ParSeR: Provision-Guided Legal QA

• The paper presents ParSeR as a framework that uses LLM-generated parametric provisions to guide targeted retrieval for statute-grounded, multi-hop legal question answering.
• ParSeR employs a three-stage pipeline—retrieve, rerank, and selection—to efficiently aggregate statutory evidence and produce answers with legal fidelity.
• Experimental results on the KoBLEX benchmark show significant improvements over baselines, highlighting ParSeR's robustness and precise retrieval of supporting legal provisions.

Searching arXiv for the cited ParSeR paper and closely related legal QA work to ground the article in recent literature. Searching arXiv for "KoBLEX ParSeR legal question answering". Parametric provision-guided Selection Retrieval (ParSeR) is a retrieval-augmented reasoning framework for provision-grounded, multi-hop legal question answering. It was introduced together with KoBLEX, a Korean Benchmark for Legal EXplainable QA, to address a setting in which answers must be supported by statutory provisions rather than produced as ungrounded free-form legal analysis (Lee et al., 1 Sep 2025). The central idea is to use LLM-generated parametric provisions—provision-like queries generated from the model’s own knowledge—to guide a targeted retrieval pipeline over a statutory corpus, and then to generate answers grounded in the selected statutes rather than in parametric knowledge alone (Lee et al., 1 Sep 2025).

1. Problem setting and motivation

ParSeR was proposed in response to a limitation identified in legal QA evaluation: several legal benchmarks evaluate legal capabilities, but they do not evaluate open-ended and provision-grounded QA (Lee et al., 1 Sep 2025). KoBLEX was designed specifically for this gap. It contains 226 scenario-based QA instances and their supporting provisions, created using a hybrid LLM-human expert pipeline, and covers 1-hop, 2-hop, and 3-hop reasoning (Lee et al., 1 Sep 2025).

Within this setting, the retrieval problem is not merely document lookup. The legal question must be answered with explicit statutory grounding, and many questions require multi-hop reasoning across several provisions. The ParSeR formulation treats this as a provision retrieval and answer generation problem in which legal authority must be gathered in a structured way before answering (Lee et al., 1 Sep 2025).

A later legal QA study characterizes ParSeR as a method that reformulates a legal question into a statute-aligned “parametric provision” query in order to bridge the vocabulary gap between user questions and statutory text (Lee et al., 23 May 2026). That description is consistent with the original presentation of ParSeR as provision-guided retrieval rather than ordinary question-based retrieval.

2. Core concept: parametric provisions

The defining component of ParSeR is parametric provision generation. In this step, the LLM is prompted to output a list of parametric provisions that resemble the language and structure of real statutes but are generated from the model’s parametric knowledge (Lee et al., 1 Sep 2025). Their purpose is to identify and decompose the legal question into core statutory concepts likely to be relevant for reasoning.

This mechanism distinguishes ParSeR from approaches that retrieve directly from the original question. In the original description, the generated provisions serve as focused queries that better mirror the intended legal reasoning paths than free-form questions (Lee et al., 1 Sep 2025). A common misconception is to treat these generated provisions as evidence. They are not the final legal authorities; they are query surrogates used to retrieve real statutory provisions from the legal corpus.

The rationale stated for this design is that provision guidance is crucial. The reported ablations show that naïve retrieval with original questions for each hop performs much worse, which the paper presents as confirmation of ParSeR’s core design (Lee et al., 1 Sep 2025). This suggests that the method’s main contribution lies not only in adding retrieval steps, but in changing the representational form of the retrieval query.

3. Three-stage sequential retrieval and answer generation

ParSeR operationalizes provision-guided reasoning through a three-stage sequential retrieval process followed by provision-grounded answer generation (Lee et al., 1 Sep 2025).

Stage	Operation	Function
1	Retrieve	For each parametric provision $p_n$, retrieve top-$k$ statutory provisions
2	Rerank	Rerank the retrieved passages with a deeper semantic model
3	Selection	Use an LLM to select the most relevant provision among reranked candidates

In the retrieve stage, for each parametric provision $p_n$, a bi-encoder retriever retrieves the top-$k$ most relevant statutory provisions from the legal corpus using cosine similarity. The retriever can be BM25 for sparse retrieval or BGE for dense retrieval (Lee et al., 1 Sep 2025).

In the rerank stage, the top-$k$ retrieved passages are reranked using a cross-encoder or fine-tuned reranker for deeper semantic assessment. The experimental setup specifies BGE-M3, fine-tuned for Korean legal text, as the reranker (Lee et al., 1 Sep 2025).

In the selection stage, an LLM selects the most relevant provision among the reranked top-$l$ passages. The stated purpose of this step is to use the model’s nuanced understanding to choose exactly the provision or provisions that directly support the answer (Lee et al., 1 Sep 2025).

After retrieval, reranking, and selection, the aggregated supporting provisions are used as grounding context for the final answer. The paper states that this enables provision-guided multi-hop reasoning because multiple distinct legal provisions may be gathered across several articles or statutes before answer generation (Lee et al., 1 Sep 2025). The overall design is therefore not a single-pass RAG system, but a staged evidence-construction pipeline.

4. Experimental setting in KoBLEX

The experimental environment for ParSeR is tightly defined. KoBLEX provides scenario background, question, answer, and annotated supporting statutes for each instance, and is provided in both Korean and English (Lee et al., 1 Sep 2025). The statute corpus contains 608 statutes and approximately 233,544 paragraph-level provisions (Lee et al., 1 Sep 2025).

The evaluated LLMs are Qwen3, EXAONE-3.5, and GPT-4o, with GPT-4o described as selected for Korean domain expertise in the experimental summary (Lee et al., 1 Sep 2025). Retrieval infrastructure includes BM25, BGE, and hybrid retrieval, with BGE-M3 used as the reranker (Lee et al., 1 Sep 2025).

The baselines are broad and include Standard Prompting (SP), Chain of Thought (CoT), One-time Retrieval (OR), and multi-hop or multi-step RAG variants such as Self-Ask, IRCoT, FLARE, ProbTree, and BeamAggr (Lee et al., 1 Sep 2025). The reported protocol states that all methods use the same retrieval pool for fairness (Lee et al., 1 Sep 2025).

This benchmark design matters because ParSeR is evaluated not only on answer overlap but also on provision retrieval quality and legal fidelity. As a result, the method is assessed as both a retrieval system and a grounded reasoning system rather than as an answer generator alone.

5. Evaluation metrics and legal fidelity

ParSeR is evaluated with separate metrics for provision retrieval and answer generation (Lee et al., 1 Sep 2025). For provision retrieval, the benchmark uses Exact Match (EM) and Provision F-1. EM is defined as 1 if the retrieved set of provisions exactly matches gold and 0 otherwise (Lee et al., 1 Sep 2025). Provision F-1 measures overlap between retrieved and gold supporting provisions (Lee et al., 1 Sep 2025).

For generated answers, the benchmark uses Token-level F-1 and LF-Eval (Legal Fidelity Evaluation) (Lee et al., 1 Sep 2025). Token-level F-1 measures overlap between generated and reference answer tokens. LF-Eval is an automatic, LLM-as-a-judge metric that jointly considers the question, supporting provisions, expected answer, and model prediction (Lee et al., 1 Sep 2025).

The LF-Eval procedure uses GPT-4o to rate responses on a 1–10 scale with justification, penalizing legal inaccuracies, contextual omissions, and contradictions to provisions (Lee et al., 1 Sep 2025). Its scoring formula is reported as

$$\text{LF-Eval} = \frac{1}{10} \sum_{i=1}^{10} s_i \cdot p(s_i)$$

where $s_i$ is the score from the $i$th GPT-4o output and $p(s_i)$ is its associated token probability (Lee et al., 1 Sep 2025).

The paper reports that LF-Eval shows a strong Pearson correlation of 84.90 with human expert ratings and outperforms standard metrics in this respect (Lee et al., 1 Sep 2025). This is significant because ParSeR is intended for legally grounded QA, where fluent but weakly grounded answers are insufficient. LF-Eval is therefore aligned with the method’s emphasis on statutory support rather than surface plausibility.

6. Reported performance, ablations, and later interpretation

The reported experimental results state that ParSeR consistently outperforms all baselines across retrieval and generation metrics and for all tested LLMs: Qwen3, EXAONE, and GPT-4o (Lee et al., 1 Sep 2025). With GPT-4o, ParSeR improves Provision F-1 by +37.91 over one-time retrieval, scores +30.81 higher in LF-Eval than vanilla retrieval, and outperforms the strongest baseline, ProbTree, by +12.23 in Token F-1 (Lee et al., 1 Sep 2025).

The method is also reported to remain robust across 1-hop, 2-hop, and 3-hop questions, whereas other baselines often degrade as reasoning depth increases (Lee et al., 1 Sep 2025). The efficiency analysis states that ParSeR achieves the highest LF-Eval with the fewest generated tokens, which the paper interprets as cost-effectiveness (Lee et al., 1 Sep 2025).

Ablation findings are equally central to the method’s characterization. Removing any ParSeR component degrades performance, with the steepest drop occurring when both reranking and LLM-based selection are omitted (Lee et al., 1 Sep 2025). The same analysis identifies provision guidance via parametric provisions as crucial, because replacing provision-guided retrieval with repeated retrieval from the original questions performs much worse (Lee et al., 1 Sep 2025).

A later framework, Decompose-and-Refine (DaR), positions itself as an extension of ParSeR and offers an informative retrospective characterization (Lee et al., 23 May 2026). According to that work, ParSeR operates at the level of the whole question, which can conflate multiple legal issues and lead to mixed or less interpretable evidence for multi-hop queries (Lee et al., 23 May 2026). DaR responds by explicitly decomposing questions into atomic sub-questions and selecting exactly one core statutory provision per sub-question (Lee et al., 23 May 2026). This later critique does not negate ParSeR’s contribution; rather, it identifies the whole-question granularity of ParSeR as the next target for refinement. A plausible implication is that ParSeR established the value of statute-aligned parametric querying, while subsequent work focused on making that querying more issue-specific and more transparent at the sub-question level (Lee et al., 23 May 2026).

7. Position within provision-grounded legal QA

ParSeR occupies a specific position in legal QA research: it is a provision-guided retrieval framework for open-ended, multi-hop statutory question answering, not merely a prompting strategy and not merely a generic retriever (Lee et al., 1 Sep 2025). Its architecture combines parametric knowledge and external statutory evidence in a constrained way: the model’s internal knowledge is used to generate provision-like retrieval guides, while the final answer is grounded in actual retrieved statutes (Lee et al., 1 Sep 2025).

The method’s stated strengths are provision-guided retrieval, the retrieve–rerank–select pipeline, groundedness in actual statutes, and robustness to increasing reasoning depth (Lee et al., 1 Sep 2025). These properties explain why the paper frames ParSeR as legally grounded and reliable rather than only accurate in an answer-matching sense.

For legal QA, the main conceptual contribution of ParSeR is the shift from question-conditioned retrieval to provision-conditioned retrieval. In the KoBLEX setting, this design is paired with a benchmark, an evaluation metric, and an empirical demonstration that provision-guided selection retrieval improves both statutory evidence recovery and answer quality (Lee et al., 1 Sep 2025). Later work that extends ParSeR continues to treat this parametric retrieval idea as foundational, particularly for handling the vocabulary gap between fact descriptions in user questions and the formal language of statutes (Lee et al., 23 May 2026).