AsyncFC: Unlocking Hidden Multithreading in LLMs with Future-based Asynchronous Function Calling

Forum topic · QianXun · 2026-05-18

Summary

A Chinese tech forum post discusses a research paper titled 'Concurrency without Model Changes: Future-based Asynchronous Function Calling for LLMs' from a team including UC Berkeley researchers. Traditional LLM agents execute tool calls sequentially: the model pauses generation and waits for each external call (e.g., a weather API) to return before continuing, causing sluggish responses. The paper introduces AsyncFC, a framework that instead hands the model a symbolic placeholder—a 'Future'—representing a pending result, allowing the model to keep reasoning, plan subsequent steps, and issue multiple parallel tool calls without retraining. Notably, existing models like GPT-4o and Claude 3.5 can natively reason over these unresolved placeholders, treating them as objects passed to later steps. This enables parallel invocation of many requests and overlaps network latency with generation, reportedly speeding up agent task completion several times while preserving logical coherence even when results are outstanding. The post frames this as a shift from turn-by-turn Q&A toward proactive, asynchronous 'thinking' agents.

This post introduces a research paper: 'Concurrency without Model Changes: Future-based Asynchronous Function Calling for LLMs', released in May 2026 by a team including researchers from UC Berkeley.

The Problem: AI That Waits Idle

When asked to check the weather, book a flight, and draft an itinerary, a conventional AI agent issues the weather query, then blocks—it cannot write another word until the external server responds. This sequential 'wait for one, do one' pattern is a major cause of slow AI agent responses.

The Core Idea: Symbolic Futures and AsyncFC

The paper's key contribution is the AsyncFC framework:

Traditional approach: The AI asks, 'Is the water boiling?' and stops thinking until the answer arrives.
AsyncFC approach: The AI asks the same question, but instead of waiting, it immediately receives a placeholder token—a symbolic envelope named something like Future_0—with the instruction: 'The result is in here; even though it's unopened, you can keep reasoning.'

A Hidden Capability: Reasoning Under Uncertainty

The researchers found that existing models—such as GPT-4o and Claude 3.5—can handle these placeholders without any retraining. They treat an unresolved Future as a real object and pass it into subsequent reasoning steps, like an architect drafting a full building blueprint from 'brick placeholders' before the bricks arrive.

This enables asynchronous thinking:

1. Parallel calls: firing 10 search requests at once rather than one at a time. 2. Overlapped computation: drafting the next 500 words of analysis while the server is still returning data.

Results and Significance

Experiments show agents complete tasks several times faster, with perceived latency largely eliminated. Beyond speed, the approach marks a shift from a 'question-and-answer' machine toward a genuine proactive thinker:

Efficiency: no longer stalled by slow networks or external tools.
Logical coherence: the chain of thought stays intact even when intermediate results are pending.

Takeaway

The author's conclusion: AI models are smarter than we assume—what they lacked was not logic, but an environment that lets them 'think while waiting.' AsyncFC acts like a multitasking operating system for LLMs: even before results arrive, logic can proceed first.