Research map / Large language models / Agents, code and math

Unbounded: A Generative Infinite Game of Character Life Simulation

ICLR 2025 · Large language models › Agents, code and math · approach: LLM-driven embodied and driving agents · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Unbounded, a generative infinite character life-simulation game where an LLM game engine (Gemma-2B distilled from GPT-4o data) and a regional IP-Adapter with DreamBooth/LCM SDXL generate all mechanics, narrative and graphics in near real time.

Problem

Traditional video games are finite, with hard-coded mechanics and pre-designed assets; generating open-ended game worlds requires consistent characters and environments plus stable game-state logic in real time.

Approach

Fully generative game where a distilled small LLM serves as game engine trained on multi-agent (world LLM and user LLM) interaction data, and a regional IP-Adapter with dynamic cross-attention masks and block drop separates character and environment conditioning in SDXL with LCM and DreamBooth LoRAs.

Key results

• Regional IP-Adapter with block drop achieves CLIP-I^C 0.676 and DreamSim^C 0.488 versus StoryDiffusion's 0.629 and 0.545 for character consistency.
• Distilled Gemma-2B scores overall 7.44 pairwise versus base Gemma-2B 6.22 and is close to GPT-4o (7.68 for GPT-4o vs 7.76 overall, judged by GPT-4).
• Block drop raises environment CLIP-I^E from 0.123 (no IP-Adapter) and 0.414 (multi-IP-Adapter without block drop) to 0.563 with the full method on the 1.0-scale ablation.

Measured effects reported in the paper

Surprising findings

• Adding IP-Adapter in down-sample blocks harms layout; dropping it there improves both prompt alignment and consistency.
• Distilled 2B model reaches near-GPT-4o judged quality, yet GPT-4o still slightly edges it on overall score.

Limitations

• LLM evaluation relies on GPT-4 pairwise judging over only 100 samples.
• Character appearance remains influenced by the surrounding environment style despite the regional method.

Experimental setup

Datasets: 5,000 topic-character pairs generated by GPT-3.5 5,000 user-simulator interaction samples generated by GPT-4o 5,000 (character image, environment, prompt) evaluation triplets 100 user-simulator interaction samples for LLM evaluation

Benchmarks: Unbounded image consistency evaluation set (5,000 triplets) Unbounded LLM game-engine evaluation (100 interaction samples)

Metrics: CLIP-I DINO DreamSim CLIP-T Grounding-DINO detection GPT-4 pairwise overall score (0-10) state update, environment relevance, story coherence, instru

Baselines: IP-Adapter IP-Adapter-Instruct StoryDiffusion Gemma-2B Gemma-7B Llama3.2-3B GPT-4o

Builds on: IP-Adapter DreamBooth latent consistency models StoryDiffusion Vicuna

Closest papers by meaning

• VirtualEnv: A Platform for Embodied AI Research (AAAI 2026 · similarity 0.88)
• From Zero to Hero: Training-Free Custom Concept Spawning in World Models (NeurIPS 2026 · similarity 0.87)
• GameGen-X: Interactive Open-world Game Video Generation (ICLR 2025 · similarity 0.86)
• Learning Interactive Real-World Simulators (ICLR 2024 · similarity 0.86)
• NarrativePlay: An Automated System for Crafting Visual Worlds in Novels for Role-Playing (AAAI 2024 · similarity 0.85)
• GameVerse: Can Vision-Language Models Learn from Video-based Reflection? (ICML 2026 · similarity 0.85)