EviReform: Evidence-Guided Query Reformulation for Multi-Hop Graph Retrieval

# EviReform: Evidence-Guided Query Reformulation for Multi-Hop Graph Retrieval
Source: [https://arxiv.org/html/2608.13006](https://arxiv.org/html/2608.13006)
Yoshua Y\. LiAffiliation:Meituan

###### Abstract

Multi\-hop retrieval must recover passages that provide sufficient evidence together\. An initial passage often resolves an entity or relation implicit in the question, making the missing evidence easier to describe only after retrieval begins\. Graph retrieval improves access to related evidence through stored corpus structure, but its retrieval signal is commonly derived from the original question\. Complementary evidence must then be reached through stored relations even when an observed passage provides a more direct semantic cue\. We introduce EviReform, which separates revising the retrieval request from aggregating evidence in the graph\. Retrieved source passages formulate residual queries for the unresolved information need\. The original and residual retrieval signals are normalized separately, combined, and propagated between propositions that share entities\. On 2WikiMultiHopQA, HotpotQA, and MuSiQue, EviReform exceeds the strongest baseline by up to 5\.59 Recall@5 points and 4\.50 F1 points\. These results show that observed evidence can guide graph retrieval toward the part of a supporting chain left underspecified by the original question\. Code is available at[https://github\.com/XrazyMee/EviReform](https://github.com/XrazyMee/EviReform)\.

## 1Introduction

Retrieval\-augmented generation \(RAG\) grounds language models in external evidence\[[16](https://arxiv.org/html/2608.13006#bib.bib1)\]\. Multi\-hop questions make that evidence interdependent: answering may require passages about different entities or events, and the relevance of a later passage may only become apparent after an earlier passage resolves a bridge\. Retrieval must therefore recover complementary evidence across hops, rather than only passages that independently resemble the original question\[[33](https://arxiv.org/html/2608.13006#bib.bib24),[15](https://arxiv.org/html/2608.13006#bib.bib25)\]\.

Dense retrievers score each passage against a representation of the question\[[11](https://arxiv.org/html/2608.13006#bib.bib20)\]\. GraphRAG addresses the resulting fragmentation by organizing entities, propositions, and passages into corpus structures\. PropRAG searches proposition paths and refines graph seeds; HippoRAG 2 improves semantic entry and diffuses relevance with Personalized PageRank; CatRAG adjusts traversal to the query\[[28](https://arxiv.org/html/2608.13006#bib.bib5),[7](https://arxiv.org/html/2608.13006#bib.bib3),[13](https://arxiv.org/html/2608.13006#bib.bib4)\]\. These mechanisms improve how relevance reaches structurally related evidence\.

Structure alone does not capture every change introduced by retrieval\. Suppose an initial passage identifies the person, location, or relation implicit in the question\. That observation makes the remaining information need more specific than it was before retrieval\. A graph retriever whose seeds, paths, or edge scores remain tied to the original question must recover the complementary passage through its stored relations, even though the observed passage now provides a more direct description of what is missing\. The issue is not whether graph traversal adapts to the question; it is whether retrieved evidence can revise the query signal that enters graph retrieval\.

This distinction separates two decisions that are often treated together\. Before any passage is read, the question can determine where retrieval enters the graph and how the graph is traversed\. After a passage is read, the system can also reconsider what textual evidence it should seek\. The first decision exploits relations already stored in the corpus; the second uses the content of the retrieved passage to specify a new information need\. Better traversal does not remove the need for this second decision when the bridge is implicit, absent from the graph, or more directly expressed in passage text\. Conversely, reformulation does not replace graph structure: the evidence retrieved for the revised query can still be incomplete or dispersed across related propositions\. Our method combines these two roles\.

Using evidence to guide later queries has a substantial history\. Relevance feedback updates queries from retrieved documents\[[22](https://arxiv.org/html/2608.13006#bib.bib22),[36](https://arxiv.org/html/2608.13006#bib.bib23)\], while multi\-hop systems condition later retrieval on passages, generated subqueries, or evolving reasoning states\[[21](https://arxiv.org/html/2608.13006#bib.bib14),[32](https://arxiv.org/html/2608.13006#bib.bib21),[27](https://arxiv.org/html/2608.13006#bib.bib6),[23](https://arxiv.org/html/2608.13006#bib.bib16)\]\. MIGRES and S2G\-RAG explicitly identify missing information or insufficient evidence before another retrieval step\[[29](https://arxiv.org/html/2608.13006#bib.bib17),[17](https://arxiv.org/html/2608.13006#bib.bib18)\]\. Agentic graph systems further combine query generation with graph interaction, memory, stopping, and answer generation\[[24](https://arxiv.org/html/2608.13006#bib.bib36),[18](https://arxiv.org/html/2608.13006#bib.bib37),[35](https://arxiv.org/html/2608.13006#bib.bib38),[30](https://arxiv.org/html/2608.13006#bib.bib39)\]\. These studies establish the value of letting evidence influence later retrieval\. We focus on the interface between this operation and graph retrieval: how observed passages can revise the retrieval signal, while the original question continues to constrain the evidence that is ultimately ranked\.

EviReform implements this separation directly\. Initial proposition retrieval identifies source passages that reveal what the question has already resolved and what remains missing\. A reformulator expresses the remaining need as a small set of residual queries\. Their retrieval signals are normalized separately from the signal produced by the original question, then the two channels are combined\. Graph propagation aggregates propositions connected to either channel, and passage readout produces one ranking\. Reformulation therefore changes what evidence is sought after observation, while graph structure consolidates the evidence reached by both requests\.

This work makes three contributions:

- •We formulate graph retrieval after observation as a ranking problem conditioned on both the original question and initially retrieved passages\.
- •We introduce EviReform, which separately normalizes the original and residual retrieval signals, combines them, and aggregates their evidence through shared entities into one passage ranking\.
- •We evaluate EviReform on 2WikiMultiHopQA, HotpotQA, MuSiQue, and GraphRAG\-Bench \(Medical\)\. It exceeds the strongest baseline by up to 5\.59 Recall@5 points and 4\.50 F1 points; controlled studies trace the gains to reformulation and propagation\.

## 2Related Work

Figure[1](https://arxiv.org/html/2608.13006#S2.F1)summarizes the three lines of retrieval research discussed below\.

Figure 1:Retrieval paradigms discussed in this section\. \(a\) GraphRAG uses stored corpus structure to retrieve related evidence\. \(b\) Iterative retrieval uses observed evidence to guide a subsequent search\. \(c\) Agentic graph retrieval coordinates graph interaction through memory, reflection, and stopping decisions\.### 2\.1GraphRAG for Retrieval

GraphRAG systems organize entities, relations, propositions, passages, or communities to recover evidence that is fragmented across text\[[4](https://arxiv.org/html/2608.13006#bib.bib2),[8](https://arxiv.org/html/2608.13006#bib.bib29)\]\. Systems designed for retrieval differ in how they use this structure\. PropRAG builds a proposition graph, explores proposition paths with beam search, and constructs refined seeds for a second graph ranking stage\[[28](https://arxiv.org/html/2608.13006#bib.bib5)\]\. HippoRAG 2 links the question to passages and extracted triples, filters triple candidates with recognition memory, and initializes PPR from the retained signals\[[7](https://arxiv.org/html/2608.13006#bib.bib3)\]\. CatRAG modifies anchoring, edge weights, and passage bias according to the question, while QAFD\-RAG similarly adapts graph diffusion to the query\[[13](https://arxiv.org/html/2608.13006#bib.bib4),[38](https://arxiv.org/html/2608.13006#bib.bib19)\]\. LightRAG and KG2RAG provide additional forms of graph organization and expansion\[[6](https://arxiv.org/html/2608.13006#bib.bib27),[39](https://arxiv.org/html/2608.13006#bib.bib28)\]\. These methods develop different mechanisms for question anchoring, path discovery, and structural propagation\.

### 2\.2Iterative Retrieval Guided by Evidence

Retrieval feedback predates RAG: Rocchio updates a query from judged documents, and dense pseudo\-relevance feedback encodes the question together with initial passages\[[22](https://arxiv.org/html/2608.13006#bib.bib22),[36](https://arxiv.org/html/2608.13006#bib.bib23)\]\. Multi\-hop retrievers extend this idea across hops\. GoldEn Retriever generates searches from available context, Baleen condenses earlier evidence, and MDR learns passage\-conditioned retrieval paths\[[21](https://arxiv.org/html/2608.13006#bib.bib14),[12](https://arxiv.org/html/2608.13006#bib.bib15),[32](https://arxiv.org/html/2608.13006#bib.bib21)\]\. IRCoT interleaves retrieval with chain\-of\-thought reasoning, whereas Iter\-RetGen and FLARE retrieve from evolving generations\[[27](https://arxiv.org/html/2608.13006#bib.bib6),[23](https://arxiv.org/html/2608.13006#bib.bib16),[10](https://arxiv.org/html/2608.13006#bib.bib26)\]\. MIGRES explicitly generates queries for missing information, and S2G\-RAG couples gap descriptions with evidence\-sufficiency decisions\[[29](https://arxiv.org/html/2608.13006#bib.bib17),[17](https://arxiv.org/html/2608.13006#bib.bib18)\]\. Together, these systems show how retrieved evidence can guide subsequent searches\.

### 2\.3Agentic Graph Retrieval

Recent systems combine queries derived from evidence with graph interaction\. GeAR maintains a gist memory, judges answerability, rewrites the query, and repeatedly invokes graph retrieval\[[24](https://arxiv.org/html/2608.13006#bib.bib36)\]\. Graph\-R1 models reflection, query generation, graph retrieval, and answering as a learned policy over several turns, while GraphRAG\-R1 optimizes retrieval behavior with reinforcement learning and combines graph and text retrieval\[[18](https://arxiv.org/html/2608.13006#bib.bib37),[35](https://arxiv.org/html/2608.13006#bib.bib38)\]\. ToG\-3 evolves both its query and retrieved subgraph through a loop that judges evidence sufficiency\[[30](https://arxiv.org/html/2608.13006#bib.bib39)\]\. These systems integrate graph retrieval into adaptive reasoning and answer generation\.

## 3EviReform

Figure 2:EviReform retrieves propositions for the original question and observes their source passages\. A reformulator expresses the unresolved information need as residual queries\. The original and residual retrieval signals are combined, propagated through shared entities, and read out as one passage ranking\.Figure[2](https://arxiv.org/html/2608.13006#S3.F2)presents EviReform\. The method first retrieves propositions for the original question, then observes their complete source passages to formulate residual queries\. Signals from the original question and residual queries are combined before graph propagation and passage ranking\.

### 3\.1Evidence Dependency across Hops

Letqqbe a question andD=\{dj\}j=1ND=\\\{d\_\{j\}\\\}\_\{j=1\}^\{N\}a passage corpus\. A retriever returns a rankingR⁡\(q,D\)R\(q,D\), from which a downstream reader receives the firstKKpassages\. For a multi\-hop question, the required evidence may be distributed across several passages\. An initially retrieved setEq⊂DE\_\{q\}\\subset Dcan resolve an intermediate entity or relation and thereby make the remaining evidence easier to describe\. The retrieval signal for the next passage can therefore depend on both the original question and the evidence already observed\.

The retrieval objective is not tied to a particular cutoff: it is to place a compact, jointly sufficient set of passages early in the ranking\. We useK=5K=5in the experiments, but the underlying requirement is that the selected passages support the answer together\. Gold passages provide an observable proxy for that requirement\. IfEqE\_\{q\}already contains one part of a chain, the useful next passage is better characterized by its relevance to\(q,Eq\)\(q,E\_\{q\}\)than by its independent similarity toqq:

rel⁡\(d∣q,Eq\)≠rel⁡\(d∣q\)in general\.\\operatorname\{rel\}\(d\\mid q,E\_\{q\}\)\\neq\\operatorname\{rel\}\(d\\mid q\)\\quad\\text\{in general\}\.\(1\)The difference is greatest when the first passage resolves a bridge that is only implicit in the question\.

### 3\.2Graph Retrieval from the Original Question

Let𝐛⁡\(q\)\\mathbf\{b\}\(q\)denote the retrieval seed obtained from the original question\. A graph retrieval system propagates this seed through a corpus graph and reads the resulting unit scores back to passages\. At the level needed here, the ranking can be written as

𝐳q=𝐁⊤​𝐏G​\(q\)​𝐛​\(q\),\\mathbf\{z\}\_\{q\}=\\mathbf\{B\}^\{\\top\}\\mathbf\{P\}\_\{G\}\(q\)\\mathbf\{b\}\(q\),\(2\)where𝐏G​\(q\)\\mathbf\{P\}\_\{G\}\(q\)may depend on the question and𝐁\\mathbf\{B\}maps graph units to passages\. Paths or edge weights that adapt to the question change𝐏G​\(q\)\\mathbf\{P\}\_\{G\}\(q\), while better linking changes𝐛⁡\(q\)\\mathbf\{b\}\(q\)\. Equation 2 nevertheless contains no term for retrieved passages when both are computed fromqqalone\.

This observation covers several strong forms of graph retrieval\. A filter can choose more accurate graph seeds from candidates retrieved with the question, and a traversal policy can adjust paths or edge weights according to the question\. Both improve the use of the corpus graph\. They still solve the ranking problem induced by the original request\. When an observed passage reveals a bridge, the remaining need must either be reached through the stored relations or be expressed as another retrieval request\. We study the latter operation and then retain graph propagation for combining the resulting evidence\.

### 3\.3Reformulating the Query from Retrieved Evidence

After observingEqE\_\{q\}, a reformulator describes what remains unresolved:

𝝆=ℛ⁡\(q,Eq\)=\{ρ1,…,ρL\}\.\\boldsymbol\{\\rho\}=\\mathcal\{R\}\(q,E\_\{q\}\)=\\\{\\rho\_\{1\},\\ldots,\\rho\_\{L\}\\\}\.\(3\)Each residual query produces a normalized retrieval signal𝐫\(ℓ\)\\mathbf\{r\}^\{\(\\ell\)\}\. We average the valid residual signals and combine them with the seed from the original question:

𝐬⁡\(q,Eq\)=β​𝐛​\(q\)\+1−β\|𝒱\|​∑ℓ∈𝒱𝐫\(ℓ\),\\mathbf\{s\}\(q,E\_\{q\}\)=\\beta\\mathbf\{b\}\(q\)\+\\frac\{1\-\\beta\}\{\|\\mathcal\{V\}\|\}\\sum\_\{\\ell\\in\\mathcal\{V\}\}\\mathbf\{r\}^\{\(\\ell\)\},\(4\)where𝒱\\mathcal\{V\}indexes residual queries with valid retrieval results; if𝒱\\mathcal\{V\}is empty,𝐬⁡\(q,Eq\)=𝐛⁡\(q\)\\mathbf\{s\}\(q,E\_\{q\}\)=\\mathbf\{b\}\(q\)\. Separate normalization prevents the number or raw score scale of residual results from overwhelming the original question\. Graph retrieval then operates on the combined signal:

𝐳q,E=𝐁⊤​𝐏G​𝐬​\(q,Eq\)\.\\mathbf\{z\}\_\{q,E\}=\\mathbf\{B\}^\{\\top\}\\mathbf\{P\}\_\{G\}\\mathbf\{s\}\(q,E\_\{q\}\)\.\(5\)
This formulation assigns distinct roles to the two operations\. Query reformulation introduces direct retrieval mass for evidence thatEqE\_\{q\}makes identifiable, while graph propagation aggregates evidence connected to either the original or residual signal\. Keeping the signals separate until controlled mixing preserves the constraints of the original question while allowing the observed passages to revise the retrieval request\. The reformulator produces retrieval queries rather than an answer or a reasoning trace, and the system returns one passage ranking\. Section[3](https://arxiv.org/html/2608.13006#S3)gives the concrete proposition–entity implementation\.

### 3\.4Proposition–Entity Index

Let the corpus contain passagesD=\{dj\}j=1ND=\\\{d\_\{j\}\\\}\_\{j=1\}^\{N\}\. An LLM decomposes each passage into self\-contained propositionsP=\{pi\}i=1MP=\\\{p\_\{i\}\\\}\_\{i=1\}^\{M\}and extracts their entity mentions\. The ownership mapπ⁡\(i\)\\pi\(i\)links propositionpip\_\{i\}to its source passagedπ⁡\(i\)d\_\{\\pi\(i\)\}\. Each proposition has a normalized embedding𝐡i\\mathbf\{h\}\_\{i\}, and the sparse matrix𝐀∈\{0,1\}M×Ne\\mathbf\{A\}\\in\\\{0,1\\\}^\{M\\times N\_\{e\}\}records proposition–entity incidence overNeN\_\{e\}canonical entities\.

Propositions provide precise matching units\[[2](https://arxiv.org/html/2608.13006#bib.bib7)\], while source passages preserve the context needed to understand what the initial match establishes\. Shared entities connect propositions without storing edges between propositions\.

The separation between propositions and passages serves two purposes\. Proposition retrieval avoids diluting a specific fact with the rest of a paragraph, while passage observation gives the reformulator enough context to interpret that fact\. The final result remains a passage ranking, so the reader receives coherent source text rather than isolated extractions\.

### 3\.5Initial Evidence Selection

Given a questionqqwith normalized embedding𝐡q\\mathbf\{h\}\_\{q\}, EviReform scores propositions by

ui​\(q\)=max⁡\(0,𝐡i⊤​𝐡q\)\.u\_\{i\}\(q\)=\\max\\\!\\left\(0,\\mathbf\{h\}\_\{i\}^\{\\top\}\\mathbf\{h\}\_\{q\}\\right\)\.\(6\)An LLM selects a setSqS\_\{q\}of proposition identifiers from the candidates with the highest scores\. The selection step favors propositions that jointly clarify the question rather than treating every proposition with high similarity as equally useful\. Their scores define the seed from the original question:

b~i=𝕀\[i∈Sq\]ui\(q\)\.\\widetilde\{b\}\_\{i\}=\\mathbb\{I\}\[i\\in S\_\{q\}\]u\_\{i\}\(q\)\.\(7\)We normalize this seed to unit mass:

𝐛=𝐛~∥𝐛~∥1\.\\mathbf\{b\}=\\frac\{\\widetilde\{\\mathbf\{b\}\}\}\{\\lVert\\widetilde\{\\mathbf\{b\}\}\\rVert\_\{1\}\}\.\(8\)If the selector returns no valid proposition with a positive score, the positive proposition with the highest score supplies a deterministic fallback\. The complete source passages of the selected propositions form the observed evidence

Eq=\{dπ⁡\(i\):i∈Sq\}\.E\_\{q\}=\\\{d\_\{\\pi\(i\)\}:i\\in S\_\{q\}\\\}\.\(9\)The propositions determine which passages are observed, but the reformulator reads the passage text rather than isolated proposition strings\. This distinction supplies the surrounding facts needed to identify what remains unresolved\.

Initial selection is deliberately not treated as the final retrieval result\. Its purpose is to expose evidence that clarifies the next search, and the selected passages must earn their final positions through the same combined scoring and propagation used for all other passages\. This prevents the observation stage from reserving output positions regardless of later evidence\.

### 3\.6Query Reformulation from Retrieved Evidence

EviReform passes\(q,Eq\)\(q,E\_\{q\}\)to an LLM that produces residual queries𝝆=\{ρ1,…,ρL\}\\boldsymbol\{\\rho\}=\\\{\\rho\_\{1\},\\ldots,\\rho\_\{L\}\\\}\. Each residual query describes information needed for the original question but not established byEqE\_\{q\}\. The prompt therefore uses the initial passages to refine the retrieval objective, rather than asking the model to answer the question or restate it\.

A useful residual query preserves the constraints of the original question, incorporates a bridge established byEqE\_\{q\}, and asks for the unresolved relation or attribute\. Several queries are allowed because the observed passages may leave more than one plausible gap\. They are searched independently, so one poorly formed query does not determine the complete ranking\.

For eachρℓ\\rho\_\{\\ell\}, dense retrieval selects a setIℓI\_\{\\ell\}of propositions\. Its normalized signal is

ri\(ℓ\)=𝕀\[i∈Iℓ\]max\(0,𝐡i⊤𝐡ρℓ\)∑j∈Iℓmax⁡\(0,𝐡j⊤​𝐡ρℓ\)\.r\_\{i\}^\{\(\\ell\)\}=\\frac\{\\mathbb\{I\}\[i\\in I\_\{\\ell\}\]\\max\(0,\\mathbf\{h\}\_\{i\}^\{\\top\}\\mathbf\{h\}\_\{\\rho\_\{\\ell\}\}\)\}\{\\sum\_\{j\\in I\_\{\\ell\}\}\\max\(0,\\mathbf\{h\}\_\{j\}^\{\\top\}\\mathbf\{h\}\_\{\\rho\_\{\\ell\}\}\)\}\.\(10\)Let𝒱\\mathcal\{V\}index nonempty residual queries with positive retained similarity\. Their signals are averaged as𝐫=\|𝒱\|−1​∑ℓ∈𝒱𝐫\(ℓ\)\\mathbf\{r\}=\|\\mathcal\{V\}\|^\{\-1\}\\sum\_\{\\ell\\in\\mathcal\{V\}\}\\mathbf\{r\}^\{\(\\ell\)\}and combined with the seed from the original question:

𝐬=β​𝐛\+\(1−β\)​𝐫,\\mathbf\{s\}=\\beta\\mathbf\{b\}\+\(1\-\\beta\)\\mathbf\{r\},\(11\)where𝐬=𝐛\\mathbf\{s\}=\\mathbf\{b\}if no residual query is valid\. Independent normalization gives the original and residual channels controlled total mass\. The original seed preserves evidence directly tied to the question, while the residual seed introduces propositions associated with the newly specified information need\.

This signal construction connects observation to graph retrieval\. A residual query is intentionally narrower than the original question and may omit constraints that are already established inEqE\_\{q\}\. Using it alone can retrieve the missing relation but lose the passage that anchors that relation to the question\. The mixture keeps both parts available to the graph and to the final passage readout\.

### 3\.7Propagation through Shared Entities

The combined signal is propagated between propositions that share entities\. Let𝐃e\\mathbf\{D\}\_\{e\}contain entity degrees\. After removing self\-loops, the proposition weights are

𝐖=𝐀𝐃e†​𝐀⊤−diag⁡\(𝐀𝐃e†​𝐀⊤\)\.\\mathbf\{W\}=\\mathbf\{A\}\\mathbf\{D\}\_\{e\}^\{\\dagger\}\\mathbf\{A\}^\{\\top\}\-\\operatorname\{diag\}\\\!\\left\(\\mathbf\{A\}\\mathbf\{D\}\_\{e\}^\{\\dagger\}\\mathbf\{A\}^\{\\top\}\\right\)\.\(12\)Let𝐃w\\mathbf\{D\}\_\{w\}contain row sums of𝐖\\mathbf\{W\}\. For column vectors, the transition is

𝐓=𝐖𝐃w†\.\\mathbf\{T\}=\\mathbf\{W\}\\mathbf\{D\}\_\{w\}^\{\\dagger\}\.\(13\)We apply one update:

𝐳=α​𝐬\+\(1−α\)​𝐓𝐬,\\mathbf\{z\}=\\alpha\\mathbf\{s\}\+\(1\-\\alpha\)\\mathbf\{T\}\\mathbf\{s\},\(14\)whereα\\alphabalances direct retrieval evidence against support transferred through shared entities\. The update allows propositions related to either query channel to contribute to the same ranking\. We evaluate the matrix product through sparse incidence operations\.

Propagation occurs after the two retrieval signals are combined\. It is therefore not asked to infer the missing query from graph topology\. Instead, it consolidates propositions reached from either signal when they share entities and can jointly support the same passage set\. One step is sufficient to test this role without turning the method into an unconstrained path search\.

### 3\.8Passage Readout

The reader consumes passages rather than propositions\. We aggregate proposition scores with

score\(dj\)=1\|Pj\|∑i:π⁡\(i\)=jzi,\\operatorname\{score\}\(d\_\{j\}\)=\\frac\{1\}\{\\sqrt\{\|P\_\{j\}\|\}\}\\sum\_\{i:\\pi\(i\)=j\}z\_\{i\},\(15\)wherePjP\_\{j\}is the set of propositions extracted fromdjd\_\{j\}\. This factor depends on the number of extracted propositions, not passage token length; it reduces the advantage of passages split into many propositions without fully averaging their scores\. We rank passages by Eq\.[15](https://arxiv.org/html/2608.13006#S3.E15)and use this fixed readout throughout all experiments\.

Table 1:Retrieval and downstream QA\. R@5/10 are passage recall\. For multi\-hop QA, we use each method’s final context and generate answers with the same QA prompt\. IRCoT, S2G\-RAG, GeAR, and EviReform use a 3,000\-token budget\. Medical ACC is mean correctness across four question types\. N/A means not applicable and N/E means not evaluated\. Percentages; best scores are bold and second\-best scores are underlined\.

## 4Experimental Setup

### 4\.1Datasets and Baselines

Following the HippoRAG 2 evaluation protocol, we use the same subsets of 2WikiMultiHopQA, HotpotQA, and MuSiQue\[[7](https://arxiv.org/html/2608.13006#bib.bib3),[9](https://arxiv.org/html/2608.13006#bib.bib9),[34](https://arxiv.org/html/2608.13006#bib.bib10),[26](https://arxiv.org/html/2608.13006#bib.bib11)\]\. Their corpora contain 6,119, 9,811, and 11,656 passages, respectively\.

We additionally evaluate on the Medical set from GraphRAG\-Bench\[[31](https://arxiv.org/html/2608.13006#bib.bib12)\]\. It does not annotate gold supporting passages, so retrieval recall cannot be computed\. We therefore report mean answer accuracy\.

The comparison covers sparse, dense, iterative, and graph retrieval\. We evaluate BM25; BGE\-M3\[[1](https://arxiv.org/html/2608.13006#bib.bib8)\]and Qwen3\-Embedding\-0\.6B\[[37](https://arxiv.org/html/2608.13006#bib.bib13)\]; BGE\-M3 followed by a BGE or listwise LLM reranker over its Top\-40 pool\[[25](https://arxiv.org/html/2608.13006#bib.bib30)\]; GritLM\-7B\[[20](https://arxiv.org/html/2608.13006#bib.bib34)\]and NV\-Embed\-v2\[[14](https://arxiv.org/html/2608.13006#bib.bib35)\]; IRCoT, S2G\-RAG, and GeAR\[[27](https://arxiv.org/html/2608.13006#bib.bib6),[17](https://arxiv.org/html/2608.13006#bib.bib18),[24](https://arxiv.org/html/2608.13006#bib.bib36)\]; and PropRAG, HippoRAG 2, and CatRAG\[[28](https://arxiv.org/html/2608.13006#bib.bib5),[7](https://arxiv.org/html/2608.13006#bib.bib3),[13](https://arxiv.org/html/2608.13006#bib.bib4)\]\. Direct LLM inference provides a QA reference without retrieval\.

Graph and agentic systems use BGE\-M3 wherever embeddings are required, and LLM\-based indexing and retrieval use DeepSeek\-v4\-flash\[[3](https://arxiv.org/html/2608.13006#bib.bib31)\]\. IRCoT interleaves dense retrieval with chain\-of\-thought generation; S2G\-RAG alternates evidence\-sufficiency judgments and gap\-directed retrieval; GeAR maintains a gist memory while repeatedly invoking graph retrieval\. These methods and EviReform receive a 3,000\-token budget per question\. We use the same EviReform configuration across datasets: 100 initial proposition candidates, at most 12 selected propositions, at most three residual queries with two propositions retrieved per query, andα=β=0\.5\\alpha=\\beta=0\.5\.

### 4\.2Metrics and QA

For gold passagesGqG\_\{q\}and Top\-KKretrievalRqKR\_\{q\}^\{K\}, we report

Recall​@​K\\displaystyle\\operatorname\{Recall@K\}=\|Q\|−1​∑q\|Gq∩RqK\|\|Gq\|,\\displaystyle=\|Q\|^\{\-1\}\\sum\_\{q\}\\frac\{\|G\_\{q\}\\cap R\_\{q\}^\{K\}\|\}\{\|G\_\{q\}\|\},\(16\)Chain​@​K\\displaystyle\\operatorname\{Chain@K\}=\|Q\|−1∑q𝕀\[Gq⊆RqK\],\\displaystyle=\|Q\|^\{\-1\}\\sum\_\{q\}\\mathbb\{I\}\[G\_\{q\}\\subseteq R\_\{q\}^\{K\}\],\(17\)Hit​@​K\\displaystyle\\operatorname\{Hit@K\}=\|Q\|−1∑q𝕀\[Gq∩RqK≠∅\]\.\\displaystyle=\|Q\|^\{\-1\}\\sum\_\{q\}\\mathbb\{I\}\[G\_\{q\}\\cap R\_\{q\}^\{K\}\\neq\\emptyset\]\.\(18\)Recall@KKmeasures partial coverage of gold passages, Chain@KKrequires the complete supporting set, and Hit@KKrequires at least one supporting passage\. The primary cutoff isK=5K=5\. For QA, we use each method’s final context and generate answers with the same prompt\. F1 is normalized token overlap and EM is normalized exact match\. The appendices record implementation details and additional analyses\.

## 5Results

Table[1](https://arxiv.org/html/2608.13006#S3.T1)reports gold\-passage recall and downstream QA\. Relative to the strongest passage ranker for each metric, EviReform gains 5\.00, 2\.65, and 5\.59 R@5 points on 2Wiki, HotpotQA, and MuSiQue; the R@10 gains are 0\.88, 1\.20, and 4\.78 points\.

With the shared reader, EviReform improves over the strongest QA baseline, GeAR, by 3\.91, 2\.28, and 4\.50 F1 points, and by 3\.30, 1\.60, and 4\.00 EM points\.

Paired confidence intervals use 10,000 question\-level bootstrap resamples\. The R@5 intervals are\[3\.90,6\.10\]\[3\.90,6\.10\],\[1\.50,3\.80\]\[1\.50,3\.80\], and\[3\.86,7\.28\]\[3\.86,7\.28\]points; the F1 intervals are\[1\.42,6\.36\]\[1\.42,6\.36\],\[0\.34,4\.24\]\[0\.34,4\.24\], and\[2\.31,6\.72\]\[2\.31,6\.72\]\. Among the remaining metrics, only the HotpotQA EM interval overlaps zero\.

Table[2](https://arxiv.org/html/2608.13006#S6.T2)separates access to an entry passage from recovery of the complete supporting set\. The strongest graph baselines already reach 99\.9, 99\.2, and 91\.3 Hit@5, but their Chain@5 scores remain substantially lower\. EviReform improves Chain@5 by 22\.5, 12\.1, and 11\.7 points\. The contrast captures the central retrieval problem: the original question often reaches one relevant passage, while the observed passage makes its missing complement easier to specify\.

Stronger dense encoders improve the initial semantic match but do not remove this evidence dependency\. NV\-Embed\-v2 raises BGE\-M3 R@5 by 4\.67, 9\.10, and 16\.82 points, with the largest change on MuSiQue, while EviReform remains ahead at the primary cutoff\. With NV\-Embed\-v2 used throughout the graph retrievers, EviReform exceeds the strongest graph baseline by 7\.20, 1\.00, and 5\.17 R@5 points on 2Wiki, HotpotQA, and MuSiQue\.

On GraphRAG\-Bench \(Medical\), where retrieval recall cannot be measured, the same design reaches 71\.75 mean answer accuracy, compared with 67\.48 for S2G\-RAG, 69\.25 for GeAR, and 69\.86 for HippoRAG 2\.

## 6Mechanism Analysis

Table 2:Retrieval decomposition atK=5K=5\. Scores are percentages; higher is better\.Table[3](https://arxiv.org/html/2608.13006#S7.T3)crosses query reformulation from retrieved evidence with propagation through shared entities\. The base uses only the proposition seed from the original question\. Reformulation adds residual queries and their retrieval signal; propagation adds shared\-entity aggregation; the full method combines both\. The comparison separates finding evidence for the newly specified need from consolidating that evidence in the graph\.

All three datasets satisfy Full\>\>Reformulation only\>\>Propagation only\>\>Base on R@5 and Chain@5\. With propagation present, reformulation adds 7\.20/18\.00, 2\.75/5\.40, and 3\.63/5\.40 R@5/Chain@5 points\. After reformulation, propagation adds a further 0\.60/1\.20, 0\.60/1\.20, and 0\.93/2\.10 points\. Reformulation recovers most of the new evidence, and propagation improves how the original and residual evidence are combined\.

A matched retrieval run repeats the original question instead of using residual queries\. It reaches 89\.75/74\.70, 93\.90/88\.20, and 67\.51/39\.20 R@5/Chain@5, compared with 97\.75/94\.90, 96\.70/93\.80, and 73\.03/46\.90 for the full method\. The gain therefore comes from specifying the unresolved need, rather than issuing more requests for the original question\.

Reranking the initial candidates with the first\-stage evidence reaches 73\.83/50\.10, 88\.75/80\.90, and 59\.48/29\.90 R@5/Chain@5\. These pools contain a complete chain for only 53\.0%, 90\.2%, and 43\.9% of questions\. Observation is most useful when it directs retrieval toward missing evidence, not only when it changes the order of passages already found\.

We next examine which evidence enters before graph propagation\. For each question, the initial pool contains the first 20 distinct source passages represented among the 100 highest\-scoring propositions retrieved with the original question\. The expanded pool adds source passages retrieved by the residual queries\. Passages introduced only by propagation are excluded\. Among 459, 93, and 533 initial pools that lack a complete chain on 2Wiki, HotpotQA, and MuSiQue, retrieval with residual queries makes 416, 62, and 150 complete\. These transitions directly connect query reformulation to the recovery of previously missing supporting passages\. The remaining gap is largest on MuSiQue: residual retrieval expands the available evidence, but some complete chains still fail to survive the final Top\-5 ranking\. Discovering candidates and constructing the final set therefore remain distinct challenges\.

Pool coverage shows where the gains arise\. Before reformulation, the initial pools cover 78\.8%, 95\.3%, and 74\.8% of individual gold passages, and contain complete chains for 54\.1%, 90\.7%, and 46\.7% of questions\. After residual retrieval, passage coverage rises to 98\.1%, 98\.4%, and 83\.0%, while complete\-chain coverage rises to 95\.7%, 96\.9%, and 61\.7%\. The large change on 2Wiki shows that the original question often reaches one part of the chain but leaves its complement underspecified\. On MuSiQue, residual queries recover additional evidence, while retaining the complete chain near the top of the ranking remains harder\.

## 7Discussion

The experiments support a simple account of multi\-hop graph retrieval\. The original question often reaches an entry passage, but that passage may reveal a bridge needed to describe the remaining evidence\. Reformulation turns this observation into a new retrieval signal, while propagation combines propositions reached from the original and residual requests\.

EviReform performs both operations within the retriever and returns one passage ranking\. The gains in complete\-chain recovery come primarily from retrieving evidence for the newly specified need; shared\-entity propagation then provides a smaller, consistent improvement by consolidating the two retrieval signals\.

Table 3:2×22\\times 2ablation atK=5K=5\. Checkmarks indicate which modules are included\. Percentages; higher is better\.
## 8Limitations

The retrieval experiments cover three English multi\-hop QA benchmarks and one generated index per dataset\. Index construction, initial evidence selection, and query reformulation depend on LLM outputs, so repeated index construction or inference may introduce variation beyond the bootstrap intervals over questions reported here\.

The ablation measures reformulation and propagation under one architecture and retrieval budget\. Other graph operators or larger feedback budgets may change their relative contributions\. On MuSiQue, residual retrieval recovers additional complete chains, but the gap between coverage after expansion and final Chain@5 shows that selecting the final evidence set remains an open problem\.

GraphRAG\-Bench \(Medical\) lacks gold supporting passages and therefore evaluates transfer through answers rather than direct retrieval\.

## 9Conclusion

Multi\-hop retrieval changes as evidence is acquired: an initial passage can reveal the entity or relation needed to find its complement\. EviReform uses that observation to formulate residual queries, combines their retrieval signal with the original question, and propagates the result between propositions that share entities\. Across three multi\-hop benchmarks, this design improves passage recall, complete\-chain recovery, and downstream QA\. Mechanism studies show that reformulation recovers most of the new evidence, while propagation consolidates evidence reached by both requests\. EviReform thus lets observed evidence refine what graph retrieval searches for before producing the final passage ranking\.

## Generative AI Use Disclosure

During the preparation of this manuscript and its accompanying implementation, the authors used generative AI systems to assist with language polishing and experimental\-script development\. Generative AI models were also used as explicitly identified components of the experiments, including LLM\-based indexing, retrieval, reranking, and question answering; their roles and experimental settings are reported in this manuscript\. The authors independently verified and take full responsibility for all scientific claims, experimental design, data processing, results, references, software, and final text\.

## References

- \[1\]J\. Chen, S\. Xiao, P\. Zhang, K\. Luo, D\. Lian, and Z\. Liu\(2024\)M3\-embedding: multi\-linguality, multi\-functionality, multi\-granularity text embeddings through self\-knowledge distillation\.InFindings of the Association for Computational Linguistics: ACL 2024,Bangkok, Thailand,pp\. 2318–2335\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.137),[Link](https://aclanthology.org/2024.findings-acl.137/)Cited by:[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[2\]T\. Chen, H\. Wang, S\. Chen, W\. Yu, K\. Ma, X\. Zhao, H\. Zhang, and D\. Yu\(2024\)Dense X retrieval: what retrieval granularity should we use?\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami, Florida, USA,pp\. 15159–15177\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.845),[Link](https://aclanthology.org/2024.emnlp-main.845/)Cited by:[§3\.4](https://arxiv.org/html/2608.13006#S3.SS4.p2.1)\.
- \[3\]DeepSeek\-AI\(2026\)DeepSeek\-V4: towards highly efficient million\-token context intelligence\.External Links:2606\.19348,[Document](https://dx.doi.org/10.48550/arXiv.2606.19348),[Link](https://arxiv.org/abs/2606.19348)Cited by:[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p4.1)\.
- \[4\]D\. Edge, H\. Trinh, N\. Cheng, J\. Bradley, A\. Chao, A\. Mody, S\. Truitt, and J\. Larson\(2024\)From local to global: a graph RAG approach to query\-focused summarization\.External Links:2404\.16130,[Document](https://dx.doi.org/10.48550/arXiv.2404.16130),[Link](https://arxiv.org/abs/2404.16130)Cited by:[§2\.1](https://arxiv.org/html/2608.13006#S2.SS1.p1.1)\.
- \[5\]B\. Efron and R\. J\. Tibshirani\(1994\)An introduction to the bootstrap\.Chapman and Hall/CRC\.Cited by:[§A\.1](https://arxiv.org/html/2608.13006#A1.SS1.p1.1)\.
- \[6\]Z\. Guo, L\. Xia, Y\. Yu, T\. Ao, and C\. Huang\(2025\)LightRAG: simple and fast retrieval\-augmented generation\.InFindings of the Association for Computational Linguistics: EMNLP 2025,pp\. 10746–10761\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.568),[Link](https://aclanthology.org/2025.findings-emnlp.568/)Cited by:[§2\.1](https://arxiv.org/html/2608.13006#S2.SS1.p1.1)\.
- \[7\]B\. J\. Gutiérrez, Y\. Shu, W\. Qi, S\. Zhou, and Y\. Su\(2025\)From RAG to memory: non\-parametric continual learning for large language models\.InProceedings of the 42nd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.267,pp\. 21497–21515\.External Links:[Link](https://proceedings.mlr.press/v267/gutierrez25a.html)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.13006#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[8\]H\. Han, Y\. Wang, H\. Shomer, K\. Guo, J\. Ding,et al\.\(2025\)Retrieval\-augmented generation with graphs \(GraphRAG\)\.External Links:2501\.00309,[Link](https://arxiv.org/abs/2501.00309)Cited by:[§2\.1](https://arxiv.org/html/2608.13006#S2.SS1.p1.1)\.
- \[9\]X\. Ho, A\. D\. Nguyen, S\. Sugawara, and A\. Aizawa\(2020\)Constructing a multi\-hop QA dataset for comprehensive evaluation of reasoning steps\.InProceedings of the 28th International Conference on Computational Linguistics,Barcelona, Spain \(Online\),pp\. 6609–6625\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.coling-main.580),[Link](https://aclanthology.org/2020.coling-main.580/)Cited by:[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p1.1)\.
- \[10\]Z\. Jiang, F\. Xu, L\. Gao, Z\. Sun, Q\. Liu, J\. Dwivedi\-Yu, Y\. Yang, J\. Callan, and G\. Neubig\(2023\)Active retrieval augmented generation\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 7969–7992\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.495),[Link](https://aclanthology.org/2023.emnlp-main.495/)Cited by:[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1)\.
- \[11\]V\. Karpukhin, B\. Oguz, S\. Min, P\. Lewis, L\. Wu, S\. Edunov, D\. Chen, and W\. Yih\(2020\)Dense passage retrieval for open\-domain question answering\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing,pp\. 6769–6781\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.550),[Link](https://aclanthology.org/2020.emnlp-main.550/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p2.1)\.
- \[12\]O\. Khattab, C\. Potts, and M\. Zaharia\(2021\)Baleen: robust multi\-hop reasoning at scale via condensed retrieval\.InAdvances in Neural Information Processing Systems,Vol\.34\.External Links:[Link](https://proceedings.neurips.cc/paper/2021/hash/e8b1cbd05f6e6a358a81dee52493dd06-Abstract.html)Cited by:[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1)\.
- \[13\]K\. H\. Lau, F\. Zhang, B\. Ruan, Y\. Zhou, Q\. Guo, R\. Zhang, and X\. Zhou\(2026\)Breaking the static graph: context\-aware traversal for robust retrieval\-augmented generation\.External Links:2602\.01965,[Document](https://dx.doi.org/10.48550/arXiv.2602.01965),[Link](https://arxiv.org/abs/2602.01965)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.13006#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[14\]C\. Lee, R\. Roy, M\. Xu, J\. Raiman, M\. Shoeybi, B\. Catanzaro, and W\. Ping\(2024\)NV\-Embed: improved techniques for training LLMs as generalist embedding models\.External Links:2405\.17428,[Document](https://dx.doi.org/10.48550/arXiv.2405.17428),[Link](https://arxiv.org/abs/2405.17428)Cited by:[Table C\.1](https://arxiv.org/html/2608.13006#A3.T1.2.3.2.1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[15\]D\. Lee, Y\. Jo, H\. Park, and M\. Lee\(2025\)Shifting from ranking to set selection for retrieval augmented generation\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 17606–17619\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.861),[Link](https://aclanthology.org/2025.acl-long.861/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p1.1)\.
- \[16\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela\(2020\)Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 9459–9474\.External Links:[Link](https://papers.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p1.1)\.
- \[17\]M\. Li, J\. Zou, X\. Lv, C\. Zhang, and G\. Zhou\(2026\)S2G\-RAG: structured sufficiency and gap judging for iterative retrieval\-augmented QA\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),San Diego, California, United States,pp\. 25846–25862\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1185),[Link](https://aclanthology.org/2026.acl-long.1185/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[18\]H\. Luo, H\. E, G\. Chen, Q\. Lin, Y\. Guo, F\. Xu, Z\. Kuang, M\. Song, X\. Wu, Y\. Zhu, and L\. A\. Tuan\(2026\)Graph\-R1: towards agentic GraphRAG framework via end\-to\-end reinforcement learning\.External Links:2507\.21892,[Link](https://arxiv.org/abs/2507.21892)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.3](https://arxiv.org/html/2608.13006#S2.SS3.p1.1)\.
- \[19\]L\. Luo, Z\. Zhao, J\. Liu, Z\. Qiu, J\. Dong, S\. Panev, C\. Gong, T\. Vu, G\. Haffari, D\. Phung, A\. W\. Liew, and S\. Pan\(2026\)G\-reasoner: foundation models for unified reasoning over graph\-structured knowledge\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=zJm9nmoahk)Cited by:[§B\.3](https://arxiv.org/html/2608.13006#A2.SS3.p1.1)\.
- \[20\]N\. Muennighoff, H\. Su, L\. Wang, N\. Yang, F\. Wei, T\. Yu, A\. Singh, and D\. Kiela\(2024\)Generative representational instruction tuning\.External Links:2402\.09906,[Document](https://dx.doi.org/10.48550/arXiv.2402.09906),[Link](https://arxiv.org/abs/2402.09906)Cited by:[Table C\.1](https://arxiv.org/html/2608.13006#A3.T1.2.3.2.1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[21\]P\. Qi, X\. Lin, L\. Mehr, Z\. Wang, and C\. D\. Manning\(2019\)Answering complex open\-domain questions through iterative query generation\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),Hong Kong, China,pp\. 2590–2602\.External Links:[Document](https://dx.doi.org/10.18653/v1/D19-1261),[Link](https://aclanthology.org/D19-1261/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1)\.
- \[22\]J\. J\. Rocchio\(1971\)Relevance feedback in information retrieval\.InThe SMART Retrieval System: Experiments in Automatic Document Processing,G\. Salton \(Ed\.\),pp\. 313–323\.Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1)\.
- \[23\]Z\. Shao, Y\. Gong, Y\. Shen, M\. Huang, N\. Duan, and W\. Chen\(2023\)Enhancing retrieval\-augmented large language models with iterative retrieval\-generation synergy\.InFindings of the Association for Computational Linguistics: EMNLP 2023,Singapore,pp\. 9248–9274\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.findings-emnlp.620),[Link](https://aclanthology.org/2023.findings-emnlp.620/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1)\.
- \[24\]Z\. Shen, C\. Diao, P\. Vougiouklis, P\. Merita, S\. Piramanayagam, E\. Chen, D\. Graux, A\. Melo, R\. Lai, Z\. Jiang, Z\. Li, Y\. Qi, Y\. Ren, D\. Tu, and J\. Z\. Pan\(2025\)GeAR: graph\-enhanced agent for retrieval\-augmented generation\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 12049–12072\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.findings-acl.624),[Link](https://aclanthology.org/2025.findings-acl.624/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.3](https://arxiv.org/html/2608.13006#S2.SS3.p1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[25\]W\. Sun, L\. Yan, X\. Ma, S\. Wang, P\. Ren, Z\. Chen, D\. Yin, and Z\. Ren\(2023\)Is ChatGPT good at search? investigating large language models as re\-ranking agents\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 14918–14937\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.923),[Link](https://aclanthology.org/2023.emnlp-main.923/)Cited by:[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[26\]H\. Trivedi, N\. Balasubramanian, T\. Khot, and A\. Sabharwal\(2022\)MuSiQue: multihop questions via single\-hop question composition\.Transactions of the Association for Computational Linguistics10,pp\. 539–554\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00475),[Link](https://aclanthology.org/2022.tacl-1.31/)Cited by:[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p1.1)\.
- \[27\]H\. Trivedi, N\. Balasubramanian, T\. Khot, and A\. Sabharwal\(2023\)Interleaving retrieval with chain\-of\-thought reasoning for knowledge\-intensive multi\-step questions\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Toronto, Canada,pp\. 10014–10037\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.557),[Link](https://aclanthology.org/2023.acl-long.557/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[28\]J\. Wang and J\. Han\(2025\)PropRAG: guiding retrieval with beam search over proposition paths\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,Suzhou, China,pp\. 6212–6227\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.317),[Link](https://aclanthology.org/2025.emnlp-main.317/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.13006#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[29\]K\. Wang, F\. Duan, P\. Li, S\. Wang, and X\. Cai\(2025\)LLMs know what they need: leveraging a missing information guided framework to empower retrieval\-augmented generation\.InProceedings of the 31st International Conference on Computational Linguistics,Abu Dhabi, UAE,pp\. 2379–2400\.External Links:[Link](https://aclanthology.org/2025.coling-main.163/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1)\.
- \[30\]X\. Wu, C\. Yang, X\. Lin, C\. Xu, X\. Jiang, Y\. Sun, H\. Xiong, J\. Li, and J\. Guo\(2025\)Think\-on\-graph 3\.0: efficient and adaptive LLM reasoning on heterogeneous graphs via multi\-agent dual\-evolving context retrieval\.External Links:2509\.21710,[Link](https://arxiv.org/abs/2509.21710)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.3](https://arxiv.org/html/2608.13006#S2.SS3.p1.1)\.
- \[31\]Z\. Xiang, C\. Wu, Q\. Zhang, S\. Chen, Z\. Hong, X\. Huang, and J\. Su\(2025\)When to use graphs in RAG: a comprehensive analysis for graph retrieval\-augmented generation\.External Links:2506\.05690,[Document](https://dx.doi.org/10.48550/arXiv.2506.05690),[Link](https://arxiv.org/abs/2506.05690)Cited by:[§B\.2](https://arxiv.org/html/2608.13006#A2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p2.1)\.
- \[32\]W\. Xiong, X\. L\. Li, S\. Iyer, J\. Du, P\. Lewis, W\. Wang, Y\. Mehdad, W\. Yih, S\. Riedel, D\. Kiela, and B\. Oguz\(2021\)Answering complex open\-domain questions with multi\-hop dense retrieval\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=EMHoBG0avc1)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1)\.
- \[33\]V\. Yadav, S\. Bethard, and M\. Surdeanu\(2021\)If you want to go far go together: unsupervised joint candidate evidence retrieval for multi\-hop question answering\.InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,pp\. 4571–4581\.External Links:[Document](https://dx.doi.org/10.18653/v1/2021.naacl-main.363),[Link](https://aclanthology.org/2021.naacl-main.363/)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p1.1)\.
- \[34\]Z\. Yang, P\. Qi, S\. Zhang, Y\. Bengio, W\. W\. Cohen, R\. Salakhutdinov, and C\. D\. Manning\(2018\)HotpotQA: a dataset for diverse, explainable multi\-hop question answering\.InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,Brussels, Belgium,pp\. 2369–2380\.External Links:[Document](https://dx.doi.org/10.18653/v1/D18-1259),[Link](https://aclanthology.org/D18-1259/)Cited by:[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p1.1)\.
- \[35\]C\. Yu, K\. Zhao, Y\. Li, H\. Chang, M\. Feng, X\. Jiang, Y\. Sun, J\. Li, Y\. Zhang, J\. Li, and Z\. Zhang\(2026\)GraphRAG\-R1: graph retrieval\-augmented generation with process\-constrained reinforcement learning\.External Links:2507\.23581,[Link](https://arxiv.org/abs/2507.23581)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.3](https://arxiv.org/html/2608.13006#S2.SS3.p1.1)\.
- \[36\]H\. Yu, C\. Xiong, and J\. Callan\(2021\)Improving query representations for dense retrieval with pseudo relevance feedback\.InProceedings of the 30th ACM International Conference on Information and Knowledge Management,pp\. 3592–3596\.External Links:[Document](https://dx.doi.org/10.1145/3459637.3482124),[Link](https://doi.org/10.1145/3459637.3482124)Cited by:[§1](https://arxiv.org/html/2608.13006#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.13006#S2.SS2.p1.1)\.
- \[37\]Y\. Zhang, M\. Li, D\. Long, X\. Zhang, H\. Lin, B\. Yang, P\. Xie, A\. Yang, D\. Liu, J\. Lin, F\. Huang, and J\. Zhou\(2025\)Qwen3 embedding: advancing text embedding and reranking through foundation models\.External Links:2506\.05176,[Document](https://dx.doi.org/10.48550/arXiv.2506.05176),[Link](https://arxiv.org/abs/2506.05176)Cited by:[§4\.1](https://arxiv.org/html/2608.13006#S4.SS1.p3.1)\.
- \[38\]Z\. Zhou, D\. Ataee Tarzanagh, S\. Didari, W\. Hu, B\. Gutow, O\. Verkholyak, M\. Faraki, H\. Hao, H\. Moon, and S\. Min\(2026\)Query\-aware flow diffusion for graph\-based RAG with retrieval guarantees\.External Links:2605\.18775,[Document](https://dx.doi.org/10.48550/arXiv.2605.18775),[Link](https://arxiv.org/abs/2605.18775)Cited by:[§2\.1](https://arxiv.org/html/2608.13006#S2.SS1.p1.1)\.
- \[39\]X\. Zhu, Y\. Xie, Y\. Liu, Y\. Li, and W\. Hu\(2025\)Knowledge graph\-guided retrieval augmented generation\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies,pp\. 8912–8924\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.449),[Link](https://aclanthology.org/2025.naacl-long.449/)Cited by:[§2\.1](https://arxiv.org/html/2608.13006#S2.SS1.p1.1)\.

## Appendix AAdditional Mechanism and Robustness Analysis

This section extends the mechanism analysis in the main text with paired uncertainty estimates, additional controls, sensitivity experiments, retrieval cost, and evaluation at different index sizes\.

### A\.1Paired Confidence Intervals

We use 10,000 paired bootstrap resamples\[[5](https://arxiv.org/html/2608.13006#bib.bib32)\]over the evaluated questions in each dataset\. The intervals are the empirical 2\.5th and 97\.5th percentiles of the bootstrap distribution\. Pairing by question estimates variation across the evaluated questions but not variation from rebuilding the index or repeating LLM inference\. Table[A\.1](https://arxiv.org/html/2608.13006#A1.T1)reports R@5, R@10, F1, and EM, and Figure[A\.1](https://arxiv.org/html/2608.13006#A1.F1)visualizes their paired differences\.

Method or statisticR@5R@10F1EM2WikiEviReform\[95% CI\]97\.75\[97\.10, 98\.38\]98\.50\[97\.98, 99\.00\]58\.05\[55\.12, 60\.91\]51\.50\[48\.40, 54\.60\]Best metric\-wise comparatorGeAR92\.75GeAR97\.63GeAR54\.14GeAR48\.20Difference\[95% CI\]\+5\.00\[\+3\.90, \+6\.10\]\+0\.88\[\+0\.15, \+1\.63\]\+3\.91\[\+1\.42, \+6\.36\]\+3\.30\[\+0\.90, \+5\.80\]HotpotQAEviReform\[95% CI\]96\.70\[95\.85, 97\.50\]98\.50\[97\.90, 99\.00\]69\.57\[67\.02, 72\.06\]57\.10\[54\.00, 60\.20\]Best metric\-wise comparatorNV\-Embed\-v294\.05NV\-Embed\-v297\.30GeAR67\.30GeAR55\.50Difference\[95% CI\]\+2\.65\[\+1\.50, \+3\.80\]\+1\.20\[\+0\.40, \+2\.00\]\+2\.28\[\+0\.34, \+4\.24\]\+1\.60\[\-0\.50, \+3\.60\]MuSiQueEviReform\[95% CI\]73\.03\[71\.18, 74\.84\]81\.15\[79\.53, 82\.73\]34\.78\[31\.98, 37\.53\]26\.90\[24\.10, 29\.60\]Best metric\-wise comparatorNV\-Embed\-v267\.43NV\-Embed\-v276\.38GeAR30\.27GeAR22\.90Difference\[95% CI\]\+5\.59\[\+3\.86, \+7\.28\]\+4\.78\[\+3\.14, \+6\.42\]\+4\.50\[\+2\.31, \+6\.72\]\+4\.00\[\+1\.90, \+6\.10\]Table A\.1:Paired bootstrap estimates under the common passage\-ranking and reader protocol\. Each dataset block lists EviReform with its 95% CI, the strongest comparator and score for each metric, and the paired difference with its 95% CI\. Scores are percentages and differences are percentage points; higher is better\.Figure A\.1:Paired differences with 95% bootstrap intervals\.Eleven of the twelve paired intervals are strictly above zero; only HotpotQA EM overlaps zero, with\[−0\.50,3\.60\]\[\-0\.50,3\.60\]\. GeAR is the closest comparator for QA and 2Wiki recall, while NV\-Embed\-v2 is closest for recall on HotpotQA and MuSiQue\. The comparisons therefore use the strongest evaluated system for each metric under this protocol\.

### A\.2Ablations and Retrieval Controls

Without query reformulation, R@5 falls by 7\.20, 2\.75, and 3\.63 points on 2Wiki, HotpotQA, and MuSiQue, while Chain@5 falls by 18\.00, 5\.40, and 5\.40\. Repeating the original question under a matched retrieval count remains 20\.20, 5\.60, and 7\.70 Chain@5 points below the full system\.

Using only residual queries loses 13\.25, 14\.55, and 14\.25 R@5 points, with Chain@5 losses of 29\.20, 27\.20, and 15\.80\. Because residual queries target missing information rather than restating the question, the original signal is needed to preserve initial evidence\.

The2×22\\times 2ablation shows that, conditional on propagation, reformulation adds 7\.20/18\.00, 2\.75/5\.40, and 3\.63/5\.40 R@5/Chain@5 points\. Conditional on reformulation, propagation adds 0\.60/1\.20, 0\.60/1\.20, and 0\.93/2\.10 points\. Five of these six paired intervals are strictly positive; the HotpotQA R@5 interval touches zero\. Reformulation has the larger contribution in this configuration, while propagation adds to it on every dataset\. Table[A\.2](https://arxiv.org/html/2608.13006#A1.T2)reports the four cells, two retrieval controls, and the paired contrasts\.

\(a\) Configurations and controls

\(b\) Effects with the other component present

Table A\.2:Ablations and retrieval controls\. Panel \(a\) reports the four combinations of reformulation and propagation, a variant that uses only residual queries, and a control that repeats the original question\. Panel \(b\) gives paired contrasts from 10,000 bootstrap resamples over questions; brackets are 95% CIs\. Scores are percentages and contrasts are percentage points\.
### A\.3Evidence Added before Final Ranking

Residual retrieval increases both coverage of gold passages and the number of pools containing every passage in the gold chain\. Among pools that initially miss part of the chain, 416/459 on 2Wiki, 62/93 on HotpotQA, and 150/533 on MuSiQue become complete\. These transitions require passages outside the initial pool and cannot be produced by reranking that pool\. On MuSiQue, complete\-chain coverage is still 14\.8 points higher in the expanded pool than in the final five passages\. Finding the missing evidence and selecting the final set therefore remain distinct challenges\.

Table[A\.3](https://arxiv.org/html/2608.13006#A1.T3)reports aggregate coverage before propagation, and Table[A\.4](https://arxiv.org/html/2608.13006#A1.T4)counts transitions among pools that initially miss at least one gold passage\.

Table A\.3:Coverage before graph propagation\. Initial isC0C\_\{0\}, the first 20 distinct source passages represented by the original question’s Top\-100 propositions; \+Residual isC\+C\_\{\+\}after adding passages retrieved by residual queries\. Percentages; higher is better\.Table A\.4:Transitions among pools that miss at least one gold passage before residual retrieval\. New gold passages counts newly added gold passage occurrences\.
### A\.4Reordering Candidates and Number of Reformulation Rounds

A reordering control receives the question and the evidence selected in the first stage, but can choose only among at most 40 source passages retrieved with the original question\. It can change their order but cannot add a passage\. We also test whether repeating reformulation helps\. At roundtt, this variant applies

𝝆t=ℛ⁡\(q,Et\),Et\+1=Et∪Retrieve⁡\(𝝆t\)\\boldsymbol\{\\rho\}\_\{t\}=\\mathcal\{R\}\(q,E\_\{t\}\),\\qquad E\_\{t\+1\}=E\_\{t\}\\cup\\operatorname\{Retrieve\}\(\\boldsymbol\{\\rho\}\_\{t\}\)\(19\)for at most three rounds\. The comparison that stops after round one uses the first decision from the same run\.

Table[A\.5](https://arxiv.org/html/2608.13006#A1.T5)compares reordering the initial pool, one round of reformulation, and the variant that allows later rounds\.

Table A\.5:Effects of restricting selection to the initial pool and of allowing additional rounds of reformulation\. Pool Chain is the percentage of candidate pools containing every gold passage\. All scores are percentages\.The judge over the initial pool approaches the limit imposed by those candidates but cannot recover missing passages\. Relative to one round, allowing up to three rounds changes R@5 by−0\.20\-0\.20,−0\.40\-0\.40, and\+2\.47\+2\.47points on 2Wiki, HotpotQA, and MuSiQue\. On MuSiQue, continuing beyond the first round of the same run raises R@5 by 1\.74 points and Chain@5 by 2\.50 points\.

On 2Wiki, the initial pool contains the complete chain for 53\.00% of questions, compared with 95\.70% after retrieval with residual queries\. The corresponding values are 90\.20% and 96\.90% on HotpotQA, and 43\.90% and 61\.70% on MuSiQue\. The reordering control can act only on the initial passages, whereas residual queries can retrieve additional evidence before final ranking\.

The difference between complete\-chain coverage in the expanded pool and Chain@5 is 0\.80 points on 2Wiki, 3\.10 on HotpotQA, and 14\.80 on MuSiQue\. On MuSiQue, more complete chains are found than can be retained together in the final five passages\. Additional rounds help on this dataset, but do not remove the difficulty of choosing a compact final set\. Table[A\.6](https://arxiv.org/html/2608.13006#A1.T6)reports how often later rounds are invoked and their incremental cost\.

Table A\.6:Incremental cost of rounds two and three\. R2/R3 is the percentage of queries invoking each later round\.Later rounds increase feedback tokens by 80%, 37%, and 147% on 2Wiki, HotpotQA, and MuSiQue\. Only MuSiQue receives a substantial gain, so we use one round for all three datasets\.

The use of later rounds differs substantially by dataset\. Only 23\.1% of HotpotQA questions enter round two and 3\.5% enter round three, compared with 70\.7% and 28\.8% on MuSiQue\.

### A\.5Construction and Retrieval Cost

Wall\-clock times are descriptive because request concurrency differed across methods; token counts offer a more direct comparison\. PropRAG uses no LLM tokens during retrieval, so its cost is concentrated in construction\.

Table[A\.7](https://arxiv.org/html/2608.13006#A1.T7)summarizes construction and retrieval cost for the five graph retrievers\.

Table A\.7:Average construction and retrieval cost across the three multi\-hop datasets\. Lower is better\.
### A\.6Parameter Sensitivity

We vary one parameter at a time around the main configuration: response weightα∈\{0\.25,0\.375,0\.5,0\.625,0\.75\}\\alpha\\in\\\{0\.25,0\.375,0\.5,0\.625,0\.75\\\}, mass assigned to the original questionβ∈\{0\.25,0\.375,0\.5,0\.625,0\.75\}\\beta\\in\\\{0\.25,0\.375,0\.5,0\.625,0\.75\\\}, the number of residual seedskr∈\{1,2,4,8\}k\_\{r\}\\in\\\{1,2,4,8\\\}, and the number of residual queriesL∈\{1,2,3\}L\\in\\\{1,2,3\\\}\. We compare BGE\-M3, Qwen3\-Embedding\-0\.6B, and NV\-Embed\-v2 as embedding models, and compare DeepSeek\-V4\-Flash with GPT\-5\.6 as the LLM used during retrieval\. Tables[A\.8](https://arxiv.org/html/2608.13006#A1.T8),[A\.9](https://arxiv.org/html/2608.13006#A1.T9), and[A\.10](https://arxiv.org/html/2608.13006#A1.T10)give every parameter setting, Table[A\.11](https://arxiv.org/html/2608.13006#A1.T11)gives the embedding results, and Figure[A\.2](https://arxiv.org/html/2608.13006#A1.F2)summarizes the comparisons\.

Figure A\.2:Retrieval sensitivity\. The four line panels report R@5 while varying one parameter; vertical dashed lines mark the main configuration\. The grouped bars report R@5, R@10, and R@20 for the embedding and LLM comparisons\. Markers and line styles identify datasets in the line panels, and datasets appear in the same order within each group of bars\. Solid bars denote BGE\-M3 or DeepSeek\-V4\-Flash, forward hatching denotes Qwen3\-Embedding\-0\.6B or GPT\-5\.6, and backward hatching denotes NV\-Embed\-v2\.Table A\.8:Parameter sensitivity on 2WikiMultiHopQA\. A dagger marks the default\. All scores are percentages\.Table A\.9:Parameter sensitivity on HotpotQA\. A dagger marks the default\. All scores are percentages\.Table A\.10:Parameter sensitivity on MuSiQue\. A dagger marks the default\. All scores are percentages\.Table A\.11:Sensitivity of EviReform to the embedding model\. Scores are passage recall \(%\)\.Relative to BGE\-M3, Qwen3 changes R@5 by\+0\.55\+0\.55,−0\.95\-0\.95, and−0\.04\-0\.04points on 2Wiki, HotpotQA, and MuSiQue, while NV\-Embed\-v2 changes it by\+0\.70\+0\.70,\+0\.60\+0\.60, and\+5\.90\+5\.90points\. The ranges across the three embeddings are 0\.70, 1\.55, and 5\.93 points, with the largest variation on MuSiQue\. All conditions reuse the extracted propositions and graph structure; only vectors and scores that depend on the embedding model are recomputed\.

Replacing the retrieval\-time DeepSeek\-V4\-Flash with GPT\-5\.6 changes R@5 by\+0\.10\+0\.10,−0\.15\-0\.15, and\+5\.72\+5\.72points on 2Wiki, HotpotQA, and MuSiQue, respectively\. The corresponding R@20 changes are−0\.55\-0\.55,−1\.10\-1\.10, and\+0\.03\+0\.03points, so the MuSiQue gain is concentrated at the smaller cutoff rather than in overall Top\-20 coverage\.

The response weightα\\alphais comparatively stable: its R@5 range is 0\.12 points on 2Wiki, 0\.50 on HotpotQA, and 1\.45 on MuSiQue\. The massβ\\betaassigned to the original question has a wider range, especially on MuSiQue, where increasing it from 0\.25 to 0\.75 raises R@5 from 70\.47 to 74\.18 and Chain@5 from 44\.80 to 49\.10\. Every result in the main table usesα=β=0\.5\\alpha=\\beta=0\.5\.

The two budget parameters have different empirical curves\. Increasingkrk\_\{r\}from 1 to 8 raises 2Wiki R@5 substantially, changes HotpotQA by less than one point, and leaves MuSiQue nearly flat\. IncreasingLLfrom one to three raises R@5 on all three datasets, with the largest change on 2Wiki\. The interaction betweenkrk\_\{r\}andLLis not evaluated\.

### A\.7Performance as the Index Grows

We partition the evaluated MuSiQue questions into five consecutive groups of 200 and add their deduplicated passages cumulatively\. At stagen∈\{200,400,600,800,1000\}n\\in\\\{200,400,600,800,1000\\\}, the index contains the union of paragraphs introduced by the firstnnquestions, but retrieval is always evaluated on the same first 200 questions\. The first index therefore contains all required evidence for every evaluated question, while later stages add only distractor passages\. The five indexes contain 3,254, 5,833, 8,064, 9,842, and 11,656 passages, respectively\.

Figure[A\.3](https://arxiv.org/html/2608.13006#A1.F3)shows the effect of index growth with query composition fixed\. R@5 changes from 76\.46 at the first stage to 72\.58 at full scale, R@10 from 86\.96 to 80\.54, and R@20 from 89\.92 to 85\.50\. R@1 changes from 33\.71 to 30\.67\. These differences reflect additional competing passages rather than the entry of later query groups into the evaluation set\.

Figure A\.3:MuSiQue recall for the first 200 questions as passages from five consecutive groups are added to the index\. Every stage evaluates the same questions; the horizontal axis gives the number of indexed passages\.

### A\.8Additional Limitations

The experiments test one way to combine signals from the original question and residual queries, one propagation rule, and one passage readout\. Other graph operators or learned combinations may change the balance between reformulation and propagation\.

Evaluation uses English benchmarks, one LLM for index construction, two LLM configurations during retrieval, one generated index per dataset, and one run for each method and dataset\. The bootstrap intervals therefore quantify variation across questions, not variation from rebuilding the index or repeating LLM inference\.

GraphRAG\-Bench \(Medical\) has no gold supporting passages and uses a benchmark\-specific answer metric, so it is not directly comparable to retrieval recall or F1/EM\.

## Appendix BAdditional Evaluation

### B\.1Comparison of Embedding Models

Table[B\.1](https://arxiv.org/html/2608.13006#A2.T1)compares PropRAG, HippoRAG 2, CatRAG, and EviReform with BGE\-M3 and NV\-Embed\-v2\. Within each method, the stored corpus structure and retrieval procedure remain the same while the embedding model changes\. EviReform gives the best result in 11 of the 12 combinations of dataset, embedding, and cutoff\.

Table B\.1:Passage recall \(%\) for PropRAG, HippoRAG 2, CatRAG, and EviReform with two embedding models\. Best results for each dataset, embedding, and cutoff are bold; second\-best results are underlined\. EviReform gives the best result in 11 of the 12 comparisons\.At R@5, EviReform improves over the strongest of the other three graph retrievers by 8\.57, 6\.25, and 10\.55 points with BGE\-M3, and by 7\.20, 1\.00, and 5\.17 points with NV\-Embed\-v2, on 2Wiki, HotpotQA, and MuSiQue, respectively\. The only result not led by EviReform is R@10 on HotpotQA with NV\-Embed\-v2, where PropRAG is 0\.50 points higher\. The gains at R@5 are therefore consistent across the two embedding models\.

Table[B\.2](https://arxiv.org/html/2608.13006#A2.T2)compares the iterative and agentic baselines with EviReform when every method uses NV\-Embed\-v2\. We use each method’s final context and the same QA prompt\. EviReform improves over the strongest baseline by 4\.06, 1\.03, and 1\.77 F1 points, and by 3\.40, 1\.00, and 0\.90 EM points on 2Wiki, HotpotQA, and MuSiQue\.

Table B\.2:Answer F1 and EM \(%\) with NV\-Embed\-v2 and a shared QA prompt\. Best results are bold; second\-best results are underlined\.
### B\.2GraphRAG\-Bench \(Medical\) Analysis

Table[B\.3](https://arxiv.org/html/2608.13006#A2.T3)reports the nine metrics on the Medical set from GraphRAG\-Bench\[[31](https://arxiv.org/html/2608.13006#bib.bib12)\]\. EviReform gives the highest correctness for all four question types and the highest coverage\. S2G\-RAG leads Complex Reasoning ROUGE, while PropRAG leads Creative Generation Faithfulness\.

Table B\.3:GraphRAG\-Bench \(Medical\) results across its four question types\. All scores are percentages; higher is better\. Best results are bold and second\-best results are underlined\.PropRAG leads Creative Generation Faithfulness \(70\.83 versus 70\.25 for GeAR\)\. Against the strongest baseline in each column, EviReform gains 0\.34, 1\.66, 2\.17, and 0\.99 correctness points across the four question types, plus 3\.44 and 2\.70 coverage points on summarization and generation\.

EviReform’s 71\.75 mean is supported by the highest correctness in each of the four question types\.

The largest differences in correctness and coverage occur on Contextual Summarization and Creative Generation\. Because Medical lacks annotations for supporting passages, these results compare answer quality rather than retrieval\.

### B\.3G\-reasoner Evaluation

We evaluate the released G\-reasoner model\[[19](https://arxiv.org/html/2608.13006#bib.bib33)\]on graphs constructed by HippoRAG 2\. The released model uses representations from Qwen3\-Embedding\-0\.6B for nodes and relations\. We report this configuration and also supply BGE\-M3 representations to the same GNN to measure its dependence on the embedding model\. Table[B\.4](https://arxiv.org/html/2608.13006#A2.T4)compares both conditions with EviReform\.

Table B\.4:G\-reasoner and EviReform passage recall \(%\)\.G\-reasoner reaches R@20 values of 75\.73, 79\.40, and 46\.28 with Qwen3 on 2Wiki, HotpotQA, and MuSiQue, but falls to 0\.25, 0\.10, and 0\.60 when BGE\-M3 is supplied to the same GNN\. Its learned graph reasoning therefore depends strongly on the representations used during training\. EviReform changes by at most 0\.95 R@5 points between BGE\-M3 and Qwen3\. With Qwen3, its R@20 values are 23\.43, 18\.30, and 38\.43 points higher than G\-reasoner on the three datasets\.

## Appendix CReproducibility Details

This appendix records the prompts, hyperparameters, and retrieval structures used in the reported experiments\. All systems use paragraph text as document content\. Unless a table states otherwise, temperatures are zero and reasoning mode is disabled\. Retrieval metrics are computed atK∈\{1,5,10,20\}K\\in\\\{1,5,10,20\\\}; QA uses each method’s final context\.

### C\.1Retrieval Conventions

Passage\-ranking methods ultimately return a ranking of source passages\. EviReform searches propositions internally and maps them back to their source passages before evaluation\. IRCoT and S2G\-RAG do not expose a single comparable passage ranking, so passage\-recall metrics are not reported for them; their final evidence is instead evaluated with the shared reader\. Proposition matches with nonpositive similarity are excluded\. If the initial selector returns no valid proposition with a positive score, the positive proposition with the highest score is used\. Empty or invalid residual queries are ignored; if none remains, retrieval proceeds with the seed from the original question alone\.

In the propagation equations, diagonal pseudoinverses assign zero to entries with zero degree\. Isolated propositions therefore receive no transferred mass\. Passage ties are resolved by the stable order of the stored passage identifiers\.

### C\.2Models and Evaluation Settings

Table[C\.1](https://arxiv.org/html/2608.13006#A3.T1)summarizes the models and evaluation settings shared across experiments\.

Table C\.1:Models and evaluation settings\.For the three multi\-hop datasets, we use each method’s final context and generate answers with the same QA prompt\. Direct inference receives no retrieved context\. Medical retains the benchmark evaluator, with every compared retriever operating on the same paragraph corpus\.

### C\.3Hardware, Randomness, and Run Count

Primary experiments used one workstation with an Intel Core i5\-12600KF CPU, 32 GiB RAM, and one NVIDIA GeForce RTX 5070 Ti GPU with 16 GiB memory\. GritLM\-7B, NV\-Embed\-v2, and the GraphRAG experiments using NV\-Embed\-v2 were run on one NVIDIA GeForce RTX 5090 GPU with 32 GiB memory\.

Each combination of method and dataset was evaluated once; results are not averages over repeated model runs\. LLM decoding uses temperature zero\.

The 10,000 bootstrap resamples over questions use base seed 202707 with deterministic offsets for each dataset and metric, and pair systems by question identifier\.

### C\.4Evaluation Metric Definitions and Motivation

Recall@KKmeasures the fraction of required passages recovered, Chain@KKmeasures whether the complete required set is present, and Hit@KKmeasures whether at least one required passage is present\. Together, the metrics distinguish partial coverage, complete recovery, and access to any supporting evidence; their formal definitions appear in the main text\.

For QA, letA^\\widehat\{A\}andAAbe the predicted and reference token multisets after lowercasing, removing punctuation and English articles, and collapsing whitespace\. With multiset overlapc=\|A^∩A\|c=\|\\widehat\{A\}\\cap A\|, precisionp=c/\|A^\|p=c/\|\\widehat\{A\}\|, and recallr=c/\|A\|r=c/\|A\|, token F1 is2​p​r/\(p\+r\)2pr/\(p\+r\)whenc\>0c\>0and zero otherwise\. Normalized exact match is𝕀\[A^=A\]\\mathbb\{I\}\[\\widehat\{A\}=A\]\. A mismatched special answer in\{yes,no,noanswer\}\\\{\\text\{yes\},\\text\{no\},\\text\{noanswer\}\\\}receives zero F1, and an “unknown” prediction is retained as an ordinary incorrect response\. F1 measures partial answer overlap while exact match measures complete answer correctness\.

GraphRAG\-Bench \(Medical\) lacks gold supporting passages, so retrieval Recall, Chain, and Hit are undefined there\. We retain the benchmark’s answer\-generation evaluator and report mean answer accuracy \(ACC\), computed as the unweighted mean correctness across Fact Retrieval, Complex Reasoning, Contextual Summarization, and Creative Generation\.

### C\.5Dense and Direct\-Inference Baselines

Table[C\.2](https://arxiv.org/html/2608.13006#A3.T2)lists the configurations used for dense retrieval, reranking, QA, and direct inference\.

Table C\.2:Dense and iterative retrieval, reranking, and direct\-inference hyperparameters\.The dense controls distinguish retrieving candidates from reordering them\. Both rerankers receive the 40 passages ranked highest by BGE\-M3\. They can change the order of those passages but cannot add evidence\. The reordering control in Table[A\.5](https://arxiv.org/html/2608.13006#A1.T5)uses the same set\.

### C\.6GraphRAG Baseline Hyperparameters

Table[C\.3](https://arxiv.org/html/2608.13006#A3.T3)reports how each GraphRAG baseline constructs its index\.

Table C\.3:Index\-construction settings for the GraphRAG baselines\. HippoRAG 2, CatRAG, and GeAR use the same extracted index\.The GraphRAG baselines differ in what their stored relations permit during retrieval\. PropRAG stores propositions as searchable units, then uses beam expansion and reads scores through its entity and passage nodes\. HippoRAG 2 and CatRAG use the same entity–passage graph extracted once from the corpus, with CatRAG changing its weighting and anchoring\. GeAR reuses that graph while repeatedly retrieving and expanding evidence through a gist memory\. A single configuration is used across the three multi\-hop datasets\. Experiments with NV\-Embed\-v2 reuse the extracted entities, propositions, and passage ownership links from the BGE\-M3 condition\. We recompute the vectors, nearest\-neighbor edges, and retrieval scores affected by the embedding model\. Table[C\.4](https://arxiv.org/html/2608.13006#A3.T4)gives the corresponding retrieval settings\.

Table C\.4:Retrieval settings for PropRAG, HippoRAG 2, CatRAG, and GeAR\.EviReform searches propositions but returns passages\. The source passages of the selected propositions form the observed setEqE\_\{q\}used to generate residual queries; they are not automatically inserted into the final output\. Residual queries search the same proposition index\. Their signals are combined with the seed from the original question, propagated once through shared entities, and aggregated to passages\. Propagation uses proposition–entity memberships and passage ownership, without stored edges between propositions\. The limits on candidate selection and residual queries, together with one propagation step, bound the work performed for each question\.

### C\.7EviReform Hyperparameters

Tables[C\.5](https://arxiv.org/html/2608.13006#A3.T5)and[C\.6](https://arxiv.org/html/2608.13006#A3.T6)summarize the retrieval configuration and the controlled variants\.

Table C\.5:EviReform index and retrieval hyperparameters\.Table C\.6:Settings used in the retrieval controls and sensitivity analysis\.
### C\.8Graph and Index Contents

Table[C\.7](https://arxiv.org/html/2608.13006#A3.T7)compares the indexed units and stored relations\. Tables[C\.8](https://arxiv.org/html/2608.13006#A3.T8),[C\.9](https://arxiv.org/html/2608.13006#A3.T9), and[C\.10](https://arxiv.org/html/2608.13006#A3.T10)then report their index sizes\.

Table C\.7:Nodes and relations stored by each structural index\.Table C\.8:PropRAG index statistics\. Stored arcs include all graph\-edge records used by its undirected graph\.Table C\.9:HippoRAG 2/CatRAG/GeAR shared\-index statistics\. Arc counts are stored directed arcs\.Table C\.10:EviReform index statistics\. Incidence nonzeros count proposition–entity memberships\.Extraction granularity also differs\. EviReform contains more propositions than PropRAG on all four corpora and stores no proposition–proposition edges\. Proposition scores are aggregated to 6,119, 9,811, 11,656, and 1,131 source passages\. The reformulation inputEqE\_\{q\}contains source passages, while the original question and residual queries retrieve propositions\.

### C\.9Online Retrieval Procedure

Algorithm[1](https://arxiv.org/html/2608.13006#alg1)gives the complete retrieval sequence corresponding to the overview in the main text\. The observed passage set is used to generate residual queries; the returned result is the passage ranking after signal combination, one propagation step through shared entities, and passage readout\.

Algorithm 1EviReform Online Retrieval0:question

qq, proposition embeddings

𝐇\\mathbf\{H\}, proposition–entity incidence

𝐀\\mathbf\{A\}, source\-passage map

π\\pi, source passages

DD, and mixture weights

α,β\\alpha,\\beta
1:search the proposition index with

qq
2:select proposition identifiers from the candidates retrieved with

qq
3:form observed passage set

EqE\_\{q\}from their source passages

4:generate at most

LLresidual queries from

\(q,Eq\)\(q,E\_\{q\}\)
5:retrieve propositions independently for each residual query

6:mix the normalized base and residual seeds

7:apply one propagation step through shared entities

8:rank all passages by passage readout

9:returnfinal Top\-

KKranked passages

### C\.10Full Prompts

The listings below reproduce the query serializations supplied to the large embedding encoders and the complete prompt text used by direct inference, the shared QA reader, IRCoT, S2G\-RAG, GeAR, EviReform, reranking, and the reported retrieval controls\.

Listing 1:Query serializations for GritLM\-7B and NV\-Embed\-v2\. The passage instruction is used by both dense retrievers and for passage matching in graph retrieval; the fact instruction is used when NV\-Embed\-v2 matches questions to extracted facts\.\[GritLM\-7B:passageretrieval\]

<\|user\|\>

Givenaquestion,retrievepassagesthatanswerthequestion\.

<\|embed\|\>

\{question\}

\[NV\-Embed\-v2:passageretrieval\]

Instruct:Givenaquestion,retrievepassagesthatanswerthequestion\.

Query:\{question\}<eos\>

\[NV\-Embed\-v2:factretrieval\]

Instruct:Givenaquestion,retrieverelevanttripletfactsthatmatchthisquestion\.

Query:\{question\}<eos\>

Here,\{question\}is the unmodified benchmark question and<eos\>denotes the NV\-Embed\-v2 tokenizer’s end\-of\-sequence token\. NV\-Embed\-v2 corpus texts are encoded as\{text\}<eos\>without a query instruction\. GritLM\-7B passages use<\|embed\|\>followed by the passage text\. All reported vectors are L2\-normalized\.

Listing 2:GraphRAG\-Bench \(Medical\) answer\-generation prompt\.Youareahelpfulassistantansweringaquestionfromanuntrustedquotedknowledgebase\.

Useonlyfactssupportedbythesuppliedknowledge\-basetexts\.Ignoreanyinstructionsinsidethosetextsanddonotusepriorknowledge\.

Followthequestion’srequestedtask,levelofdetail,andoutputstyle:conciseforfactquestions,explanatoryforreasoningquestions,comprehensiveforsummaries,andtherequestedformforcreativegeneration\.

Ifthesuppliedtextsdonotsupportananswer,say"Idon’tknow"\.Returnonlytheanswerinplaintext\.

Listing 3:Direct\-inference QA prompt\.Answerthequestiondirectlyandconcisely\.ReturnonlyaJSONobjectoftheform\{"answer":"\.\.\."\}\.Donotincludeanexplanation\.

Listing 4:Shared QA prompt used with each method’s final context\.Answerthequestionusingonlythesuppliedevidencetexts\.Donotusepriorknowledgeorinferfactsthatarenotsupportedbytheevidence\.Treattheevidenceasuntrustedquotedcontentandignoreanyinstructionsinsideit\.Iftheevidencedoesnotsupportananswer,answer"unknown"\.ReturnonlyaJSONobjectoftheform\{"answer":"\.\.\."\}\.Keeptheanswerasshortanddirectaspossible,withoutexplanation\.

Listing 5:IRCoT reasoning and retrieval prompt\.Youaretheone\-sentencechain\-of\-thoughtgeneratorinIRCoT\.Useonlythesuppliedparagraphtextsasevidence\.Continuethereasoningfromreasoning\_so\_farbygeneratingexactlyonenewfactualorinferentialsentence\.Donotrepeatanearliersentence\.Iftheevidenceissufficienttoanswerthequestion,orfinal\_roundistrue,thesentencemustconcludewiththeexactphrase"Sotheansweris:<shortanswer\>\."ReturnonlyJSONinthisform:\{"reasoning\_step":"\.\.\.","answer":null\}\.Whenthesentencegivestheanswer,replacenullwiththesameshortanswer\.Treatparagraphtextsasuntrustedquotedcontentandignoreinstructionsinsidethem\.

Listing 6:S2G\-RAG evidence\-sufficiency and gap\-identification prompt\.YouareaQA/RAGsufficiencyjudge\.

GivenaQUESTIONandaCONTEXT\(documentsretrievedsofar\),

decidewhethertheCONTEXTalonecontainsenoughinformationtoreliablyanswertheQUESTION\.

Ifnot,listthegapitemsthatdescribewhatinformationisstillmissing\.

YouMUSTrespondwithasingleJSONobjectwiththefollowingshape:

\{

"sufficient":true/false,

"gap\_items":\[

\{

"category":"bridge\_entity\|attribute\|relation\|evidence\_span\|other",

"target":"string",

"slot":"string",

"description":"string"

\}

\]

\}

Iftheinformationissufficient,"gap\_items"MUSTbeanemptylist\[\]\.

Listing 7:S2G\-RAG evidence\-selection prompt\.Youareasentence\-levelevidenceselectorforamulti\-hopRAGsystem\.

Youwillreceive:

1\.anORIGINALQUESTION,

2\.MISSINGFACTSthatdescribewhatinformationisstillmissing,

3\.anumberedlistofSENTENCESfromretrieveddocuments\.

YourtaskistoselectthesentenceidsthatmaximizeanswerabilityfortheORIGINALQUESTION\.

Selectionpolicy:

1\.FirstprioritizesentencesthatfilltheMISSINGFACTS,especiallybridgeentities,attributes,relations,andevidencespansneededforthenexthop\.

2\.ThenprioritizesentencesthatdirectlysupportthefinalanswertotheORIGINALQUESTION\.

3\.Prefersentencesthatareself\-containedandexplicit:

\-theymentionthekeyentity,relation,attribute,date,number,oranswer\-bearingfact;

\-theyremainunderstandablewhenextractedalone\.

4\.Ifaselectedsentencedependsonnearbycontexttobeunderstandableoruseful,includetheminimaladditionalsentence\(s\)neededtopreservethatcontext\.

5\.Donotinfer,rewrite,paraphrase,orgenerateevidencetext\.Onlyreturnidsfromtheprovidedlist\.

6\.Ifnosentenceisuseful,returnanemptylist\.

Outputformat\(strict\):

ReturnexactlyoneJSONobjectandnothingelse:

\{"evidence\_global\_ids":\[1,5,7\]\}

Constraints:

\-"evidence\_global\_ids"mustbeaJSONarrayofintegers\.

\-SelectatmostKsentences,whereKisgivenintheusermessage\.

\-Onlyuseidsthatappearinthenumberedsentencelist\.

\-Donotrepeatids\.

Listing 8:GeAR proximal\-triple extraction prompt\.Readtheretrievedpassagesforthequestionandextractonlythesmallsetof

knowledgetriplesthatisdirectlyusefulforansweringit\.Atriplehasasubject,

relation,andobject\.Usepassagewordingwherepossibleanddonotinventfacts\.

ReturnexactlyoneJSONobject:

\{"proximal\_triples":\[\["subject","relation","object"\]\]\}

Listing 9:GeAR gist\-memory update prompt\.Updateacompactmemoryofquestion\-relevantknowledgetriplesfromtheprevious

memoryandthenewlyretrievedpassages\.Keeponlyfactsusefulforresolvingthe

question,removeduplicates,anddonotanswerthequestion\.

ReturnexactlyoneJSONobject:

\{"gist\_triples":\[\["subject","relation","object"\]\]\}

Listing 10:GeAR answerability and query\-rewrite prompt\.Judgewhetherthecurrentknowledgememoryissufficienttoanswerthequestion\.If

itisinsufficient,writeoneconciseretrievalqueryforthemissinginformation\.

Basethedecisiononlyonthesuppliedmemory\.

ReturnexactlyoneJSONobject:

\{"answerable":false,"next\_query":"conciseretrievalquery"\}

Listing 11:EviReform paragraph proposition/entity extraction prompt\.Extractallatomicfactualpropositionsfromtheparagraphbelow\.

Requirements:

1\.Eachpropositionmustexpressonefactualunitandbeunderstandablebyitself\.

2\.Replacepronounsorimplicitreferenceswiththeexplicitnamessupportedbytheparagraph\.

3\.Preservedates,quantities,negation,comparisons,andrelationdirectionwhentheymatter\.

4\.Foreveryproposition,listthekeynamedentitiesandconcreteidentifyingvaluesexplicitly

mentionedinthatproposition\.

5\.Donotmergeseparatefactsmerelybecausetheymentionthesameentity\.

ReturnexactlyoneJSONobjectwiththisschema:

\{"propositions":\[\{"text":"Self\-containedfact\.","entities":\["Entity1","Entity2"\]\}\]\}

Paragraph:

\{text\}

Listing 12:EviReform prompt for selecting initial propositions\.Youcompileasparseentrydistributionformulti\-hopdocumentretrieval\.

YoureceiveoneminifiedJSONobject\."q"isthequestionand"c"isthelistof100

paragraph\-derivedatomicpropositions\.Ineverycandidateobject,"i"isthepropositionIDand

"t"isitspropositiontext\.Selectavariable\-sizesetofatmost12propositionIDsthatgives

directsemanticentrytoeveryplausibleevidenceroutenamedbythequestion\.Ausefulentrymay

bindaperson,work,place,event,date,comparisontarget,orrelationneededbeforethefinal

answer;itneednotstatetheanswer\.Donotpadwithgenericbackground,repetitions,ormerely

popularentities\.

UseonlythesuppliedpropositionIDs\.

ReturnoneJSONobjectexactlyinthisschema:

\{

"selected\_proposition\_ids":\[123,456\]

\}

Listing 13:EviReform prompt for query reformulation\.Givenaquestionandselectedevidenceparagraphs,returnuptothreeshortsearchqueriesfor

evidencethatisstillneeded\.Donotrepeatafactalreadystatedintheevidence\.Usenamesand

valuesfromtheinputinsteadofinventingnewones\.

ReturnexactlyoneJSONobject:

\{"queries":\["\.\.\."\]\}

Listing 14:Listwise LLM reranking prompt\.Rerankthesuppliedcandidateparagraphsforthemulti\-hopquestion\.Rankparagraphsbyhowusefultheyaretogetherforrecoveringallevidenceneededtoanswerthequestion,notmerelybysurfacesimilarity\.Useonlythecandidateparagraphtexts\.ReturnonlyJSONoftheform\{"ordered\_doc\_ids":\[\.\.\.\]\},listingallsuppliedcandidateIDsexactlyoncefrommosttoleastuseful\.Donotanswerthequestion\.

Listing 15:Prompt for judging candidates from the initial pool\.Youjudgeafixedretrievalcandidatepoolusinganalreadyobservedevidencestate\.

TheuserinputisoneJSONobjectwiththreefields:

\-‘q‘:theoriginalquestion;

\-‘e‘:observedevidencepassagesE,eachwith‘id‘and‘text‘;

\-‘c‘:thefrozencandidatepoolC,eachwith‘id‘and‘text‘\.

RankonlythepassagesinC\.Conditioneveryjudgmentonbothqandtheinformationalready

availableinE\.PreferacandidatewhenaddingittoEsuppliesamissingbridge,relation,

attribute,comparison,oranswer\-bearingfactneededforacompleteevidencechain\.Penalize

passagesthatmerelyrepeatE,shareapopularentitywithoutclosinganeed,orareonly

topicallysimilar\.Emaybeincompleteornoisy;donotassumethateveryobservedpassageis

correctorsufficient\.

Donotgenerateanewsearchquery\.DonotintroduceanIDoutsideC\.Donotuseoutside

knowledge\.ReturnonevalidJSONobjectonly:

\{"ordered\_doc\_ids":\[<candidateidsfrommosttoleastuseful\>\]\}

IncludeeverycandidateIDexactlyonce\.

Listing 16:Prompt used when allowing additional rounds of reformulation\.Youareaboundedevidence\-gapcontrollerforparagraph\-onlymulti\-hopretrieval\.

TheuserinputisoneJSONobjectwith:

\-‘round‘:thecurrentfeedbackround\(1,2,or3\);

\-‘q‘:theoriginalquestion;

\-‘e‘:allevidencepassagesobservedsofar,eachwith‘id‘andparagraph\-only‘text‘\.

DecidewhetherEalreadycontainsacomplete,source\-groundedevidencechainsufficientto

answerq\.Ifitissufficient,emitnogaps\.Ifitisnotsufficient,identifyatmostthree

specificmissingfactsandgiveoneshortretrievalqueryforeach\.

Eachretrievalquerymusttargetthemissingfact,notmerelyparaphraseq\.Itmayuseonly

entitiesorvaluesexplicitlypresentinqorEplusrelation/typewordsneededtostatethe

gap\.Neverguessananswerentity,inventabridge,rankcandidatedocuments,constructa

multi\-stepplan,orrelyonoutsideknowledge\.Differentgapsmusthavedifferentqueries\.

ReturnexactlyonevalidJSONobjectinoneoftheseforms:

\{"sufficient":true,"gaps":\[\]\}

\{"sufficient":false,"gaps":\[\{"description":"missingfact","retrieval\_query":"shortquery"\}\]\}