Rank-Then-Act (RTA): A Two-Stage Framework

• Rank-Then-Act (RTA) is a two-stage decision framework that first ranks candidate actions or units and then selects a final action based on that ranking.
• It is applied across diverse fields such as treatment prioritization, legal citation, reward-free control, safe reinforcement learning, and ranking and selection.
• The framework emphasizes methodological decomposition to improve reliability, safety, and performance by separating candidate generation from final decision-making.

Searching arXiv for papers on "Rank-Then-Act" and closely related uses of the term. arXiv query: "Rank-Then-Act" Rank-Then-Act (RTA) denotes a two-stage decision pattern in which a system first produces an ordering, prioritization, or candidate set and then executes a downstream action conditioned on that ranked structure. The term appears in several distinct literatures rather than as a single standardized algorithm: treatment prioritization rules are assessed by how well they rank units by treatment benefit (Yadlowsky et al., 2021); statutory-citation systems retrieve and rank candidate legal provisions before selecting a citation (Asaria et al., 18 Jun 2026); and reward-free control learns an ordinal progress scorer from expert video and then acts through reinforcement learning driven by rank correlation (Maksyuta et al., 2 Jul 2026). Related work also interprets communication-efficient safe RL and fixed-budget ranking-and-selection through the same rank-first, act-later logic (Haroon et al., 11 May 2026, Zhou et al., 2024). This suggests that RTA is best understood as an architectural motif that separates candidate supply, ordering, or prioritization from commitment to a final intervention.

1. General architecture and scope

Across the cited works, the “rank” stage is not tied to a single object. It can be a prioritization score $S(X)$ over individuals, a retrieved list of statute provisions, a progress ordering over video frames, a screening decision over proposed control intervals, or a rollout-estimated value ranking over sampling actions. The “act” stage is likewise domain-specific: treat the top-ranked units, output the governing statutory citation, optimize a policy against a rank-based reward, execute a control input only if it passes a shield, or allocate the next simulation sample to the best-ranked alternative.

Setting	Rank stage	Act stage
Treatment prioritization (Yadlowsky et al., 2021)	Score $S(X)$ orders units by expected benefit	Treat top-ranked units or evaluate the rule
Statutory citation (Asaria et al., 18 Jun 2026)	Retrieve and rank candidate provisions	Output the best citation from retrieved context
Reward-free control (Maksyuta et al., 2 Jul 2026)	Rank frames by task progress	Use rank-correlation reward for RL
Safe RL with run-time assurance (Haroon et al., 11 May 2026)	Propose timing/action, then screen safety	Execute or override with LQR backup
Ranking and selection (Zhou et al., 2024)	Rank candidate allocations by downstream PCS	Sample the top-ranked alternative

A recurrent technical theme is decomposition. In the statutory-citation setting, the paper makes the decomposition explicit as

$$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$

which separates candidate supply from final choice (Asaria et al., 18 Jun 2026). In treatment prioritization, the same separation appears as “learn or choose a ranking, then decide how to intervene using that ranking” (Yadlowsky et al., 2021). In reward-free control, the separation is between learning ordinal progress and using that progress only through a correlation-based reward during policy optimization (Maksyuta et al., 2 Jul 2026). These formulations differ in notation and objective, but all treat ranking as an intermediate decision object rather than a by-product.

2. Treatment prioritization and RATE

In causal inference, RTA is formulated as treatment prioritization: a rule outputs a score $S(X)$, where larger values mean higher priority for treatment, and the scientific question is whether the rule places people who benefit most from treatment near the top (Yadlowsky et al., 2021). The paper “Evaluating Treatment Prioritization Rules via Rank-Weighted Average Treatment Effects” formalizes this through the conditional average treatment effect

$\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$

and the Targeting Operator Characteristic,

$$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$

Here $TOC(u;S)$ compares the average treatment effect among the top $u$ fraction of the ranking to the overall ATE, with $TOC(1;S)=0$.

A rank-weighted average treatment effect (RATE) is then defined as

$$\theta_\alpha(S)=\int_0^1 \alpha(u)\, TOC(u;S)\,du,$$

with equivalent weighted-average representation

$S(X)$0

where

$S(X)$1

This formulation makes the ranking itself the evaluand. A rule can come from a CATE estimator, a risk score, or a hand-crafted rule; RATE is agnostic to derivation and only tests whether higher-ranked units actually have larger treatment benefit.

The paper places several familiar uplift metrics inside the same family. AUTOC is

$S(X)$2

and the Qini coefficient is

$S(X)$3

A central contribution is inferential: using approximately unbiased scores $S(X)$4 for $S(X)$5, the paper derives estimators, an asymptotic linear representation,

$S(X)$6

and hence

$S(X)$7

Because the asymptotic variance is complicated, the recommended practical procedure is the half-sample bootstrap.

Within this literature, RTA means that a ranking can be useful even when the final intervention threshold is not fixed. The paper explicitly distinguishes ranking evaluation from thresholded policies: if a final treatment threshold is uncertain or depends on external constraints, RATE is preferable because it evaluates the ranking without committing to a single cutoff (Yadlowsky et al., 2021). The aspirin example in the International Stroke Trial illustrates the point. A prognostic random forest risk rule had a statistically significant positive AUTOC, whereas the CATE-based methods were not significant in that analysis. The interpretation is not that prognostic risk is equivalent to treatment effect, but that the resulting ranking can still be useful for prioritization.

3. Statutory citation as retrieve-rank-select

In legal QA, RTA is instantiated as a statutory-citation pipeline whose immediate target is the correct governing provision rather than fluent legal advice. The paper “Train, Retrieve, or Both? A Four-Arm Head-to-Head for Correct Statutory Citation on the Ontario Residential Tenancies Act” studies questions on the Ontario Residential Tenancies Act, 2006 and O. Reg. 516/06, with the goal of predicting the correct citation down to section and subsection (Asaria et al., 18 Jun 2026). The end-to-end pipeline is: build a citation inventory from the consolidated Ontario RTA and O. Reg. 516/06; retrieve candidate statutory provisions; rank those candidates using dense and lexical retrieval; feed retrieved candidates into the model; output a short answer and a citation; and verify the citation with a deterministic checker against the pinned statute inventory.

The corpus is parsed into 358 provisions total, with each provision acting as a single RAG chunk. The paper reports 297 RTA sections / 895 subsections / 20 Parts, plus 61 sections / 166 subsections in the regulation. Retrieval-bearing systems use a “cite-from-context guardrail” that is “effectively an existence check,” so hallucinated citations are eliminated by construction for those systems (Asaria et al., 18 Jun 2026). This makes retrieval not merely an accuracy device but a safety device.

The experimental design is a four-arm head-to-head on Qwen2.5-7B-Instruct: base zero-shot, LoRA SFT-only, RAG-only, and an SFT+RAG hybrid. The headline real evaluation set contains only 27 real, source-cited items and is flagged as human-verification-pending, while the synthetic dataset contains 2,148 question→citation pairs split into train (1,473), synthetic test (345), and unseen-section (330) (Asaria et al., 18 Jun 2026). The paper is explicit that the real evaluation is tiny, source-disjoint from training, and preliminary.

The primary metric is citation exact-match, defined over instrument + section + subsection, with section-only matches receiving partial credit of 0.5. The paper notes that the reported “exact-match” is therefore actually a partial-credit citation score in $S(X)$8, not a strict binary metric. It also reports multi-citation precision/recall/F1, deterministic hallucinated-citation rate, and recall@k.

The main reported results are as follows.

System	Reported exact-match	Additional note
Base zero-shot	0.00	81% hallucinated citations
SFT-only	0.148	Mis-recalls sections
RAG-only, $S(X)$9, cheap bge-small	0.44	recall@5 0.815, hallucination 0.00
RAG-only, $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$0, cheap bge-small	0.37	recall@10 0.889
SFT+RAG hybrid, $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$1, cheap bge-small	0.333	Retrieval-bearing
SFT+RAG hybrid final, $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$2, cheap bge-small	0.481	F1 0.574, hallucination 0.00
Bigger retrieval pipeline	0.37	Larger embedder + cross-encoder reranker

The paper’s practical RTA conclusion is that training alone is not enough, retrieval is essential, and the best system is a hybrid. Its abstract states that “the SFT+RAG hybrid scores highest at 0.481 exact-match with zero hallucinated citations” (Asaria et al., 18 Jun 2026). It further states that the hybrid’s edge comes from “SFT making provision selection more robust to the higher-recall candidate sets that hurt zero-shot RAG.” This interpretation is supported by the reported $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$3-sensitivity: increasing $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$4 from 5 to 10 hurts zero-shot RAG (0.44 \rightarrow 0.37) but helps the SFT model (0.333 \rightarrow 0.481). The paper’s recommended design principle is therefore “retrieve first, then train the selector.”

Several cautions qualify the result. The final configuration was selected as the best-performing cell on the same 27-item set, so the 0.481 is a max-over-configurations and is optimistically biased. A larger/improved training set variant with 2,645 pairs does not improve over the original, reaching 0.444 exact-match with 0.037 hallucination. The paper also states that the aspirational target was 0.70 exact-match, and the system does not reach it (Asaria et al., 18 Jun 2026).

4. Reward-free control from ordinal progress

In reward-free control, RTA is explicitly named as a framework: “Rank-Then-Act: Reward-Free Control from Frame-Order Progress” introduces a two-stage method that learns control policies from expert video demonstrations without environment rewards (Maksyuta et al., 2 Jul 2026). Stage 1 trains a Vision-LLM offline as a progress-based ordinal scorer; Stage 2 uses that scorer during online RL, but not as a scalar reward model. Instead, the learning signal is the Spearman rank correlation between predicted progress rankings and true temporal indices in an interaction window.

The training construction is designed to suppress trivial temporal shortcuts. Demonstration videos are chopped into clips of length $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$5; the first frame is fixed as an anchor and the remaining $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$6 frames are randomly shuffled. The paper writes

$$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$7

Only the shuffled non-anchor frames are scored. The VLM $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$8 predicts progress ranks $$\text{ExactMatch} \;\le\; \mathrm{recall@}k \times P(\text{select gold}\mid\text{retrieved}),$$9, and malformed outputs receive the worst reward $S(X)$0. If ranks tie, they are randomly permuted.

The core reward is

$S(X)$1

so the training reward is

$S(X)$2

with $S(X)$3 the predicted ranks and $S(X)$4 the ground-truth temporal indices. The VLM is optimized with Group Relative Policy Optimization (GRPO), and the paper emphasizes that ordinal supervision over shuffled inputs forces the model to recover temporal ordering from visual semantics rather than list position.

During Stage 2, the frozen scorer is queried on sliding windows

$S(X)$5

with the oldest state as anchor. Rewards are queried every $S(X)$6 steps; if $S(X)$7, the reward is $S(X)$8, and at query steps the method averages over $S(X)$9 random shuffles. The default values are $\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$0 and $\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$1 (Maksyuta et al., 2 Jul 2026). The resulting reward is bounded in $\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$2, scale-invariant, and robust to calibration drift because it depends only on ordinal agreement.

Empirically, the scorer can often reach $\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$3 in under 300 steps on many tasks, indicating that the model can recover progress order from shuffled expert clips (Maksyuta et al., 2 Jul 2026). The paper evaluates on PyBoy tasks—Catrap Levels 2, 4, and 6, and Kirby—and on PointMaze-UMaze and selected MetaWorld tasks with DrQv2. Reported PyBoy success values are 1.00 ± 0.00 on Catrap Level 2, 0.72 ± 0.35 on Level 4, 0.32 ± 0.27 on Level 6, and 0.07 ± 0.09 on Kirby. The paper highlights that RTA is the only method with non-zero Kirby success among several baselines. It also reports Pearson correlation between mean cumulative reward and pass@5 during training: 0.76 for Level 2, 0.87 for Level 4, 0.42 for Level 6, and 0.13 for Kirby (Maksyuta et al., 2 Jul 2026).

The main claim is not merely performance but transfer. The paper states that a scorer trained on one source can generalize to other levels in the same game, full game playthroughs, other GameBoy playthroughs from YouTube, and even other visual domains such as MetaWorld and COIN videos in some settings (Maksyuta et al., 2 Jul 2026). This suggests that the “rank” stage is learning a reusable notion of progress order rather than a task-specific scalar reward. The paper nevertheless notes limitations: dependence on diverse expert videos, sensitivity to window size in non-monotonic or cyclic trajectories, VLM latency, and out-of-distribution failure cases.

5. Safe and communication-efficient control with run-time assurance

A different use of the same abbreviation appears in “Learning When to Act: Communication-Efficient Reinforcement Learning via Run-Time Assurance” (Haroon et al., 11 May 2026). Here RTA refers to a run-time assurance layer, but the accompanying interpretation treats the method as Rank-Then-Act because the system first decides whether a proposed timing/action pair is admissible and only then executes it. The learned policy chooses both the next inter-sample interval $\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$4 and control input $\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$5, while the shield checks safety and overrides unsafe proposals with a precomputed LQR backup.

At each decision time $\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$6, the observation is

$\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$7

and the RL policy proposes

$\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$8

The shield predicts the safety-critical variable one step ahead by

$\tau(x) = \mathbb{E}[Y_i(1)-Y_i(0)\mid X_i=x]$9

and checks whether

$$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$0

If safe, the action is accepted and held with ZOH; if unsafe, the system overrides with

$$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$1

The authors contrast this with constrained MDP methods that enforce safety only in expectation and argue that the override gives a strictly stronger guarantee whenever the linearized ZOH certificate applies (Haroon et al., 11 May 2026).

The analytical basis is CARE-derived LQR and Lyapunov theory. The paper solves

$$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$2

uses $$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$3, $$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$4, and states

$$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$5

Under ZOH with backup $$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$6, the closed-loop transition is

$$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$7

and Lyapunov decrease is characterized by

$$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$8

The paper also proves that for the linear closed-loop system there exists a state-dependent admissible interval $$TOC(u;S) \equiv \mathbb{E}[Y_i(1)-Y_i(0)\mid F_S(S(X_i))\ge 1-u] - \mathbb{E}[Y_i(1)-Y_i(0)].$$9 such that

$TOC(u;S)$0

The communication-aware reward contains a stability term, a graded Lyapunov term, a communication bonus controlled by a single weight $TOC(u;S)$1, and a safety term. The paper sweeps

$TOC(u;S)$2

for lower-dimensional environments and emphasizes that MSI is the real communication-efficiency metric, not total episode reward (Haroon et al., 11 May 2026).

Experimentally, on an inverted pendulum, cart-pole, and planar quadrotor, the learned policy achieves $TOC(u;S)$3, $TOC(u;S)$4, and $TOC(u;S)$5 higher mean inter-sample interval than a Lyapunov-triggered baseline. A fixed LQR controller at the same average rate is unstable on all three plants. Removing the RTA shield reduces MSI by $TOC(u;S)$6–$TOC(u;S)$7 and degrades state norms. A preference-conditioned extension recovers the full tradeoff frontier from one model at $TOC(u;S)$8 of training compute, and robustness experiments under $TOC(u;S)$9 mass mismatch and disturbances show graceful degradation (Haroon et al., 11 May 2026). In RTA terms, the paper’s contribution is a hard separation between proposing when to act and certifying whether that proposal may be executed.

6. Fixed-budget ranking and selection

“AlphaRank: An Artificial Intelligence Approach for Ranking and Selection Problems” casts fixed-budget ranking and selection as a sequential decision problem and is explicitly described as “rank-then-act” in the sense that the system ranks candidate allocations by their expected downstream selection quality and then allocates the next sample to the best-ranked alternative (Zhou et al., 2024). The setting has $u$0 alternatives with unknown means,

$u$1

and the objective is to identify

$u$2

under a fixed total budget $u$3.

The canonical evaluation metric is the probability of correct selection (PCS),

$u$4

with expected opportunity cost (EOC) also discussed. The state is

$u$5

and the MDP action is to allocate the next sample to one of the alternatives. The Bellman-style objective is

$u$6

AlphaRank approximates action values by rollout. For candidate action $u$7,

$u$8

where $u$9 is a base policy such as EA, KG, AOAP, OCBA, or EI. Since the quantity has no closed form, the method performs $TOC(1;S)=0$0 Monte Carlo rollouts, generates Bernoulli rewards $TOC(1;S)=0$1, and estimates

$TOC(1;S)=0$2

The action actually taken is

$TOC(1;S)=0$3

Here the “rank” stage is literal: candidate allocations are ordered by rollout-estimated future PCS.

A second stage pretrains a neural network offline to approximate these rollout values. The network takes the current state as input and outputs a value vector over actions. Targets come from rollout, and online decision making becomes

$TOC(1;S)=0$4

The paper reports a complexity reduction from rollout’s $TOC(1;S)=0$5 to direct online NN use at $TOC(1;S)=0$6 (Zhou et al., 2024). For very large $TOC(1;S)=0$7, the method is embedded in a divide-and-conquer/recursion framework, with at most

$TOC(1;S)=0$8

rounds and lower-bounded PCS via Bonferroni.

A notable conceptual claim is that many classical policies emphasize means and variances but overlook induced correlation among competing sample means. The authors argue that AlphaRank’s rollout-based policy can allocate more to lower-mean or second-best alternatives in low-confidence regimes because doing so can improve PCS by better balancing mean, variance, and induced correlation (Zhou et al., 2024). In RTA language, this is a learned ranking policy over actions rather than a fixed heuristic.

7. Common themes, misconceptions, and limitations

A common misconception is that Rank-Then-Act denotes one specific algorithm. The literature summarized here does not support that reading. The term is used for treatment prioritization metrics, statutory-citation pipelines, video-based reward-free control, run-time-assured timing-and-control, and fixed-budget ranking and selection (Yadlowsky et al., 2021, Asaria et al., 18 Jun 2026, Maksyuta et al., 2 Jul 2026, Haroon et al., 11 May 2026, Zhou et al., 2024). The commonality is structural: separate the production of an ordering, candidate set, or admissibility judgment from the final intervention.

A second misconception is that ranking quality can be inferred from outcome prediction quality alone. The RATE paper explicitly rejects that equivalence: a prioritization rule should be evaluated by whether it sorts people by treatment benefit, not by whether it predicts outcomes or pointwise treatment effects accurately (Yadlowsky et al., 2021). The statutory-citation paper reaches an analogous conclusion from a different angle: training alone improved formatting but still mis-recalled exact sections, whereas retrieval determined whether the correct legal candidate was present at all (Asaria et al., 18 Jun 2026). This suggests that RTA systems often have distinct failure modes for candidate supply and final choice.

A third misconception is that a rank-based intermediate signal is necessarily weak or indirect. In reward-free control, the rank-correlation reward is the sole learning signal, yet the paper reports that it consistently matches or outperforms prior video-based reward learning methods and rank-based baselines, while showing strong cross-task reuse of a single pretrained progress scorer (Maksyuta et al., 2 Jul 2026). In communication-efficient safe RL, the intermediate screening step is not merely diagnostic; the shield materially changes the reachable stability–communication tradeoff, and its removal reduces MSI and worsens state norms (Haroon et al., 11 May 2026).

The limitations are domain-specific and substantial. The statutory-citation results are preliminary because the real evaluation set has only 27 items, is human-verification-pending, uses single-run point estimates with no confidence intervals, and the final model choice was selected on that same set (Asaria et al., 18 Jun 2026). RATE provides inference for a given prioritization rule, but a non-significant RATE can mean either that no meaningful heterogeneity exists or that the ranking rule fails to capture it (Yadlowsky et al., 2021). Reward-free control depends on diverse expert videos, is sensitive to window length on non-monotonic or cyclic trajectories, and incurs VLM latency (Maksyuta et al., 2 Jul 2026). The safe RL framework is grounded in stabilization around a known equilibrium with CARE-based LQR backups and classical Lyapunov-STC, which enables clean analytical comparison but also defines the regime in which the strongest guarantees apply (Haroon et al., 11 May 2026). AlphaRank’s rollout quality depends on the base policy and on Monte Carlo accuracy, even though the paper proves consistency under the stated conditions (Zhou et al., 2024).

Taken together, these works support a precise but non-monolithic interpretation of Rank-Then-Act. It is not a single model family, but a recurring research design in which the system first constructs an ordering, candidate set, or admissibility judgment and only afterwards commits to treatment, citation, control, or sampling. The technical value of that separation differs by domain—grounding and hallucination control in law, identification of benefit heterogeneity in causal inference, scale-invariant reward construction in control from video, pointwise safety in adaptive control, or PCS-oriented sample allocation in ranking and selection—but the core methodological move remains the same: rank first, then act.