GenSco: Ordered Passage Alignment in Multi-hop QA

• GenSco is a retrieval-augmented generation framework that decomposes multi-hop questions into sub-questions and aligns passages to mirror the latent reasoning chain.
• It employs a two-LLM architecture where GPT-3.5 handles decomposition and answer generation, while Open-LLaMA-3B scores candidate passages based on negative log-likelihood.
• Empirical results on datasets like 2WikiMultiHop and MuSiQue show significant EM improvements by ensuring that selected passages are both relevant and correctly ordered.

Searching arXiv for the target paper and a few closely related retrieval/alignment works to support citations. Searching arXiv for "GenSco question decomposition passage alignment multi-hop QA (Fazili et al., 2024)". Searching arXiv for "ChronoRAG temporal question answering chronological passage assembling (Kim et al., 26 Aug 2025)". GenSco is a retrieval-augmented generation framework for multi-hop question answering that uses question decomposition to align an ordered sequence of passages with the latent reasoning chain of a query. It targets two failure modes in large-language-model-based RAG: inadequate or distracting context in the prompt, and weak factual reasoning even when relevant evidence is present. The central claim is that decomposing a multi-hop question into sub-questions and then selecting passages stepwise for those sub-questions produces context that is both more relevant and better ordered for final answer generation (Fazili et al., 2024).

1. Problem setting and rationale

Multi-hop QA requires composing evidence across multiple passages, often in a specific order. In a standard RAG pipeline, a retriever surfaces candidate passages and an LLM receives a fixed top-$k$ set as context for answer generation. GenSco is designed for the case where this top-$k$ context is either incomplete or contaminated by distractors, and where the model fails to follow the latent reasoning chain even when relevant evidence is present (Fazili et al., 2024).

The motivating idea is that a multi-hop question implicitly contains a sequence of reasoning steps. Decomposing the question into simpler sub-questions $q_1, q_2, \ldots$ exposes that sequence. Passage alignment then serves three functions. It selects passages that are maximally relevant to each step, orders them according to the reasoning chain, and filters distractors that would otherwise inflate prompt length and increase hallucinations. In the paper’s framing, this improves truthfulness by furnishing fine-grained external knowledge for each hop rather than relying on a monolithic retrieved context (Fazili et al., 2024).

This emphasis on ordered evidence distinguishes GenSco from settings where retrieval is treated only as a relevance-ranking problem. A plausible implication is that, for multi-hop QA, passage order is not merely a presentation choice but part of the reasoning substrate itself.

2. Two-LLM architecture and alignment procedure

GenSco consists of two distinct LLMs. The Generator LLM $G$ is used for question decomposition and final answer generation; the paper instantiates it as GPT-3.5 (text-davinci-003). The Scorer LLM $S$ is an auxiliary open-source model, Open-LLaMA-3B, used to semantically guide passage selection by scoring the compatibility between a sub-question and candidate passages via negative log-likelihood (Fazili et al., 2024).

The pipeline begins with empty context $C$ and an empty list of sub-questions. At each iteration, $G$ generates the next sub-question $q_i$ conditioned on the original question $Q$, the previous sub-questions, and the already selected passages. Then $S$ scores each candidate passage relative to $k$0, the highest-scoring passage is appended to the growing context, and a stopping criterion is checked. After decomposition stops, $k$1 is called once with $k$2 and the aligned passage sequence to generate the final answer (Fazili et al., 2024).

This design is explicitly cost-oriented. Unlike decomposition-based methods that answer every intermediate sub-question, GenSco invokes the generator once for final answer generation, while using $k$3 generator calls only for decomposition and $k$4 scorer calls for passage scoring. The paper presents this as a way to reduce latency and avoid compounding errors from incremental answer generation (Fazili et al., 2024).

Two stopping variants are defined. GenSco-max stops when the generator emits an end-of-decomposition keyword or repeats a sub-question. GenSco-stop adds an NLL-based stopping criterion using the scorer. The distinction matters empirically, because the stopping policy determines how aggressively the method explores the candidate pool.

3. Formalization and scoring

For a multi-hop question $k$5, GenSco decomposes the problem into an ordered set of sub-questions:

$k$6

Let the candidate passages be $k$7. At step $k$8, with already selected context

$k$9

the scorer assigns passage $q_1, q_2, \ldots$0 the score

$q_1, q_2, \ldots$1

where $q_1, q_2, \ldots$2 is the negative log-likelihood under $q_1, q_2, \ldots$3 and $q_1, q_2, \ldots$4 denotes concatenation (Fazili et al., 2024).

GenSco uses greedy alignment:

$q_1, q_2, \ldots$5

and then updates

$q_1, q_2, \ldots$6

For GenSco-stop, decomposition halts when adding the new sub-question hurts the scorer’s likelihood of regenerating the original question:

$q_1, q_2, \ldots$7

with $q_1, q_2, \ldots$8 (Fazili et al., 2024).

Final answer generation is a single generator call:

$q_1, q_2, \ldots$9

The paper also discusses the direction of scoring. GenSco scores the sub-question given passages rather than passages given the sub-question. By Bayes’ rule, either direction is admissible under approximately uniform passage priors, but conditioning on passages is presented as more practical because passages are longer and long-sequence likelihoods can approach zero (Fazili et al., 2024).

This formalization makes GenSco a post-retrieval filter-aligner rather than a retriever replacement. Candidate passages still come from a standard retriever; the novelty lies in the decomposition-guided selection and ordering layer.

4. Experimental configuration

The system is evaluated on three established multi-hop QA datasets: 2WikiMultiHop, Adversarial HotPotQA, and MuSiQue. The paper uses 1000 development QAs from 2WikiMultiHop, 308 development instances from Adversarial HotPotQA, and 1252 2-hop questions from MuSiQue. Candidate sets are supplied by BM25, GTR, or a cross-encoder retriever, and GenSco operates over those retrieved pools (Fazili et al., 2024).

Few-shot prompting is dataset-specific. The decomposition prompt includes prior sub-questions, selected subcontexts, few-shot examples, and an end-of-decomposition token "<FIN></FIN>". The final answer prompt also uses dataset-specific few-shot examples. The few-shot configuration is 2 shots for 2WikiMultiHop, 4 shots for Adversarial HotPotQA, and 3 shots for MuSiQue. For top-$G$0 passage feeding into final prompts, the paper uses top-5 for 2WikiMultiHop, top-2 for Adversarial HotPotQA, and top-4 for MuSiQue (Fazili et al., 2024).

The implementation uses max-levels $G$1 for 2WikiMultiHop, motivated by the fact that instances have at most five supporting passages. Temperature is 0 by default, although setting the generator temperature to 0.5 yields small but consistent gains. Infrastructure requirements are modest in the sense that no finetuning is required: OpenAI API access is used for $G$2, while $G$3 can run via Hugging Face on GPU or CPU (Fazili et al., 2024).

Evaluation uses Exact Match, Precision, Recall, and F1, with Exact Match defined as

$G$4

The paper also includes a faithfulness analysis using K-Precision, defined as lexical overlap of the generated answer with the passages (Fazili et al., 2024).

5. Empirical performance and ablation findings

GenSco reports its largest gains on MuSiQue and 2WikiMultiHop. On MuSiQue, GenSco-max+FS reaches 42.7 EM and GenSco-stop+FS 39.1 EM, compared with the best baseline EM of 27.6 from Self-Ask. On 2WikiMultiHop, GenSco-max+FS reaches 43.2 EM and GenSco-stop+FS 41.4 EM, compared with the best baseline EM of 37.3 from Self-Ask. On Adversarial HotPotQA, GenSco is competitive but does not surpass the strongest baseline in EM (Fazili et al., 2024).

Dataset	Best baseline EM	GenSco EM
2WikiMultiHop	37.3	43.2 / 41.4
Adversarial HotPotQA	56.8	54.87 / 55.52
MuSiQue	27.6	42.7 / 39.1

The gains are reported as +5.9 EM on 2WikiMultiHop and +15.1 EM on MuSiQue over the strongest inference-based baselines. Statistical significance is not explicitly reported, but the paper characterizes the MuSiQue improvements as large and the 2WikiMultiHop improvements as robust (Fazili et al., 2024).

The ablations isolate the role of decomposition and ordering. Removing question decomposition by replacing $G$5 with $G$6 causes severe degradation relative to GenSco-stop and GenSco-max, though performance remains above BM25. Randomly shuffling the aligned passage sequence before final answer generation significantly degrades EM, F1, Precision, and Recall, showing that the method benefits from sequence alignment rather than selection alone (Fazili et al., 2024).

The analysis on 2WikiMultiHop shows that GenSco-stop yields higher precision and lower recall than GTR, which the paper interprets as selecting fewer distractors. The faithfulness analysis is more nuanced: GTR can obtain higher K-Precision than GenSco even when its answer F1 is lower, and the reported Pearson correlations between K-Precision and F1 are only approximately $G$7 to $G$8. The paper’s conclusion is that correctness improves when the number and order of passages are tailored to the reasoning chain, even if raw lexical overlap with passages is lower (Fazili et al., 2024).

A qualitative example illustrates the mechanism. For the question asking for the place of birth of the director of The One and Only Ivan, GenSco first generates the sub-question identifying the director, selects the passage mentioning Thea Sharrock, then generates the second sub-question asking for Thea Sharrock’s place of birth, selects the passage containing “London, England,” and finally answers “London, England.” In the same example, the GTR baseline misses a relevant passage and answers “Bucharest” because of a distractor about a Romanian “Ivan” (Fazili et al., 2024).

6. Position within the retrieval and RAG literature

GenSco is situated between two research lines: LLM-based decomposition for reasoning and retrieval-side passage selection. Relative to decomposition-based QA methods such as Self-Ask, ReAct, and Tree-of-Thought, GenSco performs decomposition only for context alignment and does not answer intermediate steps, which reduces latency and the risk of compounding generation errors (Fazili et al., 2024). Relative to Verify-and-Edit, which focuses on post-editing reasoning chains, GenSco operates pre-generation by aligning context to the expected reasoning path (Fazili et al., 2024).

Its emphasis on ordering also connects to a broader set of passage-centric methods. In narrative temporal QA, ChronoRAG constructs chronologically assembled passage neighborhoods and shows that preserving narrative order improves ROUGE-L on NarrativeQA and its Time Questions subset (Kim et al., 26 Aug 2025). In multi-hop QA reranking, later work on contextual passage utility argues that passage usefulness depends on inter-passage dependencies and reasoning order rather than independent relevance alone (Jain et al., 6 Dec 2025). These results are not part of GenSco itself, but they suggest that GenSco’s alignment principle is compatible with a wider shift from isolated passage relevance toward context-sensitive passage utility and ordering.

The method’s limitations are explicit. Performance depends on decomposition quality; if the generator emits suboptimal or redundant sub-questions, passage alignment degrades. The scorer LLM must provide reliable NLL preferences, and smaller open-source models may be noisy. Benefits increase when candidate sets are larger, with the paper stating that GenSco thrives when there are at least ten candidate passages and that gains shrink when the pool is very small, as in Adversarial HotPotQA with only four passages per question (Fazili et al., 2024).

GenSco also does not argue that long context windows alone solve multi-hop reasoning. Instead, it treats alignment as complementary to long-context modeling: more context tokens do not eliminate the need to reduce distractors and preserve ordered evidence. The practical implication is that, even with strong LLMs, multi-hop QA can benefit from a training-free post-retrieval layer that transforms a candidate set into a reasoning-aligned passage sequence (Fazili et al., 2024).