Persona-judge: Personalized Alignment of Large Language Models via Token-level Self-judgment

Trait induction and control2025arXivApproved editorial review

Authors: Xiaotian Zhang, Ruizhe Chen, Yang Feng, Zuozhu Liu

Keywords: Large Language Models, Personality, Persona, Model Evaluation, LLM Evaluation

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

4
Authors
9
Findings
32
Limitations
10
Evidence

Editorial summary

English

Persona-judge is a decoding algorithm for combining two written preference descriptions without parameter updates. The same checkpoint is loaded twice as draft and judge and receives different prefixes, for example, creative and vivid, alongside the user query. The draft proposes a window of lambda tokens from distribution q. The judge computes p under the other prefix; each token is accepted with probability min(1,p/q), and after rejection it is resampled from the normalized positive part of p−q. The appendix proves that for a step with fixed q and p, the corrected token is distributed exactly as p. Roles then alternate. Thus, self-judgment means that two instances of the same model, conditioned by different instructions, compare token distributions. There is no psychological self, learned persona embedding, or reward model during generation. A more precise label is training-free decoding-time multi-attribute steering. Personalization consists of researcher-written criteria; no users state preferences and no individual profile is constructed. Evaluation uses the 50-prompt P-Soups/Koala set and HelpSteer2, filtered by prompt length without reporting final n. Applying the released selection to the current official split yields 146 unique prompts, but no dataset revision is pinned, so this cannot be guaranteed to be the paper's exact sample. The primary backbone is Llama-3-Base-8B-SFT and model generality is tested across nine models from 0.5B to 9B. Criteria are helpful, harmless, humor, correct, informative, professional, creative, touching, and vivid. On P-Soups helpful/harmless, Persona-judge averages 0.93 versus 0.90 for Aligner and 0.90 for its fixed-role variant. That value is an arithmetic mean of four outputs from three reward checkpoints, two ArmoRM dimensions and two GPT-2 reward models, on different scales. Its weighting is not justified, and the 0.03 gap has no interval or test. GPT-4 also compares each output with direct preference prompting. Across nine models and nine criteria, win rates range from 40% to 98% on P-Soups and 45% to 99% on HelpSteer2. Many pairs favor Persona-judge, but not all: examples below 50% include 40% on correct for Llama-3.2-3B on P-Soups, 45% on helpful for Qwen 0.5B on HelpSteer2, and 47% on creative for TinyLlama. The claimed 87% advantage for Gemma 2 9B is the mean GPT-4 preference rate from 82%, 88%, and 90%, not an 87-point absolute improvement. GPT-4 has no fixed model/version, human validation, repeated judgments, or intervals. The paper acknowledges position, self-enhancement, length, and prompt biases but only says response order is randomized. Significant is descriptive, not statistical. Training-free means no fine-tuning, not low cost. Generation requires two copies of the checkpoint and both token distributions. On four A40 GPUs, reported latency is 8.82±0.45 s for vivid, 8.91±0.37 s for two objectives, and 9.13±0.39 s for three, but there is no ordinary-decoding or direct-prompt baseline, token control, or throughput, so relative computational efficiency is not established. The official repository permits sampler inspection but cannot reproduce the tables as released. Its two results files are 27- and 30-byte text placeholders rather than results. koala_eval_50.json is absent and HelpSteer2 is read from an author's absolute local path. Active prefixes are hardcoded to vivid/creative. sample_temp, mode, rm_weight, topk, and --max_new_token are parsed or passed but do not affect the active path; generation hardcodes top-k 20, top-p 0.9, temperature 0.9, seed 123, and 128 tokens. This conflicts with the paper's greedy decoding with top-k candidates. The GPT-4 evaluator and judgments are missing, and lambda 1–6 cannot be selected through a working public interface without source editing. The README runs only 10% of P-Soups. The audited commit predates arXiv v2 by six days and has no tag or release. The faithful conclusion is that a corrected sampler alternating two conditioned distributions can often improve, according to reward models and GPT-4, joint expression of style/quality attributes relative to placing both attributes in one prompt. It does not show that the system learns individual human values, personality, user satisfaction, or durable alignment; nor does it establish unlimited scalability, safety, or savings over alternatives. The claim of no identifiable risks also overlooks that the same steering could amplify harmful, deceptive, or discriminatory preferences.

Español

Persona-judge es un algoritmo de decodificación para combinar dos descripciones de preferencias sin actualizar parámetros. El mismo checkpoint se carga dos veces, como draft y judge, y recibe prefijos distintos, por ejemplo, creative y vivid, además de la consulta. El draft propone una ventana de lambda tokens desde una distribución q. El judge calcula p bajo el otro prefijo; cada token se acepta con probabilidad min(1,p/q) y, si se rechaza, se remuestrea desde la parte positiva normalizada p−q. La demostración del apéndice establece que, para un paso con q y p fijas, el token corregido queda distribuido exactamente como p. Después los papeles se alternan. Por tanto, “self-judgment” significa que dos instancias del mismo modelo, condicionadas por instrucciones diferentes, comparan distribuciones token a token. No hay una entidad psicológica que se juzgue, un embedding aprendido de persona ni un reward model durante la generación. La etiqueta más precisa es steering multiatributo training-free en tiempo de decodificación. La personalización se limita a criterios escritos por investigadores; ningún usuario expresa preferencias y no se construye un perfil individual. El estudio usa P-Soups/Koala con 50 prompts y HelpSteer2, que se filtra por longitud sin informar el n final. Aplicar el código publicado a la versión oficial vigente deja 146 prompts únicos, pero no se fija una revisión del dataset y por ello no puede asegurarse que sea exactamente la muestra del paper. El backbone principal es Llama-3-Base-8B-SFT y la generalidad se prueba en nueve modelos entre 0,5B y 9B. Los objetivos son helpful, harmless, humor, correct, informative, professional, creative, touching y vivid. En helpful/harmless de P-Soups, Persona-judge obtiene una media 0,93 frente a 0,90 para Aligner y 0,90 para la variante de papeles fijos. Esa media es el promedio aritmético de cuatro salidas procedentes de tres reward models, dos dimensiones de ArmoRM y dos GPT-2 reward models, con escalas distintas; su ponderación no está justificada y la diferencia de 0,03 no tiene intervalo ni test. GPT-4 también compara cada salida con el prompt directo. En nueve modelos y nueve criterios, las tasas de victoria van de 40% a 98% en P-Soups y de 45% a 99% en HelpSteer2. Muchos pares favorecen Persona-judge, pero no todos: hay resultados inferiores a 50%, como 40% en correct para Llama-3.2-3B sobre P-Soups, 45% en helpful para Qwen 0.5B sobre HelpSteer2 y 47% en creative para TinyLlama. Los “87% de ventaja” citados para Gemma 2 9B son una tasa media de preferencia GPT-4 de 82%, 88% y 90%, no una mejora absoluta de 87 puntos. GPT-4 no tiene modelo/version fijado, validación humana, repeticiones ni intervalos; el paper reconoce sesgo de posición, autorrefuerzo, longitud y sensibilidad de prompt, pero solo dice aleatorizar el orden. El término “significativo” es descriptivo, no estadístico. Training-free significa que no se afina un modelo, no que sea barato. La generación requiere dos copias del checkpoint y sus distribuciones. En cuatro A40, el paper reporta 8,82±0,45 s para vivid, 8,91±0,37 s para dos objetivos y 9,13±0,39 s para tres, pero no compara con decodificación ordinaria o prompt directo, no controla tokens y no da throughput; por eso no demuestra eficiencia computacional relativa. El repositorio oficial permite inspeccionar el sampler, pero no reproducir las tablas tal como está. Los dos ficheros de results son texto marcador de 27 y 30 bytes, no resultados. Falta koala_eval_50.json y HelpSteer2 se lee desde una ruta absoluta del ordenador del autor. Los prefijos activos están hardcodeados a vivid/creative. sample_temp, mode, rm_weight, topk y --max_new_token se parsean o pasan pero no afectan a la ruta activa; el generador fija top-k 20, top-p 0,9, temperatura 0,9, semilla 123 y 128 tokens. Esto contradice la descripción de greedy decoding con candidatos top-k. No se publica el evaluador GPT-4 ni sus juicios y no hay una vía funcional para seleccionar lambda 1–6 sin editar código. El README ejecuta solo 10% de P-Soups. El commit auditado precede seis días a arXiv v2 y no tiene tag o release. La conclusión fiel es que un sampler correctivo que alterna dos distribuciones condicionadas puede mejorar con frecuencia, según reward models y GPT-4, la expresión conjunta de atributos de estilo/calidad frente a escribir ambos atributos en un prompt. No demuestra que el sistema aprenda valores humanos individuales, personalidad, satisfacción de usuarios o alineamiento duradero; tampoco demuestra escalabilidad ilimitada, seguridad o ahorro frente a alternativas. Además, afirmar que no hay riesgos identificables omite que el mismo steering podría reforzar preferencias dañinas, engañosas o discriminatorias.

Research question

Can a token-by-token sampler, using two preference prefixes and two instances of the same model, combine unseen objectives without fine-tuning or reward model and outperform direct prompting?

Method

One draft instance proposes up to lambda tokens from q and another instance of the same checkpoint, conditioned by another prefix, acts as judge with distribution p. Acceptance min(1,p/q) and resampling from max(0,p-q) are applied; the roles alternate. It is compared with alignment methods and direct prompting on P-Soups and a filtered subset of HelpSteer2 using three reward checkpoints and GPT-4 pairwise.

Sample: Fifty P-Soups prompts and an unquantified subset of HelpSteer2. Reconstruction on the current official version produces 146 unique prompts, but the dataset is not fixed by review. Nine backbones are evaluated on nine attributes. No users who provide preferences participate, nor are people or longitudinal profiles studied.

Findings

  • The current source is arXiv:2504.12663v2, revised on 11 June 2025, 13 pages, Findings of ACL 2025.
  • The 13 pages were rendered and visually inspected; the current PDF matches byte for byte with the cache and has SHA-256 3081251deccf1630b90b11ddde649db200123d4f6e0568486bedac4636a262f2.
  • Persona-judge does not update parameters and does not use a reward model during generation.
  • The accept-reject correction samples exactly from the judge distribution for each fixed step.
  • The reported P-Soups helpful/harmless mean is 0.93 versus 0.90 for the next best aggregates.
  • GPT-4 rates cover 40-98% on P-Soups and 45-99% on HelpSteer2; not all pairs improve over direct prompting.
  • Lambda=4 obtains the best aggregate in the sensitivity table, but there is no test or repetition establishing robustness.
  • Latency increases from 8.82 s for one objective to 9.13 s for three, without a direct decoding baseline.
  • The code compiles syntactically, but the release does not contain datasets, results, or reproducible GPT-4 evaluation.

Limitations

  • Personalization means attributes written by authors, not observed preferences of individuals.
  • There are no users, elicitation, profiles, memory, value conflict, or human satisfaction.
  • Unseen means without method-specific training, not unknown to the model pretraining.
  • The mechanism alternates conditional distributions; it does not learn a joint representation of preferences.
  • The correctness proof demonstrates sampling from p, not human alignment or multiobjective optimality.
  • P-Soups contains only 50 prompts.
  • The final HelpSteer2 size is not reported and the dataset review is not fixed.
  • The nine dimensions are general quality and style criteria, not a sampling of diverse human values.
  • Most scalability results depend on a single unversioned GPT-4.
  • There is no human validation of the judge, agreement, repetition, interval, or bias evaluation.
  • The order is randomized, but no results by order or inverted double evaluation are reported.
  • The words significant and advantage are used without statistical inference.
  • The 0.93 mean combines outputs from distinct reward scales with arbitrary weighting.
  • Reward models may favor length or styles close to their training data.
  • No distributions, variance across prompts, intervals, or tests per model/criterion are published.
  • There are several win rates below 50%, so the benefit is not universal.
  • Adversarial, contradictory, harmful, or discriminatory preferences are not evaluated.
  • The risks section states there are no identifiable risks, an unsupported conclusion.
  • Training-free requires two copies of the checkpoint and double computation of distributions.
  • Latency is not compared with direct prompting, normal decoding, steering, or training-based methods.
  • Effective output length, throughput, GPU memory, or token acceptance is not reported.
  • Four A40s are used, but efficiency is not normalized by hardware or cost.
  • The public result files are text placeholders, not outputs.
  • koala_eval_50.json is missing and HelpSteer2 points to a local absolute path.
  • The active prefixes in the code are hardcoded to vivid and creative.
  • sample_temp, mode, rm_weight, topk, and max_new_token do not control the active path as the interface promises.
  • The code uses top-k/top-p/temperature sampling, not greedy decoding as the paper describes.
  • The GPT-4 evaluation and its data are not published.
  • The lambda 1-6 table is not reproduced through a functional argument.
  • The README example uses only 10% of P-Soups.
  • There is no tag, release, CI, or commit explicitly linked to arXiv v2.
  • Interactive deployment, persistence across turns, or longitudinal stability is not tested.

What the study does not establish

  • It does not establish personality or internal values of an LLM.
  • It does not demonstrate alignment with real people's preferences.
  • It does not demonstrate learning of individual profiles.
  • It does not demonstrate satisfaction, trust, or benefit for users.
  • It does not demonstrate that two prefixes resolve conflicts between values.
  • It does not demonstrate improvement across all models and objectives.
  • It does not demonstrate statistical significance of the win rates or of the 0.93 aggregate.
  • It does not demonstrate efficiency over direct prompting or standard decoding.
  • It does not demonstrate unlimited scalability to any human preference.
  • It does not demonstrate safety against harmful preferences.
  • It does not allow reproducing the full tables with the published release.

Traceability

Scope: Full text

Version: arXiv:2504.12663v2, revised 11 June 2025, 13 pages; Findings of ACL 2025

Consulted source: https://arxiv.org/abs/2504.12663

Review: Codex complete bilingual full-text fidelity pass, current arXiv v2 and byte-level PDF verification, all-page visual inspection, token-sampling proof interpretation, table reconstruction, live official HelpSteer2 sample reconstruction, GPT-4 and reward-evaluator validity audit, public repository static code audit, interface-versus-active-path reconciliation, and reproducibility assessment; summaries written from the complete paper, tables, official dataset, and code rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • princeton-nlp/Llama-3-Base-8B-SFT primary backbone
  • Qwen2.5-0.5B-Instruct and Qwen2.5-3B-Instruct
  • TinyLlama-1.1B-Chat-v1.0
  • Gemma-2-2b-it and Gemma-2-9b-it
  • Llama-3.2-3B-Instruct
  • Tulu-2-dpo-7b
  • Llama-3.1-Tulu-3-8B
  • RLHFlow/ArmoRM-Llama3-8B-v0.1
  • Ray2333 GPT-2 helpful and harmless reward models
  • Unversioned GPT-4 pairwise evaluator

Instruments and metrics

  • Persona-judge draft/judge accept-reject sampler
  • Direct multi-preference prompting baseline
  • Reward scores for helpful and harmless
  • GPT-4 pairwise win rate on nine written criteria
  • Arithmetic average across four heterogeneous reward outputs
  • Lambda sensitivity from 1 to 6
  • Wall-clock latency for one, two, and three objectives

Data used

  • P-Soups modified Koala evaluation set, 50 prompts
  • HelpSteer2 validation split, filtered by prompt length; final paper n unreported
  • Current official HelpSteer2 reconstruction: 146 unique prompts after released filtering and [1::2] selection
  • Official code repository commit ebd149a6097062ee0c46a7dc9c3b84e84a80276e
  • No released raw experimental outputs or GPT-4 judgments

Evidence and location

  • Version, title, authors, and venue: arXiv:2504.12663v2 metadata and page 1 checked 15 July 2026
  • Algorithm and distributive correctness: Sections 2.1-2.3, pages 2-3; Appendix C.2, page 10
  • Main comparison and win rates: Tables 1-3, pages 3-5
  • Models, evaluators, and GPT-4 protocol: Appendix D.1-D.3, pages 10-12
  • Nine criteria and results per model: Table 4, page 12
  • Inference cost: Appendix D.5 and Table 5, pages 12-13
  • Reconstructed HelpSteer2 size: Official nvidia/HelpSteer2 validation rows checked through the Hugging Face Dataset Viewer API on 15 July 2026
  • Code, arguments, datasets, and artifacts: Official repository commit ebd149a6097062ee0c46a7dc9c3b84e84a80276e audited 15 July 2026
  • Methodological and code audit: reports/verification/article-186-method-and-code-audit.json
  • Complete visual inspection: All 13 PDF pages rendered and visually inspected on 15 July 2026