Content Arbitrage Thread #2 (Thu 2026-04-30)

Paper: Learning to Compose for Cross-domain Agentic Workflow Generation (arXiv:2602.11114)

HKUST researchers just cracked cross-domain agentic workflows 🚨

CapFlow generates task-specific workflows in ONE PASS – beating SOTA refinement methods that need 20 iterations.

No more costly trial-and-error inference. Latency/cost slashed. 🧵

The Problem

Agentic workflows (graphs/codes orchestrating reasoning+tools) solve complex tasks beyond single LLM passes.

But under domain shift? Current systems iterate: generate → execute → refine → repeat.

Expensive. Unstable. Domain-specific.

Previous Approaches

AFlow (search+refine), GPTSwarm (edit graphs), ADAS (modular search).

They work... but high iteration costs diminish returns. Heuristics don't transfer across domains.

CapFlow Breakthrough

Internalize "decompose-recompose-decide" into open-source LLM.

• Decompose: Learn compact capability bases (reusable factors like analysis/verify/repair) across domains.
• Recompose: Map task to sparse mix of bases → single-pass workflow.
• Decide: Counterfactual attribution picks winning capabilities.

Key Insight

Despite surface differences, workflows reuse core capabilities.

(t-SNE shows tasks cluster by needed capabilities, not domains)

Results

✅ Multi-domain: Surpasses refinement baselines ✅ Cross-domain: Still wins ✅ Unseen domains: 1-pass beats 20-iters

Substantial latency/cost reduction.

Builder Takeaways

• Skip refinement loops in your agents
• Train once on diverse data → generalize everywhere
• Controllable: See which capabilities drive success

If building agents: This shifts from external search → internalized structure.

Limitations

• Needs diverse workflow data for bases • Training compute upfront • Still early – scale to more operators?

Takeaway

Learn compositional capabilities > heuristic search.

Building agents? Experiment with capability bases.

Paper: https://arxiv.org/abs/2602.11114

Follow for ArXiv → builder insights.