Small AI Teams Can Make Better Use of Extra Computing Power

More computing, without retraining

In Test-Time Compute Scales Better with Parallel Agent Teams, Junlin Wang, Haozhe Chen, Ang Li, Rahul Chalamala and Zhuang Liu investigate how to improve a fixed AI model by spending more computation when it works on a problem. This is called test-time scaling: extra computation during use, rather than additional training.

Their study, submitted to arXiv on November 6, 2026, is a preprint. Its findings should be treated as preliminary, not as a settled rule for organizing AI systems.

The researchers focus on tasks where candidate solutions can receive automatic scores. A verifier is the checking system that supplies those scores. That makes it possible to run several attempts and retain the highest-scoring result without asking a person to choose.

Collaboration inside, independence outside

One familiar strategy is best-of-N sampling: run N independent agents and keep the best result. Here, an agent is a model-driven system that uses tools and iteratively works on a task. Another strategy puts all available agents into one collaborating team.

The proposed method combines these approaches. With a budget of 100 agents, for example, it runs ten separate teams of ten and selects the best team's submission.

Within each team, agents have separate working environments and conversation histories but can exchange notes and solution files through a shared forum. They have no assigned roles or manager, and reading and posting are voluntary. Separate teams do not share discoveries. A team's score is its best member's verified result, not an average or a consensus answer.

The intended balance is straightforward: collaboration can improve an attempt, while multiple independent teams preserve opportunities to discover different solutions.

What the experiments found

The main experiments used Qwen3.8-27B, a model with 27 billion parameters, on 12 tasks spanning areas including groundwater modeling, machine learning, mathematical formalization and GPU kernel optimization. A GPU kernel is a program that performs a computation on a graphics processor.

Agents generally worked for 12 hours each; the browser-security task used six hours. Most tasks offered scoring during the run, but two machine-learning tasks graded the final submission only at the deadline.

At a 100-agent budget, ten teams of ten outperformed 100 independent agents on 10 of the 12 tasks. Independent agents did better on Kimi-Linear Decode, a GPU programming task, and slightly better on Active Learning, a task about choosing informative training examples.

One large improvement appeared in Carleson Formalization, where agents filled missing proofs in a Lean mathematics project. The reported score rose from 19.0 out of 100 for the best of 100 independent agents to 54.8 for ten teams of ten. That measures partial progress, not completion of the entire formalization.

Against one 100-agent team, ten smaller teams scored 8% to 30% higher on all four tasks where that comparison was run. Those four results do not establish the same advantage on the other eight tasks.

The performance curves were estimated from pools of completed runs by averaging the best score across subsets of different sizes. They were not separate, newly repeated experiments at every plotted budget.

Why several teams might help

The authors explain the results using a statistical model that separates average performance from variation between attempts. Collaboration can raise a team's average score. Sampling several teams then creates chances to select an unusually strong outcome.

Independent agents offer more separate attempts for the same agent budget, but those attempts may start from a lower average quality. One large team produces just one team-level outcome, losing the opportunity to select among independent alternatives.

This explanation is useful but approximate. It assumes independent samples with normally distributed scores—a bell-shaped distribution. It predicted the direction of the teams' advantage on nine of 12 tasks, but struggled when failures or score ceilings distorted the distributions.

Important limits

Small-team pools contained only 10 or 11 runs per task. Larger-team experiments had one to five runs per size, and the largest teams used a different forum-index display. These limits complicate conclusions about the best team size.

Results also depended on the model. With the stronger DeepSeek V4.1 Flash, teams offered less benefit on tasks individual agents already handled well.

At matched reported inference spending, Qwen teams beat five GPT-5.6-sol agents on four of eight eligible tasks. That comparison priced Qwen by GPU time and GPT by token usage; changing the accounting changed the outcome.

The practical lesson is therefore conditional: several small teams look promising for scored optimization tasks. Without a reliable verifier, choosing the best team's answer becomes another unsolved part of the problem.