Research map / Multi-agent systems and games / LLM agents
Interactive Speculative Planning: Enhance Agent Efficiency through Co-design of System and User Interface
Read the paperCheck your idea against 108,967 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 Interactive Speculative Planning, which runs a fast approximation agent alongside a slower target agent with asynchronous verification and a UI rescheduler that lets users interrupt and supply steps, reducing LLM agent planning latency on OpenAGI and TravelPlanner.
Problem
LLM-based agent planning is slow due to large backbone models and extensive intermediate reasoning (e.g., ReAct, tree-of-thought, multi-agent debate), and long silent waits degrade user experience during task delegation.
Approach
Adapt speculative decoding to agent planning: a cheap approximation agent proposes steps sequentially while the target agent verifies each step asynchronously using the provisional prefix; a rescheduling UI presents results in order and lets users actively interrupt slow steps or supply oracle steps.
Key results
- On OpenAGI Setting 1 (GPT-4-turbo DG approximation, ReAct target), total time fell about 22% (33.91s vs 43.63s) and stepwise time about 32%.
- On OpenAGI Setting 3 (multi-agent debate target), total time was cut by 42.30% and stepwise time by 38.29%.
- On TravelPlanner, commonsense micro pass rate dropped in all settings (e.g., Setting 1: 48.6% normal vs 41.7% speculative), while delivery rate stayed at 55.6%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use direct-generation GPT-4-turbo approximation with ReAct GPT-4-turbo target (Setting 1) | Agent planning · OpenAGI | total time (TT): 43.63 -> 33.91; cut by 22.27% | Helped |
| Use direct-generation GPT-4-turbo approximation with CoT GPT-4-turbo target (Setting 2) | Agent planning · OpenAGI | total time (TT): 39.96 -> 28.64; cut by 28.32% | Helped |
| Use CoT GPT-4-turbo approximation with MAD target (Setting 3) | Agent planning · OpenAGI | total time (TT): 182.70 -> 105.42; cut by 42.30% | Helped |
| Use direct-generation GPT-3.5-turbo approximation with direct-generation GPT-4-turbo target (Setting 4) | Agent planning · OpenAGI | total time (TT): 5.77 -> 4.63; cut by 20.37% | Helped |
| Use direct-generation GPT-4-turbo approximation with ReAct GPT-4-turbo target (Setting 1) | Agent planning · TravelPlanner validation | total time (TT): 176.28 -> 137.33; decreased by 21.43% | Helped |
| Use direct-generation GPT-4-turbo approximation with CoT GPT-4-turbo target (Setting 2) | Agent planning · TravelPlanner validation | total time (TT): 121.37 -> 98.09; decreased by 19.18% | Helped |
| Use CoT GPT-4-turbo approximation with MAD target (Setting 3) | Agent planning · TravelPlanner validation | total time (TT): 733.12 -> 568.10; decreased by 25.46% | Helped |
| Use speculative planning instead of normal planning | Setting 1 agent planning · TravelPlanner validation | Commonsense Constraint Micro Pass Rate: 48.6% -> 41.7% | Hurt |
| Use speculative planning instead of normal planning | Setting 2 agent planning · TravelPlanner validation | Commonsense Constraint Micro Pass Rate: 41.7% -> 34.7% | Hurt |
| Use speculative planning instead of normal planning | Setting 3 agent planning · TravelPlanner validation | Commonsense Constraint Micro Pass Rate: 54.3% -> 48.6% | Hurt |
Surprising findings
- Speculative planning showed some data points with longer stepwise time than normal planning, mainly at low approximation accuracy, attributed to API randomness and concurrency slowdowns.
- Speculative planning lowered commonsense constraint pass rate on TravelPlanner in all three settings, since soft matching accepts approximate actions.
Limitations
- Exact or soft match verification is overly strict or lenient, so speculative results may differ from normal planning on TravelPlanner.
- Substantially higher token usage and API cost, and up to 5 concurrent API calls.
- No backtracking in the UI; security risks from executing unverified approximation actions.
Experimental setup
Datasets: OpenAGI TravelPlanner validation
Benchmarks: OpenAGI TravelPlanner
Metrics: total time stepwise time total tokens maximum concurrent API calls API cost delivery rate commonsense constraint micro pass rate
Baselines: ReAct CoT multi-agent debate direct generation
Builds on: speculative decoding ReAct chain-of-thought multi-agent debate
Closest papers by meaning
- AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.87)
- Agent S: An Open Agentic Framework that Uses Computers Like a Human (ICLR 2025 · similarity 0.86)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.86)
- Dynamic Speculative Agent Planning (ICLR 2026 · similarity 0.86)
- Speculative Actions: A Lossless Framework for Faster AI Agents (ICLR 2026 · similarity 0.86)
- Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance (ICLR 2025 · similarity 0.86)
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.