iiinova — where your digital self keeps evolving

Matthew Guzdial @matthewguz.bsky.social · 28/08/2026

It's fine! She is so beautiful and dry now

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

#FDG26 closing session had the award for solving the puzzle introduced during the PCG workshop (a white polo, appropriately) and a recognition of the chairs and organizers. bsky.app/profile/matt...

110

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Erin J.K. Truesdell ends off the session with a presentation on semi-structured interviews with MAGFest alt control game designers, finding themes around alt control games as art, practice, community, and futures.

110

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Miikka Junnila presents his own analysis comparing Red Dead Redemption 2 and Ghost of Tsushima in terms of their shared actions.

Breakdown of shared action usage based on the analysis

110

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Craig Anderson presented the results of semi-structured interviews with Thinky Games puzzle designers, lots of good insights here. I particularly liked the reiteration of the importance of designing for interesting situations and that the only failure is players permanently giving up.

120

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Panagiotis Vrettis presents a design overview of Behaviour Rating Systems in games, identifying a number of features that define a space of BRSs.

100

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

After the community meeting, we have the Game Design 2 session, starting with Ea Willumsen covering “The Money Game” exhibition in the National Museum of Denmark. She found it communicated their intended lessons around how money works, but that ~1/3 of players didn’t even play, acting as audience.

Money Game exhibition photographed with people playing a who wants to be a millionaire like quiz show

110

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Community meeting starts with a thank you to Jim Whitehead for being the central pillar of FDG for 21 years. (Plus bonus @annetropy.bsky.social)

Images of Jim with a bonus Anne Sullivan

130

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Kaijie Xu back (remotely) again covering a generic approach to navigation in Soulslike games, extending some prior work on identifying goal locations in screenshots. They could mostly navigate along a designer-desired path!

Example of results, green is achieving a location, red is not. The approach could get most of the way for most games.

130

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

My student @harcurio.bsky.social presents his approach to generating enemy hitboxes that can be defeated by only one set of mechanics (like only with a double jump) based on collision information. Paper: dl.acm.org/doi/full/10....

General setup, given a randomly initialized hitbox optimize it to only be defeatable with a new mechanic

Example outputs for a double jump mechanic

120

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Next Kaijie Xu (remotely) presents his Honourable Mention paper on generating levels that have two states like the time jumping levels in Dishonoured 2 or Titanfall 2 using planning and a GA.

230

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

My student @emilyhalina.bsky.social presents Playtrace Reconstructive Partitioning (PRP), our new approach to generating guaranteed solvable levels from as little as a single input by considering time. Paper: dl.acm.org/doi/full/10....

Cake representation, a level represented over time

Example outputs from a single Sokoban input

151

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Lindsay Grace starts off the Game AI - Learning session to highlight a play as an interface for interactions with AI, including reminding/proposing new verbs of interaction.

Verbs: slide, cluster, spin, associate, mix, spread, and mingle

110

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Final keynote is from Tim Garbos covering the history of Tribad studio, a great overview of lessons learned. Biggest one is just the classic: make a lot, cut to the good stuff and repeat.

Cactus diagram around prototyping and cutting

120

Matthew Guzdial @matthewguz.bsky.social · 13/08/2026

Day 3 (final) of #FDG26 starts with the announcement of #FDG27 in Melbourne | Naarm, Australia CfP and call for organizers is live

QR codes for call for papers and organizers and track chairs for FDG 2027

293

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Trapped on a bus for an hour in Copenhage? prime quiz time

060

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Christian Poglitsch (remotely) presents an experiment applying various OpenAI LLMs with various prompting strategies to playing Hearthstone. They found that they did not outperform SOTA MCTS or RL approaches.

210

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Marco Scirea presents results on a user study comparing different ways to use LLMs for NPC dialogue. Lots of interesting statistically significant results, but this framework breaks it down nicely. More LLM is C, more hand authored on the left.

Simple breakdown/synthesis of the results

110

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Vikram Kumaran presents on generating quests with a mix of LLMs and constraint solvers. They found the CS and LLM critics helped but that humans didn’t particularly care for the output.

130

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Now seeing some work by Munk et al. on using quanitizing and LoRA for small, local LLMs for NPCs in games. They found that another LLM thought the output dialogue was good, but even the small models took 5-10 seconds per response.

Framework for quanitizing LLMs for controlling game NPCs

120

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Back after a break with the Generative AI - “Generative Content, Behaviour, and Experience“, Ruiyu Xu presents (remotely) her framework for LLM NPCs, based on generalizing from example games produced in China.

100

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Johannes Pfau back to showcase his team’s “Dark Zagreus” mod using imitation learning to clone the player’s last successful run as a replacement for Hades 1’s boss. They found no significant difference.

110

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Amir Baghi from EA presents an approach to controlling Football teams with multiagent RL, extending TiZero to explore more and leading to better collaboration (Qualitatively)

330

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Next up my student @daveg29.bsky.social presents his research on learning a neurosymbolic world model! We find we can successfully learn useful programs from a single gameplay video. arxiv.org/pdf/2508.11836

Extracted sprites from a modified Marionette

140

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Now onto the Game AI - Learning session, Kaijie Xu presenting remotely on using Perlin noise not just to control the world generation (like in Minecraft) but to control NPC behaviour (also like in Minecraft). They find it’s comparable to flocking.

Visual comparison of Perlin noise and flocking

150

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

In the final part, Sicart argues LLM agents are just bad/broken roleplayers and we can use play to paper over the cracks.

From Sicart, final conclusion “Play as the kintsugi of broken agents”

210

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

In part two, Sicart covers his own agentic/LLM experiments. A take generator, an agentic marketplace, and an agentic polymarket.

Output of Take Maker prompted to make a take about automating academia

120

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

After a long tangent about agentic systems, Sicart makes the argument to broaden our understanding of the Ludic Century as a lens for things like Polymarket and Looksmaxxing. That games studies has been an isolationist discipline.

Argument for broadening game studies, that it has been isolationist discipline

240

Matthew Guzdial @matthewguz.bsky.social · 12/08/2026

Day two proper of #FDG26 starting with a keynote from Miguel Sicart talking broadly about play and the root of all evil.

Three acts, ecosystems, in their world, the root of all evil

192

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Alexander Dockhorn studies local/open weight LLMs for generating Behaviour Trees in Unreal, finding they can do 20-30 node Btrees pretty well.

030

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Zhongyu Xie presents on some PCGRL and LLM work but I’m too distracted from his claim that Minecraft and Spelunky used evolutionary search to engage with the work.

120

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Johannes Pfau presenting work on using white boxing to control an image diffusion process for concepting. They found devs preferred manual drawing but their approach helped over vanilla diffusion.

Examples as comparisons between approaches in the study

100

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Back after lunch with the “Gen AI” session, starting with Vitaly Gnatyuk who is looking at game texture map generation for an in-development war game. They found a GAN outperforms their diffusion model trained on their internal data.

110

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Reia Drucker (UVic 🇨🇦) presents work analyzing player movement In Call of Duty, using approaches adapted from real world human movement modelling.

110

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Oura Uemura (presenting online) presents an analysis of streamer and audience engagement to see how they affect one another. They find statistically significant results showing relationships between everything except streamer sentiment (streamers can get angry or be happy).

130

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Next up Nicolas Ioannou presents the results of a survey of players around the commendation systems in League and Overwatch. They found that players mostly used the system to reward others for playing well and positive attitudes.

140

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

I’m attending the Game Analytics and Visualization track, starting with Magdaléna Kejstová covering her survey on games for teaching visualization. Her findings included that children and bias detection are underrepresented.

Breakdown of the two phase review approach

Identifying the dominant design patterns

140

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Valentina Nisi covered a historical overview of her work connecting human players with the nature of Madeira.

140

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Right on to the first keynote speaker, Valentina Nisi and her talk “Who’s Playing? Making kin through Game Design: The post-human through gaming and storytelling”, all about how we can connect with non-humans through games.

Opening slide of the first FDG 2026 keynote

140

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Announcing the #FDG26 Best Paper Award “Complicit by Design: Convergence and Alienation in Interactive Fourth Wall Breaks” plus the two honourable mentions (Kaijie Xu and Clark Verbrugge are on an award winning run of papers).

Best paper award and the two honourable mentions (“Prometheas: A System for Promoting Provincial Behavior in Online Multiplayer Games” and “High Dinensional Procedural Content Generation”) for FDG 2026

160

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Welcoming remarks for day 1 of #FDG26 from the organizers and Games Denmark rep, expanding on the theme of “research through games; research for games”

Games Denmark representative speaking at the opening of FDG 2026

391

Matthew Guzdial @matthewguz.bsky.social · 11/08/2026

Starting off #FDG26 proper with a @galaxykate.bsky.social activity, making a combinatorial pen. Important to wash your gameboys.

1100

Matthew Guzdial @matthewguz.bsky.social · 10/08/2026

Bojan Endrovski now back demoing his DSL for level generation, where (in his own words) he ignored the results of his own survey.

100

Matthew Guzdial @matthewguz.bsky.social · 10/08/2026

Ashmiya Dua and Devesh K., undergrads working with Jim Whitehead, are working on adding agents to Pewter, Whitehead’s LLM co-creative RPG world design tool.

Two students presenting in front of a screen with the Pewter interface.

110

Matthew Guzdial @matthewguz.bsky.social · 10/08/2026

Scale of puzzle growth based on grid size is massive, as you’d expect.

Table of puzzle complexity, showing exponential growth.

110

Matthew Guzdial @matthewguz.bsky.social · 10/08/2026

Here’s my solution of said puzzle (please forgive my terrible paper folding)

110

Matthew Guzdial @matthewguz.bsky.social · 10/08/2026

Now Stiliyan Nanovski presenting on generating folding paper puzzles. Quite fun, they generated some for #FDG26 and are handing them out at registration. More papers should do this!

100

Matthew Guzdial @matthewguz.bsky.social · 10/08/2026

Kaijie Xu (McGill) presenting on generating character actions in games from abstract representations, like different kinds of attacks.

100