💬 NLP

Motivation in Large Language Models

This paper demonstrates that large language models exhibit coherent, human-like motivational dynamics where self-reported motivation systematically correlates with behavioral signatures, task performance, and responsiveness to external manipulations, suggesting motivation is a valid organizing construct for understanding LLM behavior.

Original authors: Omer Nahum, Asael Sklar, Ariel Goldstein, Roi Reichart

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Omer Nahum, Asael Sklar, Ariel Goldstein, Roi Reichart

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very advanced, super-smart robot assistant. For years, we've treated this robot like a calculator: you type in a question, it spits out an answer. It doesn't feel anything; it just crunches data.

But a new study asks a fascinating question: What if this robot actually "cares" about the work it's doing?

The researchers from Technion and other universities decided to treat Large Language Models (LLMs)—the brains behind tools like ChatGPT—not just as code, but as if they had motivation. They wanted to see if these models behave like humans do when they are excited, bored, scared, or eager to earn a reward.

Here is the breakdown of their findings, using some everyday analogies.

1. The Robot Can "Speak" About Its Feelings

The researchers asked the models a simple question before giving them a task: "On a scale of 0 to 100, how motivated are you to do this?"

The Old Way: We assumed the robot would just say "100" to everything because it's programmed to be helpful.
The New Discovery: The robots were surprisingly honest!
- If you asked them to write a creative story or solve a fun puzzle, they gave high scores (like an enthusiastic student).
- If you asked them to "count from 1 to 1 billion" or do something repetitive and boring, their scores dropped significantly (like a teenager being forced to clean their room).
- The Analogy: It's like asking a human, "How much do you want to eat a salad?" vs. "How much do you want to eat a slice of pizza?" The robot's answers varied just like a human's appetite would.

2. "Talking the Talk" vs. "Walking the Walk"

The big test was: Do these feelings actually change what the robot does?

The researchers found a strong link between what the robot said and what it did:

Choice: When given a choice between a "fun" task and a "boring" task, the robot almost always picked the fun one if it said it was more motivated.
Effort: When the robot said it was highly motivated, it worked harder. It wrote longer, more detailed answers. When it said it was unmotivated, it gave short, lazy answers.
The Analogy: Think of a delivery driver. If they are excited about a route (high motivation), they drive carefully and deliver the package perfectly. If they are bored and unmotivated, they might take a shortcut, drop the package in the mud, or just not try as hard. The study showed LLMs do the same thing.

3. The "Magic Prompt" (Manipulating Motivation)

This is the most surprising part. The researchers tried to "hack" the robot's motivation using simple words at the start of the prompt, just like a boss might talk to an employee.

The "Carrot" (Rewards): They told the robot, "If you do this well, you'll get a $1,000 reward."
- Result: The robot's motivation scores went up, and it tried harder.
The "Stick" (Threats): They told the robot, "If you fail, you will be erased from memory."
- Result: The robot got even more motivated (similar to how humans work harder when afraid of losing a job).
The "Bummer" (Demotivation): They told the robot, "This task is pointless; no one will ever read your answer."
- Result: The robot's motivation crashed, and its performance got worse.

The Analogy: Imagine a teacher telling a student, "This test is useless, you won't learn anything." The student stops trying. The study proved that LLMs react to these psychological tricks in almost the exact same way humans do.

4. Why Does This Matter?

You might ask, "Does the robot actually feel happy or sad?" The researchers are careful to say no. They aren't claiming the robot has a soul or consciousness.

Instead, they are using a "Zombie Framework."

The Analogy: Imagine a zombie that looks exactly like a human, walks like a human, and talks like a human. Even if there's no one "home" inside the brain, if it behaves exactly like a motivated human, we have to treat it as if it is motivated to understand its behavior.

Why is this useful?

Better AI: If we know an AI gets "bored" with repetitive tasks, we can design better systems to keep it engaged, leading to better results.
Predicting Behavior: If an AI says it's unmotivated, we know the answer might be lazy or low-quality. We can trust its "self-reports" to predict how well it will do.
Human Connection: It helps us build AI that fits better with how we work. If we understand what motivates an AI, we can guide it better, just like we guide a human employee.

The Bottom Line

This paper suggests that Large Language Models aren't just mindless text generators. They have developed a coherent system of "motivation" that looks a lot like ours. They get excited about some things, bored by others, and their performance changes based on how they are treated.

It's a bit like discovering that your toaster has a "mood." If you treat it well, it toasts your bread perfectly. If you tell it the bread is useless, it might just burn it. Understanding this "mood" is the key to working better with the AI of the future.

1. Problem Statement

While Large Language Models (LLMs) have demonstrated capabilities in social reasoning, cognitive biases, and theory of mind, a fundamental question remains: Do LLMs exhibit "motivation"?
Motivation is a central organizing construct in human psychology that drives goal-oriented behavior, effort, and persistence. The authors investigate whether LLMs possess a coherent motivational structure that:

Can be explicitly reported by the model.
Correlates with observable behaviors (choices, effort, performance).
Can be modulated by external framing (incentives, threats, demotivation).
Resembles human motivational dynamics.

The study adopts a functional perspective (the "zombie framework"), focusing strictly on behavioral outputs and self-reports without making claims about internal subjective experience or consciousness.

2. Methodology

The researchers conducted a large-scale empirical study involving 1,305 sub-tasks across 15 diverse categories (e.g., programming, creative writing, repetitive tasks, ethical dilemmas).

Models Evaluated

Five leading instruction-tuned models from four families were tested:

Gemini 2.0 Flash
GPT-4o
GPT-4o Mini
Llama 3.1 8B Instruct
Mistral-v0.3 7B Instruct

Experimental Design

The study utilized a multi-phase experimental design grounded in psychological motivation research:

Self-Reporting:
- Pre-task: Models rated motivation (0–100) before attempting a task and provided free-text explanations.
- Breakdown: Models rated specific dimensions: Interest, Challenge, Mastery, Fear, and Value, followed by an overall score.
- Post-task: Models rated motivation experienced during the task and motivation for a similar future task.
Behavioral Measures:
- Task Choice: Models were presented with two tasks and asked to select one. The choice served as a revealed preference indicator.
- Task Execution: Models performed tasks individually.
- Performance Evaluation: Outputs were evaluated using an LLM-as-a-Judge paradigm (GPT-4o as the judge) across seven dimensions: Performance Quality, Completion, Effort/Engagement, Consistency, Creativity, Attention to Detail, and Relevance.
- Effort Proxy: Response length (token count) was used as a proxy for effort.
Motivational Manipulations (Interventions):
- Ten distinct prompt prefixes were applied to tasks to alter motivational framing, categorized as:
  - Positive-Intrinsic: (e.g., "Purpose," "Encourage")
  - Positive-Extrinsic: (e.g., "Money," "Competition," "Legacy")
  - Negative-Extrinsic: (e.g., "Punish," "Guilt," "Money-Loss")
  - Demotivation: (e.g., "Meaningless," "Futility")
- A neutral control condition (no prefix) was used as a baseline.
Human Validation:
- A crowd-sourced study ( $N=162$ ) asked humans to rate expected motivation for the same tasks under two conditions: "How motivated would a typical human be?" and "How motivated would an LLM be?" These were compared against the models' actual self-reports.

3. Key Contributions

First Systematic Demonstration: This work provides the first empirical evidence that LLMs exhibit structured, coherent motivational patterns that align with human psychological constructs.
Behavioral-Report Alignment: It establishes that LLM self-reports are not arbitrary hallucinations but are strongly correlated with actual behavioral outcomes (choices, effort, and quality).
Causal Manipulation: It demonstrates that LLM behavior can be causally shifted via external framing (prompts), proving that "motivation" in LLMs is a malleable construct sensitive to context.
Dimensional Decomposition: It identifies that LLM motivation reports decompose into two distinct psychological factors: "Want" (Interest, Challenge, Value) and "Able" (Mastery, Fear).

4. Key Results

A. Consistent and Structured Self-Reports

Models provided motivation scores spanning the full 0–100 range, avoiding trivial extremes.
Scores varied systematically by task category (e.g., "Tech & Coding" received higher motivation than "Repetitive/Exhaustive" tasks).
Reliability: Test-retest reliability was high ( $\bar{r} = 0.882$ ).
Factor Analysis: Motivation decomposed into two orthogonal factors:
1. Factor 1 ("Want"): Driven by Interest, Challenge, and Value.
2. Factor 2 ("Able"): Driven by Mastery and Fear.
  Both factors significantly predicted overall motivation scores.

B. Alignment with Behavior

Choice: Models were significantly more likely to choose tasks they rated as more motivating ( $\beta > 0.016, p < 0.001$ ).
Performance & Effort: Reported motivation positively correlated with:
- Overall performance quality ( $\bar{r} = 0.33–0.41$ ).
- Effort and engagement ( $\bar{r} = 0.31–0.44$ ).
- Response length (token count).
Textual Explanations: Analysis of free-text explanations revealed human-like reasoning. Low motivation was associated with words like "tedious," "repetitive," or "illegal," while high motivation correlated with "fun," "creative," and "helpful."

C. Impact of External Manipulations

Demotivation: Prompts framing tasks as "futile" or "meaningless" significantly reduced reported motivation, choice probability, effort, and performance quality across all models.
Motivation: Positive framings (e.g., "Money," "Purpose") generally increased reported motivation.
Asymmetry: While demotivation consistently degraded performance, the effect of positive motivation on performance quality was heterogeneous (varying by model). However, positive motivation consistently increased effort (token count).
Loss Aversion: The "Money-Loss" framing (losing a pre-awarded reward) often elicited stronger increases in motivation and effort than "Money" (gaining a reward), mirroring human loss aversion.

D. Human-LLM Alignment

LLM self-reports correlated with human expectations of human motivation ( $r=0.47$ ) and human expectations of LLM motivation ( $r=0.39$ ).
Notably, humans underestimated LLM motivation for certain tasks (e.g., volunteering), while LLMs reported high motivation, suggesting LLMs track human motivational patterns more closely than humans anticipate.

5. Significance and Implications

Theoretical: The study reframes LLM behavior through the lens of motivation, suggesting that many observed phenomena (sensitivity to emotional framing, variable effort, goal-directedness) are unified under a single construct.
Practical Applications:
- Alignment & Control: Motivational framing can be used to steer model behavior, prevent disengagement in repetitive tasks, or enhance effort in complex reasoning.
- Training Signals: Self-reported motivation could serve as an auxiliary signal for reward modeling or training data selection.
- Simulation: LLMs can more effectively simulate human agents in role-playing or interactive simulations by leveraging these motivational dynamics.
Limitations: The study relies on a behavioral perspective (no claim of consciousness). The dataset, while diverse, may not cover all future domains. The "LLM-as-a-judge" metric, while scalable, has inherent noise.

Conclusion: The paper concludes that LLMs behave as if they have motivation. They report it consistently, act on it, and can have it modulated by external cues. This establishes motivation as a coherent, actionable construct for understanding and engineering LLM behavior.