LeagueBot: Intelligent Agents for Esports

• LeagueBot is a family of intelligent agents for competitive games, utilizing LLMs, supervised learning, and behavioral modeling to address real-time gameplay challenges.
• It delivers multi-modal support including cognitive/emotional assistance, tactical tutoring, match outcome prediction, and champion recommendation in high-pressure environments.
• Empirical studies show measurable improvements in novice performance, prediction accuracy, and user satisfaction through advanced architectures and multi-layered agent systems.

LeagueBot refers to a spectrum of intelligent agents and frameworks—primarily for League of Legends (LoL) and analogous competitive games—spanning applications in cognitive/emotional support, gameplay tutoring, match outcome prediction, champion recommendation, and full-agent autonomous play. Contemporary LeagueBots integrate recent advances from LLMs, structured prompt engineering, supervised learning, collaborative filtering, and behavioral modeling to address the multifaceted demands of high-pressure, information-rich, and complex game environments (Lee et al., 1 Feb 2026, Silva et al., 2017, Ye et al., 2020, Do et al., 2021, Do et al., 2020).

1. Voice LLM Companions: Cognitive and Affective Support

The 2026 LeagueBot system is a desktop-based, real-time voice assistant deploying GPT-4.1 (temperature 0–0.5), integrated with speech-to-text and text-to-speech APIs (e.g., ElevenLabs), and modular prompt engineering specifying persona, environment, and guardrails. The system ingests game state through LoL APIs (scoreboard, objectives, itemization, team compositions) and contextualizes user utterances in live play, generating guidance and emotional support dynamically (Lee et al., 1 Feb 2026).

Quantitative evaluation of this LeagueBot configuration shows statistically significant reductions in cognitive challenge (mean difference –0.68, p = 0.002, d = 0.58), performative challenge (mean difference –0.58, p = 0.003, d = 0.56), and perceived pressure/tension (mean difference –0.47, p = 0.043, d = 0.37) for novice players (N = 33, within-subjects, paired t-tests). No significant impact was observed for decision-making challenge or gameplay enjoyment, and match outcomes did not differ across conditions (p > 0.2).

Qualitative thematic analysis reveals enhanced access to context-sensitive information, reduced information-seeking anxiety, and provision of positive reinforcement. Key limitations include excessive verbosity, inability to deliver non-verbal knowledge (e.g., item icons), occasional LLM hallucinations, and unacceptable response latencies (up to 5s in critical moments). Future design improvements propose multi-modal integration, adaptive granularity, and longitudinal profiling (Lee et al., 1 Feb 2026).

2. Tutoring and Contextual In-Game Guidance

In the plugin-based paradigm, LeagueBot operates as a two-layered in-game agent: a Movement Layer tracks a player (bottom-lane Soraka in the canonical study) and a Skill & Tutorial Layer issues support actions and real-time pedagogical feedback. The decision flow is governed by a behavior tree; context-sensitive rules fire chat messages or pings when thresholds are crossed (e.g., low HP, turret aggro, poor minion last-hitting). This LeagueBot variant delivered substantial KDA factor improvement for tip-enhanced matches and stabilized novice performance. However, scope was limited to single-champion, single-role, and template-based fixed advice (Silva et al., 2017).

3. Match Outcome and Win Probability Prediction

LeagueBot engines perform pre-match and real-time prediction by modeling structured features:

• Post-champion-select prediction: Inputs per player include champion mastery points and champion-specific ranked win rate; team-level summary statistics (mean, median, standard deviation, skewness, kurtosis) are concatenated to yield a 44-dimensional feature vector. A five-layer DenseNet (ELU activations, BatchNorm, dropout, sigmoid output) achieves 75.1% accuracy on held-out test matches. The analysis confirms strong statistical linkage between champion-specific skill and match outcome, even after skill-based matchmaking (Do et al., 2021).
• In-game (real-time) prediction: Leveraging data at four percent-elapsed-time (PET) snapshots (20%, 40%, 60%, 80%), models aggregate kills, first blood, objectives, gold, XP, player levels, and more. Gradient boosted trees (LightGBM, XGBoost), logistic regression, and random forest approaches yield maximum accuracies in the 81–85% range at PET ≥ 60%. Model architectures and deployment flows are tuned for <100 ms end-to-end latency and integrate calibration methods such as Platt scaling for reliable probability outputs. These outcomes generalize to betting, spectating, and feedback applications (Junior et al., 2023).

A confidence-calibrated winner predictor utilizes a four-layer MLP with dual output heads for prediction mean $\mu(x)$ and log variance $s(x)$, realizing Monte Carlo cross-entropy loss and achieving an expected calibration error (ECE) of 0.57% (compared to 1.11% for temperature scaling), thus ensuring probabilistic outputs are trustworthy (Kim et al., 2020).

4. Recommendation Systems and Behavioral Personalization

LeagueBots with champion recommendation capabilities implement collaborative filtering with truncated singular value decomposition (SVD). Ratings $r_{u,i}$ are proportional to Champion Mastery Points, normalized per player. The model factorizes the user-champion matrix $R \approx PQ^T$ (with per-user and per-item biases) and minimizes regularized squared error. The microservice design enables cold-start fallbacks, real-time inference, and UI integration with role and playstyle filters. In user studies, SVD-driven recommendations deliver significantly higher player satisfaction scores than random baselines (mean(system) = 6.46 vs. mean(random) = 5.18, p = 0.01257) (Do et al., 2020).

5. Full Autonomous Agents and Supervised/Hybrid Play

Supervised learning leagues (SL-LeagueBots) for LoL combine vector, image, and local state encodings. Model inputs include concatenated vector features (champion state, cooldowns, gold, deltas over K frames), global minimap and local skill-zone images, all normalized appropriately. The end-to-end network comprises Conv/MLP encoders, macro-intent and micro-action heads, and multi-task losses (cross-entropy, regularization). Macro heads output softmax probabilities over map regions, which are then injected into the micro decision module for refined sequential play. Training utilizes hundreds of thousands of high-Elo replays, per-champion scene segmentation, and rigorous action labeling. Ablations confirm that omitting macro-intent, scene-sampling, or key modalities substantially degrades win rate against top players (Ye et al., 2020).

Complementing supervised pipelines, classical two-layer agent architectures leverage Influence Maps (IMs) for tactical spatial analysis and decision-making (e.g., pathfinding, threat avoidance, tactical target selection). These IM-based LeagueBots, though subpar in high-skill ARAM, demonstrate strong solo performance (92% farming efficiency, perfect survival in simple contexts) (Silva et al., 2017).

6. Statistical Ranking and Tournament Analysis

LeagueBot frameworks for open-ended model ranking (e.g., LLM chatbots) utilize factored-tie and Thurstonian covariance-enhanced paired-comparison models. The Rao–Kupper model with low-rank competitor-specific tie thresholds and low-rank covariance on latent skill enables robust leaderboard inference and tier clustering, even with abundant ties and latent group structure. Parameter identifiability is secured with mean, scale, and translation constraints on $(\mu, \Sigma, G, \Lambda)$. The open-source leaderbot package operationalizes this framework in arena settings, supporting future LeagueBot deployments in both conversational AI and competitive agent domains (Ameli et al., 2024).

7. Meta-Agent League Training and Generalization

The LeagueBot paradigm is broader than LoL, extending to agent populations trained in adversarial or tournament settings (e.g., StarCraft II). Architectures such as TStarBot-X train a league of policies under distributed league management, exploiting roles (main agent, exploiters, evolutionary exploiters), prioritized fictitious self-play (PFSP), rule-guided policy search (RGPS), and divergence-augmented policy optimization (DAPO). This configuration yields superhuman agents (ELO +750 over elite bots) at an order-of-magnitude lower compute cost than previous systems, demonstrating that diverse and stabilized agent leagues are crucial for high-level generalization (Han et al., 2020). A plausible implication is that similarly stratified and regularized LeagueBot frameworks may benefit MOBA agent robustness on adversarial ladders.

---

In summary, LeagueBot, as a term, encompasses a technologically rich, modular family of systems fusing LLM/NLP guidance, behavioral modeling, supervised and self-play agent training, recommendation, prediction, and human-in-the-loop support. These agents are being executed for both player-facing and developer-facing applications across esports, robotics, and AI research, with accumulating evidence for efficacy in facilitating learning, reducing cognitive and affective load, and delivering robust gameplay analytics (Lee et al., 1 Feb 2026, Ye et al., 2020, Silva et al., 2017, Do et al., 2021, Do et al., 2020, Han et al., 2020, Ameli et al., 2024, Junior et al., 2023, Kim et al., 2020).