Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

A tabular foundation model, TabPFN-v2, can be adapted to solve the Travelling Salesman Problem in minutes with a single training sample.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

TabPFN-v2, fine-tuned on one 500-node TSP sample in about two minutes, constructs TSP tours reaching 2-5% of Concorde optimality after 2-opt post-processing, across instance sizes 50 to 1000.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection First real test of TabPFN on TSP; one-sample adaptation works well enough to be worth a referee, but the paper oversells comparative gains before adding a 2-opt baseline and measuring the train/test feature shift it admits. the 4 major comments →

arxiv 2511.05872 v1 pith:KYKEKCQB submitted 2025-11-08 cs.LG cs.AImath.CO

Adaptation and Fine-tuning with TabPFN for Travelling Salesman Problem

Nguyen Gia Hien Vu , Yifan Tang , Rey Lim , Yifan Yang , Hang Ma , Ke Wang , G. Gary Wang This is my paper
classification cs.LG cs.AImath.CO
keywords TabPFNTravelling Salesman Problemcombinatorial optimizationfoundation modelnode-based encodingfew-shot adaptationin-context learning2-opt
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that TabPFN-v2, a foundation model designed for tabular data, can be repurposed to solve combinatorial optimization problems, using TSP as the test case. The central claim is that one adapted model, fine-tuned for two minutes on a single 500-city instance, produces competitive tours on instance sizes from 50 to 1000 nodes. The authors argue this bypasses the need for millions of training samples and days of GPU training required by current deep-learning solvers. If true, it suggests pretrained tabular foundation models are a practical starting point for routing problems under tight data and compute budgets.

Core claim

The paper demonstrates that TabPFN-v2, originally a supervised model for small-to-medium tabular data, can be adapted to construct TSP tours via a node-based encoding scheme. Each step of route construction is cast as a regression problem: given the current node's coordinates and those of its five closest neighbors, the model predicts the next node's coordinates. A distance-based softmax then converts predictions into edge probabilities, and a greedy decoding builds the tour, optionally followed by 2-opt post-processing. Across TSP-50, TSP-100, TSP-500, and TSP-1000, the same adapted model achieves solution gaps of 13.27%, 15.36%, 19.66%, and 22.90% without post-processing, and 2.07%, 2.59%,

What carries the argument

The methodological engine is the node-based encoding: each tabular row represents a current node as its 2-D coordinates plus the coordinates and distances of its five closest neighbors, with the target being the next node's coordinates. This fixed-size input allows the same adapted model to handle any instance size, avoiding whole-instance encoders that grow with the graph. The decoding step translates the model's predicted coordinates into a probability matrix via softmax over median-normalized distances, then greedily assembles a feasible tour, with 2-opt as an optional refinement.

Load-bearing premise

The adapted model is trained using neighbor sets computed around the known next node, but evaluated using neighbor sets computed around the current node, and the paper assumes this distribution shift is negligible without measuring it.

What would settle it

Re-evaluate the adapted model using evaluation inputs built exactly as the adaptation inputs (i.e., neighbors of the true next node) and compare the resulting gaps to those reported; if the gaps change materially, the train/test feature mismatch is not benign and the headline numbers overstate the method's capability.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single adapted TabPFN-v2 model spans instance sizes from 50 to 1000 nodes without per-size retraining, mitigating the generalization breakdown common in whole-instance encoders.
  • Training data requirements drop from millions of samples to one, and adaptation time from days to minutes, enabling rapid deployment in data-scarce settings.
  • With simple 2-opt post-processing, solution quality becomes comparable to specialized learning-based solvers, closing the gap to within about 2-5% of exact solutions.
  • The fixed per-node input size keeps inference lightweight and allows all node predictions to run in parallel, making the method suitable for real-time routing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The neighbor-feature train/test mismatch — neighbors of the ground-truth next node during adaptation versus neighbors of the current node during evaluation — is asserted to have minimal impact but is not measured; if the shift is material, the reported no-2-opt gaps would be optimistic.
  • The same node-based encoding recipe could plausibly transfer to other routing CO problems such as vehicle routing, since the tabular formulation does not depend on TSP-specific structure.
  • The method's practical value lies in data- and compute-constrained environments, not in settings where exact or near-exact optimality is mandatory; it does not beat strong heuristics like LKH on solution quality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes the first application of TabPFN-v2, a tabular foundation model, to a combinatorial optimization problem, specifically the Traveling Salesman Problem (TSP). The method uses a node-based encoding: for each node, the 2-D coordinates of the current node and of its k nearest neighbours, plus distances, are used as features; the model is adapted on a single 500-node tour generated by OR-Tools to predict the coordinates of the next node. Two separate TabPFN-v2 regressors predict x and y. Routes are decoded from a probability matrix via softmax over normalized distances and constructed greedily, optionally improved by 2-opt. The adapted model is evaluated on TSP-50, TSP-100, TSP-500, and TSP-1000 using the Concorde-based datasets of Joshi et al. (2021) and Fu et al. (2021). Reported gaps range from 13.27% to 22.90% without 2-opt and 2.07% to 5.31% with 2-opt, all from a single model adapted in about two minutes on one sample. The paper claims that TabPFN-v2 can be successfully adapted to CO problems with extreme data and compute efficiency and generalizes across instance sizes.

Significance. If the claims hold, the paper would demonstrate an unusual and potentially valuable result: a tabular foundation model with a lightweight, node-based adaptation can produce competitive TSP solutions across a wide range of instance sizes using a single training sample and only minutes of adaptation. This is a genuinely novel application direction. The paper uses external Concorde-based test datasets, quotes many established baselines, and the reported gap arithmetic is internally consistent. However, the significance is conditional on two issues that directly affect the empirical claims: the neighbour-selection hyperparameter k is tuned on the same datasets used for the headline small-size results, and the input feature distribution differs between adaptation and evaluation without any measurement of the effect. Until these are addressed, the reported improvements and the comparison with prior methods are not fully trustworthy.

major comments (4)
  1. [§5.1, Table 1] The value k=5 is selected by evaluating the adapted model on 30 instances 'randomly sampled from the dataset by Joshi et al. (2021)'. These are the same test datasets used for the reported TSP-50 and TSP-100 results in Table 1. Thus the small-size gaps (13.27% and 15.36% without 2-opt; 2.07% and 2.59% with 2-opt) are in-sample for hyperparameter selection, not independent test-set results. The claim that the model generalizes 'across varying instance sizes' is weakened for the sizes where k was tuned. Please either use a separate validation set for k selection, report results across several choices of k, or clearly state that the small-size numbers are post-selection.
  2. [§4.2.1] The paper defines neighbour features differently in adaptation and evaluation. During adaptation, the k-closest neighbours are those of the ground-truth next node; during evaluation, they are those of the current node, because the next node is unknown. This changes every coordinate and distance feature in the input row. The paper asserts this 'has minimal impact' and cites Zhao & Wong (2025), but provides no measurement of the distribution shift or its effect on performance. Since the core contribution is the node-predicting adaptation strategy, the model may be evaluated on inputs it never saw during adaptation, making the without-2-opt gaps and the learned component's contribution upper bounds. Please quantify the shift and, ideally, evaluate a variant where the features are kept consistent between training and inference (e.g., always using current-node neighbours) and report the resul
  3. [§4.3] The 2-opt post-processing is said to run 'until the time limit is reached', but no time limit is specified anywhere in the paper. Since the comparison with other 2-opt-enhanced baselines depends directly on the post-processing budget, the absence of this parameter makes the reported 2-opt gaps (e.g., 2.07% on TSP-50, 5.31% on TSP-1000) not reproducible. Please state the time limit (or iteration budget) and, ideally, report the actual 2-opt runtime for each instance size. The claim of a 'two-minute adaptation phase' may also be misleading if substantial 2-opt time is excluded from the comparison.
  4. [§5.2] The central claim that a single adapted model works across all four instance sizes is based on one adaptation run using one 500-node OR-Tools solution. Because TabPFN-v2 fine-tuning can be sensitive to the training sample and initialization, a single run does not establish the robustness of the method. Please repeat the adaptation with several independently generated 500-node tours and different random seeds, and report mean and variance of the resulting gaps for at least TSP-100 and TSP-500. This is needed to support the 'rapid adaptation' and 'generalization' claims.
minor comments (6)
  1. [§4.2.3] The notation 'n2' for the matrix of distances should be 'n^2' or 'n×n'. Also, Equation (9) uses the softmax over the whole n^2 vector after setting diagonal entries to zero; this should be stated more clearly.
  2. [Figure 1] The figure described in §4.2.1 is not visible in the manuscript text as provided. If the figure is missing from the submission, it needs to be included; the description is otherwise helpful.
  3. [Table 1 and Table 2] The training time '2m' for TabPFN is ambiguous: does it include both the x-coordinate and y-coordinate models, and does it include the hyperparameter search over k? Please clarify. Also, the 2-opt time is not included in this number, which should be stated explicitly.
  4. [§4.2.1] The phrasing 'the input for the next node prediction is set to be the information of the current node position' is slightly inconsistent with the later description that neighbour distances are computed to the next node during adaptation. Please make the two phases of feature construction precise and consistent in notation.
  5. [§2.1.1] The definition 'undirected, edge-directed graph' is contradictory. A TSP graph is usually undirected; 'edge-directed' should be removed unless the authors intend an asymmetric variant.
  6. [§5.1] The claim that the authors 'do not evaluate the effect on larger instance sizes to reduce risk of bias' is good, but the k-selection on the small test sets is itself a source of bias, as noted in the major comments. Please rephrase to acknowledge this trade-off.

Circularity Check

0 steps flagged

No circularity found; the core claim is an empirical evaluation against external Concorde baselines, and the single self-citation is peripheral.

full rationale

The paper's claimed derivation chain is an empirical feasibility claim: fine-tune TabPFN-v2 on a single OR-Tools tour (§4.1: 'this solver is run for 30 minutes to solve the 500-point sample') and evaluate on held-out Concorde-generated test sets (§4.4: 'the solutions in all the above-mentioned test datasets ... are created using Concorde exact solver'). The teacher signal is external, the evaluation baselines are external, and no 'prediction' in the paper is constructed from a fitted constant. The k=5 neighbor choice is selected by scanning average gaps on evaluation instances (§5.1), which is test-set tuning/overfitting rather than a definitional circularity, and the paper does not relabel this selection as a prediction. The only self-citation (Vu et al. 2025, §2.2) is a peripheral example of foundation-model applications and is not load-bearing. The train/eval neighbor-feature mismatch in §4.2.1 is an acknowledged assumption ('such a difference in the distance calculation has minimal impact'), but it is not a circular reduction: evaluation uses current-node-centered features, so the final measured gaps do not reintroduce the adaptation target into the output. Therefore no step reduces to its own input by construction; score 0.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

No free parameters beyond k and the 2-opt stopping rule; no invented entities — the method reuses TabPFN-v2 as-is and introduces no new postulates. The main hidden load sits in the domain assumptions: the OR-Tools teacher's quality, the locality principle, and the tolerance to the train/test neighbour-feature mismatch.

free parameters (2)
  • k (number of closest neighbours) = 5
    Chosen by scanning k=1..40 on the TSP-50/TSP-100 evaluation test sets (§4.2.2, §5.1, Figure 2); 'the best performance is achieved when k is five.' This is test-set-based hyperparameter selection for the small-instance results.
  • 2-opt stopping time limit = unspecified
    Post-processing 'continues until the time limit is reached' (§4.3), but the limit is never stated, so the effort behind the with-2-opt gaps (2.07-5.31%) is not reproducible and could differ between instance sizes.
axioms (5)
  • domain assumption The OR-Tools 30-minute solution for the single 500-node teacher sample is near-optimal enough to serve as the training signal.
    §4.1: the solver runs for 30 minutes and 'the best-found solution is recorded'; its quality relative to Concorde/LKH is never verified, yet the entire adaptation learns from this one tour.
  • domain assumption Neighbour features computed around the ground-truth next node (adaptation) and around the current node (test) are distributionally compatible.
    §4.2.1 explicitly describes the difference and asserts 'minimal impact' via citation to Zhao & Wong (2025); no experiment in the paper measures the shift.
  • domain assumption Nodes close in space tend to be connected in the optimal tour (locality principle).
    §4.2.1: 'nodes in proximity are more likely to be connected in the optimal solution (Zhao & Wong, 2025)'; this justifies both the k-NN features and the tolerance to the train/test mismatch.
  • domain assumption Spatial k-NN edges suffice to construct complete tours, with the full-matrix edge set as a feasibility fallback.
    §4.2.3: candidate edges are added from highest probability via spatial k-NN edges first; only if these fail is the full-matrix set used, so final tour quality depends on the spatial k-NN edge set.
  • standard math Averaging over 30 test instances is sufficient for statistically reliable comparison.
    §4.4: justified via the central limit theorem cited to Chang & Lee (2017); reasonable in spirit, but no variance or confidence intervals are reported, so the 30-instance averages cannot be assessed.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptation and Fine-tuning with TabPFN for Travelling Salesman Problem." pith.science (2026). https://pith.science/paper/KYKEKCQB

@misc{pith2026251105872,
  author       = {Pith},
  title        = {Pith review of: Adaptation and Fine-tuning with TabPFN for Travelling Salesman Problem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KYKEKCQB}},
  note         = {Machine review of arXiv:2511.05872}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Tabular Prior-Data Fitted Network (TabPFN) is a foundation model designed for small to medium-sized tabular data, which has attracted much attention recently. This paper investigates the application of TabPFN in Combinatorial Optimization (CO) problems. The aim is to lessen challenges in time and data-intensive training requirements often observed in using traditional methods including exact and heuristic algorithms, Machine Learning (ML)-based models, to solve CO problems. Proposing possibly the first ever application of TabPFN for such a purpose, we adapt and fine-tune the TabPFN model to solve the Travelling Salesman Problem (TSP), one of the most well-known CO problems. Specifically, we adopt the node-based approach and the node-predicting adaptation strategy to construct the entire TSP route. Our evaluation with varying instance sizes confirms that TabPFN requires minimal training, adapts to TSP using a single sample, performs better generalization across varying TSP instance sizes, and reduces performance degradation. Furthermore, the training process with adaptation and fine-tuning is completed within minutes. The methodology leads to strong solution quality even without post-processing and achieves performance comparable to other models with post-processing refinement. Our findings suggest that the TabPFN model is a promising approach to solve structured and CO problems efficiently under training resource constraints and rapid deployment requirements.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Beyond IID: How General Are Tabular Foundation Models, Really?

    cs.LG 2026-06 unverdicted novelty 7.0

    Tabular foundation models excel on tiny- to medium-sized IID data but are outperformed by traditional tree-based and deep learning models on non-IID, large, and high-dimensional datasets, based on evaluations across 1...

Reference graph

Works this paper leans on

9 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Alanzi, E., & Menai, M. E. B. (2025). Solving the Traveling Salesman Problem with Machine Learning: A Review of Recent Advances and Challenges. Artificial Intelligence Review, 58(9). https://doi.org/10.1007/s10462-025-11267-x Alves, F., Pacheco, F., Rocha, A. M. A. C., Pereira, A. I., & Leitão, P. (2021). Solving a Logistics System for Vehicle Routing Pro...

    Pith/arXiv arXiv 2025

  2. [4]

    K., Cappart, Q., Rousseau, L

    https://doi.org/10.1186/1751-0473-1-3 Joshi, C. K., Cappart, Q., Rousseau, L. M., & Laurent, T. (2021). Learning TSP Requires Rethinking Generalization. In L. D. Michel (Ed.), Leibniz International Proceedings in Informatics, LIPIcs (Vol. 210, pp. 33:1-33:21). Schloss Dagstuhl- Leibniz- Zentrum fur Informatik GmbH, Dagstuhl Publishing. https://doi.org/10....

    Pith/arXiv arXiv 2021

  3. [8]

    https://doi.org/10.7717/PEERJ-CS.972 Luo, J., Yuan, Y., & Xu, S. (2025). TIME: TabPFN -Integrated Multimodal Engine for Robust Tabular -Image Learning . http://arxiv.org/abs/2506.00813 Mazurowski, M. A., Dong, H., Gu, H., Yang, J., Konz, N., & Zhang, Y. (2023). Segment Anything Model For Medical Image Analysis: An Experimental Study. Medical Image Analysis,

    Pith/arXiv arXiv 2025

  4. [59]

    https://doi.org/10.1016/j.addma.2022.103126 Skinderowicz, R. (2022). Improving Ant Colony Optimization Efficiency For Solving Large TSP Instances. Applied Soft Computing,

    arXiv 2022

  5. [74]

    https://doi.org/10.1016/j.techsoc.2023.102260 Held, M., & Karp, R. M. (1962). A Dynamic Programming Approach to Sequencing Problems. Journal of the Society for Industrial and Applied Mathematics, 10(1), 196–210. https://doi.org/10.1137/0110015 Helsgaun, K. (2017). An Extension of the Lin -Kernighan-Helsgaun TSP Solver for Constrained Traveling Salesman an...

    arXiv 2023

  6. [89]

    https://doi.org/10.1016/j.media.2023.102918 Meng, D., Cao, Z., Wu, Y., Hou, Y., Ge, H., & Zhang, Q. (2025). EFormer: An Effective Edge -based Transformer for Vehicle Routing Problems. In J. Kwok (Ed.), Proceedings of the Thirty -Fourth International Joint Conference on Artificial Intelligence, IJCAI-25 (pp. 8582–8590). https://doi.org/10.24963/ijcai.2025/...

    arXiv 2023

  7. [120]

    https://doi.org/10.1016/j.asoc.2022.108653 Sun, Z., & Yang, Y. (2023). DIFUSCO: Graph -based Diffusion Solvers for Combinatorial Optimization. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, & S. Levine (Eds.), Advances in Neural Information Processing Systems (Vol. 36, pp. 3706–3731). Curran Associates, Inc. https://proceedings.neurips.cc/paper_...

    arXiv 2022

  8. [130]

    J., & Lee, M

    https://dl.acm.org/doi/pdf/10.5555/3648699.3648829 20 Chang, H. J., & Lee, M. C. (2017). Applying Computer Simulation to Analyze the Normal Approximation of Binomial Distribution. Journal of Computers (Taiwan), 28(5), 116–131. https://doi.org/10.3966/199115992017102805011 Crişan, G. C., Pintea, C. M., Pop, P. C., & Matei, O. (2020). Economical Connections...

    arXiv 2017

  9. [260]

    S., & Sun, L

    https://doi.org/10.1016/j.knosys.2022.110144 Zhou, C., Li, Q., Li, C., Yu, J., Liu, Y., Wang, G., Zhang, K., Ji, C., Yan, Q., He, L., Peng, H., Li, J., Wu, J., Liu, Z., Xie, P., Xiong, C., Pei, J., Yu, P. S., & Sun, L. (2024). A Comprehensive Survey On Pretrained Foundation Models: A History From BERT To ChatGPT. International Journal of Machine Learning ...

    arXiv 2022

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.