Final-Step Matching Essentials

• Final-step matching is a decisive phase in multi-step processes that resolves remaining uncertainty through binary decisions based on prior evidence.
• It applies across various domains—such as labor markets with vacancy chains and online matching algorithms—to ensure stability and optimal transitions.
• LLM-based entity matching exemplifies final-step matching by synthesizing token-level analyses into a forced binary outcome, enhancing clarity in decision-making.

Final-step matching refers, across multiple research domains, to the decisive stage in a multi-step assignment, integration, or stabilization process where an explicit, usually binary, decision is made (e.g., a definitive matching, hiring, or entity link) based on cumulative prior computations or collected evidence. In settings ranging from stable labor markets to entity resolution with LLMs, final-step matching mechanisms resolve outstanding uncertainty: this may involve resolving vacancy chains in labor markets, determining student–advisor assignments under practical constraints, or synthesizing evidence for entity coreference. The precise structure, complexity, and guarantees of final-step matching vary significantly depending on model assumptions, the format of inputs, and application goals.

1. Final-Step Matching in Stable Labor Markets

In the one-to-one labor matching framework, final-step matching models the sequence of transitions required to restore stability after a destabilizing event, such as a voluntary resignation. Suppose a worker $w^0$ resigns in pursuit of a preferred firm; the final-step matching problem is to count and describe the exact steps by which the market transitions from the initial stable matching $\mu$ to a new stable matching $\nu$ in which $w^0$ has improved her position. This re-stabilization process is formalized as a vacancy-chain algorithm, where each step corresponds to a reassignment event (i.e., a previously matched worker occupying the current vacancy, thereby creating a new vacancy) (Bonifacio et al., 2024).

The length of the vacancy chain is intrinsically linked to the underlying lattice structure of the set of stable matchings. Rotations (cycles in the reduced preference profile) form the combinatorial backbone of the transition process. Given a sequence of $k$ rotations $\sigma^1, \ldots, \sigma^k$ connecting $\mu$ to $\nu$, the total number of steps executed by the vacancy-chain algorithm is

$L = |\sigma^k| + \sum_{i=1}^{k-1} (|\sigma^i| - 1)$

with $|\sigma^i|$ denoting the cardinality of the $\mu$0th cycle. This count encapsulates both the structure of the market's preference lattice and the specific instability being resolved (Bonifacio et al., 2024).

2. Final-Step Matching in Multi-Stage Stable Matching Algorithms

Final-step matching also arises in multi-stage or online assignment problems, where interventions, attrition, or preference modifications necessitate stable re-allocation. In the online two-stage stable matching problem, the goal is to minimize disruption (i.e., the number of assignment changes) between an initial and a final stable matching, under uncertainty about agent arrivals or departures (Bampis et al., 2022). The "final step" here is the computation of $\mu$1, the final stable matching, after observing modifications to the agent population.

A key result is the identification of a dominance property: computing a men-optimal (or equivalently, university-optimal) matching $\mu$2 at the first stage minimizes, in the worst-case, the number of modifications required in the final stage. Subsequently, $\mu$3 is computed as the maximum-overlap stable matching with respect to $\mu$4, ensuring minimal reassignments (so-called "divorces") (Bampis et al., 2022). The procedure is provably optimal for two-stage markets, but no competitive online algorithm exists if more than two stages are allowed.

3. Stepwise Synthesis in LLM-based Entity Matching

In the context of entity matching using LLMs, final-step matching denotes the decisive chain-of-thought prompt where the LLM synthesizes earlier analyses—token-level comparisons and attribute importance—for a forced Yes/No entity resolution decision (Bopardikar et al., 28 Nov 2025). Here, explicit multi-step decomposition improves transparency and, in some scenarios, accuracy.

The step-3 prompt structure concatenates token match status and ranked attribute relevance (outputs of previous steps). The LLM then responds:

• "YES"/"NO": binary outcome indicating entity equivalence,
• with a brief justification sentence (implicit short-form chain-of-thought).

No explicit scoring function, probability threshold, or calibration is used for the final decision, and the framework eschews additional soft-score aggregation—output is always forced to be binary (Bopardikar et al., 28 Nov 2025). In a debate-based variant, separate prompts solicit supporting and opposing arguments, but this method was less empirically effective for end-to-end F1.

4. Mechanism Design and Stability Guarantees

In many real-world assignment systems, final-step matching algorithms are required to guarantee stability, Pareto-optimality, and, often, side-optimality (e.g., student-optimal or firm-optimal outcomes).

The PolyGS algorithm, as implemented in the ELLIS 2020 PhD matching program, embodies final-step matching in a many-to-one market by executing a deferred-acceptance process yielding a stable, student-optimal matching. After a sequence of filtering, interviews, and pre-selection, the final matching step applies PolyGS to strict, empirically constructed preferences, with stability formally defined as the absence of blocking pairs (Mordig et al., 2021). Empirically, nearly all students are matched to their top-choice advisor, and the outcome is provably unique under strict preferences.

5. Finalization under Incomplete or Evolving Preferences

Final-step matching is especially subtle when inputs are incomplete or uncertain, as in the Finalizability of Tentative Matches (FTM) problem. Here, a matching step is considered final if, for every possible extension of truncated preference lists, a specific tentative assignment persists in the resulting resident-optimal matching (Tamaki, 2016).

The FTM decision problem is coNP-complete in general, even for stable marriage with unit quotas, though tractable in restricted cases such as one-to-one, resident-minimal instances (where further truncation would alter tentative matchings). For such instances, negative certificates can be captured combinatorially (as prescriptions) or via integer programming formulations (Tamaki, 2016). Additionally, there exists a polynomial-time sufficient condition: a tentative match is finalizable if it is not endangered at any iteration of a safe-set peeling procedure.

6. Complexity, Algorithmic Structure, and Practical Behavior

Different models impose varied computational and interpretive burdens on the final-step matching subroutine.

• In the vacancy-chain model with rotations, the number of steps is governed by the combinatorics of preference cycles, which can be enumerated in polynomial time but may require non-trivial bookkeeping (Bonifacio et al., 2024).
• Multi-stage stable matching with minimum-divorce objectives reduces final-step computation to constrained optimization over stable matching polytopes, solvable in polynomial time for two stages but intractable beyond that (Bampis et al., 2022).
• LLM-based entity matching involves only prompt design and forced binary parsing, with the bulk of algorithmic complexity in prior reasoning stages or data pre-processing (Bopardikar et al., 28 Nov 2025).
• In the FTM setting, coNP-completeness for general or high-quota cases contrasts with efficient combinatorial solutions in specific structured inputs, with practical heuristics such as maximal safe set construction reliably pseudo-finalizing most matches (Tamaki, 2016).

A summary of archetypal final-step matching contexts is given below:

Setting	Stability Criterion	Complexity
Labor vacancy chain (Bonifacio et al., 2024)	Stable, rotation-based	Polytime
Online 2-stage (Bampis et al., 2022)	Stable, min divorce	Polytime (2-stage)
LLM entity matching (Bopardikar et al., 28 Nov 2025)	Forced Y/N classification	Prompt-based
FTM with truncation (Tamaki, 2016)	Persistent under completions	coNP-c (general), polytime (1-FTM-RM)
PolyGS hiring (Mordig et al., 2021)	Stable, side-optimal	Polytime

7. Illustrative Examples and Variants

Demonstrative case studies underscore subtlety in final-step matching frameworks:

• In the vacancy-chain process, the sequence and size of preference cycles directly dictate the total number of required stepwise reassignments before stability is reattained. For example, two cycles of length 2 yield three matching events for a single worker resignation (Bonifacio et al., 2024).
• In entity resolution, explicit structured prompts fed to LLMs partition evidence and focus the final, interpretive binary classification (Bopardikar et al., 28 Nov 2025).
• For incomplete preference lists, finalizability can often be ascertained for a majority of matches via a polynomial-time "safe match" test, but pathological cases may require inspection or generate computational hardness (Tamaki, 2016).

Taken together, these mechanisms illustrate final-step matching as a unifying but multifaceted subroutine: it is the locus where stability, optimality, minimal disruption, or semantic alignment is enforced, subject to the entire structure of preceding steps and the underlying domain’s combinatorial or inferential framework.