Preprint reports training method that preserves correctness while shortening responses

Researchers report a training method that improves language model correctness while rewarding shorter responses. In an arXiv preprint, their Trust-Anchored Gating method achieved 0.581 pass@4 on Qwen3-4B versus 0.220 with additive reward aggregation, while shortening responses by 44% relative to correctness-only training. The method conditions auxiliary rewards on satisfying an anchor reward, preventing credit for outcomes that meet only an auxiliary goal. Experiments across Qwen and Llama models covered mathematics and program synthesis. The researchers also prove convergence toward the primary task in a simplified analytical setting when the primary reward is the anchor; that guarantee is limited to that setting.