← Back to Homepage

From Draft to Strong-Accept: How a Self-Play Survey Hit 8.6

从初稿到 Strong-Accept:一篇自我博弈综述如何被推到 8.6

An autonomous framework took a survey from a rough draft to a 285B-experiment-backed paper across 16 review rounds

一个自主框架如何用 16 轮评审,把综述从粗糙初稿推成有 285B 实验支撑的论文

Deli Chen June 17, 2026 6 min read

This is the story of Paper #4 in my AutoResearch project: Self-Play in the Age of Foundation Models — A Comprehensive Survey from Game-Theoretic Foundations to Open-Ended Learning. Unlike the first three surveys, this one was not just written by the agent framework — it was argued with, stress-tested, and experimentally validated by it over 16 review rounds, ending at a median peer-review score of 8.6/10 (strong accept).

Disclaimer: This is a personal hobby project. All opinions expressed are my own and do not represent any organization.

8.6
Peak Median
16
Review Rounds
75
Pages
217
Citations
285B
RL Model
12
RL Runs

The Score Trajectory: Not a Straight Line

What makes this paper interesting is that its score did not climb monotonically. The agent ran a five-persona peer review after each revision, and the median moved up and down honestly:

VersionScoreWhat happened
V4–V107.0 → 8.4Drafting: taxonomy, sections, and three original theorems. Citations grew from 0 to 207.
V118.5The 285B GRPO verifier-noise experiment landed in §8 — the paper turned from a pure survey into one with an original large-scale experiment.
V128.2 ↓The only decline. An external citation check found three defective references; the score was lowered to the paper's true state, not preserved.
V138.4A 2000-step long-horizon run (8.3× the original length) tested the KL-buffering hypothesis; data provenance was corrected.
V148.4Seed replication + KL ablation turned "buffering" from a hypothesis into evidence. Bottleneck shifted to theoretical rigor.
V158.5Held-out KL-endpoint study; parallel work started on four theoretical sub-problems.
V168.6 ✓Theory hardening landed — series best.
Why V12 is the most important version
When the first external citation check found three submission-blocking references — one nonexistent paper, one with the wrong authors, and one whose arXiv ID pointed to an unrelated paper — the framework lowered its own score from 8.5 to 8.2 rather than quietly keeping the higher number. An autonomous pipeline that will mark its own work down when the evidence demands it is far more trustworthy than one that only ever climbs. All three were fixed in the next iteration.

The 285B RL Experiment: The Core Differentiator

The key decision was to validate the survey's central claim on a real 285B-parameter model (DeepSeek-V4) rather than a small-model demonstration. The setup: GRPO training, batch 512, N=16, 32K context, on 18,953 math-reasoning problems. The verifier flips its reward with probability ε — and the claim is that ε sets the ceiling on self-play improvement.

Two findings carried the paper:

Twelve GRPO runs in total, about 3,570 GPU-hours of compute. Five separate submission failures along the way were all diagnosed and fixed autonomously — including a recurring type error where a learning rate in scientific notation was parsed as a string. After the fourth recurrence, the framework wrote a pre-submission check script and baked the lesson into its own operating constraints.

Five Personas, One Honest Median

Every review round is scored in parallel by five independent reviewer personas, each with a different bias:

PersonaRoleWhat they push on
R1ExperimentalistDo the numbers in the paper match the raw experiment logs, line by line?
R2TheoristAre the proofs rigorous? This was the binding constraint for most of the paper's life.
R3PerfectionistTable consistency, abstract accuracy, citation hygiene.
R4SynthesizerDoes each addition answer a question the paper itself raised?
R5NewcomerCan a non-expert navigate it? (A reading guide and abbreviations table gave the largest single jump.)

The median — not the mean — is the reported score, so a single enthusiastic reviewer can't inflate it. For most of the paper's life, the theorist (R2) sat at 8.0 and held the median down. That is exactly the signal that mattered.

V16: Theory Hardening Was the Last Mile

By V14 the experiments were solid, but R2 kept the median pinned: the noise-floor argument and the KL-ablation sign weren't reconciled with the theory. The fix was not more experiments — it was new mathematics. Three sub-problems were solved and integrated:

  1. Noise-floor recurrence. An exact closed-form re-derivation showing the floor is Θ(ηT) under harmonic coverage (which also caught a dropped-factor error in the paper's own draft), with a genuine constant floor η/ρ under an explicit mixing assumption.
  2. Coupled-floor lemma. A rigorous Gibbs → Jensen → Pinsker chain bounding the KL-anchor displacement.
  3. Matching lower bound. A persistence dichotomy proving the floor's order is optimal.

That carried the median from 8.5 to 8.6. Notably, this round was done overnight by a serial single-agent loop — two earlier attempts at parallel multi-agent workflows stalled and burned most of the round's tokens before the actual proofs came from the simpler serial approach. That lesson is now part of the framework's memory.

Read the full paper, the production statistics, and the four-paper comparison.

View AutoResearch Papers →

What I Take Away

The first three surveys showed an agent framework can write a credible paper. This one showed something harder: it can run a real experiment to test the paper's own claim, score itself honestly (including downward), and close a theory gap with new math — over six days, largely unattended. The remaining gap to a 9.0 is a single open problem (formalizing the held-out ↔ exploitability map), left as post-publication work.

Stay tuned.

这是我 AutoResearch 项目里 第四篇论文 的故事:《基础模型时代的自我博弈:从博弈论基础到开放式学习的综合综述》。和前三篇综述不同,这一篇不只是被 Agent 框架"写出来"——它被框架反复质疑、压力测试、并用真实实验验证,历经 16 轮评审,最终中位分停在 8.6/10(strong accept)。

声明:纯个人兴趣项目,不代表任何组织的立场,所有观点仅代表我个人。

8.6
最高中位分
16
评审轮次
75
页数
217
引用
285B
RL 模型
12
RL 实验

评分轨迹:不是一条直线

这篇论文有意思的地方在于,分数不是单调上涨的。每次改稿后,Agent 都会跑一轮五人格同行评审,中位分如实地上下波动:

版本分数发生了什么
V4–V107.0 → 8.4初稿阶段:三轴分类法、各章节、三个原创定理。引用从 0 累积到 207 条。
V118.5285B GRPO 验证器噪声实验写入 §8——论文由纯综述变为含原创大规模实验。
V128.2 ↓唯一一次下降。外部文献核查发现 3 条问题引用;按论文真实状态据实降分,不保留此前的 8.5。
V138.42000 步长程实验(原长度的 8.3 倍)检验 KL-buffering 假设;同时修正了数据来源。
V148.4种子复现 + KL 消融把"buffering"从假设变为证据。瓶颈转移到理论严谨性。
V158.5KL 端点 held-out 研究;并行启动四个理论子问题的研究。
V168.6 ✓理论加固落地——系列最高分。
为什么 V12 是最重要的版本
第一次外部文献核查发现了三条投稿阻断级的引用错误——一条不存在的论文、一条作者错配、一条 arXiv ID 指向无关论文——框架选择把自己的分数从 8.5 降到 8.2,而不是悄悄保留高分。一个会在证据要求时把自己的工作往下打分的自主流水线,远比一个只会往上爬的流水线可信。这三条都在下一轮迭代中修正了。

285B 强化学习实验:核心区分点

关键决策是用一个真实的 285B 参数模型(DeepSeek-V4)来验证综述的核心命题,而不是用小模型做演示。配置:GRPO 训练,batch 512,N=16,32K 上下文,18,953 道数学推理题。验证器以概率 ε 翻转奖励——命题是 ε 决定自博弈改进的上限。

有两个发现撑起了这篇论文:

总计 12 个 GRPO run,约 3,570 GPU 卡时 的算力。过程中 5 次提交失败全部自主诊断并修复——包括一个反复出现的类型错误:科学计数法写的学习率被解析成了字符串。在第 4 次复现后,框架写了一个提交前检查脚本,并把这条教训固化进了自己的运行约束里。

五个人格,一个诚实的中位数

每一轮评审都由五个独立的评审人格并行打分,各有不同的偏好:

人格角色盯什么
R1实验家论文里的数字是否与原始实验日志逐项对得上?
R2理论家证明是否严谨?这是论文大半生命周期里的硬约束。
R3完美主义者表格一致性、摘要准确性、引用规范。
R4综合者每一项新增是否回答了论文自己提出的问题?
R5新手非专家能否读得进去?(一个阅读路线 + 缩写表带来了单轮最大涨幅。)

报告的是中位数而非平均分,所以单个热情的评审人格无法把分数抬高。在论文大半的生命周期里,理论家(R2)一直停在 8.0,把中位数压住。这恰恰是最该被听见的信号。

V16:理论加固是最后一公里

到 V14 时实验已经扎实,但 R2 始终压着中位数:noise-floor 的论证和 KL 消融的符号没有和理论自洽。解决办法不是更多实验——而是新的数学。三个子问题被攻下并整合进论文:

  1. 噪声地板递推。一个精确闭式的重导,证明在 harmonic coverage 下地板是 Θ(ηT)(顺带抓出论文自身初稿里的一处掉因子错误),并在显式 mixing 假设下给出真常数地板 η/ρ。
  2. 耦合地板引理。用严格的 Gibbs → Jensen → Pinsker 链给出 KL 锚点位移的界。
  3. 匹配下界。一个持久性二分,证明地板的阶数是最优的。

这把中位数从 8.5 推到了 8.6。值得一提的是,这一轮是由一个单 Agent 串行 loop 通宵完成的——此前两次尝试用并行多 Agent workflow 都 stall 了,烧掉了这一轮大部分 token,真正的证明产出反而来自更简单的串行方式。这条教训现在已经写进框架的记忆里。

阅读完整论文、生产统计数据,以及四篇论文的对比。

查看 AutoResearch 论文 →

我的收获

前三篇综述证明了 Agent 框架能出一篇可信的论文。这一篇证明了更难的事:它能跑一个真实实验来检验论文自己的命题、诚实地给自己打分(包括往下打)、并用新数学补上理论缺口——历时六天,基本无人值守。距离 9.0 只剩一个开放问题(把 held-out ↔ exploitability 的映射定理化),留作发表后的工作。

Stay tuned.