BayesFlow: A Probability Inference Framework for Meta-Agent Assisted Workflow Generation

Automatic workflow generation is the process of automatically synthesizing sequences of LLM calls, tool invocations, and post-processing steps for complex end-to-end tasks. Most prior methods cast this task as an optimization problem with limited theoretical grounding. We propose to cast workflow generation as Bayesian inference over a posterior distribution on workflows, and introduce \textbf{Bayesian Workflow Generation (BWG)}, a sampling framework that builds workflows step-by-step using parallel look-ahead rollouts for importance weighting and a sequential in-loop refiner for pool-wide improvements. We prove that, without the refiner, the weighted empirical distribution converges to the target posterior. We instantiate BWG as \textbf{BayesFlow}, a training-free algorithm for workflow construction. Across six benchmark datasets, BayesFlow improves accuracy by up to 9 percentage points over SOTA workflow generation baselines and by up to 65 percentage points over zero-shot prompting, establishing BWG as a principled upgrade to search-based workflow design. Code will be available on https://github.com/BoYuanVisionary/BayesFlow.

💡 Research Summary

The paper tackles the problem of automatically constructing multi‑step workflows that orchestrate large language model (LLM) calls, tool invocations, and post‑processing for complex tasks. While prior work treats workflow design as an optimization problem—using heuristics such as Monte‑Carlo Tree Search, evolutionary mutation, or linear search—these approaches lack solid theoretical foundations and typically output a single high‑scoring solution, limiting diversity and interpretability.

The authors re‑formulate workflow generation as Bayesian posterior sampling. They view the meta‑optimizer LLM as providing a prior distribution p(s₁:T) over all possible step‑wise code fragments (the workflow), and they incorporate an external reward R(s₁:T) (e.g., validation accuracy) via an energy‑based model, yielding an unnormalized posterior q(s₁:T|s₀) ∝ p(s₁:T)·exp(R(s₁:T)). This casts the task as drawing samples from a distribution that balances the LLM’s internal knowledge with observed task performance.

To approximate this posterior they propose Bayesian Workflow Generation (BWG), a two‑stage sampling framework:

1. Parallel Look‑ahead Rollouts – For each partial workflow of length t‑1, a single next step is sampled from the prior. Then K stochastic completions are generated in parallel, each scored by the reward function. The average of exp(R) over the K rollouts becomes an importance weight for that partial prefix. Normalized weights are used to resample prefixes, promoting those with higher expected downstream reward while mitigating weight‑degeneracy common in pure prior re‑weighting.

2. Sequential In‑loop Refinement – After weighting, a global refinement operator G creates M new complete workflows based on the current pool (including both partial prefixes and previously completed samples). In the implementation G is an MCTS module adapted from AFLO​W, which can modify earlier steps, thereby correcting early mistakes that pure look‑ahead cannot fix. All new and existing workflows are re‑weighted by exp(R) and resampled to form the next generation of prefixes.

The authors prove (Theorem 1) that, when the refinement step is omitted, the weighted empirical distribution of the algorithm converges asymptotically to the true posterior, guaranteeing that sampling from the posterior rather than the prior improves expected reward. With refinement enabled, the theoretical guarantee no longer strictly holds, but empirical ablations show substantial performance gains.

Experiments span six benchmark datasets covering mathematical reasoning, code generation, data analysis, and more, using both closed‑source (e.g., GPT‑4) and open‑source LLMs. BayesFlow (the concrete instantiation of BWG) consistently outperforms state‑of‑the‑art workflow generators (e.g., AFLO​W, CAMEL, MetaGPT) by up to 9 percentage points in accuracy on the math reasoning set and yields an average improvement of 4.6 % across all tasks. Compared to zero‑shot prompting, gains reach as high as 65 %. Moreover, the posterior sampling approach produces a diverse set of high‑quality workflows, offering flexibility for downstream deployment.

In summary, the paper makes three major contributions: (1) a principled Bayesian formulation of workflow generation that naturally encourages diversity and provides convergence guarantees; (2) the BWG framework that unifies parallel look‑ahead importance weighting with a global refinement step, subsuming many existing methods as special cases; and (3) BayesFlow, a training‑free algorithm that demonstrates consistent empirical superiority across multiple domains and model families. This work establishes Bayesian inference as a solid theoretical foundation for automatic workflow synthesis and opens avenues for further research on scalable, provably correct workflow design.