Research map / Reinforcement learning / MDPs and planning
EMFuse: Energy-based Model Fusion for Decision Making
Kejie He, Yi-Chen Li, Yang Yu
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 |
|---|---|---|---|
| fuse by energy sum instead of parameter averaging (Soup) | expert LLMs · OpenCompass subject-mix | accuracy: 60.88 -> 63.49 | Improved |
| fuse by energy sum instead of PackLLM logit packing | expert LLMs · OpenCompass subject-mix | accuracy: 63.15 -> 63.49 | No clear effect |
| energy-sum fusion instead of PackLLM | expert LLMs · OpenCompass finance-suite | accuracy: 88.27 -> 89.21 | Improved |
| add EMSelect per-step expert selection on top of EMFuse (ablation) | EMFuse · OpenCompass subject-mix and finance-suite | accuracy: 63.49 -> 64.80; 89.21 -> 90.39 | Improved |
| add EMSelect on top of EMFuse for math sets (ablation) | EMFuse · MGSM (math) | accuracy: -2.40 | Hurt |
| use temporal-consistency uncertainty penalty instead of zero penalty (ablation) | ADETM · D4RL HalfCheetah medium | IQM normalized return: 7.94 -> 41.83 | Improved |
| use temporal-consistency uncertainty instead of action-noise sensitivity (ablation) | ADETM · D4RL Hopper medium | IQM normalized return: 37.22 -> 49.03 | Improved |
| use action-noise local sensitivity penalty instead of no penalty (ablation) | ADETM · D4RL HalfCheetah medium | IQM normalized return: 7.94 -> 1.72 | Hurt |
| single ADETM instead of ensemble of 5 dynamics models (ablation) | dynamics model · D4RL medium (Hopper) | epoch FLOPs (T): 44.35 -> 12.32 | Improved |
| use entropy-based reweighting instead of uniform fusion weights (ablation) | EMFuse · OpenCompass | accuracy | No clear effect |
| energy fusion of 1B experts instead of an 8B model | EMFuse · AlpacaEval domains | token-level KL to 1B expert: 0.5063 -> 0.0391 (harmless) | Improved |
| use EMFuse instead of Mixed-data oracle on MuJoCo medium | EMFuse · D4RL MuJoCo medium average | IQM normalized return: 47.8 -> 50.1 | No clear effect |
Closest papers by meaning
- Offline Transition Modeling via Contrastive Energy Learning (ICML 2024 · similarity 0.86)
- Analytic Energy-Guided Policy Optimization for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Model Fusion through Bayesian Optimization in Language Model Fine-Tuning (NeurIPS 2024 · similarity 0.85)
- Reinforcement Learning with Imperfect Transition Predictions: A Bellman-Jensen Approach (NeurIPS 2025 · similarity 0.85)
- Autoregressive Language Models are Secretly Energy-Based Models: Insights into the Lookahead Capabilities of Next-Token Prediction (ICML 2026 · similarity 0.85)
- Joint Learning of Energy-based Models and their Partition Function (ICML 2025 · similarity 0.84)
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.
