Technical Summary: Mind the Gap: The Divergence Between Human and LLM-Generated Tasks

Problem Statement

While Large Language Models (LLMs) have enabled the creation of generative agents capable of simulating autonomous behavior in virtual worlds, it remains uncertain whether these models operate on the same cognitive principles as humans. Current performance benchmarks (e.g., TaskBench, AgentGen) focus on planning and automation capabilities but often overlook a fundamental question: do LLMs truly replicate the deep cognitive mechanisms driving human goal-setting, or do they merely mimic surface patterns?

Human autonomous goal-setting is theorized to be driven by two core pillars:

1. Value-Driven Motivation: Stable personal values (e.g., openness to change vs. conservation) that guide what individuals find desirable.
2. Embodied Experience: Sensorimotor constraints and accumulated interactions with the physical and social world that ground actions in reality.

The authors hypothesize that current LLMs, trained purely on text, lack the internal mechanisms for value prioritization and the sensorimotor grounding necessary to generate human-like, embodied goals. This study aims to quantify the divergence between human and LLM task generation to determine if providing psychological profiles to LLMs can bridge this gap.

Methodology

The study employed a two-experiment design to compare human task generation against that of GPT-4o.

Experiment 1: Task Generation

• Participants: 176 human participants recruited via Prolific.
• Design: A 2×2 between-subjects factorial design manipulating:
  • Environmental Complexity: High (32 living room items) vs. Low (25 items, excluding high-affordance entertainment objects).
  • Social Context: Social (presence of a "person" icon) vs. Non-social (presence of a "mannequin").
• Procedure: Participants were asked to generate at least three tasks using the provided items to "pass the time," detailing the title, objects, setup, goals, and scoring.
• Psychological Measures: Prior to generation, participants completed:
  • PVQ21: To measure personal values (Openness to Change vs. Conservation; Self-Interest vs. Self-Transcendence).
  • TWS Questionnaire: To measure cognitive style (Intuitive vs. Systematic).
  • BFI-2: To measure sociability.
• LLM Conditions:
  • Raw Condition: GPT-4o received only the task prompt.
  • Matched Condition: GPT-4o received the task prompt plus the full psychological profile (values and cognitive style) of a corresponding human participant.
  • Both LLM runs used a temperature of 1.0 to encourage diversity.

Experiment 2: Task Attribute Evaluation

• Participants: 75 independent raters.
• Stimuli: A subset of tasks generated by humans (575) and the Matched GPT-4o (176, one per profile).
• Procedure: Raters evaluated tasks on a 0–10 scale across dimensions of:
  • Creativity: Fun, Novelty.
  • Embodiment: Mental Demand, Physical Demand, Body Part Engagement, Overall Difficulty.
• Analysis: Inter-rater reliability was confirmed (ICC 0.58–0.90). Statistical comparisons used Chi-square tests for topic distribution and Mann-Whitney U tests for rated attributes.

Key Results

1. The Human Signature: Value-Driven and Environmentally Sensitive

Human task generation was systematically predicted by psychological traits:

• Values: Higher "Openness to Change" significantly predicted tasks rated as more novel and fun. "Self-Interest" showed no such relationship.
• Cognitive Style × Environment: An interaction effect was observed between thinking style and environmental complexity. Individuals with a systematic thinking style generated more fun tasks in simple environments, while those with an intuitive style thrived in complex environments, producing more diverse tasks. This suggests humans flexibly adapt cognitive strategies to environmental demands.

2. The Human-LLM Gap

Despite being conditioned on human psychological profiles, GPT-4o failed to replicate human behavioral signatures:

• Thematic Bias: LLMs exhibited a strong bias toward abstract, "Mental & Artistic" tasks (e.g., music, writing), which accounted for 74% of Matched GPT output. In contrast, humans generated a balanced mix including "Relaxation & Household" activities (32% for humans vs. ~5% for LLMs).
• Social Deficit: In social contexts, 58% of human tasks were multiplayer/social, whereas LLMs generated social tasks only 24–29% of the time.
• Embodiment Gap: LLM tasks were rated as significantly more mentally demanding and significantly less physically demanding than human tasks. They required less engagement of the arms and lower body, reflecting a lack of sensorimotor grounding.
• The Creativity Paradox: Despite these gaps, LLM-generated tasks were rated as more novel and more fun than human tasks. The authors attribute this to the LLM's ability to produce combinatorial creativity unbound by physical feasibility or social norms, likely amplified by RLHF alignment which optimizes for immediate engagement rather than mundane realism.

Key Contributions

1. Evidence of Human Signatures: The study provides behavioral evidence that autonomous human task generation is systematically shaped by stable personal values (specifically Openness to Change) and cognitive style, supporting a value-driven account of goal-setting.
2. Identification of LLM Limitations: The study demonstrates that even when provided with explicit psychological profiles, LLMs fail to exhibit core human signatures. They lack internal mechanisms for value prioritization and produce outputs that are thematically biased toward abstraction and disconnected from physical and social reality.
3. The "Ungrounded Creativity" Trade-off: The research highlights a critical disconnect where LLMs can generate text perceived as highly creative and novel precisely because they are ungrounded in the physical and social constraints that characterize human experience.

Significance and Claims

The paper concludes that there is a core gap between the value-driven, embodied nature of human cognition and the statistical patterns of LLMs.

• For AI Research: The findings suggest that creating truly human-aligned autonomous agents requires more than scaling LLMs or fine-tuning on text. It necessitates incorporating intrinsic motivation systems, world models, and sensorimotor grounding to bridge the gap between linguistic proficiency and embodied goal generation.
• For Cognitive Science: The proposed text-based task generation paradigm offers a scalable tool to study goal-setting in naturalistic contexts, enabling new comparisons between human cognition and machine outputs.

The authors maintain a modest stance, noting that their study focuses on a single architecture (GPT-4o) and a text-based paradigm, suggesting that future work should expand to broader model ranges and interactive 3D environments to test the generality of these findings.