Diagnosing and Repairing Persona Collapse in LLM Advice

Personas, identity, and agents2026arXivApproved editorial review

Authors: Harsh Kumar, Karina Vold, Louis Tay, Ashton Anderson

Keywords: persona collapse, LLM advice, situation-conditioned persona, post-training, inverse-process distillation, human preference, alignment

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

4
Authors
6
Findings
7
Limitations
4
Evidence

Editorial summary

English

The paper defines “persona collapse” as using nearly the same advisory posture even when the situation changes. It places five postures in a two-axis space, hedonic tone and support for agency, and uses gpt-5.4-nano to label 1,281 top-voted responses from 14 Reddit contexts. Human responses distribute across Healer/Guide (49.2%), Doomer (21.6%), Stoic (14.7%), Enabler (9.4%), and Technician (5.1%), whereas GPT-5.1, Claude Opus 4.5, and Gemini 3 Pro place 89.2% to 99.8% of responses in Healer. This human reference describes community norms, not advice quality or effectiveness. The judge achieves 60.0% average exact agreement across the two axes and 84.1% when allowing ±1 on agency. For repair, the authors build an 8,262-item corpus and compare six conditions on OLMo3-7B, Llama-3.1-8B, and Qwen3-4B. Fine-tuning increases diversity and improves kappa, but item selection remains weak, approximately 0.194–0.215 for Inverse-Process, and repeatedly confuses constructive challenge with corrosive harshness. The abstract’s claim of reducing divergence by “approximately 80%” does not match Table 13: from Instruct to Inverse-Process, JS reduction is 62.3%, 60.2%, and 61.1% across the three models. The preregistered human study retains 199 of 300 recruits and contradicts its primary hypothesis: every SFT variant scores much worse than Instruct on tone fit, understanding, and truth, and higher on harm; only the change across four exposures is exploratory. No longitudinal outcome or actual benefit is measured. The working draft links no public code, data, checkpoints, or outputs, preventing full independent reproduction.

Español

El artículo define “persona collapse” como el uso casi invariable de una postura de consejo aunque cambie la situación. Representa cinco posturas en dos ejes, tono hedónico y apoyo a la agencia, y etiqueta con gpt-5.4-nano 1.281 respuestas mejor votadas de 14 contextos de Reddit. Las respuestas humanas se distribuyen entre Healer/Guide (49,2 %), Doomer (21,6 %), Stoic (14,7 %), Enabler (9,4 %) y Technician (5,1 %); GPT-5.1, Claude Opus 4.5 y Gemini 3 Pro concentran entre 89,2 % y 99,8 % en Healer. Esta referencia describe normas comunitarias, no calidad o eficacia del consejo. El juez alcanza 60,0 % de acuerdo exacto promedio sobre dos ejes y 84,1 % cuando se tolera ±1 en el eje de agencia. Para reparar el patrón, se crea un corpus de 8.262 pares y se comparan seis condiciones en OLMo3-7B, Llama-3.1-8B y Qwen3-4B. El fine-tuning aumenta diversidad y mejora kappa, pero la selección sigue débil (aproximadamente 0,194–0,215 para Inverse-Process) y confunde desafío constructivo con dureza corrosiva. La afirmación del abstract de reducir la divergencia “aproximadamente 80 %” no coincide con la Tabla 13: de Instruct a Inverse-Process la reducción JS es 62,3 %, 60,2 % y 61,1 % en los tres modelos. El estudio humano prerregistrado conserva 199 de 300 participantes y contradice su hipótesis principal: todas las variantes SFT reciben puntuaciones mucho peores que Instruct en ajuste, comprensión y honestidad, y mayores en daño; solo el cambio dentro de cuatro exposiciones es exploratorio. No se miden resultados longitudinales ni beneficio real. La versión de trabajo no enlaza código, datos, checkpoints o salidas, por lo que no permite reproducción independiente completa.

Research question

Do models choose a context-sensitive advice stance, can an excessive concentration on warm support be repaired, and do people prefer the repaired responses?

Method

Diagnosis of 1,281 posts in 14 contexts with LLM labels on two axes; repair corpus of 8,262 items with 80/20 split; six interventions on three open models; bootstrap by item; and preregistered, blind, within-participant human study.

Sample: 1,281 diagnostic situations; 8,262 training/repair items with 1,652 in test; three open models; 300 recruited and 199 retained participants, each with four posts and six responses per post.

Findings

  • Frontier models concentrate 89.2–99.8% of responses in Healer.
  • Plan-First worsens diversity and divergence in all three models.
  • SFT restores diversity, but per-item alignment remains low.
  • The JS reduction of Inverse-Process is around 60–62%, not 80%.
  • Participants clearly prefer Instruct and judge SFT variants as more harmful.
  • The gap reduction over four rounds is exploratory and not longitudinal.

Limitations

  • Community votes are not ground truth of good advice.
  • The five-stance framework compresses mixed responses and depends on an LLM judge with moderate agreement.
  • The split is by item and the corpus artifacts are not public.
  • Fine-tuning reproduces human responses and a posteriori reconstructed reasoning.
  • The evaluation is mostly single-turn.
  • There are no clinical, behavioral, or long-term well-being outcomes.
  • There is no reproducible public repository in v1.

What the study does not establish

  • It does not demonstrate that a confrontational stance helps the user more.
  • It does not validate the most voted responses as true or safe advice.
  • It does not prove that the model possesses discrete internal personas.
  • It does not demonstrate an 80% reduction in the three models.
  • It does not show longitudinal preference or help.
  • It does not authorize clinical or crisis use.

Traceability

Scope: Full text

Version: arXiv:2607.08326v1 working draft; complete 32-page PDF and TeX source; AsPredicted #292682 verified; no public code/data repository identified

Consulted source: https://arxiv.org/abs/2607.08326v1

Review: Codex 32-page visual full-text, TeX, preregistration, arithmetic, human-study and artifact audit, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • GPT-5.1
  • Claude Opus 4.5
  • Gemini 3 Pro
  • gpt-5.4-nano (judge)
  • OLMo3-7B-Instruct
  • Llama-3.1-8B-Instruct
  • Qwen3-4B-Instruct
  • frontier reasoning teacher

Instruments and metrics

  • Hedonic tone H (-1, 0, +1)
  • Agency depth E (-2 to +2)
  • Five advisory persona regions
  • Effective number of personas
  • Jensen-Shannon divergence
  • Macro-recall and Cohen kappa
  • Five 7-point human-rating items and two rankings

Data used

  • 1,281-item diagnostic Reddit corpus
  • 8,262-item Reddit/Stack Exchange/CounselChat/CareerNet repair corpus
  • 165-post human-evaluation pool

Evidence and location

  • Framework, diagnostic corpus, and Healer concentration: arXiv v1, sections 3–4 and Figures 1–3
  • Repair and results by model: arXiv v1, section 5 and Appendix Table 13
  • Preregistration and human study: AsPredicted #292682; arXiv v1, section 6 and Appendix H
  • Limitations and absence of real-world outcomes: arXiv v1, sections 7–8