New AI Method Scales Long-Horizon Agent Tasks by Intelligently Merging Parallel Attempts

Parallel test-time scaling—running multiple inference attempts and combining their results—has become a proven strategy for improving AI reasoning. Yet when applied to agentic tasks that involve long chains of interactions and tool use, simple aggregation methods break down. Researchers have now developed a smarter approach to handle this challenge.

The Problem with Existing Methods#

Standard aggregation works well for brief reasoning tasks: you generate several chain-of-thought responses, pick the best answer, and you're done. Agentic tasks, however, are fundamentally different. These involve agents interacting over many turns, calling external tools, and producing outputs that rarely have a single "correct" answer. Throwing away all but the final response wastes valuable intermediate information, while concatenating every trajectory exceeds what most models can process in one go.

How AggAgent Works#

The proposed solution treats parallel agent trajectories as an environment itself. AggAgent functions as a meta-agent equipped with lightweight tools to inspect candidate solutions and search across the collection of trajectories generated from parallel runs. Rather than making a blind choice, this aggregation agent actively navigates the space of attempted solutions, synthesizing insights on demand. This allows it to extract and combine relevant information from multiple attempts in a way that fits within computational constraints.

Empirical Results#

The technique was tested across six benchmarks using three large language model families: GLM-4.7, Qwen3.5, and MiniMax-M.5. AggAgent outperformed all existing aggregation methods by up to 5.3% absolute on average, and achieved gains as high as 10.3% on two deep research tasks. Critically, the added computational cost remains bounded—the aggregation overhead stays within the expense of a single additional agentic rollout, making the approach practical at scale.

Implications#

These results suggest that agentic aggregation represents an efficient and cost-effective path to improving parallel test-time scaling. By acknowledging the unique structure of long-horizon tasks and designing aggregation around it, the work opens a new direction for how AI systems can leverage multiple attempts to tackle complex, multi-step problems.