Research map / Image generation / Text-to-image diffusion
STEVE-1: A Generative Model for Text-to-Behavior in Minecraft
Shalev Lifshitz, Keiran Paster, Harris Chan, Jimmy Ba, Sheila A. McIlraith
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add classifier-free guidance at inference (lambda>0 vs lambda=0) (ablation) | STEVE-1 · dirt collection | dirt collected: 7.5 \times | Improved |
| add classifier-free guidance at inference (lambda>0 vs lambda=0) (ablation) | STEVE-1 · wood collection | wood collected: 15 \times | Improved |
| raise guidance scale lambda beyond 5.0 to 7.0 (ablation) | STEVE-1 · log and dirt collection | items collected: 5.0 to 7.0 | Hurt |
| train with goal dropout p_uncond=0.1 instead of no guidance (ablation) | STEVE-1 · inventory collection tasks | items collected: 2 to 3 times | Improved |
| use longer, more specific prompt instead of 'get dirt' (ablation) | STEVE-1 · wood collection | wood collected: 1.6\times | Improved |
| use longer, more specific prompt instead of 'get dirt' (ablation) | STEVE-1 · dirt collection | dirt collected: 2\times | Improved |
| use longer, more specific prompt instead of 'get dirt' (ablation) | STEVE-1 · seed collection | seeds collected: 3.3\times | Improved |
| chain visual prompts (get dirt then build tower) instead of directly prompting | STEVE-1 · tower building | success rate: 50\% | Improved |
| chain visual prompts (get logs then craft planks) instead of directly prompting | STEVE-1 · wooden planks crafting | success rate: 70\% | Improved |
| fine-tune from pretrained VPT weights instead of random initialization (ablation) | STEVE-1 · programmatic tasks (wood) | items collected | Improved |
| increase training data beyond 20M frames | STEVE-1 · seed collection and travel distance | performance: 20M | No clear effect |
| increase training data beyond 60M frames | STEVE-1 · log and dirt collection | items collected: 60M | Improved |
Closest papers by meaning
- OmniJARVIS: Unified Vision-Language-Action Tokenization Enables Open-World Instruction Following Agents (NeurIPS 2024 · similarity 0.90)
- GROOT: Learning to Follow Instructions by Watching Gameplay Videos (ICLR 2024 · similarity 0.88)
- MrSteve: Instruction-Following Agents in Minecraft with What-Where-When Memory (ICLR 2025 · similarity 0.88)
- ADAM: An Embodied Causal Agent in Open-World Environments (ICLR 2025 · similarity 0.87)
- MineObserver 2.0: A Deep Learning & In-Game Framework for Assessing Natural Language Descriptions of Minecraft Imagery (AAAI 2024 · similarity 0.87)
- CrafterDojo: A Suite of Foundation Models for Building Open-Ended Embodied Agents in Crafter (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
