Hi! I have two questions and would greatly appreciate it if you could respond.
[Question 1] How personality differentiation is achieved and maintained.
From my understanding:
- Each agent has a detailed character profile (10 personality traits, 9 talents, values, background story, etc.) injected into the system prompt.
- The self-referential advantage ensures diverse personality types can be selected for training.
- But the SFT is performed on a single set of model weights using trajectories from all selected agents.
My question: If we strip the character profile from the system prompt and give only a generic instruction (e.g., "You are a person in a simulated world") to the model, does it still exhibit distinct personality behaviors? Or does it default to a generic "socially competent agent" persona?
In other words, is the personality differentiation:
- (A) Engineered entirely through prompt-level character profiles (the model acts as a universal roleplayer),
- or
- (B) Partially embedded in the model weights through the training process (the model has learned behavioral tendencies that persist even without profile conditioning)?
This distinction would greatly help my novel simulation project,
where I'm trying to break free from the "general-purpose model with
a prompt wrapper" paradigm and build characters with their own
inherent behavioral patterns.
[Question 2] Feasibility of low-cost local training
Assuming we use a local general-purpose model (7B-14B range) without access to large-scale compute, is it feasible to train a "barely passable" character model using a similar rejection-sampling + SFT approach as Agentopia, at a reasonable cost?
I'm new to AI training and don't yet have a clear sense of the cost-performance curve in this direction, so I'd appreciate any insight into the feasibility and entry barriers.
Thank you for your work!
您好！有两个问题想请教一下，如果能回复的话我会非常感谢。
【问题 1】人格差异化的本质
我的理解是：
- 每个 Agent 有详细的角色档案（10 维人格特质、9 维天赋、价值观、背景故事等），通过 system prompt 注入。
- 自参照优势机制确保了不同人格类型的 Agent 都有机会被选入训练集。
- 但 SFT 是在单一共享的模型权重上，用所有被选中的 Agent 的轨迹数据联合微调的。
我的问题是：如果从 system prompt 中移除角色档案，只给一个通用指令（比如"你是模拟世界中的一个人"），模型是否还能表现出具体的人格行为？还是会退化为一个通用的"社会适应性强的 Agent"人格？
换句话说，人格差异化是：
(A) 完全通过 prompt 的角色档案工程实现的（模型扮演的是通用角色演员），
还是
(B) 通过训练过程部分嵌入到了模型权重中（即使去掉角色档案，模型也学到了某些持久的行为倾向）？
这个区分对我正在做的小说模拟项目有很大帮助——我正在尝试让 AI 角色摆脱 "通用模型套 prompt" 的局限，真正拥有属于自己的行为模式。
【问题 2】低成本本地训练的可行性
如果假设我们用的是本地通用模型（比如 7B-14B 规模），普通人没有大规模算力资源，是否有可能通过类似 Agentopia 的拒绝采样 + SFT方案，以可承受的成本训练出一个"勉强及格"的角色模型？
我是 AI 领域的新手，对训练方面的投入产出比还没有清晰的概念，所以想了解一下这个方向的可行性和门槛在哪里。
非常感谢！
Hi! I have two questions and would greatly appreciate it if you could respond.
[Question 1] How personality differentiation is achieved and maintained.
From my understanding:
My question: If we strip the character profile from the system prompt and give only a generic instruction (e.g., "You are a person in a simulated world") to the model, does it still exhibit distinct personality behaviors? Or does it default to a generic "socially competent agent" persona?
In other words, is the personality differentiation:
This distinction would greatly help my novel simulation project,
where I'm trying to break free from the "general-purpose model with
a prompt wrapper" paradigm and build characters with their own
inherent behavioral patterns.
[Question 2] Feasibility of low-cost local training
Assuming we use a local general-purpose model (7B-14B range) without access to large-scale compute, is it feasible to train a "barely passable" character model using a similar rejection-sampling + SFT approach as Agentopia, at a reasonable cost?
I'm new to AI training and don't yet have a clear sense of the cost-performance curve in this direction, so I'd appreciate any insight into the feasibility and entry barriers.
Thank you for your work!
您好！有两个问题想请教一下，如果能回复的话我会非常感谢。
【问题 1】人格差异化的本质
我的理解是：
我的问题是：如果从 system prompt 中移除角色档案，只给一个通用指令（比如"你是模拟世界中的一个人"），模型是否还能表现出具体的人格行为？还是会退化为一个通用的"社会适应性强的 Agent"人格？
换句话说，人格差异化是：
(A) 完全通过 prompt 的角色档案工程实现的（模型扮演的是通用角色演员），
还是
(B) 通过训练过程部分嵌入到了模型权重中（即使去掉角色档案，模型也学到了某些持久的行为倾向）？
这个区分对我正在做的小说模拟项目有很大帮助——我正在尝试让 AI 角色摆脱 "通用模型套 prompt" 的局限，真正拥有属于自己的行为模式。
【问题 2】低成本本地训练的可行性
如果假设我们用的是本地通用模型（比如 7B-14B 规模），普通人没有大规模算力资源，是否有可能通过类似 Agentopia 的拒绝采样 + SFT方案，以可承受的成本训练出一个"勉强及格"的角色模型？
我是 AI 领域的新手，对训练方面的投入产出比还没有清晰的概念，所以想了解一下这个方向的可行性和门槛在哪里。
非常感谢！