KataGo Study Probes How Go Networks Learn Symmetry

KataGo has published a five-model study that challenges a basic horizon machinelearning assumption about data augmentation. Random rotations and reflections can make outputs consistent, yet the network’s internal concepts need not become uniformly symmetric.

The study examines how KataGo represents Go positions across the board’s eight spatial orientations. Go’s rules remain unchanged under four rotations and four reflected rotations. KataGo trains on randomized orientations, but its tested models do not enforce that symmetry throughout their architectures.

That distinction creates the central conflict. Data augmentation teaches a model what answers should remain stable, while architectural equivariance constrains how the model performs its calculations. KataGo’s results suggest those routes can produce very different internal organizations, even when both lead to superhuman play.

The research also compares convolutional networks with newer transformer models. Both families learn substantial symmetry, but they appear to break and recover it in different places. The transformer attention heads produced the study’s strangest result: some organized into directional families resembling a hexagon.

The work is an interpretability study, not a new KataGo release or a playing-strength benchmark. Its experiments and article were produced largely with AI under detailed human direction. That unusual provenance makes the study both interesting and unusually dependent on reproducibility.

KataGo Measured Symmetry Inside Five Released Networks

The study asks whether symmetry lives only in KataGo’s answers or also inside the representations used to produce them.

KataGo maintainer David Wu published the study on August 1, 2026. He introduced it as a small investigation into whether Go networks learn orientation-independent concepts or memorize related concepts separately for each direction.

The distinction matters because the rules provide a clean experimental setting. A Go position remains strategically identical after a valid rotation or reflection. The coordinates change, but liberties, connections, territory, captures, and legal continuations transform with the board.

KataGo uses stochastic eightfold augmentation during training. Each training batch receives a randomly selected board orientation. The model repeatedly sees equivalent positions from different directions, but ordinary layers remain free to process those directions differently.

The researchers evaluated five released checkpoints. These included two convolutional network sizes, an SGD-trained convolutional control, and two transformer models. The comparison therefore covered architecture, model scale, and one optimizer variation.

The test set contained 100 positions sampled from professional games on standard 19-by-19 boards. Researchers passed transformed versions of those positions through each network. They then compared internal channels and attention heads after accounting for how spatial coordinates should rotate or reflect.

This property is called equivariance. A representation is equivariant when transforming the input produces a corresponding, predictable transformation in the representation. Invariance is stricter because the representation itself remains unchanged.

A Go policy cannot simply be invariant at every spatial output. If the board rotates, the recommended move should rotate with it. Internal feature maps can still represent orientation-independent concepts if their spatial locations transform consistently.

The project’s symmetry study examines this behavior at multiple points inside each model. It measures individual channels in convolutional networks and individual attention heads in transformers.

This is more informative than checking only final policy agreement. Two models can recommend nearly identical moves while using very different internal algorithms. One can reuse a common representation, while another maintains several orientation-specific components that later converge.

The repository includes data, reports, scripts, figures, and checkpoint download instructions. Its documented pipeline analyzes all eight transformations and provides the material needed to regenerate the published figures.

The researchers report that a full transformer analysis reproduced 1,520 stored inner-layer arrays exactly. Deep attention and correlation measurements agreed within a stated tolerance of 0.0000002. No channel or head changed classification during that reproduction check.

Those checks do not independently validate the scientific interpretation. They do establish a stronger audit trail than a chart without code, data, or model identifiers.

The event’s importance comes from what was measured. KataGo did not merely ask whether augmentation improves generalization. It inspected where a trained system shares concepts across orientations and where it preserves directional distinctions.

Why the Horizon MachineLearning Question Extends Beyond Go

KataGo turns an abstract argument about inductive bias into a test with exact, known symmetries.

An inductive bias is a built-in preference that guides which solutions a model learns. Convolution supplies a translation-related bias through shared filters. Standard convolution does not automatically impose equivalent handling of rotations and reflections.

Data augmentation offers a softer approach. Instead of restricting the model’s computations, training presents transformed examples and penalizes inconsistent predictions. The model learns from evidence that orientation should not alter the strategic answer.

The alternative is an explicitly equivariant architecture. Such a model shares parameters across a defined transformation group, constraining features to transform according to known rules.

Cohen and Welling formalized this approach in their group-equivariant networks research. Their group convolutions extend weight sharing to transformations including quarter-turn rotations and reflections.

The advantage is not merely aesthetic. A model does not need to rediscover the same relationship independently for every permitted transformation. This can reduce sample requirements and make behavior more predictable under transformations.

The cost is reduced freedom. Exact symmetry can become the wrong constraint when the data pipeline, boundary conditions, sensors, or task contain meaningful asymmetries. A model may benefit from breaking the rule temporarily.

Go makes that tradeoff unusually visible. Its abstract rules respect the board’s dihedral symmetry, which consists of four rotations combined with reflections. Yet useful local calculations can still be directional.

A network examining a shape may distinguish north from east during an intermediate operation. A later layer can combine those directional features into an orientation-independent judgment. Symmetric outputs therefore do not require every internal operation to be symmetric.

That observation pressures two simple stories.

The first says augmentation should make a conventional model effectively symmetric everywhere. KataGo’s internal measurements do not support such a uniform conclusion.

The second says symmetry must be hard-coded to obtain reusable symmetric concepts. The networks reportedly learn substantial internal equivariance without that architectural guarantee.

The useful question lies between those claims. Researchers need to know which symmetry emerges, which symmetry fragments into feature families, and where the fragments recombine.

This horizon machinelearning issue applies wherever transformations preserve meaning. Medical images, astronomical surveys, satellite scenes, molecules, robotics, and physical simulations all contain geometric structure.

In those fields, developers often choose between larger augmented datasets and specialized equivariant layers. KataGo suggests the choice affects internal organization, not only benchmark accuracy.

That can influence reliability. A system that learns eight related shortcuts might agree on common inputs while failing differently outside the training distribution. A system with shared representations has fewer independent ways to drift.

The study does not establish that KataGo’s more symmetric features are safer. It provides measurements that make the hypothesis testable.

It also gives interpretability researchers a controlled environment. Language models contain uncertain semantic relationships and disputed evaluation labels. Go supplies exact transformations and an objective mapping between corresponding board points.

That clarity helps separate two difficult questions. Researchers can first test whether a representation respects a known symmetry. They can then investigate what strategic concept that representation carries.

Convnets and Transformers Break Symmetry Differently

The central result is not that one architecture learned symmetry and the other failed, but that they placed asymmetry in different internal components.

KataGo’s convolutional models reportedly developed many channels that behaved equivariantly across transformed positions. A channel associated with a strategic pattern could follow that pattern as the board rotated or reflected.

However, the convolutional networks were not uniformly symmetric. Some features remained orientation-sensitive. The larger convolutional model also displayed what the project describes as a mild “asymmetry haze,” rather than a simple split between perfect and broken channels.

That haze matters because learned symmetry is not necessarily binary. A feature can behave consistently on most positions while deviating slightly on others. It can also combine a dominant symmetric signal with a smaller directional component.

The SGD-trained control helps test whether this organization belongs only to a particular optimization path. Optimizers influence which solution a network reaches among many solutions with similar training loss.

The study includes that control, but it does not turn optimizer choice into the main explanation. Architecture produced the sharper reported difference.

Transformers distribute spatial computation through attention. An attention head assigns weights between board locations, allowing one location to collect information from other positions.

The KataGo transformers learned attention heads with distinct symmetry behavior. Some heads remained closely aligned under transformations. Others appeared in groups whose members corresponded to different directions.

That is a more structured outcome than eight unrelated memorized solutions. If rotating an input maps one head’s behavior onto another head, the group acts as an orbit. The individual heads are directional, while the family collectively respects the board’s symmetry.

This arrangement resembles a team of specialists. One component handles a relation in one direction, another handles its rotated counterpart, and later computation can combine them.

The study repository reports a sharp convolutional-versus-transformer difference in where symmetry becomes broken. It also documents orbit analysis for non-equivariant features.

One corrected result deserves attention. An early analysis contained a sign-cancellation error that suggested certain sign-paired channels lacked rotated counterparts. The corrected calculation found a closed, sign-carrying four-member orbit.

That correction strengthens the project’s transparency while weakening any claim that the workflow was error-free. The current repository identifies the flawed report and directs readers to the corrected version.

The distinction between a truly isolated feature and an orbit member is substantial. An isolated feature suggests separate learning without a recognizable transformation rule. An orbit suggests the model encoded symmetry across multiple components.

Sign changes add another layer. A transformed feature can represent the same structure while reversing its activation sign. Downstream weights may ignore that sign or use paired components to reconstruct the original relationship.

This is why simple channel correlation can mislead. Researchers must test permutations, signs, and spatial transformations. Otherwise, an organized equivariant representation can look inconsistent.

The finding also complicates the horizon machinelearning debate over “shared concepts.” Sharing does not require a single neuron or channel to mean the same thing in every orientation.

A concept can be distributed across a structured collection of units. The representation remains reusable because transformations map the collection onto itself.

That interpretation aligns with broader work on equivariant networks, but KataGo did not use a formally equivariant transformer architecture. The organization emerged from training signals and the model’s available computational structure.

This does not show that transformers naturally discover every relevant group. The experiment concerns five KataGo checkpoints, one game, and a finite set of transformations.

It does show that internal asymmetry can be organized rather than arbitrary. That is the study’s most consequential reversal.

The Hexagonal Attention Pattern Is the Strangest Clue

Several transformer heads reportedly arranged into directional structures resembling a hexagon, despite Go’s square grid and eightfold board symmetry.

At first glance, a hexagon seems misplaced on a square Go board. The board supports horizontal, vertical, and diagonal relationships, while its global symmetry involves quarter turns and reflections.

The study’s figures show attention heads with preferred offset patterns. These patterns describe where a head tends to look relative to a queried board location.

Across some layers, related heads occupied directional roles that formed a six-part organization. The authors describe these as hexagonally arranged attention heads.

This does not mean the transformer discovered a literal hexagonal Go board. It indicates that the learned directional basis did not mirror the most obvious human decomposition of the task.

Neural networks can choose internal coordinates that efficiently support downstream computation without matching a human label set. A basis can be useful even when its individual components resist direct semantic names.

The hexagonal pattern therefore functions as a clue, not a decoded algorithm. It suggests that attention heads coordinate through geometry more structured than independent feature learning.

The researchers examined individual positions, average offset profiles, ladders, and ko-related examples. A ladder is a forcing capture sequence that travels diagonally across the board. Ko is a repeating local fight restricted by rules against immediate board repetition.

Both situations require relationships across distance. They provide plausible tests for directional attention because local stones can depend on faraway continuations or recapture constraints.

The repository includes a per-head ladder ablation. Ablation removes or suppresses a component to observe how outputs change. Such tests move beyond visual resemblance, but they still require careful interpretation.

If removing one head changes a ladder evaluation, that does not prove the head exclusively implements ladder reading. Attention heads interact, and a damaged representation can affect several tasks.

Likewise, a head that looks along a direction may carry content unrelated to geometric direction. Attention weights show communication patterns, not the complete information transmitted through them.

This limitation is familiar across mechanistic interpretability. A clear visualization can invite a stronger semantic story than the evidence supports.

The KataGo work handles that risk by presenting correlations, orbit structure, and selected interventions. Its conclusions remain narrower than a complete reverse-engineering of play.

Still, the pattern is valuable because unexpected regularity can generate new experiments. Researchers can test whether the six-head arrangement appears across training seeds, model sizes, or other board games.

They can also train a transformer with explicit dihedral equivariance. Comparing its heads with KataGo’s learned structure would reveal whether the hexagonal basis compensates for a missing architectural constraint.

Another test would remove augmentation after a particular training stage. If the structure persists, symmetry may become self-reinforcing after the model establishes an internal algorithm.

A third test would alter the board shape. KataGo supports rectangular and nonstandard boards, but the current study focuses on professional 19-by-19 positions. Rectangular boards remove some square symmetries while retaining reflection and half-turn relations.

If the same six-part organization survived that change, it would argue against a direct encoding of the board’s symmetry group. If it reorganized, the geometric interpretation would gain support.

The attention result is therefore interesting because it opens a tractable research program. It is not evidence that researchers now understand KataGo’s complete reasoning process.

Reproducible Code Does Not Remove the Verification Gap

The study is unusually inspectable, but its AI-heavy production process demands more independent replication, not less.

Wu says AI produced almost all experiments, implementation, and writing under significant human oversight. The public repository identifies Claude as the primary generation system.

That disclosure changes how readers should evaluate the work. It does not automatically discredit the findings. It does raise the importance of independent code review, fresh experiment design, and replication outside the original workflow.

AI-generated analysis can produce internally consistent errors. The same mistaken assumption can enter the code, documentation, figures, and prose, creating the appearance of agreement.

The sign-cancellation correction is a concrete example. The initial pipeline produced a false conclusion about rotated counterparts. An adversarial review found the problem, and the corrected analysis changed the interpretation.

The project reports five independent reviewers recomputed article values from raw data. It also reports end-to-end reproduction against released checkpoints and a pure Python input featurizer.

Those are meaningful safeguards. However, the reviewers appear to be part of the project’s AI-assisted process, rather than independent research groups. The claim should therefore be read as an internal adversarial audit.

The sample also limits generalization. One hundred professional positions can reveal strong recurring structure, but they do not represent every legal Go state.

Professional games concentrate on strategically plausible positions. Unusual board states, adversarial arrangements, small boards, large boards, and artificial symmetric positions may produce different internal behavior.

Model selection creates another boundary. Five checkpoints support architecture comparisons, yet they cannot isolate every factor separating convolutional networks and transformers.

Training histories, parameter counts, optimization schedules, feature pipelines, and checkpoint maturity can all affect representation. A controlled experiment would train matched architectures across repeated random seeds.

The equivariance score itself also embodies analytical choices. Researchers must select normalization, matching thresholds, spatial alignment rules, and methods for handling signs or permutations.

A different metric could split borderline channels differently. The reported “asymmetry haze” makes that sensitivity especially relevant.

Playing strength provides another complication. KataGo’s public project describes its models as among the strongest open-source Go systems, but strength does not guarantee broad internal reliability.

Prior researchers trained adversarial policies that achieved a reported win rate above 97 percent against KataGo at superhuman settings. The attacks exploited unusual cyclic structures rather than ordinary professional play.

That history does not contradict the symmetry findings. It shows why clean average behavior should not be confused with comprehensive understanding.

A model can learn elegant transformation structure for common positions and still fail on rare arrangements. Internal regularity and adversarial robustness are related questions, not interchangeable ones.

Research on approximate equivariance offers another warning. A 2023 equivariance analysis found that small architectural details can break exact symmetry and reduce generalization to unseen transformations.

That work also found cases where relaxing exact equivariance helped when the data’s symmetry did not perfectly match the network’s assumptions. The broader lesson is that symmetry constraints involve measurable tradeoffs.

For KataGo, all eight tested transformations are valid under standard square-board rules. Yet input features, board boundaries, history encoding, and downstream search still deserve individual checks.

The current study should therefore be treated as a reproducible research claim. It offers code and falsifiable measurements, but independent researchers have not yet established how broadly its conclusions transfer.

That framing is more useful than either dismissing the work as AI-generated or accepting its visual patterns at face value.

What Researchers Should Watch Next

Three follow-up signals will show whether KataGo found a general property of learned representations or a checkpoint-specific curiosity.

The first signal is independent replication across training seeds. Another group should rerun the measurements on separately trained convolutional and transformer models with matched settings.

Repeated orbit structures would strengthen the claim that architecture shapes where symmetry breaks. Large variation between seeds would suggest the reported structures are only one solution among many.

The second signal is a controlled comparison with explicit equivariance. Researchers can train a group-equivariant Go network and compare its efficiency, strength, calibration, and internal organization with augmentation-only models.

That experiment would test the main opponent directly. It could show whether hard-coded symmetry eliminates redundant learning or removes useful directional computation.

The comparison should measure more than final Elo. Training sample efficiency, inference cost, transformed-position consistency, and performance on adversarial boards would expose the real tradeoff.

The third signal is causal intervention on the discovered feature orbits. Researchers should perturb complete directional families rather than deleting one attention head at a time.

If removing an entire orbit damages transformed versions of a task consistently, the geometric interpretation becomes stronger. If effects remain scattered, the orbit may describe appearance without capturing function.

Ladder and ko tests offer useful starting points, but future work should include pre-registered position sets. This would reduce the risk of selecting examples after inspecting attention patterns.

Researchers should also test rectangular boards and intentionally asymmetric rules. Those settings can reveal whether a learned feature tracks genuine task symmetry or merely reflects the augmentation pipeline.

For developers, the practical lesson is measured. Data augmentation can encourage significant internal structure, even when architecture does not enforce that structure.

However, consistent outputs do not tell developers how a model organized the problem. They need internal measurements, transformation tests, and adversarial evaluation when orientation consistency affects safety or reliability.

For interpretability teams, KataGo offers a useful model of research packaging. The project connects an accessible article with scripts, measurement data, raw reports, corrections, and reproduction instructions.

Teams evaluating AI-assisted research can borrow that pattern. Keeping searchable local reports in an engineering knowledge base can help reviewers trace claims back to experiments and revisions.

The broader horizon machinelearning question remains open. Neural networks can learn symmetry without receiving a rigid architectural rule, but they do not necessarily express it through uniformly symmetric components.

KataGo’s most interesting result is the middle ground. Its networks appear to build reusable symmetric concepts alongside organized directional specialists.

That makes the next step clear. Researchers should stop asking only whether a model is symmetric and start testing where, how, and under which inputs that symmetry holds.

Will independent replications recover the same feature orbits and hexagonal attention families? The answer will determine whether this is a compact KataGo case study or a reusable map for interpreting learned geometry.