ZIXUAN WANG*, 2026

<aside> 📌

Personal technical writeup, based on work done during an internship on the MiroMind RL team, written for personal learning only

</aside>

Figure 1: MiroVerse data engine. Two synthesis tracks: curated public datasets passed through a quality filter and a verifiability check, and raw data lifted into a concept graph and expanded, converge on an agent whose trajectories are kept only when they pass a verifier.

<aside> 🧭

Abstract

From large scale industry practice, a deep‑research agent's capability upper bound is mainly set by the questions it trains on, the environment(tool) it interacts with. The hard, under‑appreciated problem is synthesizing multi‑hop questions that are simultaneously: (i) genuinely multi‑hop (no single‑retrieval shortcut) (ii) verifiable (pairing checkable gold answer) (iii) difficulty‑controllable (iv) diverse and cheaply scalable

• Pillar 1: Formalization‑driven synthesis. Model the task as a formal set‑algebraic expression $\textcolor{blue}{Q}$ over Knowledge Projections on an entity-relation graph $\textcolor{blue}{\mathcal{G}}$, then realize it into language so the required reasoning is isomorphic to the intended structure.
• Pillar 2: Difficulty as uncertainty. Build a dense graph by random walks, then obfuscate entity references (replace a name with a definite description) to raise the agent's answer uncertainty. </aside>

Color convention

1. Why this matters

Open deep‑research agents have scaled along two visible axes: model size and context length. MiroMind's tech report (arXiv:2511.11793) names a third: interaction depth, up to ~600 tool calls inside a 256K‑token context, and shows accuracy rising monotonically as the tool‑call budget grows. But interaction depth is bounded by the same thing: the questions in the training set. A question solvable by one search teaches a single retrieval, no matter how long the context or how many GRPO steps you run on it. A shortcut‑solvable question never teaches a 40‑hop search. So the real lever is the data‑generation process. MiroMind's data engine (Figure 1) runs two synthesis tracks: curated public data passed through quality and verifiability filters, and raw data lifted into a concept graph and expanded. Both feed into an agent that runs full trajectories, keeping only those a Success/Fail verifier accepts. The resulting traces are overwhelmingly web‑grounded. Reading the shares off Figure 1's pie chart (which displays percentages; the integer counts below are computed from the 602,179 total tool calls):

A question worth training on must be four things at once:

• Genuinely multi‑hop.
• Verifiable.
• Difficulty‑controllable.
• Diverse and cheaply scalable.

<aside>

Principle. Properties (i)-(iv) are jointly achievable only when you fix the latent structure first (blue) and certify the item with a verifier (green); only then does the surface text (red) become a faithful encoding of that structure. Text‑first generation hopes the structure emerges; structure‑first guarantees it.

</aside>

1.1 The failure mode: passage‑to‑question

The naive baseline shows an LLM a passage and asks it to "write a hard multi‑hop question." It controls surface fluency and topical relevance, but nothing about structure or answer.

• Structure ≠ reasoning mismatch. WebShaper's central complaint is the "inconsistency between information structure and reasoning structure, question and answer." The passage's information layout need not match the reasoning the question demands.
• One‑search shortcuts. The gold answer often sits a single retrieval away: the "multi‑hop" question collapses to one hop.