Long-distance exploration in insightful problem-solving

The paper investigates how insight and its associated Aha experience arise during problem-solving, focusing on the dynamic search processes within a solution space. Insightful solutions are phenomenologically marked by sudden realization, conviction of correctness, and positive affect, yet the mechanisms of the search process remain debated. Two key theories are considered: constraint relaxation (de-fixation) which reduces counterproductive constraints to open new search areas, and progress monitoring which detects insufficient progress and prompts strategic shifts. The authors hypothesize that de-fixation and exploration interact dynamically and that insightful searches differ from non-insightful ones in the breadth and pattern of exploration. To test this, they use the Japanese Compound Remote Associates (RAT/CRA) tasks to control constraints and monitor search in word space, combined with a simulation model capturing de-fixation and exploration capacity.

Prior work links insight to sudden restructuring and solution activation, and associates the Aha experience with metacognitive shifts and reductions in perceived difficulty. Mechanistic accounts include constraint relaxation and progress monitoring, with exploration being a key implication of the latter. The RAT/CRA tasks, although convergent in nature, engage divergent thinking when analytical strategies are insufficient, and have been widely used to probe insight-related cognition and physiology. Semantic search models, incubation effects, and retrieval-induced forgetting have been implicated in overcoming fixation. Network and hemispheric coding theories suggest that broader, coarse semantic coding (often right-hemisphere biased) supports insight. Debate remains over whether RAT performance is primarily driven by problem difficulty (remoteness) versus individual abilities (intelligence, literacy, executive functions).

Two behavioral experiments and a simulation model were used.

• Experiment 1 (FC-RAT; fixation-controlled RAT): Participants (final n = 349; 222 male, 127 female; mean age 33.9 ± 12.7 years) completed 16 questions (8 Fixation, 8 Neutral). In Fixation, each question kanji (Q1, Q2, Q3) was paired with a non-solution fixation cue (C1, C2, C3) forming meaningful 2-kanji words; in Neutral, unrelated cues (C1', C2', C3') formed no meaningful pairs. Time limit: 45 s per item. Participants clicked when they had an answer, typed it, received correctness feedback, and rated Aha on a 1–7 scale. Aha trials were ratings >4; No-Aha ≤4. The 16 items were matched by difficulty using published accuracy norms, with balanced assignment to conditions.
• Word distance measures: Using the Tsukuba Web Corpus (NINJAL-LWP; ~1.1B words), the authors built a 2-kanji corpus of 32,688 words (3,819 unique kanji), normalizing frequencies W to [0,1] relative to the most frequent word. For each question, normalized frequencies from Q1, Q2, Q3 to the answer A (W_Q1, W_Q2, W_Q3) quantified association strength and difficulty, summarized as d_answer = mean[log(W_Q1), log(W_Q2), log(W_Q3)]. Similarly, fixation strength to cues was d_cues = mean[log(W_C1), log(W_C2), log(W_C3)] (undefined for neutral cues). Bootstrapped regressions across the 16 questions (1,000 resamples) related d_answer and d_cues to average accuracy and reaction time.
• Statistical analyses: Two-sample t-tests with Bayes factors (Cauchy prior scale sqrt(2)/2) and Cohen’s d. Mixed-effects models (random intercepts for participants): logistic regression for accuracy (correct=1/incorrect=0) and Aha rate (Aha=1/No-Aha=0), and linear mixed models for reaction time on correct trials. Fixation models included d_answer and d_cues; Neutral models included d_answer only.
• Simulation-based power analysis: Target power 80% for SESOI=0.2, varying trial counts (100–1000) and effect sizes (0.1–0.6). Reaction times simulated as linear combinations of d_answer and d_cues plus Gaussian noise. Mixed-effects models assessed fixed-effect significance at α=0.05 over 100 simulations. Minimum ~600 trials needed for 80% power at SESOI=0.2; the collected data exceeded this per category (Fixation Aha 674, Fixation No-Aha 2118, Neutral Aha 909, Neutral No-Aha 1883).
• Experiment 2 (TT-RAT; thought-tracing RAT): On-site participants (n = 105; same as Experiment 1 on-site; mean age 22.5 ± 2.4 years) entered sequential candidate kanji over 45 s for 16 distinct questions with smaller d_answer values. Only fixation cues were shown. Incorrect entries faded after 0.5 s to reduce fixation. Success feedback and Aha ratings (1–7) were collected. Explore distance analysis quantified the search: for each thought Ti, normalized frequencies linking Ti to each question kanji were computed; distances were derived using the mean of log frequencies and averaged across the thought sequence to yield Explore_dist.
• Simulation model: A semantic search model navigated the kanji corpus, driven by three parameters: Fix_factor (strength of fixation influence), Defix_factor (strength of de-fixation or suppression of incorrect thoughts), and Explore_cap (exploration capacity, i.e., how many top-activated kanji can be sampled stochastically per step). Reaction time was defined as the number of steps to reach the answer. A total of 1,975,000 simulations (250 model configurations × 7,900 runs) assessed parameter effects. The model analyzed correlations among d_answer, d_cues, reaction time, and accuracy and probed time-limit (step-threshold) dependencies.

• Fixation impaired solution finding: No-answer rate was higher with fixation cues (Fixation 33.1 ± 2.3% vs Neutral 21.8 ± 1.9%; t(696)=7.54, p=1.45e-13, BF10=4.02e10; d=0.57), while wrong-answer rates did not differ significantly (Fixation 27.2 ± 2.2% vs Neutral 24.3 ± 2.1%; t(696)=1.88, p=0.061, BF10=0.47; d=0.14).
• Reaction times increased under fixation: RT Fixation 18.5 ± 0.6 s vs Neutral 14.7 ± 0.6 s (t(695)=8.65, p=3.61e-17, BF10=1.29e14; d=0.65). Correct trials were faster than incorrect (12.4 ± 0.4 s vs 24.8 ± 1.0 s; t(664)=24.34, p=5.30e-94, BF10=1.80e91; d=1.89). Among correct trials, Aha were faster than No-Aha (12.5 ± 0.59 s vs 14.1 ± 1.0 s; t(586)=3.05, p=0.0025, BF10=7.94; d=0.25).
• Aha rate unaffected by fixation presence: Fixation 59.8 ± 4.0% vs Neutral 61.8 ± 3.8% (t(685)=-0.57, p=0.571, BF10=0.10; d=-0.04).
• Mixed-effects regressions (Experiment 1): • Accuracy: d_answer negatively predicted accuracy in Fixation (Beta=-0.48, 95% CI [-0.56,-0.40], t(2789)=-11.35, p=3.06e-29) and Neutral (Beta=-0.34, 95% CI [-0.42,-0.26], t(2790)=-8.66, p=7.78e-18). Unexpectedly, d_cues also negatively predicted accuracy in Fixation (Beta=-0.17, 95% CI [-0.24,-0.10], t(2789)=-4.75, p=2.11e-6). • Reaction time (correct trials): d_answer positively predicted RT in both Aha (Fixation Beta=1.62, 95% CI [0.99,2.25], t(671)=5.01, p=6.91e-7; Neutral Beta=1.27, 95% CI [0.78,1.76], t(907)=5.07, p=4.89e-7) and No-Aha trials (Fixation Beta=1.83, 95% CI [1.21,2.44], t(2115)=5.80, p=7.41e-9; Neutral Beta=1.85, 95% CI [1.16,2.53], t(1881)=5.29, p=1.35e-7). d_cues positively predicted RT in Aha trials (Beta=1.25, 95% CI [0.73,1.78], t(671)=4.67, p=3.71e-6) but negatively in No-Aha trials (Beta=-1.22, 95% CI [-1.76,-0.68], t(2115)=-4.44, p=9.49e-6). • Aha rate (correct trials): d_answer negatively predicted Aha rate (Fixation Beta=-0.40, 95% CI [-0.49,-0.30], t(2789)=-8.26, p=2.18e-16; Neutral Beta=-0.25, 95% CI [-0.34,-0.17], t(2790)=-6.04, p=1.77e-9). d_cues negatively predicted Aha rate in Fixation (Beta=-0.13, 95% CI [-0.21,-0.05], t(2789)=-3.24, p=1.20e-3).
• Explore distance predicts Aha (Experiment 2): Accuracy was 51.8 ± 14.3%. Correct trials involved fewer thoughts than incorrect/no-answer trials (1.82 ± 0.59 vs 3.61 ± 1.12; t(104)=20.90, p=5.13e-39). Explore_dist positively correlated with Aha ratings (Spearman r=0.139, 95% CI [0.070,0.205], p=3.84e-5). Aha trials had greater Explore_dist than No-Aha (5.7 ± 0.1 vs 5.4 ± 0.1; t(868)=3.53, p=0.0004, BF10=34.22; d=0.24). The number of thoughts did not correlate with Aha (r=-0.002, p=0.95) and did not differ between Aha and No-Aha (t(868)=0.17, p=0.87, BF10=0.08; d=0.01).
• Simulation model: RT increased with higher Fix_factor and decreased with stronger Defix_factor and larger Explore_cap. Correlations between d_answer and log(RT) depended on Explore_cap: negative with small Explore_cap (1–3) and positive with larger (≥4), mirroring opposite d_cues–RT effects in Aha vs No-Aha. Optimal Explore_cap increased with problem difficulty (Spearman r=0.637, 95% CI [0.464,0.773], p=4.63e-10).
• Time-limit dependent accuracy effects: Model predicted that the sign of the d_cues–accuracy correlation flips with step/time thresholds; empirical re-analysis of FC-RAT supported positive correlations under short time limits and negative correlations under longer limits.

Findings support the hypothesis that insightful problem solving is characterized by broader, longer-distance exploration within the solution space, rather than by de-fixation alone. While de-fixation is vital to suppress misleading paths, it is not the determining factor for insight occurrence. Mixed-effects analyses and TT-RAT show that stronger Aha experiences are associated with longer exploration distances and quicker successful solutions, especially for easier problems or when strong impasses are resolved. The simulation model accounts for unexpected empirical patterns: opposite relationships of cue distance with reaction time in Aha vs No-Aha, and the negative association of cue distance with accuracy under typical time limits. It suggests that Aha moments involve greater exploration capacity and that each problem has an optimal exploration capacity tuned to difficulty; overly broad exploration can slow easy problems by adding distractions, whereas insufficient exploration hampers difficult problems. Conceptually, the work aligns with associationistic search over semantic networks and clarifies how exploration capacity, de-fixation, and time constraints jointly shape insight. The study distinguishes fixation from biased search, proposing future tests (e.g., instructions to ignore cues, introducing new cues or hints) to separate true fixation from mere bias. Broader implications span creativity enhancement and AI, emphasizing adaptive exploration strategies and representational changes not yet captured in the current model.

The study combines controlled variants of the Japanese RAT with a simulation model to reveal that insight and the Aha experience are marked by long-distance exploration, expanding the range of candidate solutions considered. It quantifies problem difficulty and fixation via corpus-based distances, demonstrates how cue proximity and time limits modulate accuracy and reaction times, and links exploration distance to Aha strength. The model explains distinctive fixation effects across Aha vs No-Aha and accuracy patterns, and predicts optimal exploration capacities by difficulty. Future research should: increase question and cue diversity; augment Aha ratings with impasse reports; capture representational change and multiple strategies; examine individual differences (e.g., cognitive flexibility, working memory, knowledge) and their neural bases via combined computational-neuroimaging approaches; and develop models that tailor exploration capacity to individual vocabularies and brain dynamics.

• Limited number and variety of RAT questions and cues, as tasks were bundled with others.
• Reliance on the Aha rating alone to index insight may miss cases where insight follows prolonged impasse; stronger fixation can produce more Aha despite difficulty.
• The model does not incorporate representational change mechanisms, focusing instead on search dynamics within a fixed semantic space.
• Distinguishing true fixation from biased search was not directly tested; future work should assess persistence of cue influence under new cues or explicit instructions to ignore cues.
• The model explains aggregate behavior rather than individual differences; the roles of intelligence, literacy, and executive functions in Aha remain unclear and warrant individualized modeling and neural measurement.