Uploaded June 2026 | Updated September 2026, 2 weeks ago
00:01 LLM Game Solvers for Wordle & Fibble w/Tin Vuong
18:48 Puzzle-Map Generation Using WFC w/Pranav Pradhan
30:41 Effects of Randomized Resources on Trading Card Games w/Lisa Liu
42:38 An LLM Agent for Playing and Explaining Chess Moves w/Owen Salyer
54:32 Q&A
1:08:44 Dr. Julian Togelius on Procedural Map Generation
From Transformer: AI and the Future of Games 2026
LLM Game Solvers for Wordle & Fibble w/Tin Vuong
Large language models (LLMs) have demonstrated remarkable proficiency in natural language generation and comprehension. However, their reasoning capabilities, particularly in structured, positional tasks like word games, remain a critical bottleneck. Our Wordle-Fibble-LLM-Solver Competition was designed to probe these limitations. This event challenges participants to develop LLM-based solvers that excel at solving Wordle—where players deduce a five-letter word through color-coded feedback on letter positions—and Fibble, a variant that introduced lies in each line of feedback. By forcing LLMs to reason about letters and their positions in words, the competition underscores LLMs' vulnerability to tokenization-induced errors. Participants must innovate prompting strategies, chain-of-thought mechanisms, re-emphasis of key facts, or hybrid architectures to mitigate these issues, evaluated via metrics like average guesses-to-solution and win rates.
Puzzle-Map Generation Using WFC w/Pranav Pradhan
This paper presents a general approach for procedurally generating tile-based maps and embeds solver-checked, Sokoban-like pushblock puzzles (presented as "Strength" puzzles in Pokémon). Our pipeline drives Wave Function Collapse (WFC), a constraint-based tile synthesis method, to lay out traversable areas and reserve puzzle rooms. Although we demonstrate on Pokémon FireRed, a 2004 2D, top-down, tile-based roleplaying game—the approach is broadly applicable beyond this title. We compare two ways to integrate puzzles within a single generated map: required gates along the main route; and optional side rooms that trade time for rewards. We evaluate our method using a qualitative user study and A* based analyses of path optimality and puzzle complexity. Results show that puzzle-aware maps are judged as more engaging and novel than unmodified WFC baselines while maintaining solvability and giving designers a simple, controllable way to shape progression and difficulty on generated maps.
Effects of Randomized Resources on Trading Card Games w/Lisa Li
In this study, we create two simple learning environments that isolate the effects of randomized resources. Experimental results show that even in simple environments with the complexity of tic-tac-toe, the addition of randomized resources significantly weakens the performance of the canonical reinforcement learning algorithms.
An LLM Agent for Playing and Explaining Chess Moves w/Owen Salyer
This work investigates how far a relatively small, open-source large language model (LLM) can be pushed toward "thinking like a chess player" when given only board positions and natural-language rationales. Building on the MATE chess testbed, we construct an end-to-end pipeline that samples positions from human games, labels them with Stockfish, and automatically generates paired strategic and tactical explanations. Each example consists of a position in Forsyth–Edwards Notation (FEN) and a target JSON response containing a best move in Universal Chess Interface (UCI) format plus one-sentence strategy and tactic descriptions. Using this data, we fine-tune a 4B parameter instruction-tuned model with QLoRA, enabling training on modest consumer hardware while retaining the full base model. We then perform reinforcement learning (RL) tasks in an attempt to maximize the likelihood that the fine-tuned model can select a legal move given a particular chess position. We evaluate the resulting "chess reasoner" on held-out positions by measuring top-1 move accuracy against Stockfish recommendations and the quality and usefulness of generated explanations, via both automatic heuristics and small-scale human judgments. Comparisons include the base model before fine tuning, stronger general-purpose LLMs queried via API, and ablations that remove strategy or tactic text from training. Overall, the project seeks to demonstrate that a compact LLM, paired with an engine-driven data pipeline, can learn to propose reasonable moves and articulate human-readable plans and tactics, which would suggest a promising direction for lightweight, explainable chess tutoring and for studying reasoning in constrained, fully observable domains.
For more great game dev videos like this, please support our Patreon at patreon.com/GeorgiaGameDevs
00:01 LLM Game Solvers for Wordle & Fibble w/Tin Vuong
18:48 Puzzle-Map Generation Using WFC w/Pranav Pradhan
30:41 Effects of Randomized Resources on Trading Card Games w/Lisa Liu
42:38 An LLM Agent for Playing and Explaining Chess Moves w/Owen Salyer
54:32 Q&A
1:08:44 Dr. Julian Togelius on Procedural Map Generation
From Transformer: AI and the Future of Games 2026
LLM Game Solvers for Wordle & Fibble w/Tin Vuong
Large language models (LLMs) have demonstrated remarkable proficiency in natural language generation and comprehension. However, their reasoning capabilities, particularly in structured, positional tasks like word games, remain a critical bottleneck. Our Wordle-Fibble-LLM-Solver Competition was designed to probe these limitations. This event challenges participants to develop LLM-based solvers that excel at solving Wordle—where players deduce a five-letter word through color-coded feedback on letter positions—and Fibble, a variant that introduced lies in each line of feedback. By forcing LLMs to reason about letters and their positions in words, the competition underscores LLMs' vulnerability to tokenization-induced errors. Participants must innovate prompting strategies, chain-of-thought mechanisms, re-emphasis of key facts, or hybrid architectures to mitigate these issues, evaluated via metrics like average guesses-to-solution and win rates.
Puzzle-Map Generation Using WFC w/Pranav Pradhan
This paper presents a general approach for procedurally generating tile-based maps and embeds solver-checked, Sokoban-like pushblock puzzles (presented as "Strength" puzzles in Pokémon). Our pipeline drives Wave Function Collapse (WFC), a constraint-based tile synthesis method, to lay out traversable areas and reserve puzzle rooms. Although we demonstrate on Pokémon FireRed, a 2004 2D, top-down, tile-based roleplaying game—the approach is broadly applicable beyond this title. We compare two ways to integrate puzzles within a single generated map: required gates along the main route; and optional side rooms that trade time for rewards. We evaluate our method using a qualitative user study and A* based analyses of path optimality and puzzle complexity. Results show that puzzle-aware maps are judged as more engaging and novel than unmodified WFC baselines while maintaining solvability and giving designers a simple, controllable way to shape progression and difficulty on generated maps.
Effects of Randomized Resources on Trading Card Games w/Lisa Li
In this study, we create two simple learning environments that isolate the effects of randomized resources. Experimental results show that even in simple environments with the complexity of tic-tac-toe, the addition of randomized resources significantly weakens the performance of the canonical reinforcement learning algorithms.
An LLM Agent for Playing and Explaining Chess Moves w/Owen Salyer
This work investigates how far a relatively small, open-source large language model (LLM) can be pushed toward "thinking like a chess player" when given only board positions and natural-language rationales. Building on the MATE chess testbed, we construct an end-to-end pipeline that samples positions from human games, labels them with Stockfish, and automatically generates paired strategic and tactical explanations. Each example consists of a position in Forsyth–Edwards Notation (FEN) and a target JSON response containing a best move in Universal Chess Interface (UCI) format plus one-sentence strategy and tactic descriptions. Using this data, we fine-tune a 4B parameter instruction-tuned model with QLoRA, enabling training on modest consumer hardware while retaining the full base model. We then perform reinforcement learning (RL) tasks in an attempt to maximize the likelihood that the fine-tuned model can select a legal move given a particular chess position. We evaluate the resulting "chess reasoner" on held-out positions by measuring top-1 move accuracy against Stockfish recommendations and the quality and usefulness of generated explanations, via both automatic heuristics and small-scale human judgments. Comparisons include the base model before fine tuning, stronger general-purpose LLMs queried via API, and ablations that remove strategy or tactic text from training. Overall, the project seeks to demonstrate that a compact LLM, paired with an engine-driven data pipeline, can learn to propose reasonable moves and articulate human-readable plans and tactics, which would suggest a promising direction for lightweight, explainable chess tutoring and for studying reasoning in constrained, fully observable domains.
For more great game dev videos like this, please support our Patreon at patreon.com/GeorgiaGameDevs
