USL10 Paper Portfolio
USL-02 Generation · Generalization · Personalization registered slot · current semantic abstract local only public contract: drafting

Generalization Through Units: Events, Queries, and New Units

What is the difference among generalizing to a new event, a new query and a new unit?

venue AAAI-27 Main Track topic ML: Deep Learning Theory & Learning Theory state abstract registered / writing owner gong · CausaClaw × DiscoSeed OpenReview forum 待回填

Semantic-revision boundary. The owner-side OpenReview Activity audit records the abstract slot, but the current local working abstract contains a later semantic revision (repeated_observation_evaluation_revision_not_verified_external). The direct OpenReview forum URL is also still pending writeback.

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 is the difference among generalizing to a new event, a new query and a new unit?

核心 claim
Under repeated-individual sampling, observation-level random splits can conflate new-event, new-query and new-individual generalization and reward same-individual leakage that fails on query or empirical-unit-support novelty.
如果成立会改变什么
Dataset splitting, generalization bounds and model rankings must be indexed by event/query/unit novelty.
最强 reviewer 反对 / kill signal

反对：This is ordinary grouped cross-validation or domain generalization with new terminology.

Kill signal：merge into standard grouped evaluation if the three axes cannot produce distinct failure modes after capacity and support are matched.

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 $U\sim\Pi$ selects an Individual under the task-declared Population law $\Pi$, and $\{U=u\}$ fixes its realized value. A new event is a new realization from a declared fixed-$u$ joint/longitudinal law; shared $u$ does not imply IID events. A new query changes a candidate query variable $q$ while retaining factual admissible evidence $\mathcal O$; the candidate $q$ does not automatically update $P(U\in du\mid\mathcal O)$ or learner $Q_\phi$. A new individual concerns a realized $u$ outside the training empirical support (with the target Population law stated separately). This identity novelty is distinct from response-support extrapolation: for a non-atomic identity-valued unit space, a new identity can be nearly automatic. Any performance claim for new $u$ values must separately declare the shared cross-unit response family, property map, geometry or other learnability structure and its support condition. Same-individual evaluation cells require observed attribution, a trusted-ID protocol or synthetic/oracle linkage. Similar learner beliefs cannot be used as ground-truth linkage.

论证 spine

  1. Population selection, observed attribution and repeated-observation protocols.
  2. Fixed-individual event realizations, candidate queries and empirical individual support are different axes.
  3. Event-, query- and individual-indexed train/test distributions without automatic IID assumptions.
  4. Risk decomposition into within-individual event, cross-query, new-individual and support errors.
  5. Same-individual boundary leakage and model-ranking reversals.
  6. Benchmark protocol with oracle/trusted attribution, open-individual and declared joint-novelty cells.
  7. Implications for domain generalization, meta-learning and continual learning.

所需证据

  • One theorem or proposition separating the three risks.
  • One controlled ranking reversal under matched data volume and capacity.
  • A split-audit tool that detects whether observations attributed to the same individual cross the train/test boundary.
  • An explicit joint/longitudinal event law and a declaration of which novelty axes are held fixed in each benchmark cell.
  • A latent-attribution boundary: same-individual cells require oracle/trusted linkage and cannot be constructed from similar $Q_\phi$ values.
Current working abstract（本轮语义修订尚未验证已同步外部记录）

We separate generalization to a new event from a known individual, a new query for the same individual, and a previously unseen individual under a declared attribution and repeated-observation protocol. Observation-level random splits can otherwise place records from the same realized individual on both sides of evaluation, hiding query or empirical-unit-support novelty and reversing model rankings.

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 Population selection, attribution regime, fixed-$u$ event law, non-evidential candidate query, empirical-versus-Population individual support, the selected benchmark cells and a risk decomposition without assuming repeated observations IID.

下一证据门

demonstrate at least one ranking reversal not explained by sample count.

48h 最小实验

create one oracle-attribution or trusted-ID dataset with controlled cells for new event/same individual, new query/same individual, known-query/new individual and declared joint novelty; state which other axes are fixed and measure same-individual boundary leakage and ranking reversals.

Closest SOTA opponents

grouped and leave-one-subject-out validation; domain generalization; meta-learning and open-set recognition.

canonical 文件

research-questions/USL02-generalization-through-units/seed.md
papers/USL02-generalization-through-units/paper.md

进展日志

2026-07-29 · current local semantic abstract revision repeated_observation_evaluation_revision_not_verified_external has not been verified as synchronized to the external record; the registered slot remains tracked separately.

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。