Research map / Multi-agent systems and games / LLM agents
DreamPhase: Offline Imagination and Uncertainty-Guided Planning for Large-Language-Model Agents
Shayan Mohajer Hamidi, Linfeng Ye, Konstantinos N. Plataniotis
Read the paperCheck your idea against 137,217 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.
Proposes DreamPhase, an offline imagination and uncertainty-guided planning framework for LLM agents to improve closed-loop decision-making.
Problem
LLM agents struggle with closed-loop decision-making in multi-step tasks due to static pretraining and limited temporal grounding.
Approach
Uses offline imagination (simulated future rollouts) combined with uncertainty-guided planning so LLM agents can plan over imagined trajectories; the abstract is truncated so specifics are unavailable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| plan with offline latent imagination plus uncertainty gate instead of ARMAP-M token search | ARMAP-M, Llama-2-7B · Eight agent tasks average | success rate avg: 42.3 -> 50.1 | Improved |
| use DreamPhase instead of ARMAP-M | ARMAP-M, Llama3-70B · WebShop/SciWorld/Game-of-24 average | success rate avg: 51.2 -> 53.6 | Improved |
| use DreamPhase instead of ARMAP-M | ARMAP-M, Mistral-7B · WebShop/SciWorld/Game-of-24 average | success rate avg: 28.9 -> 30.7 | Improved |
| use DreamPhase instead of ARMAP-M | ARMAP-M, Llama-8B · WebShop | API calls per episode: 39.8 -> 9.3 | Improved |
| use DreamPhase instead of ARMAP-M | ARMAP-M, Llama-8B · WebShop | success rate: 60.2 -> 61.8 | Improved |
| add uncertainty safety gate and imagination | ARMAP-M, Llama-8B · WebShop | critical irreversible incidents per 1000: 7.1 -> 1.4 | Improved |
| use DreamPhase instead of ARMAP-M | ARMAP-M, Llama-8B · ALFWorld-std | success rate: 0.31 -> 0.39 | Improved |
| use DreamPhase instead of AgentGym | AgentGym, Llama-2-7B · BabyAI | success rate: 82.7 -> 82.3 | No clear effect |
| use DreamPhase instead of ARMAP-M | ARMAP-M, Llama-2-7B · MAZE/BIRD tasks | success rate: MAZE 12.0 -> 14.0; BIRD 10.5 -> 13.5 | Mixed |
Closest papers by meaning
- From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.89)
- RIG: Synergizing Reasoning and Imagination in End-to-End Generalist Policy (ICLR 2026 · similarity 0.88)
- Agent Planning with World Knowledge Model (NeurIPS 2024 · similarity 0.88)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.87)
- Introspective Planning: Aligning Robots' Uncertainty with Inherent Task Ambiguity (NeurIPS 2024 · similarity 0.87)
- Planning from Imagination: Episodic Simulation and Episodic Memory for Vision-and-Language Navigation (AAAI 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.
