三件套没有一个可学习参数。这个设计的赌注是:前沿多模态模型已经具备足够的感知与推理能力,缺的只是把"该看的东西"以足够保真度送到眼前的机制。赌注的另一面同样重要:harness 是模型无关的机械件——修一次,当前与未来的所有模型立刻受益(论文用开源 GLM-5.3 Flash 320B
验证了这一点:官方 harness 1.89 → 完整 VISTA 66.93,见 One More
Thing)。
读分数时要记住三件事:其一,RHAE
的分子分母都是动作数,它不度量样本效率、计算成本或学习效率(VISTA
一局消耗约 30.7M
token,人类是零样本的);其二,"人类"在四个基准里的口径完全不同(首次玩家
vs 2 名新手 vs 众包中位数 vs 16 名成人);其三,所有未完成记 0
的规则意味着完成度是分数的前提——这正是为什么后面消融阶梯里"放宽动作限额"能单独提分。
评测能力与评测有效性需分开。和Agent's Last Exam所提示的评测覆盖问题相呼应,25
局满分只界定此公开样本上的表现;不代表广泛交互世界泛化,也不替代私有集、分布外及重复运行验证。
证据边界:上述跨论文链接仅用于概念对照;本文结论仍以 VISTA arXiv v1
与其报告实验为准,未做独立复现,也没有据此声称因果机制、私有集成绩或安全保证。
完整图谱
官方 arXiv v1 源码中识别到的 15 张独立 figure PDF
均在此呈现;主文已使用的图在这里重复收录,便于逐图核查。
fig_teaser.pdf:VISTA
的设计对比:从一次性视觉编码转向模型可迭代检索与重组视觉输入。
fig_trace.pdf:VISTA 在 ARC-AGI-3 与 GameWorld
中的轨迹示例。
fig_designs.pdf:VISTA
的关键设计组件及逐回合处理流程。
fig_traj_ar25.pdf:ARC-AGI-3
示例轨迹:镜像移动、目标覆盖及逐步操作。
fig_envs.pdf:评测环境示例:交互式 GameWorld、AI
GameStore 与静态 BabyVision。
fig_results.pdf:论文报告的跨环境评测结果。
fig_harness.pdf:ARC-AGI-3 上逐步增加 harness
组件的对照结果。
fig_repr.pdf:ARC-AGI-3
中文本网格与图像观测的表示对比及结果。
fig_toolsweep.pdf:ARC-AGI-3
上不同检视工具与配置的消融结果。
fig_glm.pdf:GLM-5.3 Flash 320B 在最小
harness、删减版 VISTA 与完整 VISTA 下的结果。
fig_repr3d.pdf:ARC-AGI-3 的 3D 渲染示例及 2D/3D
表现对比。
fig_memory.pdf:模型主动使用视觉记忆的行为分析。
fig_traj_bp35_l1.pdf:BP35 第 1 关的完整 VISTA
轨迹。
fig_traj_bp35_l3.pdf:BP35 第 3 关的完整 VISTA
轨迹。
fig_traj_beyond.pdf:GPT-5.6 Sol 在 ARC-AGI-3
之外任务中的四个 VISTA 轨迹示例。
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 proposes VISTA, a
training-free visual harness for interactive environments
consisting of three mechanical components: rendered-image
observations, a lossless archive of every returned frame, and
model-directed inspection tools (region zoom + numeric pixel
readout). On the 25 public ARC-AGI-3 games it reaches RHAE 100.00
with Claude Opus 5.0 (7,302 actions vs 17,135 human) and 99.00
with GPT-5.6 Sol (three runs, 99.04 ± 0.17), and it improves over
official and best-reported baselines on GameWorld, AI GameStore,
and BabyVision. The authors claim it is the first system at
perfect/near-perfect ARC-AGI-3 level without program synthesis.
Strengths. (1) Clean variable isolation: the
six-rung ablation ladder attributes 70.05 → 99.00 to the
memory/inspection channel specifically, under one model and
budget. (2) Unusually broad evidence for a systems paper: 4
benchmarks, 3 backends including a fully open-source stack
(GLM-5.3: 1.89 → 66.93), context/scale sweeps, run-to-run
repetition, per-game tables. (3) Honest self-assessment:
contamination risk and model dependence are stated by the authors
themselves; official baselines were re-run rather than merely
cited. (4) Reproducibility: MIT-licensed code and official ARC
Prize scorecards exist.
Weaknesses. (1) All headline results are on the
public 25 games; the contamination threat the authors acknowledge
is not probed (no temporal split, no private-set evaluation), so
"100.00" carries an unquantified risk of training-data overlap.
(2) Budgets are not held constant across the ladder: rungs
(i)–(ii) change stopping criteria (5× human + 12h → 2,000 actions
+ 48h); the authors disclose and partially isolate this, but
headline shorthand "13.33 → 99.00" conflates budget and channel.
(3) Human references are heterogeneous across benchmarks (486
first-time players vs 2 novice grad students vs 106 crowdworkers
vs 16 adults), so "beats humans" claims differ in strength. (4)
Token cost (30.7M/game) is reported but not scored; RHAE measures
action efficiency only. (5) No failure analysis of the one losing
game (ls20, 75.03).
Questions. Do you plan private-set evaluation,
and does the public-set 100.00 survive it? Can a contamination
probe (e.g., models released before vs after the benchmark) bound
the overlap effect? What is the failure mode on ls20? How does
performance degrade when intermediate frames are removed from the
archive (i.e., is the archive's value in final states or animation
frames)?
Rating: 7 — Accept (poster). A narrow claim
backed by dense, well-structured evidence; the main open risk
(public-set contamination) is inherited by all leaderboard entries
and explicitly acknowledged. Confidence: 4 (high;
I read the full source and reproduced the reading of all headline
numbers from the TeX tables; I did not re-run the system).
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.
更早的谱系:ReAct(Yao et al., ICLR 2023,
arXiv:2210.03629,被本文引言直接引用)。ReAct 确立了"推理 ↔ 行动"交替的文本智能体范式:thought → action →
observation 循环。VISTA
的每回合管线是这个循环的直接后代——提示词里"play
之前简述预期、之后简述可见变化"就是 ReAct 的 thought
槽位。但两者在"observation"这一半发生了决定性分叉:ReAct
的世界是文本的,观察免费且无损;VISTA
的世界是视觉的,观察有损且会被上下文管理悄悄丢弃。VISTA
的历史贡献可以概括为:把 ReAct
循环带进视觉交互世界时,发现了"observation
通道本身需要工程"——并给出了无损存档 +
主动检视的补丁。这层继承关系在论文的 related work 里被归入 agent
基础设施一节(同时引用的还有 Voyager 的终身技能库——VISTA 的
GUIDE.md 笔记与之同构)。
更晚的引用:截至
2026-10-15(检索日),未能核验到任何已发表的引用本文的论文。论文列出仅两周,这是正常的时间窗。目前能核验的"后续痕迹"只有三类:ARC
Prize 官方平台上 VISTA 两次运行的成绩记录页(GPT-5.6 Sol 99、Opus
5.0 100)、官方开源仓库(joshhhhhan/VISTA,MIT
协议)与项目博客(vista-research.github.io),以及同一 scoreboard
上的程序合成系统(Retrodict 99.86、Tycho 100.00
等——它们是同期并存者而非引用者)。我把这个空缺如实记为"谱系未闭合":VISTA
在引用网络中的位置,要等第一批以"非程序合成的视觉路线"为基线的后续工作出现才会显形。
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.
Follow-up 只在 VISTA 存在且成功的前提下才可能:(1) VISTA
提供了接口语义(帧存档 + 区域检视 +
数值读出)与"通道修复足够"的成功证据,使内化目标有明确定义;(2)
VISTA 的运行日志提供了检视策略的示教数据(无检视的对照 13.27
也现成);(3) VISTA 的机械 harness
提供了同一环境下的可回滚对照面。做法:把 inspect/read_pixels
调用序列作为轨迹的一部分,微调多模态模型在推理时原生地选择"回看哪帧、看哪个区域、放多大"——把显式工具调用变成隐式注意力偏置。评价指标在
VISTA 的坐标系内定义:同分下的 token 预算(目标:10×
压缩)、弱模型上的增益(GLM 档从 66.93
起步)、以及私有集上的分数(这是 VISTA 自己没能做的检验)。
诚实的风险声明:检视策略可能不可迁移(它是模型特定的问题求解副产品,不是通用技能);内化后失去显式
inspect 日志意味着失去 VISTA
附带的可审计性——那个收益要另想办法保住。
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.
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?