Interaction-as-a-Judge works as an RL reward without needing access to code. We fine-tune Qwen-3.5-9B with an interaction-only reward and it improves over the base model by +6.7pp win rate (~+50 Arena points). Compared to a code-only reward, it builds more successful applications (81.3% to 88.5%) and generates 35.7% less code bloat.
Interaction-as-a-Judge works as an RL reward without needing access to code. We fine-tune Qwen-3.5-9B with an interaction-only reward and…
Interaction-as-a-Judge works as an RL reward without needing access to code. We fine-tune Qwen-3.5-9B with an interaction-only reward and it improves over the base model by +6.7pp win rate (~+50 Arena points).
Arena · X
Content notes
Original publication and rights belong to the source.