juicios, telemetría, transcripciones, 161.280 predicciones, scripts, seeds ni el workbook citado. Por tanto, la contribución defendible es un experimento amplio que demuestra una señal incremental y dependiente del contexto en ejemplos históricos de evaluadores. No demuestra una personalidad estable, una copia fiable de la persona, generalización fuera de estas tareas ni capacidad para sustituir evaluadores humanos.

Research question

To what extent can an LLM reproduce the decisions of a specific human evaluator using their historical judgments, interface telemetry, and retrospective explanations, which signals help, and which properties predict the difficulty of simulation?

Method

4×4×4 factorial study with 32 internal professional annotators, 4,200 human judgments on Helpful and Harmless tasks, per-person split of 40 demonstrations and 60 validation items, four judge models, four demonstration contents, and four numbers of examples. Each label is simulated through a binary cascade preference-neutrality and direction. The audit read the complete PDF, rendered and inspected its 22 pages, and checked tables, figures, appendices, prompt, TeX, hashes, baselines, multiplicity, and artifact availability.

Sample: Thirty-two professional annotators from an internal team, 21 per task and 10 in both. Each completes 100 items in the assigned task; each conversation receives three labels. No country, demographics, compensation, detailed contractual relationship, or representativeness analysis are reported.

Findings

  • PERSONAJUDGE averages 0.480 versus 0.452 for the Base Judge on harmlessness and 0.510 versus 0.496 on helpfulness.
  • On the paired subset, demonstrations from the target person outperform the other evaluator control by 0.028 and 0.044 points.
  • Judgment plus retrospective reasoning is the best family marginally, with 0.505 and 0.537.
  • Interface telemetry reduces the mean relative to using only labels on both tasks.
  • The highlighted configuration improves 9.9 points on harmlessness and 5.8 on helpfulness, but it is a selected best case.
  • After FDR per configuration, 3 of 64 improvements survive on harmlessness and none on helpfulness.
  • The mean does not significantly exceed the per-person majority label baseline.
  • Accuracy on individual deviations is 0.367 and 0.360; captured direction is 0.617 and 0.632.
  • Neutral is the most difficult class, with recall 0.356 and 0.242.
  • Simulatability does not correlate across tasks in the ten common persons, but their neutrality rate does.
  • Retrospective reasoning requires approximately five times more time than the initial judgment.
  • Public artifacts do not allow reproducing the simulations or the analyses.

Limitations

  • The tasks are restricted to helpfulness and harmlessness comparisons from the same benchmark.
  • The 32 participants are internal professional annotators, not a population sample.
  • Only ten persons appear in both tasks.
  • No demographics, country, compensation, or working conditions are reported.
  • Consent is declared, but no ethical review, exemption, or institutional identifier.
  • No delayed test-retest of the human judgments is performed.
  • The explanations are retrospective and may be post hoc rationalizations.
  • Textualized telemetry may be too granular or noisy to represent attention.
  • Each condition uses a single sample of demonstrations and does not estimate variation due to example selection.
  • Two helpfulness evaluators use a special case without the neutrality round.
  • The results average across highly heterogeneous models, demonstration types, and shots.
  • The best result of 9.9 points is a selection among multiple configurations.
  • The mean improvement is small and does not exceed the per-person majority baseline.
  • Neutral has low recall and ambiguous cases are precisely the most difficult.
  • No immutable snapshots or complete inference parameters are published.
  • No judgments, telemetry, transcripts, predictions, scripts, seeds, or statistical workbook are published.
  • The arXiv source package only contains TeX, bibliography, figures, and static tables.
  • There is no out-of-domain, language, modality, or population evaluation.
  • No downstream impact on reward modeling, auditing, or safety decisions is measured.
  • The use of interaction traces and reasoning raises privacy and workplace surveillance risks.

What the study does not establish

  • It does not demonstrate human or stable personality in the simulated evaluator.
  • It does not demonstrate that the model reconstructs the causal decision process.
  • It does not demonstrate a reliable copy of a person beyond these two tasks.
  • It does not demonstrate generalization to unseen evaluators, ordinary users, or diverse populations.
  • It does not demonstrate that telemetry improves simulation.
  • It does not demonstrate that retrospective reasoning is a truthful explanation of the original judgment.
  • It does not demonstrate that PERSONAJUDGE outperforms a simple personal frequency predictor.
  • It does not demonstrate independent reproducibility without the data and outputs.
  • It does not justify substituting human evaluators in real decisions.

Traceability

Scope: Full text

Version: arXiv:2607.05742v1 [cs.HC], submitted 7 Jul 2026; 22 pages; complete PDF and TeX source package

Consulted source: https://arxiv.org/abs/2607.05742v1

Review: Codex full-text, visual, statistical, human-subjects, artifact-availability and claim-boundary audit, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • Claude-3.5-Sonnet-V2
  • Claude-3.7-Sonnet-V1
  • DeepSeek-R1
  • Amazon Nova Premier

Instruments and metrics

  • Three-class evaluator judgment: Prefer A, Neutral, Prefer B
  • Two-round preference-detection and direction cascade
  • Structured interface telemetry event schema
  • Replay-cued retrospective think-aloud protocol
  • Three-class accuracy against each evaluator own held-out label
  • Cross-evaluator matched-demonstration control
  • Per-evaluator and global majority-class baselines
  • Leave-one-out oracle consensus reference
  • Repeated-measures ANOVA, Friedman tests, Wilcoxon signed-rank contrasts, Holm and FDR corrections
  • Deviation accuracy, alignment lift and captured direction

Data used

  • Anthropic Helpful and Harmless conversational response pairs
  • 700 unique helpfulness conversations and 700 unique harmlessness conversations
  • 4,200 human judgments: 2,100 per task
  • Interface event traces and replay-cued retrospective reasoning from 32 in-house annotators
  • 161,280 reported simulated judgments across 64 configurations per task
  • arXiv TeX source package with prompt, static tables and figures but no raw study data or executable analysis

Evidence and location

  • Identity and version: arXiv:2607.05742v1, title page and metadata, 7 Jul 2026
  • Complete PDF inspected: .cache/editorial-sources/article-079/source.pdf; 22 pages; sha256 dc73b3d8909bd059b29a5fa6048ccce88eade22ea8b8c6914c9f098666073753
  • Sample and tasks: Paper §§4.1–4.2 and Appendix A
  • 4×4×4 design and 161,280 simulations: Paper §§4.3–4.4
  • Mean results and cross-control: Paper §5.1.1 and Figure 2
  • Effect of reasoning and telemetry: Paper §5.1.2 and Table 1
  • Best configuration and multiplicity: Paper Table 2 and Appendix H.4
  • Majority baseline and deviations: Paper §5.1.3, Tables 9 and 13–16
  • Transfer across tasks: Paper §5.1.3 and Figure 3; n=10
  • Temporal cost: Paper Appendix B and Table 3
  • Prompt and telemetry schema: Paper Appendices C–E and TeX source package
  • Absence of executable artifacts: Complete arXiv source-package inventory audited 18 Jul 2026; no raw data, code, outputs or workbook
  • Comprehensive visual inspection: All 22 pages rendered and visually inspected, including Figures 1–8 and Tables 1–16, 18 Jul 2026