Interpolative Decoding: Exploring the Spectrum of Personality Traits in LLMs

Trait induction and control2025arXivApproved editorial review

Authors: Eric Yeh, John Cadigan, Ran Chen, Dick Crouch, Melinda Gervasio, Dayne Freitag

Keywords: Interpolative Decoding, Contrastive Decoding, Big Five, HEXACO, Economic Games, Human Twinning, Behavioral Simulation

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

6
Authors
6
Findings
9
Limitations
5
Evidence

Editorial summary

English

The paper studies interpolative decoding: rather than authoring a separate prompt for every trait level, it defines two endpoint prompts and combines their token distributions with a scalar lambda. It tests two variants. Mixture decoding averages the distributions with lambda in [0,1], while contrastive decoding anchors one endpoint and amplifies its difference from the other. The clearest evidence concerns output control, not human personality. In the Big Five study, three of each trait's six facets supply 12 few-shot examples and the remaining three supply 12 assessment items. Random answers to the examples define the target score, and all 20 facet partitions are evaluated. The table reports correlations of 0.69 for openness, 0.83 for conscientiousness, 0.83 for extraversion, 0.70 for agreeableness and 0.73 for neuroticism. However, its caption calls them Spearman correlations between answers and target levels, whereas the text says Pearson correlation between lambda and trait score; the statistic and unit are not consistently described. Mixture decoding also switches almost discontinuously near lambda 0.5, while contrastive decoding changes more gradually. This shows that questionnaire output can be steered with related descriptions and examples; it does not validate an internal psychological trait. The second study uses a dictator game. It interpolates HEXACO honesty-humility, agreeableness and emotionality from -30 to 30 while permuting four prompt sentences. Coworker payout correlates 0.74/0.78 with honesty-humility, 0.49/0.56 with agreeableness and -0.03/-0.08 with emotionality in Pearson/Spearman terms. The qualitative ordering resembles the selected human literature, but no matched human sample is collected or reanalyzed, and no equivalence test, interval or significance result is given. In Pandemic, twelve conflicting social/tactical recommendation pairs are tested at eight lambdas, 96 described decisions. Following the co-player falls with lambda, Pearson -0.82 and Spearman -0.94, and author-inspected lexical counts move in corresponding directions. The human-twinning experiment is far narrower than the label suggests. It observes one person over five games and 25 Pandemic turns. An external planner first proposes only the top three or five move sets; Gemma 3 4B and 12B re-rank them through A/B comparisons in both orders. The authors acknowledge that order changes the result in nearly 50% of cases. Four mixture lambdas, four contrastive lambdas and a baseline are compared by perplexity on those same 25 turns, with no train/test separation or second person. Interpolation slightly lowers perplexity in three of four reported groups: 3.92 versus 3.95 for Gemma 12B top-3; 4.84 versus 5.00 for 12B top-5; and 4.93 versus 5.31 for Gemma 4B top-5. For Gemma 4B top-3, the baseline wins at 3.67 versus the best shown interpolation value of 3.97. Human-action coverage does not change with decoder: the planner misses an average 1.00 moves for top-3 and 0.80 for top-5. The table shows only configurations described as most significantly different from baseline, but no test, p-value, correction or complete grid is published. This is in-sample distribution fitting over planner-covered actions, not prospective individual prediction. A final MLP uses 1,294 synthetic responses for training and 214 for validation to recover lambda; MSE is 2.96, 2.19 and 1.28 by trait. Recovering a generation setting from the same LLM's text does not estimate a human trait. Reproducibility is very low. Outside the twinning subsection, the paper does not identify the LLM; temperature, top-p, seeds, versions and raw results are absent. The contrastive formula is written over probabilities rather than logits or log probabilities, and without code its actual implementation cannot be checked. The official BSD-3 repository at commit 42cfaf8 contains only a README, license and three images; none of its seven commits ever contained code or data, despite the paper saying they will be released there. The 25 human turns, consent, demographics, expertise, ethics review and privacy policy are also absent. The faithful conclusion is that lambda controls several textual regularities and decisions under the tested prompts. A human psychological continuum, formal replication and a twin that predicts new people or unseen turns are not established.

Español

El artículo estudia interpolative decoding: en lugar de redactar un prompt distinto para cada nivel de un rasgo, define dos prompts extremos y mezcla sus distribuciones token a token mediante un parámetro lambda. Prueba dos variantes. Mixture decoding promedia las distribuciones con lambda entre 0 y 1; contrastive decoding ancla un extremo y amplifica su diferencia con el otro. La evidencia más clara es de control de salidas, no de personalidad humana. En Big Five, para cada rasgo se eligen tres de seis facetas como 12 ejemplos few-shot y se reservan las otras tres como 12 ítems de evaluación. Las respuestas aleatorias a los ejemplos fijan el score objetivo y se recorren las 20 particiones posibles. La tabla publica correlaciones de 0,69 en apertura, 0,83 en responsabilidad, 0,83 en extraversión, 0,70 en amabilidad y 0,73 en neuroticismo. Sin embargo, el pie las llama Spearman entre respuestas y niveles objetivo, mientras el texto dice Pearson entre lambda y score; la unidad y el estadístico no quedan descritos de forma consistente. Mixture decoding además salta casi de un extremo al otro alrededor de lambda 0,5, mientras contrastive cambia de forma más gradual. Esto muestra que el cuestionario puede dirigirse con prompts y few-shot relacionados; no valida un rasgo psicológico interno. El segundo bloque usa un dictator game. Interpola honestidad-humildad, amabilidad y emocionalidad de HEXACO entre -30 y 30, permutando cuatro frases del prompt. El pago al compañero correlaciona 0,74/0,78 con honestidad-humildad, 0,49/0,56 con amabilidad y -0,03/-0,08 con emocionalidad, en Pearson/Spearman. El orden cualitativo coincide con la literatura humana seleccionada, pero no se recoge ni reanaliza una muestra humana con el mismo protocolo y no hay prueba de equivalencia, intervalos o significación. En Pandemic, doce pares de recomendaciones social/táctica se prueban en ocho lambdas, 96 decisiones descritas. La probabilidad de seguir al compañero cae con lambda, Pearson -0,82 y Spearman -0,94; también cambian conteos léxicos construidos por inspección de los autores. El experimento denominado human twinning es mucho más estrecho de lo que sugiere el término. Observa a una sola persona durante cinco partidas y 25 turnos de Pandemic. Un planner externo propone solo los tres o cinco mejores movimientos; Gemma 3 4B y 12B los reordenan mediante comparaciones A/B en ambos órdenes. Los autores reconocen que el orden cambia el resultado en casi el 50% de los casos. Se comparan cuatro lambdas de mixture, cuatro de contrastive y un baseline mediante perplexity sobre esos mismos 25 turnos, sin separación train/test ni segundo sujeto. Interpolación reduce ligeramente la perplexity en tres de cuatro grupos publicados: 3,92 frente a 3,95 para Gemma 12B top-3; 4,84 frente a 5,00 para 12B top-5; y 4,93 frente a 5,31 para Gemma 4B top-5. En Gemma 4B top-3 gana el baseline, 3,67 frente al mejor 3,97 interpolado. La cobertura de acciones humanas no cambia con el decoder: el planner pierde de media 1,00 movimientos con top-3 y 0,80 con top-5. La tabla muestra solo configuraciones descritas como más significativamente distintas del baseline, pero no publica test, p-value, corrección ni grid completo. Por ello es ajuste in-sample de una distribución sobre acciones ya cubiertas por el planner, no predicción individual prospectiva. Un último experimento entrena un MLP con 1.294 respuestas sintéticas y valida en 214 para recuperar lambda; los MSE son 2,96, 2,19 y 1,28 según rasgo. Recuperar un parámetro generativo desde texto del propio LLM no estima un rasgo humano. La reproducibilidad es muy baja. Fuera del apartado de twinning no se identifica el LLM empleado; faltan temperatura, top-p, seeds, versiones y resultados crudos. La fórmula contrastive se escribe sobre probabilidades, no logits o log-probabilidades, y sin código no puede verificarse la operación real. El repositorio oficial BSD-3, commit 42cfaf8, contiene únicamente README, licencia y tres imágenes; sus siete commits nunca incluyeron código ni datos, pese a que el paper dice que se publicarán allí. Tampoco se aportan los 25 turnos humanos, consentimiento, demografía, experiencia, revisión ética o política de privacidad. La conclusión fiel es que lambda controla varias regularidades textuales y decisiones en los prompts ensayados. No queda demostrado un continuo psicológico humano, una replicación experimental formal ni un gemelo capaz de predecir a personas nuevas o turnos no observados.

Research question

Can the token-by-token combination of two extreme prompts continuously control traits and decisions of an LLM and, by inverting that search, approximate the actions of a specific person?

Method

Compares mixture and contrastive decoding between pairs of prompts. Evaluates Big Five control with 20 facet partitions, dictator game decisions under three HEXACO traits, 96 social/tactical decisions in Pandemic, action fitting for one person across 25 turns with Gemma 3, and a lambda regression trained only on synthetic responses.

Sample: Five Big Five traits with 20 facet partitions; three HEXACO traits in a dictator game; 12 Pandemic scenarios by eight lambdas; one person, five games, and 25 turns for twinning; 1,294/214 synthetic tuples for regression. The model is not identified in the majority of experiments.

Findings

  • Lambda markedly modulates Big Five scores, although Table 1 does not consistently describe whether it reports Pearson or Spearman or its exact unit.
  • In the dictator game, honesty-humility relates more to payment than agreeableness and emotionality remains close to zero, reproducing only the qualitative pattern of selected literature.
  • In 96 Pandemic decisions, favoring the tactical signal reduces following the partner and shifts the lexicon toward the risk city.
  • In a single person, interpolated settings modestly reduce in-sample perplexity in three of four published groups; the baseline wins the fourth.
  • The regressor partially recovers lambda from synthetic text, but does not estimate personality or human behavior.
  • The public release contains no implementation, data, or reproducible results.

Limitations

  • Unidentified LLM in four of five empirical sections and absence of sampling config, seeds, hardware, and versions.
  • No raw data, intervals, tests, corrections, or full grid of configurations.
  • Pearson/Spearman inconsistency and undefined unit of analysis in Big Five.
  • Semantically arbitrary controls and mixture decoding nearly discontinuous rather than continuous.
  • Only isolated dimensions; multivariate trait combination is not tested.
  • Twinning with one person, 25 turns, no holdout, and conditioned by a planner that omits actions.
  • A/B decisions sensitive to order in nearly 50% of cases and unvalidated parsing fallback.
  • No human data, consent, demographics, experience, ethical review, or privacy.
  • Repository of five files with no code or data despite the promise of future publication.

What the study does not establish

  • An internal human psychological trait or a psychometrically valid continuum in the LLM.
  • That mixture decoding smoothly traverses the spectrum.
  • A human replication with a comparable protocol and sample.
  • Significance, robustness across seeds, controls, or model families.
  • Simultaneous control of several traits without interference.
  • Prospective prediction of a person, generalization to new turns, or to other individuals.
  • Twinning of communication, reasoning, or general behavior; only a distribution of actions is fitted.
  • That the synthetic regressor recovers a human trait.
  • Safety or legitimacy for unethical experiments, surveillance, or insider threat profiling.
  • Reproducibility via the linked repository.

Traceability

Scope: Full text

Version: arXiv:2512.19937v1, submitted 2025-12-23; preprint under review; official repository history audited separately

Consulted source: https://arxiv.org/abs/2512.19937

Review: Codex 20-page visual, official-arXiv-v1, full-method, Big-Five facet-split, correlation-definition, HEXACO dictator-game, 96-scenario information-weighting, single-subject 25-turn twinning, candidate-coverage, choice-order, selective-reporting, synthetic-regression, repository-history, human-data, ethics, reproducibility and claim-boundary audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Gemma 3 4B, identified only for human-twinning experiment
  • Gemma 3 12B, identified only for human-twinning experiment
  • Unidentified LLM for Big Five, dictator game, Pandemic information-weighting and regression-data experiments
  • Unspecified Sentence-BERT-style embedding model
  • Three-layer MLP lambda regressor

Instruments and metrics

  • Big Five inventory items split by six facets per trait
  • Big Five low/middle/high website descriptions
  • Author-rewritten HEXACO endpoint descriptions
  • Dictator game payout
  • Pandemic action choice and planner candidate sets
  • Author-inspected collaboration/other-player/tactical-city lexical counts
  • Pearson and Spearman correlations
  • Observed-action perplexity
  • Lambda mean-squared error

Data used

  • Twenty Big Five three-of-six facet partitions per trait; raw responses not released
  • Dictator-game generations over three HEXACO traits and seven lambda settings; raw data not released
  • Ninety-six Pandemic social-versus-tactical decisions; raw data not released
  • One player, five Pandemic games and 25 turns; human data not released
  • Synthetic lambda-regression data: 1,294 training and 214 validation tuples; not released

Evidence and location

  • Metadata, version, authors, category, and license: arXiv:2512.19937v1
  • Method, five experiments, tables, limitations, and appendices: arXiv v1 PDF, 20 pages, sha256 8631f22db04bec67b1f1f465a1ce6264b3e2adf67e82d750331a0292e9cd45d1
  • Institutional status and publication as preprint: SRI publication page and arXiv record
  • Actual content and evolution of the public artifact: SRI-AIC/foil commit 42cfaf8a7be6569b32905795f6cdd71cc1b0abce and all seven reachable commits
  • Psychometric audit, correlations, economic games, single subject, code, data, and boundaries: reports/verification/article-250-arxiv-interpolative-decoding-psychometrics-economic-games-human-twinning-code-data-and-claim-audit.json