Interaction Scaling: Grounding the Third Axis of Te... | AI Research

Interaction Scaling: Grounding the Third Axis of Test-Time Compute explores a new way to improve AI performance after a model has been trained. While researchers typically try to make models smarter by having them "think" longer (reasoning) or by generating many attempts and picking the best one (sampling), these methods are limited because they rely entirely on the model's internal knowledge. This paper introduces a third approach called "interaction," where the model proposes an artifact—such as code or a visual layout—and then uses an external tool to measure its actual behavior. By feeding this real-world data back into the model, it can correct its own mistakes in ways that internal reasoning cannot.

The Power of Grounding

The core argument of the paper is that for interaction to work, it must be "grounded" on both sides of the loop. This means the feedback used to guide the model and the metric used to judge the final result must come from an instrument that actually observes the artifact's flaws. If a model tries to critique its own work without external data, or if an evaluator uses a flawed method—like a vision-language model looking at a screenshot that hides layout errors—the system fails to improve. True progress only happens when the feedback loop is connected to a tool that can objectively measure the artifact's performance or physical properties.

Breaking the Internal Ceiling

The researchers tested this approach on hard coding tasks and visual design projects. They found that reasoning-only and sampling methods eventually hit a "ceiling" where adding more compute provides no further benefit because the model is simply reshuffling information it already possesses. In contrast, interaction strategies consistently improved performance. On coding tasks, a proposer-reviewer harness achieved a perfect 100% pass rate. By receiving specific error messages from a test runner, the model could identify and fix edge-case bugs that it would never have corrected through internal thought alone.

Seeing What the Judge Misses

A major finding of the study is that standard evaluation methods can be misleading. When testing visual artifacts like slide layouts, a common practice is to have a vision-language model (VLM) look at a screenshot. However, the researchers discovered that these screenshots often crop out or obscure defects, leading the VLM to rate broken designs as "perfect." When they replaced this with a deterministic tool that measures the actual geometry of the layout, they were able to detect and remove 40% to 74% of defects. This highlights that if the metric used to score a model is not grounded in the reality of the task, it will fail to capture the actual gains made by the interaction loop.

Key Takeaways

The paper demonstrates that interaction scaling is a distinct and powerful way to boost performance, provided the system is designed correctly. The effectiveness of this approach depends on the "coverage principle": the feedback instrument must be capable of observing the specific property that is broken. Whether it is a test runner for code, a geometry engine for layouts, or a search tool for factual accuracy, the instrument must provide information that the model does not already have. Without this grounded connection, the loop cannot effectively guide the model toward a better result.