TacEleven: generative tactic discovery for football open play

TL;DR

TacEleven introduces a generator-critic framework for open-play football tactic discovery, coupling a language-controlled tactical generator (LTG) with a multimodal tactical critic (MTC) to produce and select long-horizon open-play tactics. By decomposing tactics into sequences of text-to-trajectory pairs and employing a tactical tree search, TacEleven enables counterfactual exploration and autoregressive multi-step planning, validated across CF, SS, and MS tasks and through expert questionnaires. Quantitative metrics (xG, xT, PC, FTE, CAE) and qualitative expert assessments show that TacEleven generates realistic, effective, and adoptable tactics, with over half of multi-step proposals rated adoptable in real-world elite scenarios. The framework leverages precise data alignment, a variational spatiotemporal model, and language-based control to bridge data-driven analysis and coaching practice, signaling potential applicability to other long-horizon, language-guided decision tasks.

Abstract

Creating offensive advantages during open play is fundamental to football success. However, due to the highly dynamic and long-sequence nature of open play, the potential tactic space grows exponentially as the sequence progresses, making automated tactic discovery extremely challenging. To address this, we propose TacEleven, a generative framework for football open-play tactic discovery developed in close collaboration with domain experts from AJ Auxerre, designed to assist coaches and analysts in tactical decision-making. TacEleven consists of two core components: a language-controlled tactical generator that produces diverse tactical proposals, and a multimodal large language model-based tactical critic that selects the optimal proposal aligned with a high-level stylistic tactical instruction. The two components enables rapid exploration of tactical proposals and discovery of alternative open-play offensive tactics. We evaluate TacEleven across three tasks with progressive tactical complexity: counterfactual exploration, single-step discovery, and multi-step discovery, through both quantitative metrics and a questionnaire-based qualitative assessment. The results show that the TacEleven-discovered tactics exhibit strong realism and tactical creativity, with 52.50% of the multi-step tactical alternatives rated adoptable in real-world elite football scenarios, highlighting the framework's ability to rapidly generate numerous high-quality tactics for complex long-sequence open-play situations. TacEleven demonstrates the potential of creatively leveraging domain data and generative models to advance tactical analysis in sports.

Figures (10)

• Figure 1: An overview of TacEleven.(A), An intuitive example of LTG. The model takes an event description and a historical spatiotemporal graph as input and outputs a future spatiotemporal graph; both graphs are visualized as tactical sketches. In the tactical sketches, the attacking direction is from left to right, with red representing teammates, blue representing opponents, and yellow representing the ball. Trajectories are illustrated from light to dark, culminating at the circles. Players performing events are highlighted with yellow hexagons and their corresponding targets are highlighted with yellow crosses. In this example, the output of the generator shows that Verratti passes to Neymar. Note that minor abrupt bends in the trajectories are attributable to noise in the data.(B), The generator–critic framework integrates three tasks with progressive tactical complexity: counterfactual exploration, single-step discovery and multi-step discovery. In counterfactual exploration, the generator use counterfactual tactical search to produce tactical proposals aligned with high-level event descriptions. In single-step discovery, the critic selects the most effective proposals with a high-level instructions. Multi-step discovery selects proposals iteratively, with the generator producing successive steps in an autoregressive manner. Together, these tasks create a counterfactual tactical tree search, enabling long-sequence tactic generation.
• Figure 2: Different counterfactual full-filed trajectories generated by LTG. (a) shows the historical trajectory used as part of the LTG input. (b)–(l) present the generated tactical proposals given different event descriptions under the same historical trajectory. The subcaptions of (b) to (l) indicate either the intended pass target or the player carrying the ball for each event.
• Figure 3: Generalizability of different scaling models and datasets.Factual Trajectory Error (m) and Counterfactual Alignment Error (m) are shown on the horizontal and vertical axes, respectively. Solid lines connect models trained on datasets of different sizes, color-coded from blue (108K) to green (1076K). Marker size indicates model scale, ranging from 5M to 1,373M parameters. Dashed grey lines connect models of identical size, highlighting the influence of data scale under fixed model capacity.
• Figure 4: Historical / Factual / Discovered trajectories with explanations below each scenario. In Scenario 1, the result illustrates how MTC identifies a pass from Verratti to Hakimi as the optimal choice, balancing the tactical continuity and spatial exploitation. In Scenario 2, MTC selects a pass to Neymar considering his creative style. In Scenario 3, within the middle-third build-up, MTC favors a pass to Mbappe considering his speed and positioning.
• Figure 5: Multi-step counterfactual exploration with an instruction: "Design a football tactic that focuses on collaboration among midfield players to execute localized passing sequences, break through the opposing defense, gain offensive advantages, and create potential breakthroughs and attacks. The objective is to create numerical superiority in central areas, enabling the team to outplay opponents in tight spaces and progress the ball forward."