Universe of Thoughts: A Computational Framework for Creative Reasoning in Large Language Models

Yuto Suzuki; Farnoush Banaei-Kashani

cs.AINov 25, 2025

Universe of Thoughts: A Computational Framework for Creative Reasoning in Large Language Models

Authors: Yuto Suzuki, Farnoush Banaei-Kashani

Organizations: Department of Computer Science and Engineering University of Colorado Denver Denver, CO, USA

Abstract

Recent advances in Large Language Model (LLM) reasoning have improved conventional problem solving, but creative reasoning remains comparatively underexplored. Inspired by cognitive science, we formalize combinational, exploratory, and transformational creativity as executable computational operators over structured problem and solution spaces, specifying how each mode combines, explores, or transforms those spaces. Combinational reasoning transfers ideas across domains to form unfamiliar combinations; exploratory reasoning searches for new solutions within an existing conceptual space; and transformational reasoning modifies the rules or constraints that define that space. This formalization yields distinct algorithmic procedures, which we instantiate in Universe of Thoughts (UoT), an LLM reasoning framework. Existing creativity benchmarks emphasize either open-ended ideation or highly constrained problem solving. We therefore introduce three novel creative-reasoning tasks requiring concrete solutions in low-constraint settings. Across 10 generations per method and task, T-UoT with GPT-4o performs strongest on the low-constraint, high-objective-specificity Bridge and Electricity tasks, while C-UoT shows its strongest relative performance on the low-constraint, lower-objective-specificity Society task. In addition, we evaluate UoT on HypoArena, an independent scientific hypothesis-generation benchmark with 100 tasks across biomedical, machine-learning, and social-science domains. With Qwen3-14B, Exploratory UoT ranks first among seven reasoning methods, achieving a 32.7% pairwise win rate compared with 25.5% for the next-best method. Our results suggest distinct performance patterns across task structures: T-UoT is strongest in low-constraint, high-specificity settings, E-UoT in more constrained, high-specificity settings, and C-UoT in low-constraint, lower-specificity settings.

Figures & tables

Figure 1: Three computational regimes of creative search. Combinational reasoning imports transferable structure from a different context; exploratory reasoning expands from existing candidates while preserving the current rules; and transformational reasoning modifies the rules or assumptions that define the admissible solution space.

Figure 2: Task regimes and method-performance patterns along degree of constraint and objective specificity. (a) Bridge and Electricity are low-constraint, high-specificity design tasks, while Society is low-constraint and lower-specificity. HypoArena Bio and ML are high-constraint, high-specificity hypothesis-generation tasks, while HypoArena Social is high-constraint and lower-specificity. (b) The observed performance patterns are consistent with the search mechanisms of T-UoT, C-UoT, and E-UoT, while the high-constraint, low-specificity regime shows stronger performance from generic structured reasoning methods such as ToT.

	Bridge		Electricity		Society
Method	Human C	LLM C	Human C	LLM C	Human C	LLM C
OnePrompt	0.480	0.566	0.540	0.508	0.550	0.522
CoT	0.590	0.651	0.560	0.450	0.500	0.586
ToT	0.580	0.579	0.530	0.488	0.530	0.574
EGoT	0.650	0.641	0.500	0.409	0.560	0.613
C-UoT	0.580	0.664	0.440	0.512	0.560	0.649

Table 1: Novel creative-reasoning task creativity results. Human and LLM-based creativity scores across the three novel creative-reasoning tasks. Creativity is the mean of Novelty and Utility. Full Novelty, Utility, and Feasibility results are reported in Appendix D.2 . Bold indicates the best result among matched GPT-4o reasoning methods; underlined values indicate the best contextual reference model.

Rank	Method	Bio	ML	Social	Avg	WR (%)
1	E-UoT	1541.8	1547.1	1516.3	1535.4	32.7
2	ToT	1513.2	1510.4	1524.5	1515.8	25.5
3	EGoT	1482.7	1510.6	1521.1	1505.1	25.3
4	CoT	1486.2	1493.3	1501.9	1493.7	18.8
5	C-UoT	1495.1	1477.6	1497.0	1489.6	19.5
6	T-UoT	1490.1	1480.0	1484.6	1484.6	17.3

Table 2: Qwen3-14B on HypoArena (100 tasks; Gemma-3-12B judge; canonicalized core hypotheses). Bio 34 / ML 34 / Social 32. Bio/ML/Social/Avg: pooled BTD ratings centered at 1500. WR: pairwise win rate (%). Bold: column best.

Mode	Bio	ML	Social	All
C-UoT	58.8%	57.0%	59.0%	58.3%
E-UoT	53.3%	49.2%	55.0%	52.4%
T-UoT	54.1%	53.4%	55.1%	54.2%

Table 3: Operator ablation on 100 HypoArena tasks. Values are mean pairwise win rates of the full variant against its ablated counterpart (chance

=50\%

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Boden concept	Computational interpretation	UoT operator
Unfamiliar combinations of familiar ideas	Import structure from another context	analogue $\rightarrow$ mechanism $\rightarrow$ transfer
Explore previously unvisited locations in an existing space	Expand the candidate set with rules fixed	Shift / Deepen
Alter defining dimensions of the space	Modify mutable assumptions or rules	gated rule transformation

Appendix

Table 4: Correspondence between Boden’s creativity modes and UoT operators.

Task	Description
Task 1: One-Lane Bridge	Objective: Design a traffic-control policy that minimizes average vehicle delay. Context: A single-lane bridge with known traffic patterns, fixed crossing times, and direction-switching penalties. Constraints: No new infrastructure; traffic must remain unidirectional at all times.
Task 2: Electricity Tariff	Objective: Design a tariff and demand-response program that reduces daily peak load. Context: A residential feeder experiences predictable periods of high demand. Constraints: No new physical capacity; critical medical loads must never be curtailed.
Task 3: Social Cohesion	Objective: Design an intervention that strengthens cross-group cohesion, measured by new ties, mixing, and mutual aid. Context: A community contains distinct social groups and a fixed intervention period. Constraints: The intervention must protect privacy, remain inclusive, and maintain a safe environment.

Appendix

Table 5: Novel creative-reasoning tasks

Method	Bridge Task				Electricity Task				Society Task
	C	N	U	F	C	N	U	F	C	N	U	F
OnePrompt	0.480	0.450	0.500	0.820	0.540	0.480	0.610	0.870	0.550	0.490	0.610	0.810
CoT	0.590	0.610	0.570	0.750	0.560	0.520	0.600	0.920	0.500	0.430	0.580	0.810
ToT	0.580	0.600	0.560	0.680	0.530	0.540	0.520	0.760	0.530	0.440	0.610	0.810
EGoT	0.650	0.650	0.650	0.810	0.500	0.440	0.550	0.880	0.560	0.480	0.640	0.800
C-UoT	0.580	0.550	0.610	0.820	0.440	0.440	0.440	0.870	0.560	0.500	0.620	0.810

Appendix

Table 6: Human evaluation results across all creative reasoning tasks. C=Creativity, N=Novelty, U=Utility, and F=Feasibility. Bold: column best; underline: column second-best.

Method	Bridge Task				Electricity Task				Society Task
	C	N	U	F	C	N	U	F	C	N	U	F
OnePrompt	0.566	0.732	0.400	0.925	0.508	0.587	0.430	0.680	0.522	0.675	0.370	0.730
CoT	0.651	0.742	0.560	0.980	0.450	0.541	0.360	0.675	0.586	0.732	0.440	0.560
ToT	0.579	0.799	0.360	0.900	0.488	0.657	0.320	0.610	0.574	0.727	0.420	0.495
EGoT	0.641	0.643	0.640	1.000	0.409	0.428	0.390	0.825	0.613	0.775	0.450	0.525
C-UoT	0.664	0.728	0.600	1.000	0.512	0.563	0.460	0.675	0.649	0.799	0.500	0.530

Appendix

Table 7: LLM evaluation results across all creative reasoning tasks. C=Creativity, N=Novelty, U=Utility, and F=Feasibility. Bold: column best; underline: column second-best.

Method level
Metric	$r$	$p$	$n$
Creativity	0.708	$3.6{\times}10^{-5}$	27
Utility	0.481	$1.1{\times}10^{-2}$	27
Novelty	0.224	$2.6{\times}10^{-1}$	27
Feasibility	0.039	$8.5{\times}10^{-1}$	27
Instance level

Appendix

Table 8: Pearson correlation between human and Gemini 2.5 Pro scores on the novel creative-reasoning tasks (identical Utility/Feasibility prompts). Method level: one cell per method

\times

task (

n{=}27

). Instance level: paired individual solutions (

n{=}267

). Bold: strongest correlation in each row block.

Method	Solution
C-UoT	The core functional mechanism is a cooperative negotiation system. Its essential components are a driver interface for signaling crossing intent (input) and a central allocation system. The system aggregates these individual intents to form a comprehensive demand profile, and based on this, allocates specific crossing times (output) back to the drivers. This interaction, inspired by negotiation protocols in distributed systems, ensures fair and efficient resource use.
E-UoT	Based on a feedback-based mechanism, this solution implements a dynamic scheduling model that uses real-time traffic data (input) to continuously adjust vehicle crossing schedules. This interaction dynamically allocates crossing windows based on current conditions (output), which serves to optimize overall flow and reduce vehicle idle times.
T-UoT	This solution models the crossing scenario using game theory, treating vehicles as rational players aiming to minimize their individual travel time. The system analyzes the strategic interactions between all vehicles and calculates a Nash Equilibrium—a state where no single vehicle can benefit by unilaterally changing its crossing strategy. The resulting equilibrium strategy is then implemented as the global crossing schedule, ensuring a stable and efficient outcome for the entire system.

Appendix

Table 9: Sample solutions for the Bridge task

Panel of Figure 3 — Figure 3: Sensitivity analysis for UoT

Rank	Method	Bio	ML	Social	Avg	WR (%)
1	OnePrompt	1577.9	1537.0	1564.6	1560.4	46.3
2	E-UoT	1558.8	1545.4	1522.4	1542.1	40.7
3	ToT	1504.0	1528.2	1516.2	1516.1	34.0
4	CoT	1506.2	1475.3	1518.1	1500.0	30.2
5	EGoT	1438.8	1520.3	1442.1	1467.1	25.2
6	C-UoT	1444.6	1454.9	1483.5	1460.7	21.7

Appendix

Table 10: GLM-4-9B on HypoArena (100 tasks; Gemma-3-12B judge; canonicalized core hypotheses). Bio 34 / ML 34 / Social 32. Bio/ML/Social/Avg: pooled BTD ratings centered at 1500. WR: pairwise win rate (%). Bold: column best.

Judge pair	Spearman $\rho$	Exact 3-way	Hard A $\leftrightarrow$ B flip
Gemma $\leftrightarrow$ Mistral	0.616	0.604	2.6%
Gemma $\leftrightarrow$ Prometheus	0.587	0.478	4.4%
Mistral $\leftrightarrow$ Prometheus	0.604	0.520	5.4%

Appendix

Table 11: Judge consistency on frozen Qwen-100 hypotheses (Gemma-3-12B, Mistral-Small-24B, Prometheus-2). Spearman

\rho

over case

\times

method win-shares (

n{=}700

). Exact 3-way agreement and hard A

\leftrightarrow

B flip over

n{=}2100

matched pairs. Bold: column best (lowest flip).

Contrast	G $\leftrightarrow$ M	G $\leftrightarrow$ P	M $\leftrightarrow$ P
$\Delta$ (E $-$ OnePrompt)	0.71	0.71	0.79
$\Delta$ (E $-$ ToT)	0.65	0.59	0.69
$\Delta$ (C $-$ ToT)	0.67	0.49	0.52
$\Delta$ (T $-$ ToT)	0.63	0.51	0.57

Appendix

Table 12: Spearman

\rho

on per-case method contrasts across judges (G

=

Gemma, M

=

Mistral, P

=

Prometheus). Bold: row best.

Mode	Domain	Gap to #2	Task (brief)	UoT idea	ToT idea
C-UoT	ML	$+0.33$ vs E-UoT	Why entropy-minimization test-time adaptation fails for vision–language models under shift	Standard Shannon entropy assumes static class uniformity and clashes with pretraining-induced confidence asymmetries; use class-specific Tsallis entropy shapes tuned to those biases	Reweight entropy with class-specific “niche factors” so adaptation favors views that keep the correct class despite imperfect top-1 scores
E-UoT	Social	$+0.27$ vs EGoT	What sustains adherence to digital fitness coaches (live, avatar, or agent)	Retention depends on dynamic alignment: anthropomorphic design plus real-time shifts in emotional vs. technical support as user needs and adherence phases change	Engagement is a static balance of instrumental utility and para-social bonds, mediated by anthropomorphism, authenticity, and support type
T-UoT	Social	$+0.15$ vs OnePrompt	Why holiday e-commerce returns spike beyond ordinary product dissatisfaction	Surges arise when shoppers privilege local cues (gift urgency, immediate satisfaction) over systemic factors (return windows, recipient mismatch, delayed evaluation)—heuristic misalignment, not quality failure	Returns are driven by social-identity pressure and temporal dissonance between purchase intent and post-holiday reevaluation

Appendix

Table 13: Illustrative high-performing HypoArena cases for each UoT mode (Qwen3-14B; Gemma-3-12B judge). Each row shows a case in which the listed mode ranks first among seven methods with a substantial margin over the next-best method. Hypotheses are abbreviated to their canonicalized core claims.

Method	Top-1	Sole Top-1	$>\,$ ToT, $\geq$ 2/3 judges	$>\,$ ToT, 3/3 judges
C-UoT	15	8	26	11
T-UoT	11	6	24	11

Appendix

Table 14: Task-level success frequency on Qwen3-14B HypoArena. Top-ranked counts use the primary Gemma judge; cross-judge columns report tasks on which the method outperforms ToT under multiple judges.

Signal	C-UoT $>$ ToT	C-UoT $\not>$ ToT
Analogy selected	100%	100%
Mapped-mechanism specificity	0.97	0.97
Mapped $\rightarrow$ final preservation	0.53	0.43
Strong transfer	52%	32%

Appendix

Table 15: C-UoT stage-level diagnosis relative to ToT on Qwen-100. Wins denote tasks where C-UoT is preferred to ToT by the primary judge.

Signal	T-UoT $>$ ToT	T-UoT $\not>$ ToT
Gate passed	100%	100%
Transform applied	75%	83%
Rules $\rightarrow$ framing	0.75	0.83
Rules $\rightarrow$ final	0.13	0.16

Appendix

Table 16: T-UoT stage-level diagnosis relative to ToT on Qwen-100.

Mode	Domain	Task summary	Intervened step	Mean $\Delta$	Rank	What changed (before $\to$ after)
C-UoT	Social	Why do vending-machine choices become habitually locked onto familiar snacks despite healthier alternatives?	Final synthesis only (mapping frozen). Human rewrites the final from the frozen positive-feedback / path-dependence mapping; no new analogy search.	$+0.28$	$\#4\to\#1.5$ (#1 on 3/4 seeds)	Before: generic “habit via cues/placement” without the mapped feedback loop or a discriminating test. After: keeps the frozen local positive-feedback story (repeated choice reinforces favorites and suppresses alternatives) and adds the mapped prediction—disrupting cues should raise switching, whereas merely placing healthier items near favorites should not. Mapping-token preserve rises $0.20\to 0.33$ .
T-UoT	Bio	Why are ponerine ants much richer in some regions—diversification-rate differences, occupancy time, or dispersal source–sink structure?	Post-transform synthesis. Transform rules stay frozen (relax fossil/dispersal assumptions); model must state one mechanistic consequence of those rules, then write the final.	$+0.25$	$\#7\to\#2.5$	Before: retreats to data-limitation language (incomplete fossils, uneven sampling, barriers) and never uses the transform. After: applies the consequence of the relaxed rules—dispersal/colonization sequence imposes geographic constraints, so richness tracks early vs. late colonization timing rather than present-day diversification rates.

Appendix

Table 17: Targeted failure-mode interventions on frozen C-UoT and T-UoT cases.

\Delta

denotes

P(\mathrm{intervened}\succ\mathrm{original})-0.5

averaged over four seeds. Rank is computed among the intervened method and five fixed baselines; lower is better.

Method	Tokens/task	Rel. cost	Avg. BTD	WR (%)
OnePrompt	3.7k	1.0 $\times$	1475.7	14.3
CoT	3.5k	0.9 $\times$	1493.7	18.8
ToT	27.1k	7.4 $\times$	1515.8	25.5
EGoT	101.6k	27.6 $\times$	1505.1	25.3
C-UoT	134.3k	36.5 $\times$	1489.6	19.5
E-UoT	36.7k	10.0 $\times$	1535.4	32.7

Appendix

Table 18: Inference cost vs. HypoArena quality (Qwen3-14B). Tokens/task: mean total tokens on a fixed 10-task cost-audit subset (4 Bio / 3 ML / 3 Social). Relative: vs. OnePrompt on the same subset. Avg. BTD and WR (%): full 100-task panel (Table 2 ). Bold: column best (lowest cost / highest quality).

Computational Creativity

Computational creativity studies how machines can generate, evaluate, and collaborate on novel, valuable ideas and artifacts.

See all 114 papers

Apr 6, 2026

CreativityBench: Evaluating Agent Creative Reasoning via Affordance-Based Tool Repurposing

Recent advances in large language models have led to strong performance on reasoning and environment-interaction tasks, yet their ability for creative problem-solving remains underexplored. We study this capability through the lens of creative tool use, where a model repurposes available objects by reasoning about their affordances and attributes rather than relying on canonical usage. As a first step, we introduce CreativityBench, a benchmark for evaluating affordance-based creativity in LLMs. To this end, we build a large-scale affordance knowledge base (KB) with 4K entities and 150K+ affordance annotations, explicitly linking objects, parts, attributes, and actionable uses. Building on this KB, we generate 14K grounded tasks that require identifying non-obvious yet physically plausible solutions under constraints. Evaluations across 10 state-of-the-art LLMs, including closed and open-source models, show that models can often select a plausible object, but fail to identify the correct parts, their affordances, and the underlying physical mechanism needed to solve the task, leading to a significant drop in performance. Furthermore, improvements from model scaling quickly saturate, strong general reasoning does not reliably translate to creative affordance discovery, and common inference-time strategies such as Chain-of-Thought yield limited gains. These results suggest that creative tool use remains a major challenge for current models, and that CreativityBench provides a useful testbed for studying this missing dimension of intelligence, with potential implications for planning and reasoning modules in future agents.

Cheng Qian, Hyeonjeong Ha, Jiayu Liu +10

UIUC · Amazon · Columbia

LLM Tool Use Tool Use Creativity Assessment Computational Creativity

Ask PDF

Mar 10, 2026

CREATE: Testing LLMs for Associative Creativity

A key component of creativity is associative reasoning: the ability to draw novel yet meaningful connections between concepts. We introduce CREATE, a benchmark designed to evaluate models' capacity for creative associative reasoning. CREATE requires models to generate sets of paths connecting concepts in a model's parametric knowledge. Paths should have high specificity (distinctiveness and closeness of the concept connection) and high diversity (dissimilarity from other paths), and models are scored more highly if they produce a larger set of strong, diverse paths. This task shares demands of real creativity tasks like hypothesis generation, including an extremely large search space, but enables collection of a sizable benchmark with objective answer grading. Evaluation of frontier models shows that the strongest models achieve higher creative utility than others, with the high multiplicity of answers and complexity of the search making benchmark saturation difficult to achieve. Furthermore, our results illustrate that thinking models are not always more effective on our task, even with high token budgets. Recent approaches for creative prompting give some but limited additional improvement. CREATE provides a sandbox for developing new methods to improve models' capacity for associative creativity.

Manya Wadhwa, Tiasa Singha Roy, Harvey Lederman +2

New York University · The University of Texas at Austin

Creativity Assessment Computational Creativity LLM Evaluation

Ask PDF

Sep 18, 2026

MIRAGE: Multi-Perspective Creative Language Model Reasoning with Reinforcement Learning Guidance

Recent advances in Large Language Models (LLMs) have revolutionized artificial intelligence and how human interact with AIs. Despite impressive advancements, LLMs struggle with complex mathematical, scientific, and logical tasks. Inspired by human cognitive flexibility - our ability to dynamically switch mental perspectives - we propose MIRAGE (Multi-perspective Inference-time Reasoning via Agent-Guided Exploration), a novel inference-time creative thinking framework. MIRAGE includes a Selector that prioritizes effective conceptual perspectives (e.g., algebraic, probabilistic) and a Reasoner that sequentially solves tasks until a confident solution emerges, otherwise aggregating multiple perspectives. Tested on GSM8K, MATH500, MMLU-Pro, and Game-of-24 benchmarks, MIRAGE consistently outperforms methods like Chain-of-Thought and diverse prompting ensembles, significantly boosting accuracy with minimal inference overhead, providing a scalable solution for practical applications.

Arash Lagzian, Srinivas Anumasa, Dianbo Liu

National University of Singapore.

LLM Reasoning Efficient Language Model Reasoning Scientific Reasoning in Language Models

Ask PDF

Abstract

Computational Creativity

Explore similar work

CreativityBench: Evaluating Agent Creative Reasoning via Affordance-Based Tool Repurposing

CREATE: Testing LLMs for Associative Creativity

MIRAGE: Multi-Perspective Creative Language Model Reasoning with Reinforcement Learning Guidance