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

Speculative Actions: A Lossless Framework for Faster AI Agents

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration

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 speculative actions, a lossless framework where a fast Speculator predicts likely next API calls while a slow Actor runs, committing on match; evaluated on chess, e-commerce (tau-bench), HotpotQA, and an OS tuning lossy extension, with a cost-latency analysis.

Problem

AI agent execution is slow because each action requires a sequential API call with substantial latency, making agent runs take hours and bottlenecking RL and prompt optimization.

Approach

Treat every agent step as an API call and use a fast Speculator to predict the next actions and pre-launch the subsequent calls in parallel with the slow Actor, committing only when predictions match (lossless), with a closed-form latency/cost analysis of speculative breadth.

Key results

• Chess with 3 speculative predictions: average 19.5% time saving and 54.7% prediction accuracy across 5 runs.
• tau-bench retail: 22%-38% of API calls correctly predicted by the Speculator; HotpotQA top-3 accuracy up to 46%.
• OS tuning: Speculator-Actor converges in about 13 s versus about 200 s for Actor-only, with lower total cost (0.17 vs 2.18 cents).

Surprising findings

• Stronger speculator models often yield lower API prediction accuracy on HotpotQA under strict matching, because their more diverse queries are penalized.
• Depth-focused speculation does not grow exponentially in branches; active branches stay bounded by the ratio of actor to speculator latency.

Limitations

• Speedup depends on speculation accuracy, which is modest (22-55%) and strict matching penalizes semantically equivalent calls.
• Latency measurements from live APIs are stochastic and not perfectly reproducible.
• Cost analysis assumes fixed token rate and per-token cost; OS tuning setting is lossy rather than lossless.

Experimental setup

Datasets: TextArena chess tau-bench retail HotpotQA sysbench cpu

Benchmarks: TextArena chess tau-bench retail HotpotQA sysbench cpu

Metrics: next-action prediction accuracy time saved API prediction accuracy p95 latency token cost

Baselines: Actor-only Speculator-only single-model speculator fixed 1 or 2 speculations

Builds on: speculative decoding ReAct TextArena tau-bench interactive speculative planning

Closest papers by meaning

• Dynamic Speculative Agent Planning (ICLR 2026 · similarity 0.89)
• Interactive Speculative Planning: Enhance Agent Efficiency through Co-design of System and User Interface (ICLR 2025 · similarity 0.86)
• A Two-Tier Perspective on Inference-Time Parallelism in Multi-Agent LLM Systems (ICML 2026 · similarity 0.85)
• Scaling Small Agents Through Strategy Auctions (ICML 2026 · similarity 0.85)
• Speculative Safety Honeypot: Toward Proactive Defense Against Multi-turn Agent Attacks (ICML 2026 · similarity 0.85)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.85)