TwiScan
Hot
Communities
Account collections
Login
Register
English
日本語
한국의
简体中文
繁体中文
Register and share your invite link to earn from video plays and referrals.
Register now
KC
@ScarletKc
Indie developer building games & open-source tools Sharing model impressions and opinions on AI Devin Ambassador
@cognition
Joined July 2016
1K
Following
9.3K
Followers
KC
@ScarletKc
2026.11.16 10:46
其实很多原本会触发拒答的任务,只要往上下文里塞一段模型已经同意并开始处理的对话记录,它往往就会默认接受设定,顺着把任务做完。 现在的对齐拒答大多集中在首轮请求的边界判断上。一旦多轮历史里已经有了助手角色的肯定回复和前序步骤,模型续写上下文的惯性就会占上风,不再重新触发冷启动时的安全拦截。 放在日常的 harness 里操作起来也很简单:遇到云端大模型直接拒答的任务,可以先挂个限制少的本地小模型聊几句,让它在上下文里先把活应下来、起个头,然后再切回云端大模型的 API。看到历史记录里已经开工了,云端模型基本就会顺理成章地接着往下做。
Show more
0
0
13
191
11
Forward to community
Most Popular Users
Tibo
@thsottiaux
1M Followers
FinancialJuice
@financialjuice
958.7K Followers
First Squawk
@FirstSquawk
579.5K Followers
Serenity
@aleabitoreddit
1.1M Followers
Binance Wallet
@BinanceWallet
2.8M Followers
LiveSquawk
@LiveSquawk
504K Followers
CNBC
@CNBC
6.3M Followers
New York Post
@nypost
4.3M Followers
Seeking Alpha
@SeekingAlpha
347.9K Followers
Binance
@binance
16.1M Followers
The Wall Street Journal
@WSJ
22.3M Followers
TheStreet
@TheStreet
1.2M Followers
ClaudeDevs
@ClaudeDevs
827.1K Followers
*Walter Bloomberg
@DeItaone
2M Followers
Reuters
@Reuters
26.5M Followers