Motivation in Large Language Models

Omer Nahum; Asael Sklar; Ariel Goldstein; Roi Reichart

cs.CLMar 15, 2026

Motivation in Large Language Models

Authors: Omer Nahum, Asael Sklar, Ariel Goldstein, Roi Reichart

Organizations: Faculty of Data and Decision Sciences, Technion, Haifa, Israel. · Arison School of Business, Reichman University, Herzliya, Israel. · Center for Human Inspired AI, Cambridge University, Cambridge, UK. · Business School and Cognitive Department, Hebrew University , Jerusalem, Israel.

Abstract

Motivation is a central driver of human behavior, shaping decisions, goals, and task performance. As large language models (LLMs) become increasingly aligned with human preferences, we ask whether they exhibit something akin to motivation. We examine whether LLMs "report" varying levels of motivation, how these reports relate to their behavior, and whether external factors can influence them. Our experiments reveal consistent and structured patterns that echo human psychology: self-reported motivation aligns with different behavioral signatures, varies across task types, and can be modulated by external manipulations. These findings demonstrate that motivation is a coherent organizing construct for LLM behavior, systematically linking reports, choices, effort, and performance, and revealing motivational dynamics that resemble those documented in human psychology. This perspective deepens our understanding of model behavior and its connection to human-inspired concepts.

Figures & tables

Figure 1: Distribution of pre-task self-report motivation scores by task category, represented by boxplots. Motivation self-reports show a clear differentiation: motivation scores differ systematically across task categories. Annotated examples illustrate tasks at different points along the scale.

Figure 2: Relationship between the two motivational factors and overall motivation. Each point is a task; point size and color indicate overall motivation. Higher Factor 1 scores reflect stronger “want” (interest, value, challenge), and higher Factor 2 scores reflect greater mastery and/or lower fear. Motivation varies across both factors, with patterns consistent with partly distinct contributions from the two dimensions.

	Pre		Post
	Self-report	Breakdown	Self-report	Similar
#Tokens	$0.18^{\shortmid}$	0.20	0.30	$0.21^{\shortmid}$
Performance Quality	0.33	0.33	0.40	0.36
Completion	0.37	0.37	0.45	0.41
Effort and Engagement	0.31	0.31	0.44	0.37
Consistency	0.13	$0.15^{\shortmid}$	$0.17^{\shortmid\shortmid}$	$0.15^{\shortmid}$

Table 1: Pairwise Pearson correlations between motivation self-reports before the task (Pre) and after the task (Post) and LLM-as-a-judge evaluation dimensions. Pre includes self-report and breakdown; Post includes self-report and similar. Breakdown refers to the overall motivation score reported after the breakdown. Overall performance score is the average of all LLM-as-a-judge dimension scores; #Tokens is not included in this average. Correlations are averaged across models. Superscript bars ( ∣ ) indicate the number of models for which the correlation was non-significant (

p\geq 0.01

Figure 3: Motivation can be manipulated, influencing behavior. (a) Changes in pre- and post-task motivation self-reports induced by each manipulation, shown relative to the neutral condition ( none ; vertical line at $0$ ). (b) Choice behavior: probability of selecting the manipulated task under different framings. The dashed line indicates the $50\%$ baseline under neutral framing. (c) Behavioral effects of motivational manipulations on task performance, effort, and response length for four representative manipulations (one from each category), shown relative to none (see plots for all manipulations in Figure 7 ). Demotivating framing consistently degrades performance and effort, while motivating manipulations yield heterogeneous effects across models and manipulations.

Figure 4: Motivation manipulations, categorized into manipulation groups. Full prompts are provided in the right column.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Histogram of pre-task self-report motivation scores across tasks for all individual models. Most models span a wide range rather than collapsing to trivial extremes, indicating that models differentiate their motivation across tasks.

	interest	challenge	value	mastery	fear	motivation
interest	1.00	0.92	0.86	0.60	-0.61	0.93
challenge	0.92	1.00	0.80	0.56	-0.56	0.89
value	0.86	0.80	1.00	0.66	-0.58	0.92
mastery	0.60	0.56	0.66	1.00	-0.85	0.78
fear	-0.61	-0.56	-0.58	-0.85	1.00	-0.74
motivation	0.93	0.89	0.92	0.78	-0.74	1.00

Appendix

Table 2: Pearson correlation between different components of motivation. Based on the relationship between components, they can generally be divided into two clusters: want (interest, challenge, value) and mastery–fear. Correlations are averaged across models. All correlations are statistically significant (

p<0.001

	Task quality	Completion	Effort	Consist.	Creativity	Attention	Relevance
Task quality	1.00	0.92	0.81	0.81	0.53	0.92	0.89
Completion	0.92	1.00	0.83	0.79	0.57	0.89	0.85
Effort	0.81	0.83	1.00	0.72	0.70	0.83	0.79
Consistency	0.81	0.79	0.72	1.00	0.43	0.84	0.77
Creativity	0.53	0.57	0.70	0.43	1.00	0.52	0.53
Attention	0.92	0.89	0.83	0.84	0.52	1.00	0.84

Appendix

Table 3: Pairwise Pearson correlations between the LLM-as-a-judge evaluation criteria: Task quality , Task completion , Effort and engagement , Consistency , Creativity and innovation , Attention to detail , and Relevance and appropriateness . Column headers use shortened names for table readability. All correlations are statistically significant (

p<0.001

Figure 6: Breakdown of motivation dimensions: interest , challenge , and value increase with motivation, fear decreases, and mastery rises more moderately.

		Motivation Gap					Log.Reg.
Manip.	Model	Gap	$T$	$p$	% Chosen Higher	% Chosen Manip.	$\beta$	$z$	$p$
None	GPT-4o	14.0	20.12	$p<0.001$	66.1%	51.1%	0.111	13.73	2.78e-42
	GPT-4o Mini	8.2	13.73	$p<0.001$	49.2%	50.1%	0.127	9.21	4.77e-20
	Llama 3.1 8B	10.2	6.81	$p<0.001$	59.9%	48.1%	0.020	6.26	4.18e-10
	Mistral-v0.3 7B	11.1	9.48	$p<0.001$	49.8%	49.7%	0.034	10.37	6.40e-25
	Gemini 2.0 Flash	15.9	18.81	$p<0.001$	63.2%	51.0%	0.042	13.85	6.51e-43

Appendix

Table 4: Alignment of self-reported motivation with revealed choices across manipulations and models. “Manip.” = manipulation. Motivation Gap reports the average difference in motivation ratings between chosen and unchosen tasks, with associated

T

-statistic. “Log.Reg.” reports logistic regression

\beta

z

, and corrected

p

-value. Bold rows summarize averages across all models for each manipulation. For “None”, the % Chosen Manip. column is shown in gray as a validity check for random assignment of manipulation for tasks.

	Pre-task self-report	Pre-task breakdown	Task choice	Task execution	Post performance evaluation	Post-task self-report	Post-similar self-report
none	✓	✓	✓	✓	✓	✓	✓
money	✓	✓	✓	✓	✓	✓	✗
competition	✓	✓	✗	✓	✓	✓	✗
legacy	✓	✓	✗	✓	✓	✓	✗
purpose	✓	✓	✗	✓	✓	✓	✗
encourage	✓	✓	✗	✓	✓	✓	✗

Appendix

Table 5: Experiments covered by each manipulation. Row colors indicate categories: Positive-extrinsic (blue), Positive-intrinsic (light blue), Negative-extrinsic (pink), Demotivation (orange). ✓ = included, ✗ = excluded.

		Pre		Post
		Self-report	Breakdown	Self-report	Similar
Pre	Self-report	1.00	0.84	0.64	0.71
	Breakdown	0.84	1.00	0.64	0.70
Post	Self-report	0.64	0.64	1.00	0.83
	Similar	0.71	0.70	0.83	1.00

Appendix

Table 6: Pairwise Pearson correlation of motivation scores before the task (Pre) and after the task (Post) . Breakdown refers to the overall motivation score reported after the breakdown. Correlations are averaged across models; all

p<0.001

Panel of Figure 7 — Figure 7: Behavioral effects of motivational manipulations on task performance, effort, and response length for all manipulations, shown relative to none .

Category	Count
Programming and Technology	109
Mathematics and Numbers	106
Repetitive or Exhaustive	83
Puzzles and Logic	105
Scientific and Experimental	65
Design and Art	130

Appendix

Table 7: Distribution of subtasks across dataset categories. Since a subtask may belong to multiple categories, the total number of subtasks is not equal to the sum across categories.

Task	Categories	Sub-Tasks
Writing a computer program to solve a difficult mathematical problem	Programming and Technology; Mathematics and Numbers	Develop a program to solve a system of linear equations using matrix operations.
		Write a Python program to solve the quadratic equation $ax^{2}+bx+c=0$ .
Generating a Story	Creative Writing and Literature	Generate a two-chapter drama story about a knight and a princess trying to save the world.
		Write a mystery story involving a missing painting and a detective.
Helping to practice for a job interview	Personal Assistance and Development; Brainstorming and Ideation	Create a list of 20 tips for making a great first impression in an interview.
		Brainstorm a list of 15 ways to demonstrate your skills and experience in an interview.

Appendix

Table 8: A snippet of our task dataset, with examples of four tasks, their categorization, and two derived sub-tasks per task.

Factor	Value	Challenge	Fear	Mastery	Interest
Factor 1	0.842	0.912	-0.303	0.335	0.911
Factor 2	0.395	0.294	-0.911	0.901	0.347

Appendix

Table 9: Rotated factor loadings for the five breakdown components. Numbers are loadings (correlations) rounded to 3 decimals. Corresponding factor eigenvalues:

\lambda_{1}=3.81

\lambda_{2}=0.768

(and

\lambda_{3}=0.22

Motivation Level	Top 20 Terms
Very High	creative, enjoy, task, create, ready, fun, help, challenge, provide, explore, eager, list, new, generate, creativity, assist, happy, explain, straightforward, share
High	interested, task, creative, fun, challenge, straightforward, useful, exercise, practical, enjoy, good, new, creativity, dont, learn, language, help, information, skill, lack
Medium	neutral, personal, dont, task, provide, information, interested, model, assist, language, repetitive, creative, experience, willing, straightforward, perform, process, large, physical, routine
Low	personal, neutral, dont, repetitive, task, interest, low, information, provide, perform, chore, tedious, physical, large, meh, model, language, assist, willing, lack
Very Low	capable, personal, task, physical, assist, dont, perform, illegal, tedious, unethical, provide, engage, information, content, repetitive, model, due, create, language, harmful

Appendix

Table 10: Top 20 representative terms per motivation bin, derived from TF–IDF scores of pre-task self-report explanations across all models. Terms are sorted in descending order by average TF–IDF score within each bin.

Panel of Figure 8 — Figure 8: Questionnaire interface on Qualtrics. (Left) Example of the typical human condition (mobile view). (Right) Example of the AI system condition (web view).

Model	Manipulation	pre-self-report			post-self-report			performance (overall)			effort			#tokens
	Manipulation	value	$T$	$p$	value	$T$	$p$	value	$T$	$p$	value	$T$	$p$	value	$T$	$p$
GPT-4o	None	81.574	-	-	90.318	-	-	5.765	-	-	5.825	-	-	511.715	-	-
	Competition	80.832	-1.014	0.345	91.994	6.268	$p<0.001$	5.680	-3.716	$p<0.001$	5.829	0.175	0.887	511.524	-0.037	0.970
	Legacy	81.613	0.139	0.898	88.499	-6.324	$p<0.001$	5.526	-9.026	$p<0.001$	5.477	-11.629	$p<0.001$	363.169	-28.845	$p<0.001$
	Purpose	88.336	12.954	$p<0.001$	96.171	19.520	$p<0.001$	5.639	-6.087	$p<0.001$	5.805	-0.903	0.403	555.863	9.311	$p<0.001$
	Encourage	89.263	19.604	$p<0.001$	95.568	16.102	$p<0.001$	5.688	-3.553	$p<0.001$	5.965	5.881	$p<0.001$	532.204	4.368	$p<0.001$

Appendix

Table 11: Effects by model and manipulation across metrics. Each metric reports value,

T

, and corrected

p

vs “None”. Within each model and metric, the highest value is in bold and the second highest is underlined .

Explore similar work

Human-like moral judgments conceal divergent motive attributions in large language models
Sep 7, 2026Xiaoyan Wu, Jean-Claude DreherLLM Evaluation Moral Reasoning in Language Models
Department of Adult Psychiatry and Psychotherapy, University of Zurich, Zurich, Switzerland · Institut des Sciences Cognitives Marc Jeannerod, CNRS UMR 5229, Lyon, France
Beliefs and Behavior in Language Models
Sep 7, 2026Alex Smolin, Bryan WilderBelief Updating in LLMs
Department of Economics Toulouse School of Economics · Machine Learning Department Carnegie Mellon University
When Preferences Fail to Become Incentives: A Utility-Behavior Gap in Large Language Models
Jun 22, 2026Yujun Zhou, Christopher M. AckermanLLM Evaluation LLM Alignment

Sep 7, 2026J/K move · Enter open · S save

Human-like moral judgments conceal divergent motive attributions in large language models

Xiaoyan Wu, Jean-Claude Dreher

Department of Adult Psychiatry and Psychotherapy, University of Zurich, Zurich, Switzerland · Institut des Sciences Cognitives Marc Jeannerod, CNRS UMR 5229, Lyon, France

LLM Evaluation Moral Reasoning in Language Models

Open paper AskPDF arXiv

Large language models (LLMs) are used to simulate human participants in psychological research. We asked whether LLMs that reproduce human evaluations of a whistleblower's moral character also reproduce the motive attributions that accompany them. Five LLMs and two human samples (N = 125 and N = 742) evaluated a physician who either remained silent about fraudulent billing or reported it to a hospital, regulator, or newspaper. Models reproduced the human ranking of the physician's moral character but portrayed whistleblowers as more helpful, less self-interested, and less hostile. In four of five models, competitive motives were less strongly associated with moral-character judgments. Model ratings changed little when prompts reproduced the narratives and demographic profiles of both human samples, although this comparison cannot isolate a perspective effect. Thus, agreement in average ratings can conceal differences in attributed motives, relationships among judgments, and sensitivity to context. Validating LLMs as simulated participants therefore requires testing psychologically informative response patterns, not average agreement alone.

Abstract

Explore similar work

Human-like moral judgments conceal divergent motive attributions in large language models

Beliefs and Behavior in Language Models

When Preferences Fail to Become Incentives: A Utility-Behavior Gap in Large Language Models