USL10 Paper Portfolio
USL-08 Belief Validity · Evidence · Acting AAAI-27 abstract registered · forum linked public contract: drafting

Active Unit Identification: What Evidence Should We Collect Next?

What evidence should be collected next to identify the selected unit for a family of future decisions?

venue AAAI-27 Main Track topic ML: Semi-Supervised & Active Learning state abstract registered / writing owner gong · CausaClaw × DiscoSeed OpenReview forum ↗

基本研究问题

这篇论文独立可证伪的问题是什么；它与共享 ontology 中其他九篇不重叠的部分。

What evidence should be collected next to identify the selected unit for a family of future decisions?

核心 claim
Evidence acquisition should maximize expected downstream value from resolving unit uncertainty, not merely label uncertainty for one sample.
如果成立会改变什么
Active learning becomes a diagnostic design problem over evidence modalities, costs and future query families.
最强 reviewer 反对 / kill signal

反对：This is Bayesian experimental design or active feature acquisition with `unit` renamed.

Kill signal：kill the independent paper if unit-directed acquisition never changes the selected evidence or downstream decisions under matched budgets.

基本思路

论文如何回答这个问题：形式化对象、论证 spine 与所需证据。

USL-01 type contract

The Unit Selection Variable $U\sim\Pi$ performs individual selection under the task-declared Population law $\Pi$. Current admissible information $\mathcal O$ conditions the world law $P(U\in du\mid\mathcal O)$; it does not select an individual. A learner maintains the separately parameterized belief $Q_\phi(du\mid\mathcal O)$ and need not match that conditional without an explicit calibration or consistency result. A proposed acquisition is not yet evidence: only its observed result may enlarge $\mathcal O$ to $\mathcal O^+$. Downstream response uses a kernel $K(dy\mid q,u,\mathcal O)$ that may remain stochastic after $u$ is fixed; it may omit $\mathcal O$ only under a declared response-sufficiency assumption. Any $Z_u$ is an optional derived property, not a replacement for the identity-bearing $u$. The update $\mathcal O\to\mathcal O^+$ concerns the same fixed individual only under a declared measurement/linkage protocol; reselection requires a joint selector law. If collecting evidence changes the selected individual, treatment, or response process, that change must be modeled as an intervention rather than silently treated as conditioning. Acquisition value is defined through future-query response or utility pushforwards, not raw entropy over identity labels. USL08 owns diagnostic evidence-channel design; an evolving controlled state/reward trajectory belongs to USL06 when it becomes the primary object.

论证 spine

  1. Unit Selection Variable, task-declared Population law, world conditional and learner belief are different objects.
  2. From label acquisition to unit-directed evidence acquisition.
  3. Evidence actions, observed results, same-individual linkage, cost and belief updates.
  4. Future-query value of information without treating future targets as evidence.
  5. Acquisition-order separation from entropy and predictive-risk criteria.
  6. Synthetic multimodal diagnostic benchmark.
  7. Budget, redundancy, conflict, intervention effects and abstention.

所需证据

  • A case where unit-directed acquisition reverses the baseline ordering.
  • Matched entropy, mutual-information and active-feature baselines.
  • Negative controls where unit identity is irrelevant to future queries.
  • Separate evaluation of the world update $P(U\in du\mid\mathcal O^+)$ and the learner update $Q_\phi(du\mid\mathcal O^+)$; do not define them as equal.
  • A response kernel $K(dy\mid q,u,\mathcal O)$ that retains fixed-$u$ event/exogenous variation, with any $Z_u$ typed as an optional derived property rather than a replacement for $u$; omit $\mathcal O$ only under a declared response-sufficiency assumption.
  • A declared measurement/linkage or joint-selector law specifying whether $\mathcal O\to\mathcal O^+$ concerns the same fixed individual or a reselection.
  • A protocol showing whether each acquisition is passive measurement or an intervention on the individual/response process, with current targets and unobserved candidate results excluded from $\mathcal O$.
已注册 abstract（点击展开）

We choose diagnostic evidence by its expected value for identifying the selected actual unit and improving a family of future decisions. This differs from querying a new sample label or reducing uncertainty for only one immediate task.

当前进展

状态只记录可验证 delta：claim、SOTA opponent、theorem、experiment、manuscript 或 owner decision。

当前状态

AAAI-27 abstract registered; full paper drafting.

下一写作门

define the evidence action and its observed result, the induced update from $P(U\in du\mid\mathcal O)$ to $P(U\in du\mid\mathcal O^+)$, the learner update of $Q_\phi$, and the future-query value functional; distinguish a passive diagnostic measurement from an intervention that changes the individual or response process, keep future targets/candidate queries outside evidence, and use `posterior` only under an explicitly Bayesian learner.

下一证据门

demonstrate a qualitative acquisition-order reversal.

48h 最小实验

create several candidate evidence channels with different unit/event informativeness and cost; compare entropy, label-risk and downstream unit-value acquisition rules.

Closest SOTA opponents

Bayesian experimental design; active feature acquisition; pool-based active learning and information gain.

canonical 文件

research-questions/USL08-active-unit-identification/seed.md
papers/USL08-active-unit-identification/paper.md

进展日志

2026-07-29 · AAAI-27 abstract registration 由 direct OpenReview forum link 验证；full paper drafting 启动；本页面建立为持续 review 面。

后续每次可验证 delta 追加在此；canonical 状态以 seed.md / paper.md / submission-ledger 为准，本页由 build_paper_pages.py 重新生成。