From Search to Synthesis: Training LLMs as Zero-Shot Workflow Generators

Bin Dong; Gan Luo; Wotao Yin; Zihan Qin

arxiv: 2606.30704 · v1 · pith:MHPY3KBKnew · submitted 2026-06-29 · 💻 cs.LG · cs.AI· cs.CL

From Search to Synthesis: Training LLMs as Zero-Shot Workflow Generators

Gan Luo , Zihan Qin , Bin Dong , Wotao Yin This is my paper

Pith reviewed 2026-07-01 07:07 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CL

keywords workflow generationmeta-learningzero-shot generalizationreinforcement learningLLM trainingtask-level patternssynthetic dataexecution feedback

0 comments

The pith

MetaFlow trains LLMs to generate reusable task-level workflows that generalize zero-shot to new tasks and operator sets.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces MetaFlow to train large language models as generators of workflows that encode recurring algorithmic patterns at the task level rather than producing one-off solutions for each problem instance. It frames workflow generation as a meta-learning problem where the model receives a task description and an operator set, then composes a strategy. Training proceeds in two stages: supervised fine-tuning on synthetic workflow examples, followed by reinforcement learning that scores complete workflows by their execution success rate across multiple instances of the same task. If the claim holds, this yields models that match strong baselines on familiar tasks with a single forward pass while transferring to entirely new tasks and operator collections without further training. Readers would care because the resulting workflows supply structural consistency, easier debugging through explicit traces, and reuse across variations where pure instance-by-instance generation often fails.

Core claim

MetaFlow casts workflow generation as a meta-learning problem: given a task and an operator set, the model learns to compose solution strategies. It trains in two stages—supervised fine-tuning on synthetic workflow data, followed by reinforcement learning with verifiable rewards (RLVR) that uses execution feedback across problem instances in the task to improve end-to-end success. The resulting model produces effective workflows for trained tasks and exhibits strong generalization to untrained tasks and novel operator sets. Across benchmarks in question answering, code generation, and mathematical reasoning, MetaFlow achieves performance comparable to state-of-the-art baselines on in-domain

What carries the argument

MetaFlow, the two-stage procedure of supervised fine-tuning on synthetic workflows followed by reinforcement learning with verifiable rewards from execution feedback across task instances, which treats workflow generation as meta-learning to compose reusable strategies.

If this is right

Workflows supply robustness across instance variations within a task because they encode recurring patterns rather than per-example fixes.
Single-inference generation produces performance on trained tasks that matches state-of-the-art baselines in question answering, code generation, and mathematical reasoning.
The same trained model transfers to entirely new tasks and previously unseen operator sets without additional fine-tuning.
Explicit workflow traces improve interpretability and debugging compared with opaque instance-by-instance outputs.
Automatic workflow synthesis removes the need for manual expert design of reusable strategies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The separation of task strategy from instance execution could be tested by measuring how often the generated workflow reuses the same high-level structure on problems that differ in surface details.
If execution feedback during RLVR is the main driver of generalization, replacing it with other verifiable signals such as partial credit or intermediate checks should produce similar transfer.
The method opens the possibility of maintaining a library of learned workflows that can be composed or adapted when new operators become available.
Because the model learns from execution outcomes across instances, it may discover strategies that remain stable even when the underlying operators are swapped for functionally equivalent ones.

Load-bearing premise

That training on synthetic workflow data and then reinforcing workflows by their success rate across many instances of the same task will cause the model to learn general task-level patterns instead of memorizing solutions to individual examples.

What would settle it

Measuring whether workflows generated for tasks and operator sets held out from training achieve success rates close to those on the original training distribution, or whether performance collapses to levels no better than prompting an untuned model to solve each instance separately.

Figures

Figures reproduced from arXiv: 2606.30704 by Bin Dong, Gan Luo, Wotao Yin, Zihan Qin.

**Figure 1.** Figure 1: Illustration of instance-level, task-level, and meta-level (ours) workflow generation approaches. Unlike instance-leve methods that generate workflows for individual problems or task-level methods that require costly search for each new task, our meta-learning approach enables zero-shot workflow generation across tasks. The first paradigm comprises task-level approaches, exemplified by ADAS (Hu et al., 202… view at source ↗

**Figure 2.** Figure 2: Overview of MetaFlow architecture. Left: Basic operators and their functionalities, including textprocessing operators (Generate, Summarize, Revise, Ensemble) and basic tool-invocation operators (Programmer). Right: The MetaFlow training framework showing how the planner model generates workflows conditioned on domain (C, Ops), which are then evaluated on problem instances to compute rewards for GRPO opt… view at source ↗

**Figure 3.** Figure 3: Training dynamics of MetaFlow. Left: SFT training loss across different LoRA ranks (with the same learning rate lr = 0.00002). Middle: SFT gradient norms showing convergent behavior. Right: RLVR average reward demonstrating rapid policy improvement and stabilization. 5.2 Low-Cost Operator Integration: Continuous Library Expansion A key advantage of MetaFlow over task-level search methods (Zhang et al., 202… view at source ↗

**Figure 4.** Figure 4: Zero-shot generalization with novel VectorSearch operator on HotpotQA. Left: Individual workflow performance across 100 generations shows high variance, 31 workflows fail completely while the best achieves 0.74 accuracy. Middle: Average score converges to 0.29, reflecting the exploration cost of zero-shot generation. Right: Best-of-k score reaches 0.74 (60.9% above baseline 0.46), surpassing baseline at k… view at source ↗

read the original abstract

Large language models (LLMs) excel across a wide range of tasks, yet their instance-specific solutions often lack the structural consistency needed for reliable deployment. Workflows that encode recurring algorithmic patterns at the task level provide a principled framework, offering robustness across instance variations, interpretable traces for debugging, and reusability across problem instances. However, manually designing such workflows requires significant expertise and effort, limiting their broader application. While automatic workflow generation could address this bottleneck, existing methods either produce instance-specific solutions without learning task-level patterns, or cannot generalize beyond their training configurations. We present MetaFlow, which casts workflow generation as a meta-learning problem: given a task and an operator set, the model learns to compose solution strategies. MetaFlow trains in two stages: supervised fine-tuning on synthetic workflow data, followed by reinforcement learning with verifiable rewards (RLVR) that uses execution feedback across problem instances in the task to improve end-to-end success. The resulting model produces effective workflows for trained tasks and exhibits strong generalization to untrained tasks and novel operator sets. Across benchmarks in question answering, code generation, and mathematical reasoning, MetaFlow achieves performance comparable to state-of-the-art baselines on in-domain tasks with single inference, while demonstrating remarkable zero-shot generalization capabilities on out-of-domain tasks and operator sets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MetaFlow frames workflow generation as meta-learning with a two-stage SFT then RLVR pipeline, but the abstract states generalization claims without any numbers or details to assess them.

read the letter

This paper's main contribution is a two-stage training recipe for LLMs to generate reusable workflows that generalize zero-shot, though the abstract provides no results to evaluate if it works.

MetaFlow casts the problem as learning to compose solution strategies from a task and operator set. The first stage does supervised fine-tuning on synthetic workflow data. The second applies reinforcement learning with verifiable rewards drawn from execution feedback across multiple problem instances in the same task. The goal is to move past instance-specific outputs toward task-level patterns that transfer to untrained tasks and new operator sets.

The motivation is clear and practical. Instance-specific answers from LLMs often lack consistency across variations, which hurts reliability in QA, code generation, and math reasoning. Workflows could give reusable structure and interpretable traces. The paper explains why prior automatic methods fall short on generalization and positions the cross-instance RL stage as the fix.

The argument holds together on its own terms. The stress-test note is accurate that no internal contradiction appears in the described setup.

The soft spot is the complete absence of evidence. The abstract claims performance comparable to baselines on in-domain tasks and remarkable zero-shot results on out-of-domain ones, yet supplies no quantitative results, ablations, or implementation details. Without those, the key assumption—that the RLVR stage will produce task-level abstractions rather than fit the synthetic data—remains untested. Soundness cannot be judged from what is shown.

This is for researchers working on LLM agents and structured reasoning systems who want training methods for reusable outputs. A reader interested in workflow automation would find the pipeline description useful as an idea to build on.

It deserves peer review. The problem is well-defined and the proposed method is specific enough that referees can check the experiments and see whether the generalization actually materializes.

Referee Report

2 major / 2 minor

Summary. The paper introduces MetaFlow, a two-stage training approach (supervised fine-tuning on synthetic workflows followed by reinforcement learning with verifiable rewards, RLVR) that casts LLM workflow generation as a meta-learning problem. Given a task and operator set, the model learns to compose reusable, task-level solution strategies rather than instance-specific outputs. The central empirical claim is that the resulting model matches SOTA baselines on in-domain tasks (question answering, code generation, mathematical reasoning) using single inference while exhibiting strong zero-shot generalization to out-of-domain tasks and novel operator sets.

Significance. If the reported generalization results hold under rigorous evaluation, the work would be significant for shifting LLM usage from brittle per-instance solutions toward robust, interpretable, and reusable workflows. The RLVR stage, which uses cross-instance execution feedback, offers a concrete mechanism for encouraging task-level abstraction; successful validation would provide a scalable alternative to manual workflow engineering.

major comments (2)

[Abstract and §3] Abstract and §3 (method): the claim that RLVR produces task-level patterns rather than instance-specific solutions rests on the weakest assumption identified in the reader's note. No derivation, pseudocode, or ablation is referenced showing how the reward signal across instances forces abstraction; without this, the generalization results cannot be attributed to the proposed mechanism.
[§4] §4 (experiments): the abstract asserts 'performance comparable to state-of-the-art baselines' and 'remarkable zero-shot generalization' yet supplies no numerical results, baseline names, or dataset splits. Load-bearing claims require the specific tables or figures reporting these metrics to be evaluated for statistical significance and controls.

minor comments (2)

[Abstract] Notation: 'MetaFlow' is used before being defined; introduce the name explicitly on first appearance.
The abstract does not mention limitations, failure modes, or the scope of operator sets tested; adding a brief limitations paragraph would improve completeness.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on MetaFlow. We address each major comment below, providing clarifications and indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (method): the claim that RLVR produces task-level patterns rather than instance-specific solutions rests on the weakest assumption identified in the reader's note. No derivation, pseudocode, or ablation is referenced showing how the reward signal across instances forces abstraction; without this, the generalization results cannot be attributed to the proposed mechanism.

Authors: The RLVR objective computes a verifiable reward based on end-to-end execution success across multiple instances of the same task. Because a workflow that succeeds on one instance but fails on others receives a lower aggregate reward, the optimization pressure favors compositions that capture task-level structure rather than instance-specific shortcuts. Section 3 describes this cross-instance feedback mechanism, but we agree that an explicit derivation, pseudocode, and supporting ablation would make the attribution clearer. We will add a short formal argument and pseudocode to §3 in the revision and include a targeted ablation comparing cross-instance versus per-instance rewards. revision: partial
Referee: [§4] §4 (experiments): the abstract asserts 'performance comparable to state-of-the-art baselines' and 'remarkable zero-shot generalization' yet supplies no numerical results, baseline names, or dataset splits. Load-bearing claims require the specific tables or figures reporting these metrics to be evaluated for statistical significance and controls.

Authors: Abstracts are conventionally limited to high-level claims to preserve brevity; the concrete numerical results, baseline names, dataset splits, and statistical controls appear in §4 with the associated tables and figures. The claims in the abstract are directly supported by those results. No change to the abstract is required, though we can add a parenthetical reference to the relevant tables if the editor prefers. revision: no

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents an empirical two-stage training procedure (SFT on synthetic workflows followed by RLVR using execution feedback) and reports benchmark results on in-domain and out-of-domain tasks. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text that would reduce any claimed result to its inputs by construction. Generalization claims rest on held-out evaluation rather than definitional equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Review performed on abstract only; the central claim rests on the domain assumption that execution feedback across instances can train task-level rather than instance-level strategies. No free parameters or invented physical entities are mentioned.

axioms (1)

domain assumption Workflows that encode recurring algorithmic patterns at the task level provide robustness across instance variations, interpretable traces, and reusability.
Stated explicitly in the opening of the abstract as the motivation for the work.

invented entities (1)

MetaFlow no independent evidence
purpose: The two-stage training procedure that turns an LLM into a zero-shot workflow generator.
Introduced as the presented system; no independent evidence outside the paper is supplied in the abstract.

pith-pipeline@v0.9.1-grok · 5766 in / 1358 out tokens · 45569 ms · 2026-07-01T07:07:37.597507+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

11 extracted references

[5]

"" I m p l e m e n t the core problem - solving logic here

Return final n u m e r i c a l answer only # ### Input F o r m a t ‘‘‘ --- ** QU ES TIO N :** [ Com pl ete word problem text ] --- ‘‘‘ Mu lt ipl e p rob le ms follow the same s t r u c t u r e if pr ov ide d . Listing 4: CompleteTask Type Descriptionof GSM8K including domain overview, key requirements and the format of problem instance. (2)Operator Descri...
[6]

Extract all n u m e r i c a l values and their context
[7]

I de nti fy what the q ues ti on asks for
[8]

Build step - by - step c a l c u l a t i o n chain
[9]

Show i n t e r m e d i a t e results e x p l i c i t l y
[10]

"" I m p l e m e n t the core problem - solving logic here

Return final n u m e r i c a l answer only # ## 2. A v a i l a b l e O p e r a t o r s & B u i l d i n g B l o c k s All o p e r a t o r s follow a c o n s i s t e n t i n t e r f a c e pattern and are i n i t i a l i z e d with the problem text . # ### ** Core O p e r a t o r s ** ** Ge ne rat e : CREATE new i n f o r m a t i o n ** - ** S i g n a t u r ...
[11]

** Problem E x t r a c t i o n ** - Uses ‘ generate ‘ to parse the problem , e x t r a c t i n g all n u m e r i c a l values , the goal , and r e l a t i o n s h i p s between q u a n t i t i e s
[12]

** Pa ral le l So lu tio n G e n e r a t i o n ** - G e n e r a t e s three i n d e p e n d e n t s o l u t i o n s using d i f f e r e n t a p p r o a c h e s : - S ta nda rd step - by - step forward c a l c u l a t i o n - A l t e r n a t i v e v e r i f i c a t i o n method - Working b a c k w a r d s from the goal
[13]

** So lut io n Re vi sio n ** - Each so lut io n is revised to catch a r i t h m e t i c errors , v e r i f y i n g each c a l c u l a t i o n step
[14]

** En sem bl e De ci sio n ** - C om par es all three v er ifi ed s o l u t i o n s to d e t e r m i n e the correct answer through ma jo ri ty vote or error a nal ys is
[15]

" " I m p l e m e n t the core problem - solving logic here

** Answer E x t r a c t i o n ** - Uses ‘ summarize ‘ to extract just the final n u m e r i c a l value ** Key Fe at ur es :** - Uses ‘ asyncio . gather ‘ for par al lel e x e c u t i o n ( g e n e r a t i n g and r evi si ng s o l u t i o n s s i m u l t a n e o u s l y ) - M ult ip le i n d e p e n d e n t s o l u t i o n s reduce the chance of s y s t ...

2022

[1] [5]

"" I m p l e m e n t the core problem - solving logic here

Return final n u m e r i c a l answer only # ### Input F o r m a t ‘‘‘ --- ** QU ES TIO N :** [ Com pl ete word problem text ] --- ‘‘‘ Mu lt ipl e p rob le ms follow the same s t r u c t u r e if pr ov ide d . Listing 4: CompleteTask Type Descriptionof GSM8K including domain overview, key requirements and the format of problem instance. (2)Operator Descri...

[2] [6]

Extract all n u m e r i c a l values and their context

[3] [7]

I de nti fy what the q ues ti on asks for

[4] [8]

Build step - by - step c a l c u l a t i o n chain

[5] [9]

Show i n t e r m e d i a t e results e x p l i c i t l y

[6] [10]

"" I m p l e m e n t the core problem - solving logic here

Return final n u m e r i c a l answer only # ## 2. A v a i l a b l e O p e r a t o r s & B u i l d i n g B l o c k s All o p e r a t o r s follow a c o n s i s t e n t i n t e r f a c e pattern and are i n i t i a l i z e d with the problem text . # ### ** Core O p e r a t o r s ** ** Ge ne rat e : CREATE new i n f o r m a t i o n ** - ** S i g n a t u r ...

[7] [11]

** Problem E x t r a c t i o n ** - Uses ‘ generate ‘ to parse the problem , e x t r a c t i n g all n u m e r i c a l values , the goal , and r e l a t i o n s h i p s between q u a n t i t i e s

[8] [12]

** Pa ral le l So lu tio n G e n e r a t i o n ** - G e n e r a t e s three i n d e p e n d e n t s o l u t i o n s using d i f f e r e n t a p p r o a c h e s : - S ta nda rd step - by - step forward c a l c u l a t i o n - A l t e r n a t i v e v e r i f i c a t i o n method - Working b a c k w a r d s from the goal

[9] [13]

** So lut io n Re vi sio n ** - Each so lut io n is revised to catch a r i t h m e t i c errors , v e r i f y i n g each c a l c u l a t i o n step

[10] [14]

** En sem bl e De ci sio n ** - C om par es all three v er ifi ed s o l u t i o n s to d e t e r m i n e the correct answer through ma jo ri ty vote or error a nal ys is

[11] [15]

" " I m p l e m e n t the core problem - solving logic here

** Answer E x t r a c t i o n ** - Uses ‘ summarize ‘ to extract just the final n u m e r i c a l value ** Key Fe at ur es :** - Uses ‘ asyncio . gather ‘ for par al lel e x e c u t i o n ( g e n e r a t i n g and r evi si ng s o l u t i o n s s i m u l t a n e o u s l y ) - M ult ip le i n d e p e n d e n t s o l u t i o n s reduce the chance of s y s t ...

2022