Research map / Multi-agent systems and games / LLM agents

Dynamic Speculative Agent Planning

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

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 Dynamic Speculative Planning (DSP), which uses an online-trained DistilBERT TD-learning predictor to choose per-episode speculation steps for draft-and-verify LLM agent planning, with expectile-based or offset-based knobs to trade latency against cost.

Problem

Speculative planning for LLM agents accelerates latency but with fixed speculation steps wastes tokens on mismatches, and existing acceleration methods either lose performance, need offline router training, or give users little control over the latency-cost trade-off.

Approach

Formulate speculation-step prediction as a state-value prediction problem trained online with TD(lambda) returns while the agent runs asynchronously, and expose the latency-cost trade-off via expectile regression (biased training) or a post-hoc integer offset on the predicted step.

Key results

• On OpenAGI Direct-ReAct (GPT), Dyn. (offset=2) matches the fastest fixed k=6 latency (0.90 vs 0.90 time ratio) while reducing redundant cost by 60.65%.
• Dyn. (tau=0.5) reduces total cost to 0.88x of fixed k=2 on OpenAGI Direct-ReAct (GPT) while keeping latency at 1.01x.
• Predictor accuracy rises to about 0.70-0.76 in later task segments of online training.

Surprising findings

• Supervised (SFT, lambda=1) step predictors were overly conservative and in most cases slower than the fixed k=2 baseline.

Limitations

• Warm-up period of 3-10 tasks where the predictor is untrained and the system runs sequentially.
• Dynamic variants are not uniformly dominant; on TravelPlanner gains are mixed and some configurations are slower than k=2 baseline.
• Evaluated only with two agent frameworks and two API backbones.

Experimental setup

Datasets: OpenAGI TravelPlanner

Benchmarks: OpenAGI TravelPlanner

Metrics: latency ratio T(x) prompt token ratio generation token ratio cost ratio time decrease unnecessary/redundant cost increase average max concurrency average speculative step predictor accuracy

Baselines: Fixed-k speculative planning (k=2, 4, 6) Bayesian optimization heuristic Supervised (SFT) step predictor

Builds on: speculative decoding Interactive Speculative Planning TD learning expectile regression

Closest papers by meaning

• Speculative Actions: A Lossless Framework for Faster AI Agents (ICLR 2026 · similarity 0.89)
• When RL Meets Adaptive Speculative Training: A Unified Training-Serving System (ICML 2026 · similarity 0.87)
• EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts (NeurIPS 2026 · similarity 0.87)
• Interactive Speculative Planning: Enhance Agent Efficiency through Co-design of System and User Interface (ICLR 2025 · similarity 0.86)
• Learning Query-Aware Budget-Tier Routing for Runtime Agent Memory (ICML 2026 · similarity 0.86)
• Breaking the Performance Ceiling in Reinforcement Learning requires Inference Strategies (NeurIPS 2025 · similarity 0.86)