Can Induced Emotion Bias LLM Behaviors in Sequential Decision Making?

Trait induction and control2026arXivApproved editorial review

Authors: Minh Khoi Ho, Zihao Zhu, Runchuan Zhu, Levina Li, Zhiwen Fan, Zhangyang Wang, Junyuan Hong

Keywords: Persona conditioning, Behavioral control, Longitudinal behavior

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

7
Authors
4
Findings
5
Limitations
3
Evidence

Editorial summary

English

Induction is validated with six models and 40 runs per emotion; three human annotators review 80 vignettes. The main experiment runs the Iowa Gambling Task for 100 rounds with four models, three agent variants, three scenarios, six deck permutations, and 18 matched seeds per configuration.

There are no human players. Validation covers six models; the main analysis uses GPT-OSS-20B, Llama-3.1-8B, Qwen2.5-7B, and Qwen3-4B with Query, ReAct, and Reflexion agents. No stable average emotion effect appeared on long-run performance. For Qwen2.5 Query, anger advanced lock-in by about 11 rounds, d=-1.29 and p=.0003. Exploration effects changed sign across models. A strong deck-label and position bias persisted.

The prompt induces a contextual self-description, not a felt emotion. Only the Iowa Gambling Task is used. Cell-level effects are numerous and mixed. There is no human baseline collected under the same protocol. Generalization depends on specific models, prompts, and endpoints. It does not demonstrate that LLMs feel emotions. It does not establish a general causal effect of anger on decisions. It does not validate equivalence with human affective dynamics.

Español

Se valida la inducción con seis modelos y 40 ejecuciones por emoción; tres anotadores humanos revisan 80 viñetas. El experimento principal ejecuta Iowa Gambling Task durante 100 rondas con cuatro modelos, tres variantes de agente, tres escenarios, seis permutaciones de mazos y 18 seeds emparejadas por configuración.

No hay jugadores humanos. La validación abarca seis modelos; el análisis principal usa GPT-OSS-20B, Llama-3.1-8B, Qwen2.5-7B y Qwen3-4B con agentes Query, ReAct y Reflexion. No apareció un efecto medio estable de emoción sobre el rendimiento a largo plazo. En Qwen2.5 Query, la ira adelantó la fijación unas 11 rondas, d=-1.29 y p=.0003. Los efectos de exploración cambiaron de signo entre modelos. Persistió un sesgo fuerte por etiqueta y posición de los mazos.

El prompt induce una autodescripción contextual, no una emoción vivida. Solo se usa Iowa Gambling Task. Los efectos por celda son numerosos y mixtos. No hay baseline humano recogido bajo el mismo protocolo. La generalización depende de modelos, prompts y endpoints concretos. No demuestra que los LLM sientan emociones. No establece un efecto causal general de ira sobre decisiones. No valida equivalencia con dinámica afectiva humana.

Research question

Can emotion induced through imagined context alter learning, exploration, or early decision lock-in for LLM agents on a sequential task?

Method

Induction is validated with six models and 40 runs per emotion; three human annotators review 80 vignettes. The main experiment runs the Iowa Gambling Task for 100 rounds with four models, three agent variants, three scenarios, six deck permutations, and 18 matched seeds per configuration.

Sample: There are no human players. Validation covers six models; the main analysis uses GPT-OSS-20B, Llama-3.1-8B, Qwen2.5-7B, and Qwen3-4B with Query, ReAct, and Reflexion agents.

Findings

  • No stable average emotion effect appeared on long-run performance.
  • For Qwen2.5 Query, anger advanced lock-in by about 11 rounds, d=-1.29 and p=.0003.
  • Exploration effects changed sign across models.
  • A strong deck-label and position bias persisted.

Limitations

  • The prompt induces a contextual self-description, not a felt emotion.
  • Only the Iowa Gambling Task is used.
  • Cell-level effects are numerous and mixed.
  • There is no human baseline collected under the same protocol.
  • Generalization depends on specific models, prompts, and endpoints.

What the study does not establish

  • It does not demonstrate that LLMs feel emotions.
  • It does not establish a general causal effect of anger on decisions.
  • It does not validate equivalence with human affective dynamics.

Traceability

Scope: Full text

Version: arxiv; 25-page full text reviewed 2026-07-18

Consulted source: https://arxiv.org/abs/2607.12631

Review: Codex full-text and visual 25-page methodological, statistical and claim-boundary review, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • GPT-OSS-20B
  • Llama-3.1-8B
  • Qwen2.5-7B
  • Qwen3-4B

Instruments and metrics

  • Iowa Gambling Task
  • Valence-arousal validation
  • Query, ReAct and Reflexion agents

Data used

  • Matched-seed IGT trajectories

Evidence and location

  • Research question, method, results, and discussion: Full text, pp. 1-25, visually reviewed on 18/07/2026
  • Figures, tables, results, and limitations: Primary PDF sha256 4c95a4f6cdc77f96e9cbf9c17c844daaa98d16dd35a561d908b6f30b21c5f1da; methods, results, limitations, and appendices
  • Editorial decision and claim boundary: Critical record article-407, complete cross-check of 25 pages