# A Regime Theory of Controller Class Selection for LLM Action Decisions

Source: [A Regime Theory of Controller Class Selection for LLM Action Decisions](<https://www.sota2.com/research/paper/69fd6e565497b3b670c896a1>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-09T05:00:12.752Z

Authors: Zhaoyang Jiang, Zhizhong Fu, Yunsoo Kim, Jiacong Mi, Zicheng Li, Xuanqi Peng, Honghan Wu

Published: 2026-05-07T14:28:17.000Z

## Abstract

Deployed language and vision-language models must decide, on each input, whether to answer directly, retrieve evidence, defer to a stronger model, or abstain. Contrary to the common monotonicity intuition, greater per-input expressivity is not uniformly beneficial in finite samples: under identical strict cross-validation, different benchmarks prefer different controller classes. This reflects a finite-sample limitation of instance-level uncertainty signals, which can be exhausted at a distribution-dependent scale. We organize controllers into a nested lattice of four classes: fixed actions, partition routers, instance-level controllers, and prior-gated controllers, ordered by complexity. We prove a regime theory that turns three data-estimable bottlenecks into a class choice: how much improvement is possible beyond the best fixed action, whether there are enough samples for instance-level controllers to make reliable decisions, and how much improvement a coarse partition router can recover when instance-level signal is unreliable. The resulting Bernstein-tight threshold has a matching information-theoretic lower bound, and strict nested cross-validation provably selects a near-best class. Across SMS-Spam, HallusionBench, A-OKVQA, and FOLIO, the predicted class matches the empirical winner; the prior-gated controller wins on TextVQA when OCR tokens supply a label-free prediction-time prior. Code is available at https://github.com/Anonymous-Awesome-Submissions/Regime-Theory.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Visual Question Answering on TextVQA OCR n=5000 (val)](<https://www.sota2.com/research/sota/visual-question-answering-on-textvqa-ocr-n-5000-val.md>): Visual Question Answering; TextVQA OCR n=5000 (val); Answer Loss: 0.8212.

- [Knowledge-based Visual Question Answering on A-OKVQA n=1145 (held-out)](<https://www.sota2.com/research/sota/knowledge-based-visual-question-answering-on-a-okvqa-n-1145-held-out.md>): Knowledge-based Visual Question Answering; A-OKVQA n=1145 (held-out); Per-Class Loss: 0.3805.

- [Logical Reasoning on FOLIO n=203 (held-out)](<https://www.sota2.com/research/sota/logical-reasoning-on-folio-n-203-held-out.md>): Logical reasoning; FOLIO n=203 (held-out); Per-Class Loss: 0.7195.

- [Spam Detection on SMS-Spam n=1114 (held-out)](<https://www.sota2.com/research/sota/spam-detection-on-sms-spam-n-1114-held-out.md>): Spam Detection; SMS-Spam n=1114 (held-out); Per-Class Loss: 0.059.

- [Visual Hallucination Detection on HallusionBench n=920 (held-out)](<https://www.sota2.com/research/sota/visual-hallucination-detection-on-hallusionbench-n-920-held-out.md>): Visual Hallucination Detection; HallusionBench n=920 (held-out); Per-Class Loss: 0.897.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.06339>)

[PDF](<https://arxiv.org/pdf/2605.06339.pdf>)