TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics

Xu, Sheng; Liu, Guiliang; Kharrat, Tarak; Luo, Yudong; Aloulou, Mohamed; Lopez Pena, Javier; Sofeikov, Konstantin; Reid, Adam; Roberts, Paul; Spencer, Steven; Carnall, Joe; McHale, Ian; Schulte, Oliver; Zha, Hongyuan; Zheng, Wei-Shi

TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics

TacticGen reframes football analytics from predicting what may happen to generating what should happen. It models coordinated player and ball movements with a multi-agent diffusion transformer, then guides generation toward tactical objectives expressed as rules, natural language, or learned value functions.

Sheng Xu¹ Guiliang Liu^1† Tarak Kharrat² Yudong Luo¹ Mohamed Aloulou² Javier López Peña²

Konstantin Sofeikov² Adam Reid² Paul Roberts² Steven Spencer³ Joe Carnall³

Ian McHale⁴ Oliver Schulte⁵ Hongyuan Zha¹ Wei-Shi Zheng⁶

¹ The Chinese University of Hong Kong, Shenzhen

² Real Analytics

³ Birmingham City Football Club

⁴ University of Liverpool

⁵ Simon Fraser University

⁶ Sun Yat-sen University

^† Corresponding author.

Paper Code Partner

3.3M+ annotated events for training

100M real tracking frames

1432 matches from 2018-2025

80% of experts favored TacticGen

Highlights

Why this paper matters

TacticGen is positioned as a foundation-oriented model for football tactical design: accurate enough to model realistic movement, adaptable enough to respond to different objectives, and scalable enough to improve with larger models and more data.

From prediction to generation

The paper reframes football analytics from forecasting likely play evolution to generating coordinated tactical alternatives.

Multi-agent football modeling

TacticGen jointly models the ball and 22 players, explicitly handling cooperative and competitive interactions.

Zero-shot flexible tactical guidance

During inference, a single trained generator can be steered by rule functions, natural language prompts, or learned value models.

Validated with practitioners

Case studies with football experts show generated tactics are hard to distinguish from real play and tactically valuable.

Motivation

Why football tactics should be generated, not only predicted

This figure sets up the problem shift behind TacticGen: instead of only forecasting how play may unfold, the system should also produce coordinated tactical alternatives that align with a coach's intent.

TacticGen motivation figure showing the shift from prediction to tactical generation — Motivation Figure
From descriptive football analytics to tactical decision support

The motivation figure introduces the paper's core question: how to move from passive trajectory prediction toward controllable generation that can assist tactical planning.

Framework

The TacticGen pipeline

After the method overview, this figure gives a cleaner system-level view of how context encoding, diffusion generation, and tactical guidance fit together inside the complete framework.

TacticGen framework figure showing the pipeline, conditioning inputs, and guidance modules — Framework Figure
Conditioning, generation, and guidance in one model stack

The framework figure ties the paper together by showing how event context, player-ball history, and guidance signals are fused into a unified generation pipeline for tactical control.

Method

A controllable diffusion framework for football tactics

TacticGen builds on diffusion transformers and adapts them to football with explicit agent interaction modeling, context-aware conditioning, and guidance at sampling time.

Architecture

The core generation pipeline

These three components define the base system: a multi-agent denoising backbone, a context encoder that summarizes match state, and an inference-time objective module that makes the generator steerable.

Backbone

DiT Core

Multi-agent Diffusion Transformer

TacticGen extends DiT into a multi-agent setting so that noisy future trajectories can be denoised while preserving team-level coordination and opponent response.

Joint denoising across players and ball keeps team structure coherent.

Context

Temporal Fusion

Self-attentive context encoding

Past trajectories, ball history, event type, goal difference, and timestamps are fused into context embeddings through MLP-based encoding, self-attention, and adaptive conditioning.

Rich match context is compressed into controllable conditioning signals.

Guidance

Inference Control

Objective-driven generation

Classifier guidance steers samples at inference time, letting one pretrained generator adapt to tactical goals without retraining.

One generator, multiple tactical objectives, no extra training pass required.

On top of the shared architecture, TacticGen exposes multiple control interfaces so the same pretrained model can respond to different tactical design goals.

Guidance

Three ways to steer tactical generation

The guidance layer supports explicit football rules, coach-friendly language prompts, and learned value signals, covering both interpretable control and performance-oriented optimization.

Rule-based guidance

Differentiable tactical rules encode concepts like support width, zone occupation, or pitch control value.

Best for explicit tactical priors and interpretable control.

Language-based guidance

Natural language instructions are turned into executable guidance functions, enabling high-level tactical prompting.

Best for coach-friendly prompts and rapid what-if iteration.

Value-based guidance

A learned value model injects long-term utility into generation, guiding trajectories toward strategically stronger outcomes.

Best for optimizing downstream tactical quality over longer horizons.

The same architecture is deployed in two practical operating modes, letting TacticGen cover both realistic trajectory prediction and more deliberate tactical design workflows.

Operating Modes

Two practical operating modes

Both modes share the same football-native generator, but differ in how much future ball information is available at inference time and, as a result, what kind of coaching workflow they support best.

TacticGen-P Predictive

Short-context trajectory prediction

Predicts future player and ball trajectories from a short observed context window, making it a natural fit for realistic generation.

Uses recent match context without assuming the future ball path is known.
Best suited to realistic future trajectory prediction and model benchmarking.

TacticGen-C Conditional

Ball-conditioned tactical generation

Conditions on a full ball trajectory so the model can coordinate player movement around an anticipated ball path.

Provides stronger control when analysts or coaches already have a target ball path in mind.
Best suited to tactical planning, design exploration, and controlled what-if generation.

Dataset

Large-scale football dataset with aligned event and tracking data

The dataset combines analyst-annotated event logs and full-pitch optical tracking, then standardizes them into event-centric training examples with consistent spatial coordinates, temporal context, and future trajectory targets.

League coverage

1,432 matches across 2018-2025

The dataset spans multiple top-tier competitions, led by the EFL Championship and Premier League.

EFL Championship 829

Premier League 489

MLS, Eredivisie, and others 114

Event data Tracking data Aligned sequences

From raw match feeds to aligned multi-agent decision points

Each sample starts from two synchronized sources: Opta-style play-by-play annotations for on-ball actions, and optical tracking for the ball plus all players. An adapted Needleman-Wunsch alignment process links the two streams, producing structured football events that can directly condition trajectory generation.

25 Hz to 10 Hz tracking is down-sampled for stable sequence modeling

105 x 68 m all positions are mapped onto one standardized pitch

Attack to the right coordinates are flipped accordingly so tactical direction is consistent

Matches 1,432

Top-tier matches used to build the aligned dataset.

Events 3,374,599

Annotated football events after alignment.

Frames 97,760,895

Processed timesteps containing agent positions.

Sequence Setup 10 + 54

10 context frames followed by 54 subsequent frames.

Data schema

What one training example contains

Every event stores match metadata, tactical context, agent states, and the future trajectory target needed for training.

Metadata

Game ID, event ID, episode ID, timestamps, and match bookkeeping.

Match state

Goal difference, possession length, event outcome, and ball-control status.

Action labels

30 unified action types plus the next-event ball destination.

Team indicators

Home vs. away, attacking vs. defending, and episode termination flags.

Context tensors

Past agent positions and features over the fixed context window.

Trajectory tensors

Variable-length future player trajectories with aligned agent features.

Event Visualizer

Explore the football event data distribution

Visible actions 10

Trajectory Visualizer

Inspect spatial density on the normalized pitch

The heatmap is built from 1,692,800 coordinates from a single game, while the overlay dots show a smaller sampled subset for readability.

Low density High density

Results

Accurate, adaptable, and scalable

On the paper's large-scale football dataset, TacticGen improves prediction quality over prior baselines and then uses the same modeling foundation to support controllable tactical generation.

Trajectory Error 0.29 Best ADE reported for TacticGen

Final Position 0.52 Best FDE reported for TacticGen

Joint Accuracy 0.92 Best joint FDE in the comparison table

Miss Rate 10.66% Best joint miss rate in the comparison table

Scaling

Scaling behavior

From 1.74M to 311.50M parameters, up to 600K training steps, and up to 78M training examples, TacticGen shows consistent scaling gains across model size, training time, and data scale.

Scaling curves showing Joint ADE and Joint FDE improving with more training steps across model sizes from S to XXL.

Model Scale Scaling from 1.74M to 311.50M parameters consistently lowers error.

Data Scale Performance continues to improve as data grows, up to full 78M examples.

Human Study

Expert validation

In case studies with five experts, generated trajectories achieved a realism F1 of 0.50 ± 0.07, and guided TacticGen samples received a utility score of 0.81 ± 0.04.

Case study plots showing expert utility ratings concentrated near high scores and rating distributions across five raters.

Realism Experts achieved an average realism classification F1 of 0.50 ± 0.07.

Preference Guided TacticGen was preferred over ground truth in 20 of 25 pairs.

Videos

Main Paper Videos

These clips show how TacticGen compares with real trajectories and how generation changes under different tactical objectives, including rule guidance, pitch-control guidance, language prompts, and value-based guidance.

Ground-Truth Event

Start With the Reference Play

All videos below are based on the same Pass event. The interactive view here shows the ground-truth play first, so viewers can get familiar with the player layout, ball location, and overall structure before comparing generated tactics.

Red markers indicate the attacking team, blue markers indicate the defending team, and green indicates the ball.

Conditional Prediction

TacticGen-C Prediction

This clip shows conditional trajectory prediction for a pass event, where player movement is generated while following the observed ball trajectory.

Rule Guidance

Attacking Rule Guidance

The attacking team is guided by composed tactical rules: support the ball carrier, maintain width, create passing angles, and occupy Zone 14.

Rule Guidance

Defending Rule Guidance

The defending team is guided to press the ball carrier, collapse toward the ball, deepen the block, and reduce passing lanes.

Pitch Control Map

Ground-Truth PCV

This is the reference pitch-control map at the final frame, showing the baseline spatial dominance before any PCV guidance is applied.

Pitch Control Guidance

Attack High PCV Trajectory

This trajectory view shows pitch-control guidance for the attacking team, encouraging runners to move forward and create more attacking space.

Pitch Control Map

Attack High PCV Map

This final-frame pitch-control map shows how attacking control expands into more dangerous offensive areas under PCV guidance.

Pitch Control Guidance

Defense High PCV Trajectory

This trajectory view shows PCV guidance for the defending team, increasing pressure on the carrier and improving central coverage.

Pitch Control Map

Defense High PCV Map

This final-frame pitch-control map highlights stronger defensive coverage, with the defending side reclaiming more key space.

Language Prompt

LLM Prompt: Move Forward

Prompt: "Make the attacking team move forward more aggressively." The generated players respond by pushing forward faster toward the defending goal.

Language Prompt

LLM Prompt: Stretch the Defense

Prompt: "Make the right winger drift into the corner to stretch the defense and open up more space." The right winger follows the instruction by moving wide.

Value Guidance

Attack High Value

The attacking side is guided by a learned value model, leading players to push higher and closer to goal-scoring areas.

Value Guidance

Defense High Value

The defending side is guided by the value model to reduce attacking payoff, producing faster pressure on the ball carrier.

Human Evaluation

Interactive Expert Evaluation

We study TacticGen with two complementary human-facing tasks: a Realism Test that asks whether a clip looks generated or realistic, and a Utility Preference task that asks which clip demonstrates better tactical value. The clips and pairs shown on this webpage are only showcase examples, not the full set of human evaluation samples.

Realism Test

Can You Tell Generated Plays From Real Ones?

We added 10 expert case-study clips on realism. Watch each play, decide whether it looks generated or realistic, and the interface will reveal the correct answer immediately after your choice.

Clip 1 / 10

Clip 01 Clearance

Select your answer

Clip 2 / 10

Clip 02 Pass

Select your answer

Clip 3 / 10

Clip 03 Interception

Select your answer

Clip 4 / 10

Clip 04 Take On

Select your answer

Clip 5 / 10

Clip 05 Clearance

Select your answer

Clip 6 / 10

Clip 06 Save

Select your answer

Clip 7 / 10

Clip 07 Ball Touch

Select your answer

Clip 8 / 10

Clip 08 Pass

Select your answer

Clip 9 / 10

Clip 09 Pass

Select your answer

Clip 10 / 10

Clip 10 Pass

Select your answer

Utility Preference

Which Clip Shows Better Tactical Utility?

These expert case studies compare a Generated clip against a Realistic clip. For each pair, decide whether the left clip, the right clip, or equal shows better utility.

Left Clip

Right Clip

Pair 1 / 6

Pair 01 Attacking Team Pass_2367635_58331

Which clip looks better for the attacking team?

Left Clip

Right Clip

Pair 2 / 6

Pair 02 Attacking Team Pass_2370870_101489

Which clip looks better for the attacking team?

Left Clip

Right Clip

Pair 3 / 6

Pair 03 Attacking Team Pass_2450488_130479

Which clip looks better for the attacking team?

Left Clip

Right Clip

Pair 4 / 6

Pair 04 Defending Team Pass_2367605_27626

Which clip looks better for the defending team?

Left Clip

Right Clip

Pair 5 / 6

Pair 05 Defending Team Pass_2367781_146072

Which clip looks better for the defending team?

Left Clip

Right Clip

Pair 6 / 6

Pair 06 Defending Team Pass_2367824_24038

Which clip looks better for the defending team?

Paper

Paper and Citation

Paper

Access the manuscript PDF and citation information below.

Open Paper PDF Back to Top

BibTeX

@article{xu2026tacticgen,
  title   = {TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics},
  author  = {Xu, Sheng and Liu, Guiliang and Kharrat, Tarak and Luo, Yudong and Aloulou, Mohamed and Lopez Pena, Javier and Sofeikov, Konstantin and Reid, Adam and Roberts, Paul and Spencer, Steven and Carnall, Joe and McHale, Ian and Schulte, Oliver and Zha, Hongyuan and Zheng, Wei-Shi},
  journal = {arXiv preprint arXiv:2604.18210},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.18210}
}

TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics

Why this paper matters

From prediction to generation

Multi-agent football modeling

Zero-shot flexible tactical guidance

Validated with practitioners

Why football tactics should be generated, not only predicted

From descriptive football analytics to tactical decision support

The TacticGen pipeline

Conditioning, generation, and guidance in one model stack

A controllable diffusion framework for football tactics

The core generation pipeline

Multi-agent Diffusion Transformer

Self-attentive context encoding

Objective-driven generation

Three ways to steer tactical generation

Rule-based guidance

Language-based guidance

Value-based guidance

Two practical operating modes

Short-context trajectory prediction

Ball-conditioned tactical generation

Large-scale football dataset with aligned event and tracking data

1,432 matches across 2018-2025

From raw match feeds to aligned multi-agent decision points

What one training example contains

Metadata

Match state

Action labels

Team indicators

Context tensors

Trajectory tensors

Explore the football event data distribution

Inspect spatial density on the normalized pitch

Accurate, adaptable, and scalable

Scaling behavior

Expert validation

Main Paper Videos

Start With the Reference Play

TacticGen-C Prediction

Attacking Rule Guidance

Defending Rule Guidance

Ground-Truth PCV

Attack High PCV Trajectory

Attack High PCV Map

Defense High PCV Trajectory

Defense High PCV Map

LLM Prompt: Move Forward

LLM Prompt: Stretch the Defense

Attack High Value

Defense High Value

Interactive Expert Evaluation

Can You Tell Generated Plays From Real Ones?

Are the post-event player movements generated or realistic?

Which Clip Shows Better Tactical Utility?

Which side demonstrates better tactical utility?

Paper and Citation

Paper

BibTeX