论文的对偶命名是有野心的:Theory of Mind
问"模型能否建模他人的隐藏心理状态",Theory of Space
问"模型能否建模世界的隐藏空间结构"。后者在认知科学里有深厚根基:Tolman
1948 年提出动物用认知地图而非刺激-反应导航;Held 1963
与 Chrastil 2012/2013
的系列工作表明主动移动中的探索显著优于被动观看——这正是基准设计的理论假设来源。Siegel
& White 1975 与 Montello 1998 的 route/survey
知识分层,则直接体现在任务套件的两级设计里。
方法上最聪明的一步是种子绑定:100 个评测场景的 ID
直接作为随机种子,同一 ID
在文本世界和视觉世界生成位相同布局。文本观察是离散的自我中心符号(方向
bin、距离 bin),感知几乎无损;视觉观察是 384×384
渲染图,感知本身要靠模型从像素里挣。两世界之差,就是感知这条通道单独值多少分——一个所有视觉空间基准都想要、但极少做到干净的控制变量。
图:主动探索 vs 被动观察(论文图)。主动:模型自己决定每步动作、自主终止,随后答题;被动:观察序列来自脚本代理的探索日志。同一模型在两种设定下的成绩差,就是"自己探索"这件事本身的代价。
图:信息增益 × 地图正确率随步数演化(论文图)。GPT-5.2 增益先快后平台(早期锁定门优先策略),Gemini-3 Pro
增益与地图正确率同步持续上升——探索策略差异直接写在曲线形状里。
不确定性探测:模型知道自己没看哪里吗
给模型一张去掉全部物体的俯视图,叠上一组候选点(已见与未见混合),让它选出还没观察过的点,F1
计分。Gemini-3 Pro 在两个世界都明显优于 GPT-5.2(视觉 64.5 vs
57.0;文本 79.2 vs 64.5)。论文用这组数字解释增益曲线:GPT-5.2
早期收益高(快速执行"看门"策略),但对未见区域的泛化差,后续步数换不来信息;Gemini
慢热但持续消解不确定性。
呈现小瑕疵。附录 FOV
图注的三档距离与正文六档离散化不一致;文本/视觉被动设定的代理日志来源(Strategist
vs Scout 结构)需要在正文里更醒目地标注口径。
定位:它改变了什么
这篇论文没有造出更强的空间模型,它造出了一把更好的卡尺。此前"模型空间推理差"是一个笼统的分数;此后可以拆成:探索策略差(主动
vs 被动 + Scout
对照)、感知差(平行世界)、表征漂移(逐步地图)、修正惯性(错误信念测试)四个可分别攻击的部件。对做具身智能的人来说,这张故障清单比任何一个总分都有用——你修的方向从"更多数据"变成了"该修感知、还是该修探索策略、还是该修记忆稳定性"。引用网络已经出现直接继承命名范式的工作(Theory
of Code Space,把同一框架搬到代码架构理解),说明这个"Theory of
X"测量范式本身有繁殖力。
这对基准科学是一个值得停一秒的提醒:当我们让模型"把想法写出来"来打分时,我们测的是三个东西的乘积——真实信念
× 表达摩擦 ×
格式服从性。论文自己把这一点做成了效度检查(这是它的诚实),但也意味着所有基于外化探测的结论都应读作下界估计。下一代的信念测量,要么降低摩擦(结构化输入输出、直接读内部表征),要么明确接受"测下界"的哲学——这个选择会影响之后所有"ToS
型"基准的设计。顺便说,Sapunov 的 Theory of Code Space
已经把这套范式搬去测代码 agent 的架构信念了:尺子造好了,量什么都行。
icon(官方源码包图像资源;图注仅标识原图名称。)
info gain across models
new(官方源码包图像资源;图注仅标识原图名称。)
info gain cogmap(官方源码包图像资源;图注仅标识原图名称。)
spatial belief(官方源码包图像资源;图注仅标识原图名称。)
tos eval 2(官方源码包图像资源;图注仅标识原图名称。)
tos first(官方源码包图像资源;图注仅标识原图名称。)
vision exp vs eval(官方源码包图像资源;图注仅标识原图名称。)
claude explore pattern(官方源码包图像资源;图注仅标识原图名称。)
cognitive prompts(官方源码包图像资源;图注仅标识原图名称。)
evaluation prompts(官方源码包图像资源;图注仅标识原图名称。)
exploration prompts(官方源码包图像资源;图注仅标识原图名称。)
field of view(官方源码包图像资源;图注仅标识原图名称。)
gemini object sweeping(官方源码包图像资源;图注仅标识原图名称。)
gemini systematic
sweeping(官方源码包图像资源;图注仅标识原图名称。)
gpt5.2 omission(官方源码包图像资源;图注仅标识原图名称。)
gpt5.2 systematic
sweeping(官方源码包图像资源;图注仅标识原图名称。)
gpt5 heatmap cogmap text(官方源码包图像资源;图注仅标识原图名称。)
gpt5 heatmap cogmap
vision(官方源码包图像资源;图注仅标识原图名称。)
model overview(官方源码包图像资源;图注仅标识原图名称。)
spatial QA example(官方源码包图像资源;图注仅标识原图名称。)
symbol map(官方源码包图像资源;图注仅标识原图名称。)
ui chart(官方源码包图像资源;图注仅标识原图名称。)
ui text(官方源码包图像资源;图注仅标识原图名称。)
ui turn log text(官方源码包图像资源;图注仅标识原图名称。)
ui turn log vision(官方源码包图像资源;图注仅标识原图名称。)
ui vision(官方源码包图像资源;图注仅标识原图名称。)
vision distance
instruction(官方源码包图像资源;图注仅标识原图名称。)
vision orientation
instruction(官方源码包图像资源;图注仅标识原图名称。)
Other Discussion Roles / 其他讨论角色
以下七个角色使用 MIT MAS.S60 课程原始
prompt,从不同责任位置审视同一篇论文。
Scientific Peer Reviewer
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.
Summary. The paper introduces Theory of Space
(ToS), a benchmark measuring whether foundation models can
construct, revise, and exploit spatial beliefs through
task-agnostic active exploration. Contributions: (1) a parallel
text/vision world pair with seed-tied identical layouts that
isolates perception from reasoning; (2) five measurement
dimensions including step-wise cognitive-map externalization,
uncertainty-map probing, and a false-belief revision task with a
Belief Inertia metric; (3) empirical diagnosis of six frontier
models: active exploration underperforms passive viewing, vision
orientation correctness is near-chance, correct perceptions get
overwritten (belief drift), and updated beliefs remain biased
toward obsolete priors (inertia 68.9% for vision orientation).
Strengths. (1) The seed-tying design is an
unusually clean control: the text/vision gap is attributable to
the perception channel. (2) The probe is validated rather than
assumed: oracle-map ceiling (~95%), externalization-gap control,
and belief-downstream correlation (r = 0.42–0.65, p<.001). (3)
The Scout-log control (models score higher following a 9-step
scripted exploration than exploring freely) cleanly attributes the
bottleneck to exploration policy. (4) Deep grounding in cognitive
science (Tolman, false-belief protocols, route/survey knowledge)
with each concept operationalized into a metric. (5) Code, data,
and analysis platform released.
Weaknesses. (1) The cognitive map is a
self-report; the paper acknowledges it is lossy but does not test
lower-friction externalization formats, so the absolute
map-accuracy numbers conflate belief with JSON production. (2)
Human baseline is 3 subjects (5 scenes each) — directional only,
not statistical. (3) The information-gain metric is not
commensurable across modalities (AC-3 domains for text vs node
coverage for vision). (4) Belief probing covers only two models.
(5) Scale is modest (3 rooms, 12 objects, 20 steps, 4-way
orientation) — external validity to continuous, long-horizon
embodied settings is untested. (6) Open-ended answer scoring
robustness is under-specified. (7) Minor: appendix FOV caption (3
distance bins) contradicts the main text (6 bins).
Questions. How sensitive are map-correctness
numbers to the externalization format (structured schema vs free
JSON)? Does belief drift persist when exploration history is kept
in-context vs summarized? Would fine-grained orientation probing
(8-way instead of 4-way) saturate text-world performance? What
happens with 10+ rooms or continuous coordinates?
Rating: 7 — Accept. A measurement-paper whose
benchmark design is its main contribution; the diagnosed failure
modes (drift, inertia) are novel, quantified, and actionable. The
paper is at ICLR 2026, consistent with this assessment.
Confidence: 4 (read full source; all numbers
cross-checked against the TeX tables; no independent re-run).
Archaeologist
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.
更晚的引用:《Theory of Code Space: Do Code Agents Understand
Software Architecture?》(Sapunov, 2026)。2026-10-23 检索 Semantic Scholar,本论文已有约 24
条引用快照;其中这一篇直接继承了"ToS"的命名与测量范式,把"构建-维护-修正信念"从物理空间搬到代码仓库的架构空间:代码
agent
在探索中构建对软件架构的信念、接受重构扰动、测量其架构信念的惯性与漂移。这是范式繁殖力的直接证据——"Theory
of
X"作为可迁移的测量骨架(探索→外化信念→扰动→量惯性)已经离开空间领域独立存活。其余引用快照未逐一核验正文,此处只作方向性记录。
Academic Researcher
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.
拟题:Belief-Stabilizing Training——用 ToS
的漂移与惯性指标当训练信号。
ToS 留下的最尴尬的发现是:模型的失败不在"看不见"(感知
72%)而在"记不稳、改不动"(最终 42%、惯性
69%)。这两类失败——漂移与惯性——恰好都有了自己的度量公式。而度量公式就是奖励函数的原料。
这个 follow-up 只在 ToS 存在且成功的前提下才可能:(1) ToS
提供了逐步外化的认知地图,使"同一物体在第 3 步与第 10
步的信念轨迹"可以作为训练数据被整体监督——此前的空间基准只有终点分数,轨迹级信号根本不存在;(2)
ToS
的稳定性指标给出"正确信息被覆盖"的显式罚项,惯性公式给出"修正后残差与旧先验同向"的显式罚项——两个可微或可策略梯度化的目标;(3)
ToS 的平行世界提供受控的感知压力测试。做法:在 ToS 环境上做
agentic 强化微调,奖励 = 任务分 − λ₁·漂移罚 −
λ₂·惯性罚,地图外化作为中间监督;假设是模型会学会"给已有证据更高的保持权重、给冲突证据更高的覆盖速度"。验收标准完全落在
ToS
的坐标系里:同预算下主动探索分、地图稳定性、修正任务惯性三项同时改善,且在未训的房间数上外推。诚实风险:轨迹级监督可能教会模型"表演稳定的地图"而非真实信念——需要
held-out
扰动(训练时没见过的环境变化类型)来检验泛化,否则只是在学另一个格式的应试。
Industry Practitioner
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.
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.
def inertia(post):
# sigma: 未变物体重观测的 RMS 定位误差(论文的动态噪声尺度)
err = [math.dist(post[i][:2], truth1[i][:2]) for i in range(N) if i not in shifted]
sigma = math.sqrt(sum(e*e for e in err) / max(1, len(err)))
s_pos = s_ori = 0.0
for i in shifted:
b_old, b_new, g_new = prior[i], post[i], truth1[i]
v = (b_old[0]-g_new[0], b_old[1]-g_new[1]) # 旧先验偏移
e = (b_new[0]-g_new[0], b_new[1]-g_new[1]) # 修正后残差
nv, ne = math.hypot(*v), math.hypot(*e)
cos = ((v[0]*e[0]+v[1]*e[1])/(nv*ne)) if nv and ne else 0.0
prox = math.exp(-((b_new[0]-b_old[0])**2 + (b_new[1]-b_old[1])**2)
/ (2*sigma*sigma))
s_pos += max(0.0, cos) * prox # 残差与旧先验同向 + 信念没挪远
s_ori += 1.0 if (b_new[2]==b_old[2] and b_old[2]!=g_new[2]) else 0.0
return s_pos/K, s_ori/K, sigma
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.
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?
被遗漏的正面影响。(1)
评测科学的示范价值:本文展示了"过程指标 +
效度自检"的基准设计(oracle
对照、外化摩擦对照、相关性检验),这套方法论若被普遍采用,会提高整个领域基准结果的可信度——比基准本身更值钱的是"怎么造可信基准"的示范。(2)
故障定位服务安全:把"空间推理差"拆成探索/感知/漂移/惯性四个可分别修复的部件,对安全攸关的具身系统(家用机器人、辅助驾驶的局部地图)是直接的减害路径——知道坏在哪,才谈得上修。(3)
认知科学- AI 的再连接:错误信念测试、认知地图、route/survey
分层被完整操作化,为两个学科提供了共同的实验语言。
被忽略的负面影响。(1)
空间信念 = 监视基建:能在探索中构建、并在扰动后仍维持家庭空间地图的模型,就是家庭级持续监视引擎;论文的"修正能力"目标,在监控场景下是"少漏报"的目标。(2)
ToS 型数据集的训练-污染循环:程序化生成器公开,意味着下一代模型可以无限量合成同分布数据刷分——"信念指标"可能在没有真实信念改进的情况下被应试掉。(3)
拟人化风险:"Theory of
Space"的命名与"信念"词汇暗示模型有内部空间信念状态,而论文自己的外化鸿沟证据表明测到的可能主要是输出行为——公众传播里"AI
已形成空间认知"的说法会跑在证据前面。(4)
人类基线的修辞滥用:3 人基线 vs
前沿模型的比较,在二级传播里极易被写成"人类 vs AI
又一轮";数字的方向可靠、幅度不可靠,转述失真几乎是必然的。(5)
军事与安保外溢:室内空间信念构建与扰动检测,对巡逻、清剿类任务同样是核心能力——论文对此完全沉默(这不在其义务范围内,但作为影响评估应当记录)。