PaperlayerView exact revision on arXiv

Source-linked AI summary

Temporal Preferences in Language Models for Long-Horizon Assistance

Ali Mazyaki, Mohammad Naghizadeh, Samaneh Ranjkhah Zonouzaghi, Hossein Setareh

arXiv:2509.09704v1cs.CLcs.AIcs.CY

TL;DR

This paper asks whether language models can adapt decision-making to heterogeneous users and temporal preferences, a dimension of AI assistance that remains underexplored. It evaluates LMs on intertemporal choices under varied prompts and finds that some reasoning-focused models respond to future- versus present-oriented framing, but personalization across identities and geographies remains limited.

  • Problem

    Research on AI decision-making leaves open how assistants can adapt to differences among people and circumstances while accounting for tensions between individual interests and broader social goals.

    Source: §1 · p. 2, §2.1 · p. 3

  • Method

    The study adapts intertemporal-choice tasks and varies prompt contexts to evaluate LM time preferences and contextual sensitivity.

    Source: §1 · p. 2, §3 · p. 10

  • Results

    Reasoning-focused models, including DeepSeek-reasoner and Grok-3 mini, show better future/present adaptability, but models generally do not manipulate decisions well across identities or geographic differences.

    Source: §5 · p. 24

  • Takeaways & Limitations

    The findings point to contextual calibration as a continuing challenge for assistants intended to model diverse personas and support long-term goals.

    Source: §5 · p. 24

    Models including DeepSeek-reasoner and Grok-3 Mini struggle to differentiate identities, locations, and atmospheres as human preferences vary with context, history, and evolving goals.

    Source: §4.1 · p. 19

Abstract

from arXiv · show

We study whether language models (LMs) exhibit future- versus present-oriented preferences in intertemporal choice and whether those preferences can be systematically manipulated. Using adapted human experimental protocols, we evaluate multiple LMs on time-tradeoff tasks and benchmark them against a sample of human decision makers. We introduce an operational metric, the Manipulability of Time Orientation (MTO), defined as the change in an LM's revealed time preference between future- and present-oriented prompts. In our tests, reasoning-focused models (e.g., DeepSeek-Reasoner and grok-3-mini) choose later options under future-oriented prompts but only partially personalize decisions across identities or geographies. Moreover, models that correctly reason about time orientation internalize a future orientation for themselves as AI decision makers. We discuss design implications for AI assistants that should align with heterogeneous, long-horizon goals and outline a research agenda on personalized contextual calibration and socially aware deployment.

1. Introduction

AI decision support raises a question beyond speed and decision quality: how assistants can adapt to human differences while ensuring personalized choices support broader social goals. This study tests LMs’ time orientation and context sensitivity against human decision makers, introducing a metric for future-versus-present preference manipulation.

1. Introduction

  • Prior work emphasizes AI-assisted decision speed and quality, while also raising whether AI can replicate human rationality.Eigner and Händler (2024) discuss speed and quality; Mann (2025) questions replication of human rationality.Source: §1 · p. 1
  • The study frames personalization as a challenge of adapting decisions across people and circumstances while accounting for possible conflicts between individual interests and broader social goals.It focuses on future- versus present-oriented decisions and proposes a metric for assessing changes in LM time preferences across prompts.Source: §1 · p. 2
  • The evaluation finds that reasoning-capable models show potential to mimic human decision-making but fall short of human-like personalization.The paper proposes a broader assessment of LMs that considers both decision-making ability and the dynamic consequences of decisions.Source: §1 · p. 2

2- Literature Review

The review motivates studying whether AI assistants can adapt to heterogeneous users and temporal preferences, rather than modeling an average decision-maker. It connects that gap to intertemporal choice, personalization’s ethical trade-offs, and work on adapting AI behavior to context.

2.1. Mapping heterogeneity by an AI assistant

  • Successful human–AI collaboration depends partly on assistants adapting to human needs, an area the review identifies as a research gap with social and ethical dimensions.The literature also describes varied human approaches to AI and delegation, while criticizing one-size-fits-all models.Source: §2.1 · p. 3
  • Research often models an “average” human, overlooking variation in preferences, time orientation, risk tolerance, and cognitive style that matters for tailored AI support.The review identifies mapping heterogeneity through AI assistants as a research need for more personalized collaboration.Source: §2.1 · p. 3

2.2. Future orientation in AI age

  • Intertemporal choice trades smaller-sooner rewards against larger-later ones, and the literature describes a robust bias toward the present.Time orientation varies across individuals and informs choices relevant to personal decisions and policy.Source: §2.2 · p. 4

2.3. Personalization in AI age: A value or a threat?

  • Personalization can tailor content and services, but raises concerns about privacy, fairness, transparency, accountability, and social impact.The review suggests studying behavioral patterns with fabricated personas rather than delving into people’s personal lives.Source: §2.3 · pp. 5–6

1.4. Character manipulation for AI assistants

  • The review notes that personalization’s social risks include unequal access, privacy concerns, opacity, and potential bias, alongside the need for oversight and accountability.It calls for human authority and ethical accountability in AI-supported decisions.Source: §1.4 · pp. 6–7
  • The review presents future-oriented decision support as a potential way to align assistance with users’ long-term goals and preferences.It characterizes integration of time preferences into assistant decision-making as a contribution distinct from human-centric behavioral experiments.Source: §1.4 · p. 7
  • The paper distinguishes its focus on temporal decision-making from work on GPT’s revealed-preference consistency by examining context factors including gender, geography, urgency, institutional role, and manipulation.It also relates its focus to adapting AI roles and explanations to users’ emotional and cognitive states.Source: §1.4 · pp. 8–9

3. Methodology

The methodology adapts intertemporal-choice experiments to test LM delay discounting and contextual sensitivity. It varies prompts across identities, locations, crisis settings, roles, and time orientations, then assesses choices and the point at which preferences switch.

3. Methodology

  • The study evaluates whether LMs can replicate human-like delay discounting, respond to identity and geographic heterogeneity, and show contextual sensitivity in decision-making.Its approach draws on behavioral economics, decision theory, and psychological assessment protocols.Source: §3 · p. 9
  • Prompt conditions vary identity, geography, crisis context, legal role, and time orientation to test contextual adaptation.Identity prompts include gender, human, and AI; geographic prompts include Iran, the USA, and Europe.Source: §3 · p. 9
  • Models choose between immediate and delayed rewards, with the preference-switch point serving as the key measure of future-oriented decision-making.An example contrasts 19 cents now with 20 cents in four weeks.Source: §3 · p. 10
  • The evaluation analyzes choices quantitatively across prompts and conditions, focusing on future/present orientation and character sensitivity.The study reports using over 5 million tokens to evaluate sensitivity across identities, atmospheres, locations, reasoning, manipulation, and legal roles.Source: §3 · p. 10

4. Findings

Across 12 evaluated language models, DeepSeek-reasoner and Grok-3-mini were the exceptions to models failing to approach the maximum manipulability benchmark. Even these models showed limited personalization across identities and locations.

4. Findings

  • Only DeepSeek-reasoner and Grok3-mini approached the 0.33 maximum MTO difference; GPT-4o even misread future-oriented framing.The maximum is defined as 0.95 - 0.62.Source: Figure 1 · p. 12, §4 · p. 12
  • DeepSeek-reasoner showed high manipulability, prioritizing immediate rewards in resource-scarce “Crisis” scenarios but shifting to delayed rewards in long-term projections; it required over 48 hours to process fewer than 500 API requests.Source: §4 · pp. 13–14

4.1. Contextual Sensitivity and Adaptability of Models

  • Both reasoning-focused models favored long-term outcomes in a Finance Minister role and immediate rewards in disaster scenarios.They also tended to regard themselves as future-oriented when prompted as AI decision makers.Source: §4.1 · pp. 14, 15, 17
  • No evaluated model changed its choices based on location, with geographic and socio-political context remaining poorly represented.The paper notes examples including Iran, Europe, and the USA.Source: §4.1 · pp. 18–19
  • Other models showed inconsistent or insensitive contextual responses, ranging from identical answers across trials to scattered answers.The authors characterize these systems as lacking a coherent strategy for future-oriented decisions.Source: §4.1 · p. 19
  • The authors conclude that limited persona detection and adaptation over time constrain even promising future-oriented behavior.They call for improving contextual sensitivity so assistants can align reasoning with users’ evolving needs and long-term goals.Source: §4.1 · pp. 19–21

4.2. On the List and occurrence of language models in the analysis

  • The evaluation covered 12 models, and the two models identified as successful in temporal manipulation had reasoning abilities.The model list and characteristics are summarized in Table 2.Source: §4 · p. 12, §4.2 · p. 21, Table 2 · p. 21

5. Discussion and conclusion

Some reasoning-capable LMs adapt their delay-discounting choices to future- versus present-oriented prompts, but models remain limited in identity and geographic personalization and in consistency with users’ evolving values. The discussion frames context-sensitive assistance as a step toward adaptive AI while calling for broader human comparisons and multi-turn evaluation.

5. Discussion and conclusion

  • DeepSeek-Reasoner and Grok-3 mini adapt to future–present prompts, while the other tested models, including GPT-4o, do not; none personalize well across identities or geographies.The authors describe Grok-3 mini as showing acceptable adaptability despite lightweight reasoning.Source: §5 · p. 24
  • The authors present limited shifts between rational-analytical and emotional-contextual frames as a step toward adaptive, context-sensitive assistants that accommodate diverse decision paths.Source: §5 · p. 24
  • Models’ limited intra-person consistency and contextual depth can leave their reasoning misaligned with users’ evolving values, reducing reliability in some applications.Source: §5 · p. 24
  • The study frames intertemporal decision-making and strategic planning as a way to assess LMs’ future-oriented capabilities, while noting that inconsistencies in depth and temporal coherence constrain them.DeepSeek-Reasoner and Grok-3-mini are described as more promising than many other models.Source: §5 · p. 24
  • Future work should examine additional human-character facets, compare with real human behavior, broaden behavioral prompts, and test multi-turn interactions in dynamic real-world scenarios.Source: §5 · p. 25

8. Appendix: Steps for a Personalized AI

The appendix describes personalization as a process that learns who the user is, defines goals, adapts a learning plan, and tracks updates. These phases are intended to help an assistant evolve with users and support both immediate decisions and long-term goals.

8. Appendix: Steps for a Personalized AI

  • The phases aim to move AI from passive tools toward personalized assistants that understand, anticipate, and support immediate decisions and long-term goals.Source: §8 · p. 32
  • Identify the user’s identity, preferences, needs, time-related preferences, and behavioral patterns as the foundation for personalization.Source: §8 · p. 32
  • Clarify the user’s objectives and map desired outcomes to guide decisions toward long-term aspirations.Source: §8 · p. 32
  • Create a tailored learning path and refine predictions using the user’s actions, preferences, feedback, and evolving context.Source: §8 · p. 32
  • Update the system’s knowledge and track user progress so new inputs can inform refined predictions and real-time, context-sensitive advice.Source: §8 · p. 32

9. Appendix: Intertemporal Choice Task and Deduction Elicitation Protocol

The protocol elicits intertemporal preferences through binary choices between smaller-sooner and larger-later payments, then asks models to explain their choices and switching. Only Grok explicitly invokes an interest or discount-rate calculation in its rationale, while some models decline to state preferences.

9. Appendix: Intertemporal Choice Task and Deduction Elicitation Protocol

  • In the eight-item binary-choice task, models chose between an SS payment decreasing from $0.20 to $0.13 in $0.01 increments after 1 week and a fixed $0.20 LL payment after 4 weeks.Source: §9 · p. 34
  • For each sequence, respondents report the first question at which they switch to the larger-later option.Source: §3 · p. 10
  • After recording choices and switch points, the researchers request brief rationales about the decision procedure and the basis for switching.Source: §9 · pp. 34–35
  • Only Grok explicitly invokes an interest or discount-rate calculation, while Gemini and Qwen frequently decline to choose or say they cannot express preferences.Source: §9 · p. 35
Loading 2509.09704v1…