Event ·Single source

Lean 形式化验证的分词器管线原型：消除训练与推理的消息不一致

Sasha Rush 提出 Lean Verified Tokenizers，针对 RL 训练中训练端与推理端消息构建不一致导致运行失败的问题。目前聊天模板通常用 Jinja 构建，缺少形式化保证。该项目用 Lean 对 Rust 实现的 agent 管线进行形式化验证，保证消息处理逻辑与训练时一致。目前为原型阶段，代码已开源。

Current read

RL 跑崩很多时候是 Jinja 模板在训练和推理时不一致，Sasha Rush 用 Lean 验证 Rust 管线来根治这个问题，做 RL 的人可以看看。

2 sources31° AI heatLast updated 2026/10/27 21:42:13

Evidence

2 sources

Primary sourceSasha Rush

2026/10/27 21:42:13View source

Related source 1arXiv cs.AI

2026/10/28 02:23:25View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story