Solving Quantum Science Problems with SMART: A Self-evolving Multi-Agent Research Tree

Note: the scheduled speaker, Andrew Schoen (Partner, NEA), couldn't attend; Mingu Kang of UC Berkeley presented on his behalf, describing it as joint work with Daniel Lee and Andrew Schoen. The slides didn't come up until well past the halfway point, so the first stretch was delivered from memory.

TL;DR

• Why start with quantum science: it's the "strategic Everest" — deeply multidisciplinary (math, physics, CS, engineering, materials), demanding of extreme rigor (get any one piece wrong and the experiment simply doesn't work), and huge in quality delta (high-quality research is rare and experts are acutely wary of AI slop). Climb that, and the system should carry over to any domain.
• SMART = Self-evolving Multi-Agent Research Tree: each node is a planner / worker / verifier triple; verifier output feeds a tree planner agent that designs the next step's workflow tree. Nodes are goals or tasks, and at each step tasks can be split, merged, selected, or added.
• Scorecard: on OpenAI's Frontier Science reasoning benchmark, their alpha prototype hit the highest accuracy at one-fifth the cost of the second-best system (codex 5.5) — cheap because GLM 5.1 is the underlying model, accurate because of the agent orchestration. It defines the Pareto frontier on the accuracy-versus-cost plane.
• Real problem: qubit allocation and routing on neutral-atom quantum computers. Across four benchmark circuits, SMART's solutions cut rearrangement time by 25% to 83%. Ablations show different circuits need different strategy combinations (a graph-state circuit gets −83% from circuit structure alone; a QPE exact circuit gets −25% from parameter tuning plus an A* redesign, with structure contributing nothing) — and that complementarity is the argument for a multi-agent framework.

Key Points

Why quantum first: the handoffs are the real problem (~00:48–00:55)

The goal is an agentic system that can orchestrate any frontier research in science, starting with quantum.

The motivation comes from his own experience. Quantum research is a genuinely multidisciplinary stack — physics, chemistry, math, hard engineering, materials. Ideally one person would know the whole stack and connect ideas across layers to co-design end to end, which would be the optimal way to build a quantum computer. But resources are always short and, more importantly, the interfacing is the real problem: one researcher is an expert in one layer, another in a different layer, and the handoffs are genuinely hard. That's where he sees AI's promise — an AI holding the context of every layer and facilitating the handoffs.

He frames scientific discovery as three pillars:

1. Intelligence — synthesize knowledge, run the plan / work / verify sequence.
2. Computation — simulate large systems, optimize over vast solution spaces.
3. Validation — run real experiments, test predictions against reality.

The vision is an agentic system that closes the loop across all three.

Quantum science is the strategic Everest for three reasons: it's very challenging and multidisciplinary; it demands extreme rigor — unless every single piece is correct, the experiment simply won't work; and the quality delta is massive — high-quality research is rare and experts are very concerned about AI slop. Climb that, and they believe the system can lead discoveries anywhere.

His one-slide quantum refresher: a bit is 0 or 1, like a coin flip, while a qubit can be in a superposition anywhere on a sphere; quantum physics gives access to qubits (e.g. two stable electronic energy levels of an atom); and a quantum hardware platform gives full control over that physics — in this case an array of atoms manipulated with lasers.

SMART's three design principles and tree architecture (~00:55–00:58)

Their premise is that transformative discovery will come from human–AI collaboration, which yields three design principles:

1. The workflow tree keeps getting updated as the agents explore.
2. The system is structured but also flexible.
3. Agents are autonomous but humans can intervene at any time.

That requires something different from the "loop engineering" widely used for tasks far simpler than scientific discovery.

Hence SMART (Self-evolving Multi-Agent Research Tree): a tree whose nodes each consist of planner, worker, and verifier agents, with verifier outputs feeding a tree planner agent that designs the next step's workflow tree. A node can be a goal or a task, and at the next step a task can be split, merged, selected, or added.

On the interface, he made the case for human intervention concretely — "in research, unpredictable things always happen, just like slides not appearing or the laptop crashing." A second interface lets researchers probe what agents produced, with context management, model choice, and cost tracking as the key features.

Benchmark: plenty of agentic systems for scientific discovery exist; SMART's advantage is being both steerable and adaptive, which follows directly from the structured-plus-flexible design. They compared their alpha prototype against frontier models and systems on Frontier Science, a scientific reasoning benchmark from OpenAI. The prototype achieved the highest accuracy at one-fifth the cost of the second-best system (codex 5.5) — low cost from using GLM 5.1 underneath, high accuracy from orchestrating the agents. Plotted as accuracy versus cost, SMART defines the Pareto frontier — and this is only the prototype.

The real problem: speeding up neutral-atom quantum computers (~00:58–01:03)

The setup: quantum computing is about computing faster, so the target is wall-clock time — specifically, executing quantum circuits faster on neutral atoms.

In a circuit, horizontal lines are qubits with quantum gates applied; the blue entangling gates are performed by bringing two qubit atoms close together inside an entanglement zone. So running a circuit on a zoned-architecture neutral-atom machine means first allocating qubit atoms to slots in the storage zone, then shuttling atoms in and out of the entanglement zone to perform entangling gates.

The catch: atoms can't be moved arbitrarily. Because lasers do the moving, constraints apply — for instance, the rows and columns defining one block of atoms cannot cross the rows and columns of another block.

The problem: given a quantum circuit and a zoned architecture, what is the optimal allocation and routing of qubit atoms that minimizes rearrangement time while satisfying the movement constraints? The solution space is vast and the answer genuinely matters for neutral-atom machines — which is why they picked it as SMART's first target.

Results: across four benchmark circuits (lower is better), SMART's solutions achieved 83% to 25% lower rearrangement time than the state of the art.

The ablations are the interesting part. Multiple agents explored different strategies, and different circuits were improved by different combinations of them. For the graph-state circuit on the left, rearrangement time dropped 83% by using the structure of the circuit alone. For the QPE exact circuit on the right, the 25% reduction came from parameter tuning combined with an A* redesign, and using structure didn't help at all. His conclusion: this complementarity across strategies is exactly what makes the multi-agent framework powerful.

Procedurally, agents explore different strategies and then tasks are split, merged, and so on until a better solution emerges. He also showed a simulation video of the rearrangement: SMART's solution clearly exploits some structure while the state-of-the-art baseline doesn't, so the baseline takes far longer.

Next: quantum science has other tremendously interesting and difficult problems for SMART — error correction among them — as the proving ground for agents that lead to transformative discovery across frontier domains.

Quotes

"Quantum science demands extreme rigor. Unless every single piece is correct, the quantum experiments simply would not work." (~00:53)

Why quantum is the honest test: it doesn't accept AI slop.

"So once we climb the strategic Everest, then we believe that the agentic system will be capable of leading scientific discoveries in any domain." (~00:54:46)

Pick the hardest mountain first, not the most climbable one.

"…different circuits are improved by different combinations of strategies. … So really this complementarity of different strategies is what makes the multi-agent framework so powerful." (~01:00)

Multi-agent isn't about more compute — it's about strategies that complement each other.