Middle Layer releases The Trust Table benchmark

middlelayer.substack.com5 min read

The author created a game-based benchmark called The Trust Table to test whether AI models know when to answer, when to pass, and when to judge other models accurately. Using reinforcement learning on Tinker with LoRA, training improved Qwen's ability to selectively pass on questions it would have answered incorrectly and improved Inkling's true and false judging accuracy. The author argues that product managers can shape and test model behavior by aligning training rewards with product requirements. Open article →