Study Pits Human Intuition Against Neural Nets on the Traveling Salesman

A new preprint tackles a deceptively simple question: when humans try to solve the traveling salesman problem by eye, how close do they get to the mathematically optimal route, and can a neural network learn to think the same way?

The paper, posted to arXiv on July 26, 2026, runs a large-scale behavioral and computational comparison between human solvers and neural policies built on Pointer Networks. These are recurrent neural networks equipped with an attention-based pointing mechanism, designed to output probability distributions over possible tours through a set of points.

The traveling salesman problem, or TSP, asks for the shortest possible route that visits every point on a map exactly once. It's a classic hard optimization problem, and humans are surprisingly good at approximating it visually without doing any formal calculation.

Humans land in a "near-optimal basin"

The researchers found that human-drawn tours consistently cluster in what they describe as a near-optimal geometric basin. That means people's routes share many structural features with the mathematically best solution, even though they aren't identical to it.

Crucially, human tours also carry systematic, repeatable deviations from optimality, patterns of error that show up again and again rather than random noise. That consistency is what made the behavior tractable to model computationally.

Testing four ways to train a network

To figure out which training approach best reproduces human-like routing, the team tested networks under four regimes: reinforcement learning alone, supervised learning on optimal tours, supervised learning on tours humans actually drew, and a hybrid where a network pretrained on optimal solutions was then fine-tuned with reinforcement learning.

The hybrid approach won out. The best match to genuine human tours came from a model first pretrained on optimal solutions, then refined with RL, and finally decoded using Best-of-N sampling, a technique that generates multiple candidate solutions and selects the strongest one.

The research suggests that human-like problem-solving can emerge from combining structured supervised learning, reinforcement learning, and search at inference time, rather than from any single training method in isolation.

For AI researchers, the finding reframes what "human-like" reasoning might mean in optimization tasks: not a fundamentally different algorithm, but a layered process where prior structure, learned adjustment, and multiple-attempt search each contribute a piece. It also gives cognitive scientists a computational model to test against real human behavioral data on spatial reasoning tasks, rather than relying on optimal-solution baselines alone.