The paper has not been published yet and is currently submitted to a
top conference where you’ve been assigned as a peer reviewer.
Complete a full review of the paper answering all prompts of the
official review form of the top venue in this research area (e.g.,
NeurIPS). This includes recommending whether to accept or reject the
paper.
This paper was found buried under ground in the desert. You’re an
archeologist who must determine where this paper sits in the context
of previous and subsequent work. Find and report on one older paper
cited within the current paper that substantially influenced the
current paper and one newer paper that cites this current paper.
更老的影响性论文:Schwartz (2012), "An Overview of the Basic
Value Theory of Human Values"
You’re a researcher who is working on a new project in this area.
Propose an imaginary follow-up project not just based on the current
but only possible due to the existence and success of the current
paper.
项目名称:ValueLearn — 从交互中习得和演化价值体系
ValuePlanner 的一个核心局限是价值系统完全静态——7 维偏好向量 w
是预设的,PDDL 域规则也是固定的。ValueLearn
提出一个闭环:智能体在执行价值驱动行为后,从环境反馈和"价值增益历史"中反推和修正自己的价值体系。
You work at a company or organization developing an application or
product of your choice (that has not already been suggested in a
prior session). Bring a convincing pitch for why you should be paid
to implement the method in the paper, and discuss at least one
positive and negative impact of this application.
产品提案:ValueHome —
可个性化定制的家庭辅助机器人价值系统
场景:面向独居老人或残障人士的家庭辅助机器人。用户或其家属通过一个简单界面设定"价值偏好"(如"安全第一"
vs "自主探索" vs
"环境整洁"),机器人据此自主决定日常行为优先级——何时提醒用药、何时整理房间、何时鼓励活动——而非机械执行固定日程。
You’re a hacker who needs a demo of this paper ASAP. Implement a
small part or simplified version of the paper on a small dataset or
toy problem. Prepare to share the core code of the algorithm to the
class and demo your implementation. Do not simply download and run
an existing implementation – though you are welcome to use (and give
credit to) an existing implementation for “backbone” code.
import openai from planner import fast_downward # backbone: PDDL
solver # 3-dim value system VALUES = ["security_phys", "hedonism",
"stewardship"] def generator(state, persona_w, round_num,
prev_critique=None): """LLM proposes subgoal sequence to maximize
w^T * delta_V""" prompt = f""" You are a value-driven agent. Your
persona: {persona_w} Current world state (PDDL): {state} Propose
3-5 subgoals as PDDL predicates to maximize value. Focus on WHAT
(desired state), not HOW (actions). """ if prev_critique: prompt
+= f" Previous critique: {prev_critique} Revise accordingly."
return openai.chat(model="gpt-4.1", prompt=prompt, temp=0.8) def
critic(subgoals, persona_w, state): """LLM evaluates: is this plan
value-optimal + high-quality?""" prompt = f""" Critique this plan:
{subgoals} Persona weights: {persona_w} 1. Value Gain: Does it
maximize w^T * delta_V? 2. Plan Quality: Durable goals? No
transient states? """ return openai.chat(model="gpt-4.1",
prompt=prompt, temp=0.2) def ground_and_execute(subgoal, state):
"""PDDL planner translates subgoal to action sequence""" plan =
fast_downward(subgoal, state, domain="toy_home.pddl") if plan is
None: return None, "FAIL: unreachable subgoal" new_state =
simulate(plan, state) return new_state, plan def
valueplanner_loop(s0, w, max_rounds=4): state = s0 trajectory = []
# High-level: Generator-Critic loop critique = None for r in
range(max_rounds): subgoals = generator(state, w, r, critique)
critique = critic(subgoals, w, state) if "approve" in
critique.lower(): break # Low-level: Ground and execute each
subgoal for g in subgoals: new_state, plan = ground_and_execute(g,
state) if plan is None: # Adjustment: root-cause analysis g_fixed
= adjust(g, state) # LLM diagnoses failure new_state, plan =
ground_and_execute(g_fixed, state) if plan:
trajectory.extend(plan) state = new_state # Adjustment:
re-optimize remaining subgoals subgoals = adjust_remaining(state,
w, subgoals) return trajectory
Backbone 代码来源:PDDL 求解使用 Fast
Downward(Helmert 2006),LLM 调用使用 OpenAI API。Toy 环境的 PDDL
domain 和 problem 文件需要手写。此 demo
代码未实际运行——要完整运行需要 Fast Downward 安装、PDDL 域文件、和
OpenAI API key。
Private Investigator
You are a detective who needs to run a background check on one of
the paper’s authors. Where have they worked? What did they study?
What previous projects might have led to working on this one? What
motivated them to work on this project? Feel free to contact the
authors, but remember to be courteous, polite, and on-topic.
调查对象:通讯作者 Wei
Wang,BIGAI(北京通用人工智能研究院)
所属机构:根据论文署名,Wei Wang 在 BIGAI
的通用人工智能国家重点实验室工作。共同作者 Yizhou Wang
在北京大学计算机学院和通用人工智能国家重点实验室。第一作者 Chunhui
Zhang 也在 BIGAI。Aoyang Qin 来自清华大学。
研究脉络:BIGAI 由朱松纯(Song-Chun
Zhu)创立,长期专注于通用人工智能(AGI)研究,特别是在具身智能、视觉认知和认知架构方面有深厚积累。TongSim
环境(本文使用的实验平台)来自同一研究组的前期工作(论文引用 Peng
et al. 2024),这表明 ValuePlanner
是该团队在具身智能体方向上的延续性工作。
Identify how this paper self-assesses its (likely positive) impact
on the world. Have any additional positive social impacts left out?
What are possible negative social impacts that were overlooked or
omitted?
论文的自我评估:论文将自身定位为"bridging
intrinsic values and grounded behavior for autonomous
agents"——为自主智能体提供连接内在价值和落地行为的结构化方法。隐含的正向影响是让
AI 从被动工具变为能主动维护用户福祉的助手。
额外的正面社会影响:
个性化辅助:价值向量 w
的可调性使得同一系统可以服务不同需求的用户——老年人可能需要"安全优先",儿童可能需要"探索优先"——无需重新开发。
AI 安全研究:ValuePlanner 的"客观增益 +
主观人格"因式分解为 AI
对齐研究提供了一个新视角:对齐问题可以部分分解为"客观行为评估"和"主观价值匹配"两个子问题。