acceptodds
Under review as a conference paper at ICLR 2027

On the Geometry of Games and their Solvers

Abstract

Understanding where learning algorithms converge, cycle, or otherwise change behaviour is fundamental to solving games in economics and multi-agent systems. Existing game taxonomies are largely derived from analytical structure, such as potential, zero-sum, or monotone regimes, and provide powerful guarantees in restricted settings, but their boundaries need not coincide with realised solver dynamics or predict finite-time behaviour. We introduce behavioural cartography, a low-dimensional representation of games organised directly by the learning dynamics they induce. Rather than assigning only a structural class, it identifies which games elicit similar solver behaviour, where algorithmic regimes change, and how solver competence varies across game space. We construct the atlas from mechanistically distinct primitives spanning gradient, optimistic/extragradient, and fictitious-play families, evaluated across solver settings and initialisations. Games with similar induced dynamics are placed nearby in an oracle atlas z ⋆ (G), while a payoff-based model learns to locate unseen games in the same space, respecting strategic invariances including action relabelling and player exchange. Crucially, the atlas generalises beyond its construction: it predicts withheld properties such as recurrence and reversal and transfers to solver primitives excluded entirely from atlas construction. Across heterogeneous solvers, the learned atlas strongly preserves which games produce similar dynamics (pairwise-distance Spearman 0.917); when a solver is withheld, its behaviour can still be inferred from limited observations, reducing error by 30% relative to a constant reference at 205 labels. The resulting cartography reveals coherent solver-specific convergence and sensitivity fields, continuous behavioural variation along controlled structural paths, and distinctions not fully captured by conventional game classifications. Together, these results provide a predictive empirical atlas of game solving that complements analytical taxonomy by showing how algorithms behave across game space, how their capabilities relate, and how newly introduced solvers can be characterised.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Equilibrium solvers, Nash equilibrium, representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jtzvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper suggests an empirical approach to structure the space of two-player normal-form games based on their solver dynamics. The authors construct a low-dimensional representation, an "atlas" with 12 dimensions, of the game space. The atlas is meant to preserve the distance structures of trajectories which were generated across 7 solvers and multiple configurations per solver. The goal of this representation is to capture structures of games and their behavior that extend beyond traditional analytical properties such as monotonicity, potentialness, zero-sum, etc. On top of that, a direct mapping between game payoff matrices and the low-dimensional representations is learned, which means that the behavior space can be directly linked to a game's definition.

The obtained representations are tested extensively based on information that can recovered from them. For example, the authors analyze if trajectory features by unseen solvers can be deduced from the atlas and if trajectory properties (exploitability, movement, recurrence, reversal) can be predicted based on neighboring games in the structure.

Critical strengths
  1. Novelty: The overall idea of the paper is very intriguing: Instead of analyzing narrow classes of games (such as potential games), the authors suggest to learn representations that encode a more general behavior across all games empirically. I agree that many solvers share common behavioral patterns (this paper even quantifies that, see Fig. 2), and that the currently provided explanations are not sufficient for understanding the overall structure of all games. This paper is, to my knowledge, the first that applies this kind of representation learning with game theory, which is a novel step. Despite the concerns stated below, this alone is sufficient reason for me to (weakly) recommend this paper for acceptance, specifically in the context of ICLR.

  2. I also appreciate the amount of empirical evidence provided. (I didn't go through all of the results in the appendix.)

  3. Presentation: The overall exposition of the paper is clear, and its core ideas are presented intuitively. In particular, the three compression steps (G→dbeh→z∗→z^G \rightarrow d_{beh} \rightarrow z^* \rightarrow \hat{z}) are easy to understand on a high level.

Critical weaknesses
  1. Motivation: Despite the interesting prospect of having an cartographical representation of all games, this paper leaves some motivational issues and contains some inconclusive results:

    1. It is unclear to me which direct benefits can be drawn from the atlas. Whereas potentialness, zero-sum, and monotonicity provide a direct convergence guarantee for certain algorithms that can be directly drawn from the game itself, the learned representation provide no such things. While the authors provide an analyis of behavior along a path of games, this does not provide a complete picture of the set of all games and why they do or do not converge. (I admit that this is a very difficult task.)
    2. The current analysis seems limited to small, square two-player games. These, by no means, represents the space of all games and possible dynamics. It would be interesting to see how the structure of games varies outside of this narrow field.
    3. The NMAE values reported in Fig. 3 do not seem convincing to me: If 1 is considered to be the baseline value, 0.8 does not appear to be particularly good.
  2. Presentation: Given the complex implementation of the atlas training and evaluation steps, some parts of the paper remain elusive unless a reader spends substantial amount of time with the paper and its appendix.

    • How was the corpus of games sampled? (This is partially addressed in the appendix.)
    • Where exactly is the TEST set used?
    • Recurrence and reversal are only explained in the appendix but are essential for understanding the paper. I would at least explain them in words in the main text.
    • The size nn of the analyzed games is not stated in the main text.
    • Figures 1 and 3 need further explanation. Many of the metrics/comparison methods used are either not explained or explained later/in the appendix.

    The written text is sometimes quite condensed and difficult to follow. In particular, the paper is written for a very knowledgable reader: Many sentences contain words that are either subject-specific or require further context, which is often only provided in the appendix. Most readers will either have a background in game theory (such as myself) or in heavy-duty machine learning - combining both topics is an important step, which nonetheless requires taking care of both groups.

Questions for authors

Clarifications:

  • Some of the solvers are quite similar. Are the LOSO results in Fig. 3 really that great?
  • What are the comparison methods used in Figure 3?
  • Can you provide any additional insights on the structure of game behavior based on your atlas?
  • Why did you always withheld the recurrence and reversal traits? Couldn't they be used to reach a more informative atlas, while still being able to perform a leave-nn-categories-out analysis?

Future work:

  • Have you considered exploring other games with more players, non-square structure (A,B∈Rm×nA, B \in \R^{m \times n}), and more actions?
  • Does your set of prototype games include any games with complex, potentially chaotic structure? The Kohlberg-Mertens game analyzed in a paper by Milionis et al. (https://doi.org/10.1073/pnas.2305349120) could be worth a look.
4 comments
AuthorsOpenReview ↗

Response to jtzv: Scope and Analytical Guarantees

We thank the reviewer for recognising the novelty of behavioural cartography and for asking how it relates to analytical results and extends to other game sizes.

What our approach adds to analytical game theory.

Properties such as potentialness and monotonicity can tell us when an algorithm has a convergence guarantee. Our aim is different: we study how algorithms actually behave when run for a fixed number of steps. Games with the same structural properties can still lead to different outcomes for a solver. Our atlas organises games by these observed differences.

The atlas also does more than visualise behaviour. It predicts recurrence (whether a solver returns near earlier strategy states) and reversal (whether its updates change direction). Neither property was used to construct the atlas. For example, the learned atlas predicts GDA recurrence with an MAE of 0.049, compared with 0.100 using structural game-category labels (Figure 1). Our controlled game-mixture experiments also show that the atlas tracks changes in behaviour as game structure varies.

Transfer to larger games.

We have also tested whether the representation works beyond the 3×33\times3 games on which it was trained. We kept the learned payoff recogniser fixed and used only labelled 3×33\times3 games as reference examples. In particular, there were no behavioural labels from the larger games and no retraining for those sizes. The resulting prediction errors are:

Representation4×44\times4 NMAE5×55\times5 NMAE
Behavioural atlas (ours)0.604 ± 0.0490.674 ± 0.032
NfgTransformer0.690 ± 0.0710.760 ± 0.062
Response graph0.7700.790

The values average over 35 solver–behaviour targets; the reported variation for the neural representations is across three seeds. Our atlas has the lowest average NMAE of these methods at both sizes, without using any labels from the larger games.

These results support two conclusions. First, the atlas captures useful differences in solver behaviour that structural game labels alone do not explain. Second, the learned representation remains informative when applied to the larger games we evaluated.

Our claims concern two-player normal-form games under the specified solver protocol. Games with unequal action sets, substantially larger games, and multiplayer settings are natural directions for extending this work.

AuthorsOpenReview ↗

Response to jtzv: Evaluation Protocol and Figure Clarifications

Corpus and held-out evaluation.

The principal experimental corpus contains 35,804 two-player normal-form games, with the main representation and transfer experiments using 3×3 games. We construct the corpus from canonical game families and interpolations between them to cover diverse structural settings. Separate cross-size experiments evaluate transfer to 4×4 and 5×5 games. TRAIN, DEV and TEST are separated by generative root so that variants of the same game do not cross splits. The main behavioural reference sets contain 2,048 TRAIN and 512 DEV games. The new evaluation uses 5,371 held-out TEST games, with the models and evaluation settings frozen beforehand. We will make the game size and split roles clearer in the main text.

Figures 1 and 3.

Figure 1 shows variation in GDA recurrence within structural categories. Recurrence measures whether a solver returns near earlier strategy states and is excluded from atlas construction. The learned atlas predicts it with MAE 0.049, compared with 0.100 for structural baselines (oracle: 0.029). Figure 3 compares Direct GNN, response graph, Candogan PH/PHN, NfgTransformer and our learned atlas across 20, 102 and 205 labelled reference games. Its rows distinguish full-library from LOSO evaluation; its columns distinguish source-derived from withheld behavioural quantities. Figure 3 uses Ridge readouts, whereas the aggregate TEST LOSO NMAE 0.694 reported in the main response uses 10-neighbour kNN. We will state this explicitly in the caption and text.

Why withhold recurrence and reversal?

The behavioural distance uses exploitability and step-to-step movement distributions at each iteration. It does not directly track whether individual runs revisit an earlier state (recurrence) or reverse their update direction (reversal). Excluding these quantities tests whether the geometry predicts behaviour beyond its construction signals. These quantities could be incorporated into a different atlas, but they would no longer provide the same test of predicting behaviour excluded from representation construction.

Interpreting NMAE.

NMAE measures prediction error relative to a constant-median predictor fitted on the same labelled reference games. An NMAE of 0.8 therefore represents a 20% reduction in prediction error, rather than an accuracy of 80%. Our main TEST result achieves NMAE 0.694, corresponding to approximately 31% lower error than the constant reference. We also report performance for individual solvers and behavioural targets, since these vary in predictability. The new neighbourhood and failure-localisation experiments provide complementary tests of whether the representation captures useful behavioural relationships.

What the atlas reveals beyond game structure.

Our additional within-potentialness analysis addresses the reviewer's question about what behavioural cartography contributes beyond analytical classifications. Even when neighbouring games are restricted to the same TRAIN-defined potentialness quartile, behaviour-supervised representations identify games with substantially more similar finite-time solver behaviour than the evaluated structural features. This demonstrates why analytical structure and empirical solver behaviour provide complementary information: games with comparable potentialness can still differ substantially in how a learning algorithm behaves over a fixed horizon.

Thank you for the clarifications and additional experiments. The cross-size transfer results are good, and the distinction between analytical guarantees and finite-time behavioural prediction is helpful. Please incorporate these clarifications into the main text. Two questions remain: How much does LOSO performance depend on similarity between the held-out and remaining solvers? And can you provide one concrete insight into game dynamics revealed by the atlas beyond improved prediction? The within-potentialness analysis is useful, but sharing a quartile does not imply equivalent analytical structure. Overall, I maintain my weak acceptance recommendation.

Thank you for revisiting our responses and for your constructive assessment. We will incorporate the cross-size results, the relationship to analytical guarantees, and the evaluation clarifications into the main text.

Dependence on solver similarity. Our paired-exclusion experiments directly test whether LOSO transfer relies on retaining a closely related solver. Excluding GDA and averaging/anchoring together yields TEST NMAE 0.536; excluding extragradient and optimistic gradient together yields 0.599. Thus, substantial predictive information remains even when these related mechanisms are jointly excluded from atlas construction.

Transfer is nevertheless heterogeneous: jointly excluding EMD and damped proximal gives NMAE 0.948. These findings show that nominal algorithmic-family similarity does not fully determine transfer, although they do not establish uniform generalisation across arbitrary solver families. We will make this distinction explicit in the manuscript.

A concrete insight into learning dynamics. Figure 5 provides an illustrative example beyond prediction accuracy. We continuously interpolate between zero-sum and common-interest payoffs while holding the endpoint games fixed. Along this path, measured GDA recurrence decreases sharply between adjacent mixtures at λ=0.4\lambda=0.4 and λ=0.5\lambda=0.5, before gradually increasing towards the common-interest endpoint.

Thus, a smooth change in game payoffs can produce a rapid, non-monotonic change in the solver's finite-time dynamics. The behavioural atlas organises this variation and identifies the region of changing recurrent behaviour. This illustrates how empirical cartography can reveal dynamical regimes and transitions that are not directly specified by the structural interpolation itself.

We also agree that membership in the same potentialness quartile does not imply analytical equivalence. Our within-quartile experiment is intended as a controlled comparison against one structural descriptor, rather than an assertion that the games share identical convergence guarantees. We will clarify this distinction and highlight the controlled-interpolation example in the main text.

We appreciate your continued engagement with the paper.

Reviewer_kDgJOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a framework for measuring similarity between games based on the learning behaviour observed across the solvers and use this similarity to construct a low-dimensional map of games. To measure the learning behaviour, they run several learning algorithms on each game and measure how close the players get to equilibrium and how much their strategies change over time. These measurements define distances between games, which they use to construct a low-dimensional map called an "atlas" that places games with similar solver behaviour closer together. Then they use "payoff-based recogniser" a graph neural network to predict a game's atlas coordinates from its payoff matrix. Then they evaluate the prediction of behavioural properties excluded during the atlas construction and transfer to excluded solvers through leave-one-solver-out evaluation. Finally, they use the atlas to visualize solver-dependent convergence and sensitivity, and examine how the behaviour changes along controlled paths between game structures. The authors argue that the map complements existing game classifications by showing how learning algorithms behave on different games.

Critical strengths
  1. A new way to group games: The core idea of using solvers' behaviour to decide the game groupings rather than their structural category is interesting. I find this a major value in this paper because Figure 1 reveals that games with the same structural category can behave very differently under the same solver this finding shows something the structural category doesn't capture.

  2. Splitting the pipeline into stages increases the transparency of the evaluation: Firstly, the behaviour distance is measured by running the solvers, and then use that distance to build the map and finally predict games' positions on the map using only payoffs (excluding the solvers). I find this useful because each stage is evaluated separately; it is easier to understand where the errors come from and also might aid future work in deciding which stage needs improvement.

Critical weaknesses

Weakness:

  1. The main results are reported on Dev set and not the held-out TEST set: The headline numbers appear to come from the Dev set, which is used for model selection and hyperparameter selection. Line 323 states the geometry result “on DEV,” and the leave-one-solver-out result at line 374 matches Table 7, which is also labelled DEV. I consider this a major concern because using data used for model selection as the main results might overestimate the performance.

  2. Transfer to the excluded solvers is uneven, so the generality claim is not established: The authors claim that the atlas transfers to excluded solvers and captures reusable structure across solver mechanisms (lines 24–25, 109–111). However, the detailed Table 8 reveals that for 3 of the 7 solvers, entropic mirror descent (0.93), damped proximal (0.88) and fictitious play (0.86), the atlas is only slightly better than the constant baseline. The reported 30% improvement is an average driven by the remaining solvers. Since all seven solvers come from a few related families, the results do not show that the atlas can characterise a genuinely new kind of solver, which is one of the paper’s headline claims.

Questions for authors
  1. Could the authors confirm which split the main results are reported on?. If these are DEV results, please report the main results (Tables 4, 7 and 8 and the corresponding main-text numbers) on the held-out TEST split. If the TEST results are consistent with those reported, this concern would be resolved.

  2. Why does transfer fail for entropic mirror descent, damped proximal and fictitious play? One possibility is that these solvers are less similar to the rest of the panel, while solvers that transfer well have close relatives remaining (e.g, GDA and averaging/anchoring correlate at 0.96 in Figure 2). Could the authors test transfer to a solver from a family not used in building the atlas, or with a whole solver family removed? Clear gains over the constant baseline there would support the generality claim. Otherwise, I would ask that the claim be narrowed to solvers similar to those in the panel.

I could not find a code release commitment in the paper; given the pipeline’s complexity, a release would substantially aid reproducibility.

4 comments
AuthorsOpenReview ↗

Response to kDgJ: Held-out TEST Evaluation

We thank the reviewer for highlighting this. We agree that our original presentation did not sufficiently distinguish DEV results from held-out TEST evaluation. To address this, we evaluated the original protocol on the held-out TEST split. The trained representations, TRAIN-labelled support sets, 10-neighbour readout, and evaluation protocol had all been fixed before TEST evaluation.

Main findings on held-out games.

The main findings remain consistent on TEST. At 205 target labels, the learned atlas achieves an aggregate LOSO NMAE of 0.6936 on TEST, compared with 0.6991 on DEV (lower is better; 1.0 denotes the constant-median baseline). Its advantage over the matched Ordinary128 baseline also persists, with approximately 12.0% lower prediction error on TEST, compared with 12.9% on DEV.

The learned representation also retains strong geometric fidelity, achieving a pairwise-distance Spearman correlation of 0.900 on TEST, compared with 0.917 on DEV.

These results confirm that our central findings regarding behavioural geometry and aggregate cross-solver prediction extend to held-out games. We have also identified solver-specific variations, which we address separately in our response on transfer generalisation.

We will update Tables 4, 7 and 8, together with the corresponding main-text results, to clearly report held-out TEST performance and distinguish it from DEV-based model selection.

AuthorsOpenReview ↗

Response to kDgJ: Cross-Solver Transfer and Paired Exclusion

We thank the reviewer for highlighting this. We agree that the original leave-one-solver-out (LOSO) evaluation does not fully rule out the possibility that transfer is driven by closely related solvers remaining in the source panel. We therefore conducted a stricter paired-exclusion evaluation to directly test this hypothesis.

Paired-exclusion experiments.

We jointly exclude three pairs of related solvers: (i) GDA and averaging/anchoring, (ii) entropic mirror descent (EMD) and damped proximal (PD), and (iii) extragradient and optimistic gradient. For each pair, both solvers are removed from behavioural-distance construction, atlas embedding, recogniser training, and source supervision for the matched controls. Representations are retrained using only the remaining five solvers and evaluated on all TEST games, with identical target TRAIN-label subsets and a fixed 10-neighbour readout.

Transfer beyond closely related mechanisms.

The results show that cross-solver transfer is not solely explained by retaining closely related mechanisms. The learned atlas achieves pair-average NMAEs of 0.5363, 0.9482, and 0.5992 for the three exclusion settings, corresponding to 46.4%, 5.2%, and 40.1% lower prediction error than the constant reference, respectively.

In particular, substantial predictive information remains for GDA/AA and EG/Optimistic even when both related mechanisms are excluded. This provides stronger evidence that behavioural structure learned from other solvers can generalise beyond the individual mechanisms used to construct it.

Understanding weaker transfer.

We acknowledge that transfer remains uneven, particularly for EMD and PD. Their TEST NMAEs under the original LOSO protocol are 0.9800 and 0.9325, respectively, and remain similar under joint exclusion (0.9697 and 0.9267). These results indicate that weak transfer persists under paired exclusion, rather than being resolved by removing the closely related counterpart.

Our staged framework also helps identify where predictive information is lost. For example, EMD's time-averaged exploitability can be predicted with NMAE 0.761 using raw source distances and 0.727 using oracle atlas coordinates, compared with 0.907 using payoff-predicted coordinates. This indicates that useful cross-solver information exists in the source behavioural geometry, but is not fully preserved by the current payoff-to-atlas mapping.

Scope of generalisation.

Our objective is not unconditional prediction of arbitrary out-of-distribution solvers. Rather, the atlas provides a shared behavioural geometry of games that can be reused to characterise newly introduced solvers from limited target-solver observations, without reconstructing the representation. Our seven primitives span several foundational mechanisms of iterative game learning, including gradient, mirror, proximal, predictive, and response-based updates. While this set is not exhaustive, it provides a mechanistically diverse basis for investigating shared behavioural structure.

The paired-exclusion results suggest that transfer is not determined solely by nominal algorithmic-family overlap. A potentially more relevant consideration is whether a new solver's behavioural response varies sufficiently regularly over the source-induced game geometry. This offers a hypothesis for understanding when transfer may succeed, rather than a guarantee. Indeed, the weaker EMD and PD results demonstrate that transfer is not uniform across mechanisms.

We acknowledge that transfer is uneven, for example EMD and PD, our contribution is therefore not universal superiority in average prediction error, but the construction of an explicit, behaviourally grounded geometry that separates the information contained in solver-induced similarities from the effects of dimensionality reduction and payoff-based recovery, while providing a reusable basis for characterising additional solvers.

We will include the full paired-exclusion results, per-solver breakdowns, and additional transfer diagnostics in the appendix.

Thank you for the additional experiments. The TEST results are consistent with DEV, so my first concern is resolved. The paired-exclusion experiment partly addresses my second concern. So I have raised my score to 3. I request the authors to include these results, state the narrowed transfer claim as a limitation, and commit to a code release.

AuthorsOpenReview ↗

Response to kDgJ: Follow-up on Transfer Limitations and Reproducibility

Thank you for revisiting our responses and for your constructive assessment.

We will ensure that the revised manuscript incorporates the held-out TEST results and paired-exclusion experiments, including the per-solver breakdowns. We will also explicitly state the limitation concerning uneven cross-solver transfer, while distinguishing it from the demonstrated transfer across several excluded mechanisms.

Additional diagnostics on cross-solver transfer.

Our additional diagnostics help clarify this variation. In particular, EMD’s recurrence and reversal targets are nearly constant, while FP recurrence is rare and difficult to recover using the small-label median readout. Nevertheless, FP’s overall exploitability remains predictable (NMAE 0.59), and the atlas identifies PD failure regions more effectively than the evaluated structural representations (AP 0.50 versus 0.27–0.36). These results clarify which behavioural properties transfer and where the current approach is less effective.

Code availability and reproducibility.

We also commit to releasing the full experimental pipeline, including game generation, solver implementations and configurations, behavioural-distance construction, atlas training, and the strict LOSO and paired-exclusion evaluation scripts.

Thank you again for engaging with the additional evidence.

Reviewer_QX71OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a framework called behavioural cartography for learning representations of games based on the behaviour of learning algorithms, rather than relying only on traditional structural properties such as being zero-sum or potential games. The authors construct a behavioural distance between games using measurements of exploitability and strategy movement obtained from several learning algorithms. These distances are used to construct a low-dimensional representation of games through multidimensional scaling (MDS). Finally, a neural network is trained to predict these representations directly from payoff matrices, without requiring the full collection of solver experiments on new games.

The authors evaluate their framework on a large collection of synthetic two-player games. Their experiments suggest that the learned representations can predict several properties of learning dynamics, such as convergence, recurrence, and sensitivity to initialization. They also demonstrate transfer to learning algorithms excluded from the construction of the representation.

Critical strengths

1. An interesting approach to represent/categorize games. The paper proposes representing games based on how learning algorithms behave on them, rather than only using structural properties such as zero-sum or potential games. This provides an alternative way to compare and group games, which could be useful for understanding differences in learning dynamics.

2. Solid empirical evaluation. The paper evaluates the approach using several learning algorithms, configurations, and initializations across a large collection of synthetic games. The experiments include comparisons with existing representations and evaluations on withheld behavioural properties and solvers. In particular, the leave-one-solver-out results provide some evidence that the learned representations can capture information that transfers across different learning algorithms.

Critical weaknesses

1. Limited discussion of closely related work and the novelty of the proposed representation. The related-work discussion seems quite limited given the existing literature on game representations and learning across games. In particular, Omidshafiei et al. (2020), Navigating the Landscape of Multiplayer Games, appears closely related in motivation, as it also constructs a geometric representation of games and connects game properties to the computational difficulty of solving them. Other relevant works include Marris et al. (2022), Turbocharging Solution Concepts: Solving NEs, CEs and CCEs with Neural Equilibrium Solvers, and Harris et al. (2023), Meta-Learning in Games. While these works address different objectives, I think they provide important context for the proposed framework. A more comprehensive discussion would help clarify what is fundamentally different about representing games through observed solver behaviour and how this complements existing approaches.

2. Unclear practical utility beyond predicting solver behaviour. Although the empirical results suggest that the proposed atlas is useful for predicting several properties of learning dynamics, I am somewhat unsure what concrete algorithmic benefits these predictions provide. For instance, it is not clear whether the representation could be used to select a suitable learning algorithm or its hyperparameters for a given game, resulting in better equilibrium quality or computational efficiency. The experiments primarily evaluate prediction accuracy on synthetic games, rather than demonstrating improvements in such downstream tasks. Since the framework is motivated by understanding and exploiting differences in learning dynamics, I feel that demonstrating a concrete application would make the overall contribution more stronger.

Additional References

  1. Omidshafiei, S., Tuyls, K., Czarnecki, W. M., Santos, F. C., Rowland, M., Connor, J., Hennes, D., Muller, P., Pérolat, J., De Vylder, B., Gruslys, A., & Munos, R. (2020). Navigating the landscape of multiplayer games. Nature Communications, 11, Article 5603. https://doi.org/10.1038/s41467-020-19244-4

  2. Marris, L., Gemp, I., Anthony, T., Tacchetti, A., Liu, S., & Tuyls, K. (2022). Turbocharging solution concepts: Solving NEs, CEs and CCEs with neural equilibrium solvers. Advances in Neural Information Processing Systems (NeurIPS), 35, 5586–5600. https://arxiv.org/abs/2210.09257

  3. Harris, K., Anagnostides, I., Farina, G., Khodak, M., Wu, Z. S., & Sandholm, T. (2023). Meta-learning in games. International Conference on Learning Representations (ICLR). https://arxiv.org/abs/2209.14110

Questions for authors

1. Relationship to existing game representations and learning approaches. Could the authors discuss how the proposed behavioural atlas relates to earlier approaches such as Omidshafiei et al. (2020), Marris et al. (2022), and Harris et al. (2023)? In particular, what additional information does the proposed behavioural representation capture compared with existing game representations, and what advantages does it offer? I recognize that these works address somewhat different problems, but a clearer discussion would help me better understand the novelty and significance of the proposed approach.

2. Practical benefits of predicting solver behaviour. Could the authors demonstrate how the predicted behavioural properties can be used to improve an actual game-solving task? For instance, could the representation be used to select a learning algorithm or its hyperparameters for a new game, and would this lead to lower exploitability or faster convergence under a fixed computational budget? A comparison against selecting a fixed solver or using existing structural representations would help establish the practical utility of the framework beyond predicting trajectory statistics.

Additional References

  1. Omidshafiei, S., Tuyls, K., Czarnecki, W. M., Santos, F. C., Rowland, M., Connor, J., Hennes, D., Muller, P., Pérolat, J., De Vylder, B., Gruslys, A., & Munos, R. (2020). Navigating the landscape of multiplayer games. Nature Communications, 11, Article 5603. https://doi.org/10.1038/s41467-020-19244-4

  2. Marris, L., Gemp, I., Anthony, T., Tacchetti, A., Liu, S., & Tuyls, K. (2022). Turbocharging solution concepts: Solving NEs, CEs and CCEs with neural equilibrium solvers. Advances in Neural Information Processing Systems (NeurIPS), 35, 5586–5600. https://arxiv.org/abs/2210.09257

  3. Harris, K., Anagnostides, I., Farina, G., Khodak, M., Wu, Z. S., & Sandholm, T. (2023). Meta-learning in games. International Conference on Learning Representations (ICLR). https://arxiv.org/abs/2209.14110

8 comments
AuthorsOpenReview ↗

Response to QX71: Relationship to Existing Game Representations

We thank the reviewer for highlighting these complementary works. Collectively, they demonstrate that regularities across games can be characterised and exploited in different ways. Omidshafiei et al. (2020) construct a game landscape from evolutionary response-graph structure and relate its properties to game characteristics; Marris et al. (2022) learn an equivariant mapping from game payoffs to approximate equilibrium solutions; and Harris et al. (2023) establish meta-learning guarantees that exploit similarity across games.

Our work addresses a complementary question: whether the finite-time behaviours induced by a diverse range of learning algorithms can themselves provide a meaningful geometry for organising and understanding games. Rather than focusing on finding the equilibrium solutions, accelerating their computation or analyzing games based on response structure, we define similarity through time-indexed distributions of exploitability and strategy movement across diverse solver mechanisms, configurations, and initialisations. We then learn to recover this compact behavioural embedding directly from game payoffs.

This representation characterises games by similarities in the learning dynamics they induce, capturing aspects of solver progress and behavioural variability not directly represented by equilibrium solutions or response-graph topology. We further demonstrate prediction of withheld trajectory properties and transfer to solvers excluded from atlas construction, using limited target-solver observations. These results support a shared, predictive organisation of game space based on solver behaviour. We view this as a complementary empirical characterisation of games, potentially providing a foundation for studying relationships and learning mechanisms further down the line.

AuthorsOpenReview ↗

Response to QX71: Behavioural Similarity and Practical Utility

We have conducted additional strict leave-one-solver-out experiments to examine what behavioural cartography captures beyond conventional structural representations and how this information can be used.

Behavioural similarity beyond game structure. Across all seven excluded solvers, the learned atlas identifies games with more similar late-stage behaviour than the evaluated structural representations. The advantage persists when neighbours are restricted to the same potentialness quartile: for GDA, averaging/anchoring and extragradient, behaviour-supervised representations reduce the discrepancy relative to random same-quartile neighbours by 51–54%, compared with 17–27% for structural representations. This demonstrates that comparable structural properties need not imply similar finite-time learning behaviour, and that our geometry captures additional solver-relevant information.

Characterising solver competence. We also tested whether these neighbourhoods can identify regions where an excluded solver is unreliable. Using only 205 labelled reference games, the atlas identifies 62.9% of damped proximal (PD) failures within the highest-risk 20% of TEST games, compared with 35.0% using response-graph features; the behaviour-supervised E2E representation identifies 71.7%. These results demonstrate that the behavioural geometry can localise solver failure regions from limited target observations, without evaluating the solver across the entire game collection.

Solver-configuration selection. We additionally evaluated the suggested per-game selection task. Under the tested 205-label, 10-neighbour policy, the payoff-based selectors did not outperform the best fixed configuration on TEST. This identifies a limitation of the evaluated selection policy, while the neighbourhood and failure-localisation experiments establish the utility of behavioural cartography for characterising solver behaviour and competence.

Together, these findings support our central contribution: a reusable, behaviour-defined geometry that reveals solver-relevant relationships beyond conventional structural descriptions and can characterise previously excluded solvers from limited observations.

I thank the authors for the discussions and for performing many new experiments. I would like to understand better the scope and the applicability of this approach in general and for the following questions, please not that I am not requesting any additional experiments and would just like a discussion from the authors.

1. Generality of the geometry. Since the geometry is constructed from the behaviour of a particular collection of learning algorithms, how sensitive is it to choices such as learning rates, initialisations, and observation horizons? For instance, games that exhibit similar behaviour under one configuration might behave quite differently under another. Moreover, non-convergent learning dynamics can exhibit qualitatively different behaviours, including cycles, more complicated recurrent behaviour, and even chaos with strong sensitivity to initial conditions. Such phenomena are known to arise in various classes of games and learning dynamics. Given this complexity, in what sense should we expect the learned geometry to capture more general properties of games, rather than properties specific to the chosen dynamical protocols?

2. Computational scalability and utility. The method involves generating solver trajectories, computing behavioural distances, constructing the geometry, and learning to recover it from payoffs. How do the authors envision this approach scaling to substantially larger games or more complex learning dynamics? In particular, given these offline costs and the current results on solver selection, what would be a realistic setting where constructing the atlas offers an advantage over directly evaluating a few candidate solvers or learning a simpler performance predictor?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Summary: New Evidence and Shared Themes

We thank the reviewers for their careful assessments, including QX71’s positive comments on the presentation and empirical evaluation, kDgJ’s recognition of the value of our staged construction, and jtzv’s assessment of the scientific idea. We have now completed several additional evaluations that speak to their shared questions: does the geometry generalise beyond the data used to build it, what does it capture beyond structural game descriptors, and what can it be used for?

1. Held-out TEST validation (new; kDgJ).

With the original protocol frozen before evaluation on 5,371 TEST games, the learned atlas achieves LOSO NMAE 0.694 (DEV: 0.699) and distance Spearman 0.900 (DEV: 0.917). It retains approximately 12% lower prediction error than Ordinary128. The original findings therefore hold on unseen games.

2. Stronger solver exclusion (new; kDgJ).

Jointly excluding related pairs from representation construction gives TEST NMAE 0.536 (GDA/AA), 0.599 (EG/Optimistic) and 0.948 (EMD/PD). Transfer remains useful for two pairs without retaining the companion mechanism, but is uneven. Our claim concerns reusing a game geometry to characterise additional solvers from limited target observations, not unconditional prediction of arbitrary unseen algorithms.

3. Behavioural neighbourhoods beyond structure (new; QX71, jtzv).

In strict LOSO, learned-atlas neighbours show 34–48% less held-out late-gap discrepancy than the strongest evaluated structural baseline for GDA, AA, EG and optimistic gradient, and 11–17% less for EMD, PD and FP. When neighbours must share the query’s TRAIN-defined potentialness quartile, behaviour-supervised representations remove 51–54% of the discrepancy relative to random same-quartile neighbours, versus 17–27% for structural representations (descriptive reductions). This gives direct evidence that finite-time learning behaviour varies meaningfully even among games of similar potentialness.

4. Competence mapping with limited labels (new; QX71).

With 205 labelled TRAIN games, the atlas identifies 62.9% of PD failures among the highest-risk 20% of TEST games, compared with 35.0% using response-graph features; E2E identifies 71.7%. This demonstrates a practical use of behavioural cartography: locating regions where an excluded solver is unreliable, without evaluating that solver across the entire game collection.

5. What the staged representation reveals (kDgJ, QX71, jtzv).

Our construction separates the geometry measured directly from solver behaviour, its compact MDS representation and its recovery from payoffs. This allows us to assess the contribution of each stage. For example, the oracle and learned atlases achieve NMAE 0.727 and 0.907, respectively, on EMD time-averaged exploitability, identifying an opportunity to improve payoff-based recovery. Both the learned atlas and E2E, which also uses behavioural supervision, provide useful cross-solver prediction (TEST LOSO NMAE 0.694 and 0.665, respectively). The contribution is therefore not tied to a single recogniser: it is the explicit behavioural definition of game similarity and the ability to construct, analyse and recover this information. Previously reported cross-size experiments further demonstrate transfer to 4×4 and 5×5 games without target-size behavioural labels.

Summary.

Taken together, the new results establish three findings. First, the principal geometric and transfer results generalise to held-out games. Second, behaviour-defined neighbourhoods identify meaningful differences in finite-time solver behaviour that conventional structural representations miss, including among games with similar potentialness. Third, these neighbourhoods can be used to identify failure-prone regions of an excluded solver from limited observations. The additional diagnostics clarify where transfer is weaker and what information is lost when recovering the geometry from payoffs. We'll incorporate the complete results and protocols into the revised manuscript and welcome further discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.