Closed-Set Autoregressive Ranking

• Closed-set autoregressive ranking is a framework that sequentially orders a fixed candidate set using autoregressive dynamics.
• It integrates ranking signals into iterative proposal generation through mechanisms like constrained docID decoding and slate generation.
• The approach spans applications from dynamic statistical models and information retrieval to autonomous driving, enhancing both prediction and control.

Closed-set autoregressive ranking denotes a family of ranking formulations in which ordering is performed over a fixed item universe or a bounded candidate set, while the ranking itself is generated, updated, or refined sequentially rather than by a single static scoring pass. In one line of work, rankings are permutations of a fixed set of items whose latent worth parameters evolve through score-driven autoregressive dynamics; in another, a causal LLM ranks documents by autoregressively generating constrained docIDs; in a third, a single rollout first generates a candidate slate and then ranks it; and in autonomous driving, a generator proposes a finite set of trajectories and a learned scorer repeatedly refines that support through closed-loop ranking feedback. Taken together, these formulations suggest a common abstraction: a closed candidate domain, sequential dependence, and a ranking signal that either defines the output directly or feeds back into subsequent proposal generation (Holý et al., 2021, Rozonoyer et al., 9 Jan 2026, Surana et al., 13 May 2026, Ang et al., 14 May 2026).

1. Closed-set semantics and problem formulations

The defining feature of the “closed-set” condition is that the ranked objects come from a fixed or explicitly bounded domain. In the statistical time-series formulation, the item universe is fixed as

$\mathcal{Y}=\{1,\ldots,N\},$

and each observation is a ranking of the same items over time. In pointwise generative IR, the output space is restricted to a finite set of valid docID tokens, so ranking is performed over a closed set of document identifiers rather than arbitrary text. In unified slate generation and ranking, the model must construct and order candidates from a provided pool $\mathcal{E}$. In CLOVER, the generator produces a set of $K$ candidate trajectories,

$$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$

and the scorer ranks only within that set (Holý et al., 2021, Rozonoyer et al., 9 Jan 2026, Surana et al., 13 May 2026, Ang et al., 14 May 2026).

The autoregressive aspect takes different mathematical forms. In the GAS ranking model, latent worths depend on lagged scores and lagged states. In ARR for IR, docIDs are generated token by token by a decoder. In F-GRPO, a single autoregressive completion emits a <[SLATE](https://www.emergentmind.com/topics/slate)> segment followed by a <RANK> segment. In CLOVER, the paper is explicit that the method is not autoregressive in the language-model sense, but is “conceptually close” because proposal generation, scoring, target selection, and distillation are iterated in a closed loop (Ang et al., 14 May 2026).

Setting	Closed set	Sequential mechanism
Time-varying ranking	Fixed items $\{1,\ldots,N\}$	Score-driven autoregressive worth dynamics
ARR in IR	Valid docID tokens / corpus docIDs	Token-by-token docID generation
F-GRPO	Provided candidate pool $\mathcal{E}$	<SLATE> then <RANK> in one rollout
CLOVER	Proposal set $\{\tau_i\}_{i=1}^{K}$	Closed-loop proposal-ranking-distillation

A plausible implication is that “closed-set autoregressive ranking” is best understood as a structural pattern rather than a single architecture: a finite candidate domain is preserved, but ranking depends on sequential state, sequential decoding, or sequential refinement.

2. Fixed-item dynamic ranking models

The most explicit statistical realization of closed-set autoregressive ranking is the model of time-varying rankings over a fixed item set (Holý et al., 2021). A ranking $y_t$ is modeled with a Plackett–Luce distribution whose worth parameters change over time. For a complete ranking $y$, the probability is

$\PP \left[ Y = y \middle| f \right] = \prod_{r=1}^{N} \frac{f_{y^{-1}(r)}}{\sum_{s=r}^N f_{y^{-1}(s)}}.$

Because the pmf is invariant to adding a constant to all $\mathcal{E}$0, the model standardizes via

$\mathcal{E}$1

The same framework accommodates partial rankings. If only the top $\mathcal{E}$2 items are observed, the probability of the partial ranking is

$\mathcal{E}$3

This preserves the closed-set assumption because unobserved items remain in the risk set even when they are not explicitly ranked.

Autoregression enters through latent worth dynamics. With exogenous covariates, the state equation is

$\mathcal{E}$4

Here $\mathcal{E}$5 is an item-specific fixed effect, $\mathcal{E}$6 multiplies exogenous covariates, $\mathcal{E}$7 weights score feedback, and $\mathcal{E}$8 controls autoregressive persistence. In the no-covariate case, the unconditional mean is

$\mathcal{E}$9

The update is score-driven in the GAS sense. For a complete ranking, the conditional score component is

$K$0

The paper emphasizes that the score has mean zero under the model, its variance equals the Fisher information, and the score is bounded. This boundedness is used to justify unit scaling and is presented as a source of robustness to outlying rankings.

Estimation is by conditional maximum likelihood,

$K$1

with early scores initialized at zero and early $K$2 initialized at the unconditional value. Standard MLE asymptotics are used, with empirical-Hessian standard errors.

The simulation study varies $K$3, $K$4 from 10 to 100, and uses 100,000 replications per design. The reported conclusion is that estimates converge to the truth as $K$5 grows, that $K$6 is hardest to estimate in small samples, and that conventional Hessian-based standard errors are usable even in medium-sized samples. In the empirical application to Ice Hockey World Championships from 1998 to 2019, the mean-reverting model fits best by AIC, with $K$7AIC exceeding 25 relative to the alternatives, and the estimated autoregressive coefficient has 95% CI approximately $K$8, rejecting both no serial dependence and unit-root behavior. In this setting, closed-set autoregressive ranking yields probabilistic forecasts for one-step-ahead rankings, medal probabilities, and partial ranking events.

3. Autoregressive ranking with constrained docID generation

In information retrieval, the closed-set autoregressive formulation is explicit: a causal LLM ranks documents by generating tokenized docIDs conditioned on a query (Rozonoyer et al., 9 Jan 2026). Given a query $K$9, the model defines

$$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$0

where $$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$1. This is called pointwise generative ranking because each example corresponds to a single docID rather than an entire list.

The output space is explicitly constrained. Let $$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$2 be the docID-token vocabulary. Under constrained decoding,

$$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$3

The model therefore ranks a fixed finite set of docIDs rather than generating free-form text. In the in-context setting, the prompt is

$$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$4

and beam search can directly produce a top-$$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$5 ranking: a beam of size $$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$6 yields the $$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$7 most likely docID sequences.

A central theoretical claim is an expressivity separation from dual encoders. The dual encoder is written as

$$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$8

The paper states that no such architecture solves a complete ranking task when

$$\mathcal{T}_\theta(o)=G_\theta(o)=\{\tau_i\}_{i=1}^{K},$$9

with $\{1,\ldots,N\}$0. By contrast, for ARR the key proposition is that an infinite-capacity model can produce any strictly positive probability distribution over tokens in $\{1,\ldots,N\}$1 via constrained decoding if and only if $\{1,\ldots,N\}$2; sequence-level and ranking-level corollaries follow by the chain rule. The paper therefore argues that ARR’s ranking capacity is tied to token-embedding rank rather than to a document-dependent embedding dimension that scales linearly with the number of documents.

To address the claim that standard next-token prediction is rank-agnostic, the paper introduces SToICaL, the Simple Token-Item Calibrated Loss: $\{1,\ldots,N\}$3 Rank dependence enters through $\{1,\ldots,N\}$4, with reported variants including $\{1,\ldots,N\}$5, $\{1,\ldots,N\}$6, and $\{1,\ldots,N\}$7. Token-level calibration is implemented by marginalizing over a trie of tokenized docIDs.

The experiments use WordNet-derived ranking tasks and ESCI shopping queries. Metrics are CVR, nDCG, and Recall@$\{1,\ldots,N\}$8. The base LLM is Mistral-7B-v0.3-it. On WordNet, the reported NTP baseline has CVR $\{1,\ldots,N\}$9, nDCG $\mathcal{E}$0, and R@1 $\mathcal{E}$1, while fractional weighting at $\mathcal{E}$2 yields CVR $\mathcal{E}$3, nDCG $\mathcal{E}$4, and strong $\mathcal{E}$5 across the board. Trie-marginalized token supervision also helps, but the paper reports that it generally does less well than the best item-level reweighting. On ESCI, trie-based token calibration improves nDCG and all $\mathcal{E}$6 for $\mathcal{E}$7, though R@1 drops. A recurrent point in this formulation is that the ranking problem remains closed-set throughout: the model ranks valid docIDs and suppresses invalid docID generations rather than producing unrestricted text.

4. Unified slate generation and ranking in one autoregressive rollout

A second modern formulation treats closed-set ranking as a coupled generation-and-ordering problem carried out within one sequence (Surana et al., 13 May 2026). For each context $\mathcal{E}$8, the model first constructs a slate $\mathcal{E}$9 from a closed candidate set $\{\tau_i\}_{i=1}^{K}$0, then outputs a permutation $\{\tau_i\}_{i=1}^{K}$1 that orders the slate. The sequence is represented with two tagged segments,

$\{\tau_i\}_{i=1}^{K}$2

The joint policy is factorized as

$\{\tau_i\}_{i=1}^{K}$3

This factorization is not a decoupled pipeline in the conventional sense, because both phases are implemented by a single shared autoregressive model and the ranker is conditioned on the generated slate.

The expected objective combines coverage and ordering utility: $\{\tau_i\}_{i=1}^{K}$4 The slate reward is order-invariant,

$\{\tau_i\}_{i=1}^{K}$5

while the ranking reward is position-sensitive,

$\{\tau_i\}_{i=1}^{K}$6

For ranking, the paper instantiates the reward with NDCG@$\{\tau_i\}_{i=1}^{K}$7: $\{\tau_i\}_{i=1}^{K}$8

The paper’s central claim is that a single scalar reward applied uniformly to all tokens causes cross-phase contamination. It therefore replaces the standard GRPO rollout-level advantage with two phase-specific group-relative advantages: $\{\tau_i\}_{i=1}^{K}$9 The <SLATE> tokens are updated only with slate-quality signal, and the <RANK> tokens only with ordering signal. The combined loss is

$y_t$0

The empirical domains are sequential recommendation on MovieLens and LastFM, and multi-hop QA evidence selection on HotpotQA and MuSiQue. The reported models are Qwen3-4B-Instruct-2507 and Qwen3.5-2B. Across both tasks, F-GRPO improves over GRPO, zero-shot prompting, SFT, and decoupled SFT, with gains most pronounced when coverage is the bottleneck and at higher cutoffs such as Recall@3/5 and NDCG@3/5. A concrete example is LastFM with Qwen3-4B, where F-GRPO reaches 72.4 Recall@3 and 81.7 Recall@5, compared to GRPO’s 55.1 and 73.9. The analysis further reports that the slate generator matures first, the ranker matures later, the slate tends to have high recall but lower precision, and the ranker sharpens the output by trading some recall for better top-position precision.

5. Closed-loop proposal ranking in autonomous driving

CLOVER extends the closed-set ranking pattern to end-to-end autonomous driving planning (Ang et al., 14 May 2026). The system follows a lightweight generator–scorer formulation. Given an observation $y_t$1, the generator $y_t$2 outputs a finite candidate set

$y_t$3

and the scorer predicts a vector of planning-metric sub-scores

$y_t$4

with components corresponding to safety, feasibility, progress, and comfort style metrics such as collision avoidance, drivable-area compliance, TTC, progress, and comfort. The final decision is made by composing these predicted sub-scores with the NAVSIM planning metric. The method is therefore a proposal-selection planner: the generator determines support, and the scorer determines ranking order within that support.

Stage 1 expands proposal support beyond single-trajectory imitation by constructing evaluator-filtered pseudo-expert trajectories. The pseudo-expert families include lateral offsets, acceleration/deceleration profiles, stop-go behaviors, approach-brake behavior, boundary/off-road cases, and overshoot-like motions. Candidates are pre-filtered by drivable-area and occupancy checks, then scored, then selected by coverage over score patterns. The appendix gives the coverage key

$y_t$5

and greedy selection cost

$y_t$6

This makes the Stage 1 training signal explicitly set-level rather than single-point imitation.

Stage 2 performs conservative closed-loop self-distillation. The scorer is fitted to evaluator sub-scores on current proposals, while the generator is refined toward teacher-selected high-value proposals. Two teacher target sets are used: a top-$y_t$7 set selected by the composed predicted score, and a vector-Pareto set consisting of non-dominated proposals in predicted sub-score space. The vector-Pareto construction preserves multi-objective tradeoffs among safety, progress, and comfort, while stability regularization keeps updates conservative.

The paper is explicit that CLOVER does not require a globally perfect scorer. The main theorem states that if the scorer-selected target distribution $y_t$8 has more mass in the high-score region $y_t$9 than the current proposal distribution $y$0,

$y$1

and the generator update is conservative,

$y$2

then high-score support increases: $y$3 A sufficient margin-style condition is also given: if true high/low-score groups are separated by margin $y$4 and scorer error is bounded by $y$5, then high-score proposals will be ranked above low-score proposals whenever

$y$6

Evaluation uses NAVSIM planning metrics. For NAVSIM v1,

$y$7

For NAVSIM v2,

$y$8

The reported results are 94.5 PDMS on NAVSIM v1, 90.4 EPDMS on navtest with the updated official implementation, 87.2 EPDMS$y$9 with the original code, and 48.3 EPDMS on the NavHard split. On supplementary nuScenes evaluation, CLOVER obtains 0.31 L2 / 0.10 collision under ST-P3 protocol and 0.65 L2 / 0.30 collision under UniAD protocol. On 12,146 scenes with 64 proposals, the selected PDMS increases from 0.9369 for the baseline to 0.9413 after Stage 1 and 0.9448 after Stage 2. The paper interprets this as evidence that closed-loop ranking improves the proposal distribution itself, not only the final reranking rule.

6. Unifying principles, distinctions, and recurrent misconceptions

Taken together, these works suggest three recurring structural principles. First, the output space is closed: fixed items in the Plackett–Luce/GAS model, valid docID sequences in ARR, a provided candidate pool in F-GRPO, and a finite proposal set in CLOVER (Holý et al., 2021, Rozonoyer et al., 9 Jan 2026, Surana et al., 13 May 2026, Ang et al., 14 May 2026). Second, ranking is sequential: by temporal state recursion, by token generation, by staged slate-then-rank decoding, or by iterative proposal-ranking-distillation. Third, the ranking signal is not merely observational. In the dynamic ranking model, it updates future worths through the conditional score; in ARR, it is embedded directly into the training loss through SToICaL; in F-GRPO, it defines phase-specific advantages; in CLOVER, it feeds back into the next generator through teacher-selected targets.

One recurrent misconception is to equate autoregressive ranking with unconstrained text generation. The IR formulation explicitly constrains output to valid docID tokens, assigning probability zero outside the docID-token vocabulary. F-GRPO restricts the model to candidates from a closed pool. The statistical ranking model assumes a fixed item universe even when rankings are partial. CLOVER confines inference to the proposal set already generated. In all four settings, closed-set structure is not incidental; it is part of the definition.

A second misconception is to treat ranking as a purely terminal readout that does not affect generation. That description matches neither F-GRPO nor CLOVER. F-GRPO is designed precisely because candidate generation and ranking are coupled and because a poor result may stem from either poor coverage or poor ordering. CLOVER likewise argues that proposal-selection performance depends jointly on candidate-set coverage and scorer ranking quality, and its Stage 2 explicitly distills the generator toward scorer-selected targets. This suggests that in closed-set autoregressive systems, ranking often serves as a control signal over future support rather than as a passive scoring layer.

A third point of distinction concerns supervision. The dynamic ranking model is likelihood-based and probabilistic. ARR is trained with rank-aware supervised losses over tokenized docIDs. F-GRPO uses sequence-level reinforcement learning with separate group-relative advantages. CLOVER uses evaluator-filtered pseudo-expert supervision plus conservative self-distillation against evaluator-style sub-scores. These are not interchangeable training regimes, even when they instantiate a similar closed-set ranking pattern.

The resulting research landscape is therefore heterogeneous but coherent. Closed-set autoregressive ranking can refer to fixed-item stochastic ranking with autoregressive worth dynamics, constrained decoder-based ranking over docIDs, single-rollout slate construction followed by ranking, or closed-loop ranking-driven refinement over finite trajectory sets. The common thread is not a shared architecture, but the combination of bounded candidate support, sequential dependence, and a ranking objective that is structurally tied to how future ranked outputs are formed.