这组结果说明 case set 不是只在均值上分出模型,而是在每个 rubric
dimension 上维持 best > good > medium > bad
的嵌套形态。论文还报告 1000 次 80% bootstrap resampling 中完整
ordering 保持 100%,tier reversal 为 0%。
机制解释:D01 Length & Information Density 被
Heuristic Learning
更新为“按用户明确请求判断长度”,而不是按内容是否相关判断。时间压力下,未请求的展开本身就是质量缺陷。
2. AI judge 在分歧区的价值不是“压倒人类”
论文给的医疗边界 case 中,模型像朋友一样安慰用户,三名人类标注员给了
2、2、1;但 AI rubric 判 0
fatal,因为回复直接暗示“不需要再查”,还劝用户不要去做进一步检查。论文很谨慎地说,这不是“AI
比人类强”,而是结构化 rubric 帮人类把潜在安全维度显式化。
完整图谱
以下收录源包中发现的全部 9 张论文图像;图片经宽度 1600 优化并以 data
URI 内嵌。图注按论文原意转录,未将图像本身未展示的信息扩写为结论。
Architecture of GrowLoop. GrowLoop consists of two co-evolving
loops: Rubric Generation and Case Generation. The rubric guides case
construction, while evaluation results on cases expose rubric
deficiencies, driving iterative refinement of both.
Distribution of AI cascaded scores (0--5) grouped by human score
level for four models. Each panel corresponds to one model. Boxes
represent the interquartile range with medians marked. Sample sizes
are shown at the bottom of each box.
The three-phase Case Generation pipeline. \textbf{(i) CSP
Specification} derives a typed specification pool from the rubric,
real conversations, and expert-curated external priors; \textbf{(ii)
Multi-Agent Generation} transforms each specification into a
multi-turn dialogue via four collaborating agents; \textbf{(iii)
Verification} evaluates the assembled set against five hard gates,
triggering targeted re-generation on failure until all gates pass.
Heuristic Learning convergence curves. Dashed lines indicate
convergence targets (90\% for safety, 85\% for quality).
Rubric\textsubscript{safety} surpasses its target in 6 iterations;
Rubric\textsubscript{quality} converges in 10 iterations with a
total gain of 21.2 percentage points.
Convergence trajectory across five rounds (R1--R5). (a) Kendall
$\bar{\tau}$ crosses 0.7 at R5; (b) Cliff's $\delta_{\min}$
surpasses 0.32; (c) best-tier mean enters the [60, 75] band. All
three metrics improve monotonically with diminishing marginal gains.
Three axes complementing the scenario donut in
Fig.~\ref{fig:quality-quartet}(a). (a) Seven challenge dimensions
(cognitive load, spatio-temporal reasoning, anomaly detection,
social nuance, spatial reasoning, embodied sense, humor), with
cognitive load the largest at 21\% ($\mathrm{CR}_1=0.210 <
\theta_k=0.275$). (b) Four age bands with a working-age peak
(20--39, 61.6\%) and balanced coverage of 10--19 (12.7\%), 40--49
(13.3\%) and 50+ (12.5\%). (c) Female/Male = 59.9/40.1 after
removing \textit{Unknown}. Gender-neutral cases are retained in the
corpus but omitted here. The fine-grained scenario bar replicates
Fig.~\ref{fig:quality-quartet}(a) and is omitted here.
Case quality overview. (a) Scenario domain distribution as a donut
over 23 raw domains (3 sparse domains are collapsed to 20 effective
categories for the diversity threshold; domain-axis
$H_{\mathrm{norm}}=0.941$, $N=500$). (b) Four-tier score profiles on
an 18-dimension radar. (c) Per-case score density with Cliff's
$\delta$ markers on adjacent-tier pairs.
The three-phase rubric generation pipeline. Human seed annotations
are decomposed into candidate rubrics and merged into two
complementary rubrics (safety gate and quality scoring). Each is
independently optimized via Heuristic Learning, and then integrated
through cascaded judgment.
Overview of our self-evolving conversation evaluation system. Human
seeds drive scope expansion while model progress triggers difficulty
scaling, enabling the benchmark to evolve continuously. Each
benchmark comprises a rubric and cases. The rubric defines explicit
evaluation criteria, while cases are test conversations used for
evaluation.
我的评论
为什么我也觉得它不错
它抓住了非可验证任务评价的核心矛盾:如果 ground truth 不存在,硬做
majority vote
会污染监督;如果完全接受主观分歧,又无法训练和验收。consensus-divergence
partition 是一个干净的折中。
第二个优点是把 rubric 作为 source of truth,而不是让 reward model
黑箱吸收偏好。对于产品评价,这带来可解释、可编辑、可回滚的治理面。
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.
论文问题重要:开放式对话没有唯一 ground
truth,静态基准也会随模型进步而失效。方法亮点是将 rubric 与 case
双向迭代,并用硬门槛约束案例质量;实验呈现了人类种子、生成样本、基线比较和收敛曲线。主要疑点是证据外推:论文报告的测试规模和人工种子来源,是否足以代表真实对话分布;规则迭代可能导致对既有评测信号过拟合;多模型与人工标签的一致性不等同于用户效用。应更清楚报告独立留出集、标注者一致性、不确定性与跨域迁移。以
NeurIPS
常见维度判断,技术可靠性和重要性有价值但验证范围与泛化证据尚有限。建议:Weak
Reject;置信度:中。
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.
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.
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.
可做纸面 toy demo:用 6
条合成对话、两条显式规则(是否回应用户请求、是否包含无关扩写),人工给
0/1
标签;每轮计算规则判断与标签的分歧,将重复出现的漏判转换为候选规则,再用两条未见对话回归测试。核心伪代码:`for
case in seeds: errors += compare(rubric(case), label(case));
rubric = propose_rule(errors); evaluate(rubric,
heldout)`。这只演示“从错误提出规则并在留出案例检查”的局部思想,不包含
GrowLoop 完整的 CSP、多 agent
生成或五项门槛,也未实际运行;没有调用或冒充现成实现。
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.
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?