The Frontier

TacticAI — Google DeepMind with Liverpool FC, published in Nature Communications on 19 March 2024 — treats the corner snapshot as a fully connected graph. Each player is a node carrying position, velocity, height, weight, a team flag and a kicker flag; GATv2 attention passes messages between all pairs, and predicting the receiver becomes node classification. Across 7,176 Premier League corners, it names the receiver in its top three 78.2% of the time.

The elegant part is the symmetry. A corner does not care which way the pitch is drawn, so TacticAI enforces D₂ group equivariance: the model averages its answer over the identity and the three reflections (flip_x, flip_y, flip_xy) — a free 4× augmentation the authors report was worth several accuracy points.

And it does not stop at prediction. TacticAI generatively adjusts player positions to raise or lower the probability that a shot follows — a tactics assistant, not a scoreboard. In blind evaluations, Liverpool's own experts preferred its suggested setups over the real, historical ones 90% of the time.

Large Events Models (Mendes-Neves et al., Machine Learning, 2024, with companion preprints) tokenize a match. An event is a tuple e_k = (t_k, x_k, y_k, a_k, Δτ_k) — type, location, outcome, inter-event time — and the model is the plain autoregressive factorization:

P(E) = ∏ₖ P(eₖ | e₁, …, eₖ₋₁)
P(eₖ | ·) = P(tₖ | hₖ) · P(xₖ, yₖ | tₖ, hₖ) · P(aₖ | tₖ, xₖ, yₖ, hₖ)

where h_k is a transformer encoding of the history. The event-type vocabulary runs to roughly 20–40 tokens, the pitch is quantized to a zone grid, and training minimizes summed cross-entropy — exactly a language-model loss. Everything downstream is Monte Carlo: win probability is the fraction of N sampled continuations that end in a win; a generalized expected threat is V(s) = E[goal within horizon | s] from rollouts, replacing Karun Singh's 2019 fixed-grid recursion outright. The family was refined into a single unified tabular autoregressive model at ECML-PKDD 2025.

The same move works one level up. RisingBALLER (Adjileye, Hudl StatsBomb Conference, 11 October 2024) makes the player, not the event, the token: BERT-style masked-player pretraining over match sheets yields player embeddings that transfer to scouting and next-match performance prediction — the clearest public blueprint for a football players foundation model.

StatsBomb 360 (launched 2021) attaches a freeze frame to every event: the positions of all camera-visible players at the moment of the action. Crucially for this lab, 360 frames sit in the free open-data repository for Euro 2020, the 2022 World Cup, Women's Euro 2022, the 2023 Women's World Cup and Euro 2024. Hudl acquired StatsBomb (reported June 2024 by The Athletic) and relaunched the company as Hudl Statsbomb at its October 2024 conference — the same event where RisingBALLER was presented.

Opta Vision (Stats Perform; announced June 2022, rolled out at scale across 2024–25) goes the other way: an AI-enriched fusion of tracking and event data across 80+ competitions, yielding off-ball runs, pressure and pass prediction as standard feeds. Stats Perform now markets 140+ AI models and 8 domain-specific foundation models.

Language models arrived through the front office first. Sevilla FC and IBM unveiled Scout Advisor in January 2024 — watsonx plus Llama doing retrieval and generation over 200,000+ human scouting reports — and Twelve Football's TwelveGPT (David Sumpter et al., 2024–25) turns match and recruitment data into written analysis. Neither replaces a scout; both compress the reading pile.

The lineage starts with DeepMind's Graph Imputer (Omidshafiei et al., Scientific Reports, 2022), which predicted the positions of players the broadcast camera could not see — generative modelling as gap-filling. GenTac (Rao et al., arXiv, 13 April 2026) completes the turn: diffusion over continuous multi-player trajectories coupled to 15 discrete tactical events, producing counterfactual, style-conditioned rollouts — it can be asked to play a sequence the way a specific team would. Alongside it sit SMGDiff (ICCV 2025, soccer motion diffusion) and Diffoot (graph-conditioned trajectory diffusion).

Video caught up in parallel. MatchVision / UniSoccer (Rao et al., CVPR 2025) is the first vision-language foundation model for soccer, trained on the SoccerReplay-1988 corpus — commentary, classification and retrieval from raw broadcast footage.

The honest caveat: evaluation is the frontier's soft spot. The LEM papers' own protocol is to compare simulated against real distributions — possession length, shots per possession, zone occupancy — by Kolmogorov–Smirnov distance. "Looks like football, statistically" is a weaker guarantee than a calibration curve, and the field knows it.