UCSB & LinkedIn Research: Agents Can Speculate Their Own Tool Calls
dair_ai · x · 2026-07-30
Agents spend a large share of time waiting on tool results. New research from UCSB and LinkedIn proposes unifying both roles in one model to hide this latency.
- Unified Architecture: The model runs in agent mode to solve tasks and speculator mode to predict its next tool call, fully reusing the prefix KV cache.
- Joint RL: Derives speculation targets from the agent's own rollouts and alternates updates, preventing dual-mode training from degrading task success.
- Results: Next tool-call Hit@1 rises from 44.1 to 61.2 for Qwen3-4B and from 48.9 to 66.3 for Qwen3.5-4B.
More from coding & agent
- LlamaIndex Founder: Humans May Stop Reviewing AI Code in 1-2 Years — dotey · 2026-07-30
- The New MCP Is Stateless: A Visual Guide to the Protocol Rewrite — dima806_dima · 2026-07-30
- ARC-AGI-3 API clarification: 'reasoning' field logs model output, not private CoTs — GregKamradt · 2026-07-30
- Claude Opus 5 Recreates Fallout Game in a Single 2.3MB HTML File — chrisfirst · 2026-07-30
- CodePilot v0.62 Adds Right-Click File Tree and Live Markdown Preview — op7418 · 2026-07-30
- A Practical Guide to Building Frontier-Lab Quality AI Evaluations — aakashgupta · 2026-07-30