Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning to Follow Instructions in Text-Based Games
Mathieu Tuli, Andrew C Li, Pashootan Vaezipoor, Toryn Q. Klassen, Scott Sanner, 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 |
|---|---|---|---|
| Represent instructions as LTL and condition the policy on the instruction and its current progressed state | GATA on TextWorld Cooking · Cooking levels 0–3 | normalized game points and game success rate | Improved |
| Remove LTL progression (ablation) | LTL-GATA · TextWorld Cooking | performance | Hurt |
| Require GATA to examine the cookbook at the first step (ablation) | GATA-D · TextWorld Cooking | performance | No clear effect |
| Use LTL reward without LTL-based episode termination (ablation) | LTL-GATA · Cooking levels 1 and 2 | performance | Improved |
| Use LTL-based episode termination without LTL reward (ablation) | LTL-GATA · Cooking levels 1 and 2 | performance | Improved |
| Remove both LTL reward and LTL-based episode termination (ablation) | LTL-GATA · Cooking level 2, 20-game training set | performance | Hurt |
| Represent LTL predicates as multiple tokens rather than single tokens (ablation) | LTL-GATA · TextWorld Cooking, 20-game training set | performance | No clear effect |
| Train GATA with instructions stripped from observations (ablation) | GATA-D · Cooking levels 1 and 2 | normalized game points and success rate | No clear effect |
| Use 100 training games instead of 20 (ablation) | LTL-GATA · TextWorld Cooking | test performance | Improved |
| Apply LTL-GATA to games with navigation | LTL-GATA · Cooking level 3 | performance relative to GATA | Mixed |
| Use GPT-3 Da Vinci for few-shot natural-language-to-LTL translation instead of GPT-3 Ada | GPT-3 · TextWorld observation-to-LTL translation | translation correctness: 93.2% absolutely correctly and another 5.6% almost correctly; Ada translated all 100% of examples incorrectly | Improved |
Closest papers by meaning
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.89)
- Motif: Intrinsic Motivation from Artificial Intelligence Feedback (ICLR 2024 · similarity 0.89)
- Zero-Shot Instruction Following in RL via Structured LTL Representations (NeurIPS 2026 · similarity 0.88)
- RLZero: Direct Policy Inference from Language Without In-Domain Supervision (NeurIPS 2025 · similarity 0.88)
- Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.88)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · 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.
