mirror of
https://github.com/deepseek-ai/awesome-deepseek-agent.git
synced 2026-11-05 01:07:37 +00:00
[GH-ISSUE #324] Definition-first agent pipeline for Chinese LLMs — sharing a framework + benchmark #108
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/awesome-deepseek-agent#108
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @llxpy on GitHub (Jul 31, 2026).
Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/324
恭喜 V4-Flash 正式版上线,正好借这个时机分享一个 Agent 方向的思路。
独立研究者,做了一个定义前置(definition-first)的 Agent 推理流水线 MoLock:
推理之前先让模型生成两份文本——"经链"(解释用户意图)+ "说链三锁"(词义锁/边界锁/规则锁),两份高歧义文本互相锁定,逼出单份文本给不出的精确。
理论依据:中文是"弱文本"语言(词边界不在书写层),精确性可以从人为纪律注入,不必等语法。墨家"经+说"双链是两千年前在松散文言上保精确的史前实践。
初步实测(100 题中文语义测试,DeepSeek-V4-Flash):裸调幻觉率 17%,CoT 11%,英文链路 13%,定义前置 8%;均分 4.74 为四组最高。
论文 + 完整实验方案 DPB-WS(评分权交给跑者,欢迎任何人复现/反驳):
https://github.com/llxpy/-MoLock/releases
仅作思路分享。祝越做越好。