Round-Trip Consistency: Bidirectional Diffusion Models Can Predict Their Own Rollout Errors

• 类型：arxiv
• 标识：2608.00675
• 链接：https://arxiv.org/abs/2608.00675
• 主分类：multimodal
• 形态：application
• 被引：1
• 被引来源：Semantic Scholar
• S2被引：1
• 影响力被引：0
• TLDR：Round-trip consistency turns reversibility into a practical trust signal for generative models, and Bidirectional training comes at negative cost, beating direction specialists in both directions, and the backward direction doubles as a fast inverse solver.
• 待LLM分类：否
• 标题中文：往返一致性：双向扩散模型可预测自身的 rollout 误差
• TLDR中文：往返一致性将可逆性转化为生成式模型一种实用的可信信号；双向训练带来负成本，在两个方向上均优于单向专家模型；其中反向还可作为快速的逆问题求解器。
• 来源文件：
• /inbox/tom/_candidates/2026-08-10-agent-rag-longcontext-candidates.json
• [S2 enrich]