Interaction-as-a-Judge beats code rewards: +6.7pp win rate, 35.7% less code bloat
arena · x · 2026-11-13
Arena reports that Interaction-as-a-Judge works as an RL reward without code access. Fine-tuning Qwen-3.5-9B with interaction-only reward lifts win rate over the base model by +6.7pp (+50 Arena points). Versus a code-only reward, it builds more successful applications (81.3% → 88.5%) and generates 35.7% less code bloat, with charts showing code-RM runs ballooning in code size late in training while interaction-RM stays lean.
More from Models
- Qwen 4 starts speaking Chinese mid-coding-session, user reports — gnukeith · 2026-11-13
- Dev jokes Opus 5.5's style feels like it was trained on ICLR submissions — tallinzen · 2026-11-13
- Polymarket puts 36% odds on next Claude Opus shipping this month — Polymarket · 2026-11-13
- Jacques Thibs: AI's 1.5-year capability jump came mostly from unhobbling — JacquesThibs · 2026-11-13
- Gemini 4 Hydrogen debuts at #4 on LLM-to-3D benchmark, third Google model in top 4 — DynamicWebPaige · 2026-11-13
- Power user calls out OpenAI's slow shipping: Anthropic released Opus 5.5 and Fable 5.2 — Angaisb_ · 2026-11-13