Part 1: No Data, No Answers: How Hyperconnect Tames LLMs

The explanation generation problem for Tinder AI-enabled Discovery is an open-ended generation task with no ground-truth data, so simple prompt tuning alone couldn't solve it. The team first had to define which attributes should be emphasized to make a good explanation, and the process of reaching consensus on that standard together was key.

Since the same information can be interpreted differently—like commonalities between users, conflicting preferences, or ambiguous inclusion relationships—the policy could easily become unstable if the PM and MLE each relied on their own intuition. So the PM first created a draft, the engineer quickly attempted model steering, and then both evaluated the results as pass/fail while leaving detailed critiques.

In subsequent meetings, they discussed cases where judgments diverged, focusing on why something was good or bad, and immediately reflected the agreed-upon standards in the policy text. Through this iteration, the team's decision boundary became increasingly clear, the policy became shorter and more precise, and the direction the LLM should follow stabilized.

This process worked well for three reasons:

• Rather than trying to create a perfect policy from the start, they failed fast and found practical standards
• Pass/fail reduced the burden of judgment compared to a 5-point scale and made consensus easier to reach
• Through critiques, individual intuition was put into words, and those standards were shared across the entire team

Once the policy converged, since relying solely on commercial LLM prompts had limitations for serving, the team moved to the stage of training a small model that internalizes the policy. After that, since humans can't directly review thousands or tens of thousands of outputs, the next stage—LLM Judge—becomes necessary. Ultimately, creating good explanations isn't a matter of building a good model, but a matter of the team fully agreeing on the standard for the desired explanation and consistently conveying that standard to the model.