Research map / Reinforcement learning / Imitation learning
Improving Policy Learning via Language Dynamics Distillation
Victor Zhong, Jesse Mu, Luke Zettlemoyer, Edward Grefenstette, Tim Rocktäschel
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 |
|---|---|---|---|
| Pretrain and distill language-grounded dynamics representations from unlabeled demonstrations | Policy learning on SILG tasks · NetHack, ALFWorld, RTFM, Messenger, Touchdown | task performance | Improved |
| Include language descriptions in dynamics-modeling observations (ablation) | LDD · NetHack | policy performance | Improved |
| Include language descriptions in dynamics-modeling observations (ablation) | LDD · SymTouchdown | policy performance | Improved |
| Train the dynamics model on expert demonstrations instead of non-expert rollouts (ablation) | LDD · SILG environments | dynamics-model accuracy | Improved |
| Pretrain to initialize the policy learner (ablation) | LDD · Messenger | policy performance | Improved |
| Distill the pretrained dynamics representations during policy learning (ablation) | LDD · Touchdown | evaluation win rate | Improved |
| Use dynamics-based reward shaping instead of LDD | Policy learning · SILG environments | task performance | Hurt |
| Use VAE pretraining instead of LDD dynamics pretraining | Policy learning · SILG environments | task performance | Hurt |
| Use inverse reinforcement learning instead of LDD | Policy learning · SILG environments | task performance | Hurt |
| Remove the initialization step (ablation) | LDD · Messenger | policy performance | Hurt |
| Remove the distillation step (ablation) | LDD · Touchdown | evaluation win rate | Hurt |
| Remove in-game messages and character-state descriptions (ablation) | LDD dynamics-modeling observations · NetHack | policy performance | Hurt |
Closest papers by meaning
- Representation-Based Exploration for Language Models: From Test-Time to Post-Training (ICLR 2026 · similarity 0.89)
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.88)
- Text-Aware Diffusion for Policy Learning (NeurIPS 2024 · similarity 0.88)
- Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation (NeurIPS 2026 · similarity 0.88)
- RLZero: Direct Policy Inference from Language Without In-Domain Supervision (NeurIPS 2025 · similarity 0.88)
- Controlling Large Language Model with Latent Action (ICML 2025 · similarity 0.88)
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.
