Large Language Models (LLMs) struggle with knowledge-intensive tasks due to hallucinations and fragmented reasoning over dispersed information. While Retrieval-Augmented Generation (RAG) grounds generation in external sources, existing methods often treat evidence as isolated units, failing to reconstruct the logical chains that connect these dots. Inspired by Complementary Learning Systems (CLS), we propose CodaRAG, a framework that evolves retrieval from passive lookup into active associative discovery. CodaRAG operates via a three-stage pipeline: (1) Knowledge Consolidation to unify fragmented extractions into a stable memory substrate; (2) Associative Navigation to traverse the graph via multi-dimensional pathways-semantic, contextualized, and functional-explicitly recovering dispersed evidence chains; and (3) Interference Elimination to prune hyper-associative noise, ensuring a coherent, high-precision reasoning context. On GraphRAG-Bench, CodaRAG achieves absolute gains of 7-10% in retrieval recall and 3-11% in generation accuracy. These results demonstrate CodaRAG's superior ability to systematically robustify associative evidence retrieval for factual, reasoning, and creative tasks.
Figures & tables
Figure 1. The CodaRAG Framework: Connecting the Dots with Associativity. The workflow begins by transforming unstructured text into a consolidated KG (Stage I), which is then traversed via multi-dimensional Associative Navigation (Stage II) to uncover semantic, contextualized, and functional evidence. Finally, Stage III applies Interference Elimination to filter out noise, ensuring only the most relevant context is used for the LLM-generated answer. Overview of the CodaRAG framework as a three-stage pipeline. Stage I performs knowledge consolidation, including Entity Type Discovery and Fragmented Entity Merging, transforming unstructured text into a structured knowledge graph. Stage II retrieves evidence by first identifying entry nodes and then applying Associative Navigation to expand into semantic, contextualized, and functional evidence subgraphs. Stage III filters irrelevant entities and relations, constructs the final context, and generates answers using a language model.
Figure 2. Overview of complementary Associative Navigation strategies. The CodaRAG framework enhances retrieval by integrating three core mechanisms: (a) Semantic Association for local entity expansion via semantic similarity; (b) Contextualized Association for query-aware importance propagation; and (c) Functional Association for identifying structurally analogous entities through topology analysis. Together, these strategies enable multi-dimensional navigation to connect disparate but relevant information. Illustration of three Associative Navigation pathways over a knowledge graph given a query. Starting from entry entities, (a) Semantic Association expands to nearby nodes based on similarity scores, (b) Contextualized Association propagates query-aware importance using scores, and (c) Functional Association retrieves structurally analogous nodes based on graph topology. Each pathway produces a distinct evidence subgraph for retrieval.
Retrieval Performance ↑
Generation Performance ↑
Method
Fact
Reason
Summary
Creation
Fact
Reason
Summary
Creation
REL
REC
REL
REC
REL
REC
REL
REC
RGL
ACC
RGL
ACC
COV
ACC
COV
FTH
ACC
Medical Dataset
LightRAG
91.00
90.65
94.00
89.61
93.00
94.06
81.50
65.32
12.52
49.04
8.63
53.43
90.17
63.61
52.40
71.51
62.16
HippoRAG 2
89.50
90.83
83.50
88.28
91.00
93.75
77.00
66.34
36.20
68.37
27.17
64.76
68.60
77.69
55.75
50.98
50.89
Ours (KG-only)
96.00
92.33
98.00
87.67
98.50
91.73
82.50
71.22
36.03
68.35
24.06
62.36
77.46
67.48
54.95
85.97
64.72
Table 1. Performance comparison across diverse retrieval and generation tasks.
Retrieval Performance ↑
Generation Performance ↑
Method
Fact
Reason
Summary
Creation
Fact
Reason
Summary
Creation
REL
REC
REL
REC
REL
REC
REL
REC
RGL
ACC
RGL
ACC
COV
ACC
COV
FTH
ACC
Stage I Indexing: Identifying the Dots by Knowledge Consolidation
Base
94.50
73.65
89.00
66.58
79.00
66.40
52.50
35.84
30.67
56.08
24.97
60.41
52.50
63.23
32.64
76.97
59.30
Ours
91.00
80.78
92.50
70.11
83.00
68.28
58.50
40.38
30.31
57.20
25.28
60.54
60.04
66.12
37.63
82.46
61.33
Stage II Retrieval: Connecting the Dots by Associative Navigation
Table 2. Ablation analysis of core components on the Medical dataset. We conduct ablation studies by comparing our full framework with five variants: (1) Base (the original LightRAG ( Guo et al., 2024 ) baseline); (2) w/o SemA, w/o ConA, and w/o FunA, stripping the Semantic, Contextualized, and Functional Association modules; and (3) w/o Elim, excluding the Interference Elimination module.
Medical Case: Discharge summary for PCNSL transitioning from Induction to Consolidation .
Ours
Completed induction with rituximab ; plans radiation ; monitored by care team using imaging to manage nausea .
LightRAG
Completed induction with rituximab ; plans radiation ; transitioning to consolidation managed by care team (missed specific symptom correlations).
HippoRAG 2
Completed induction for PCNSL; plan includes CBC monitoring and therapies like whole-brain radiation (lacks team coordination details).
Novel Case: Historical reasoning on Spanish descriptions in the progression affecting the Mayas.
Ours
Spaniards (1517) described monuments ; despite iconoclasm, the Maya language survived, preserving knowledge of figures like Chaacmol (noted by scholar Brasseur de Bourbourg ).
LightRAG
Spaniards (invaders) observed monuments and fought Mayas; despite destroying idols, the Maya language survived (missed specific scholarly links).
Table 3. Case studies illustrating generation outcomes. Key excerpts are presented here, with details provided in Appendix A.2 . Colored text indicates entities identified by different association pathways in our method: Semantic Association , Contextualized Association , and Functional Association . The same color annotations are applied to baseline outputs only for alignment and comparison, and do not imply that these methods employ the corresponding mechanisms.
Retrieval Performance ↑
Generation Performance ↑
Method
Fact
Reason
Summary
Creation
Fact
Reason
Summary
Creation
REL
REC
REL
REC
REL
REC
REL
REC
RGL
ACC
RGL
ACC
COV
ACC
COV
FTH
ACC
LightRAG
91.41
91.24
83.04
85.20
83.37
88.45
83.33
57.40
5.55
20.18
4.46
20.37
75.91
21.57
56.85
-
21.02
HippoRAG 2
83.61
86.26
74.84
74.75
82.65
83.73
52.77
55.07
32.50
19.28
25.76
21.22
60.98
22.20
51.30
-
21.13
Ours
93.04
95.02
84.47
88.99
85.71
93.68
80.56
66.00
28.75
21.20
14.72
22.20
77.55
22.32
65.48
-
22.10
Note: Evaluation metric definitions are detailed in Sec. 4 . The symbol ↑ indicates that higher scores represent better performance.
Table 4. Retrieval and generation performance on the Novel dataset with Mistral-medium-3.1 ( Mistral AI, 2025 ) . All stages use the same model. Faithfulness for Creative Generation is marked as ‘-’ due to invalid judge outputs.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Method
#Nodes ↑
#Edges ↑
Avg. CC ↑
LCC Ratio ↑
Frag. Ratio ↓
Iso. Ratio ↓
ClosedIE
727
748
0.1178
0.695
0.0702
0.099
OpenIE
691
685
0.1206
0.631
0.0955
0.097
Ours
893
880
0.1375
0.664
0.0907
0.073
Note: Evaluation Metric definitions are detailed in Sec. 4 . ↑ higher is better; ↓ lower is better.
Appendix
Table 5. Graph structural statistics under different entity discovery methods on a single document.
Figure 3. Analysis of entity merging decisions under synonym proliferation. The figure illustrates representative cases where semantically equivalent or near-equivalent entity mentions are consolidated into canonical nodes, improving graph connectivity and reducing fragmentation while preserving evidence provenance. Distribution of entity pairs by similarity and merging decisions (merged vs skipped) across Medical and Novel datasets, showing higher merge rates at greater similarity levels.
Medical Case: Discharge summary for Primary Central Nervous System Lymphoma transitioning from Induction to Consolidation .
Retrieval Trace
Ours
Entry: Entities: Primary Central Nervous System Lymphoma; Induction; … ; Relations: Primary Central Nervous System Lymphoma ↔ Induction; Consolidation ↔ Induction; … Semantic Association: Entities: Neurological Symptoms; Kidney Dysfunction; … ; Relations: Primary Central Nervous System Lymphoma ↔ Neurological Symptoms; High-Dose Methotrexate ↔ Kidney Dysfunction; … Contextualized Association: Entities: Imaging; Care Team; … ; Relations: Imaging ↔ Biopsy; Imaging ↔ Care Team; … Functional Association: Entities: Blood-Brain Barrier; Nausea; …
You completed induction with rituximab for primary CNS lymphoma and are transitioning to consolidation (e.g., radiation). Your multidisciplinary care team will coordinate this phase, using imaging and blood tests to monitor response. They will also manage side effects like nausea and fatigue to prioritize your quality of life.
LightRAG
You have completed induction chemotherapy for primary CNS lymphoma and are entering consolidation treatment to prevent relapse. Consolidation options may include radiation therapy or high-dose chemotherapy with stem cell rescue, depending on your condition and treatment response. Ongoing multidisciplinary care will focus on disease control and quality of life.
Appendix
Table 6. Case study illustrating retrieval–generation division of labor across paradigms (Medical Case). We report retrieval behaviors and the corresponding generation outcome.
Novel Case: Historical reasoning on how the Spanish are described in the progression affecting the Mayas.
The Spaniards (1517) were surprised to find well-built stone monuments and a civilized people. Although they invaded and destroyed many idols through iconoclasm, they could not destroy the Maya language; it preserved knowledge of figures like Chaacmol (noted by scholar Brasseur de Bourbourg).
LightRAG
The Spanish acted as intrusive colonial actors who observed Maya cities and monuments. They fought the Maya and attempted to erase indigenous customs by destroying idols. However, the Maya language survived, and the Conquest serves as a chronological marker for these events.
Appendix
Table 7. Case study illustrating retrieval–generation division of labor across paradigms (Novel Case). We report retrieval behaviors and the corresponding generation outcome.
Medical Case Query: Assume the role of a genetic counselor. Prepare a summary for a patient newly found to have a BRCA1 mutation, explaining cancer risk, surveillance, and preventive strategies.
Required Evidences
Generation (Detailed Summary)
The patient has tested positive for a BRCA1 mutation; A BRCA1 mutation significantly increases breast cancer risk; Having a BRCA1 mutation is a strong risk factor; The mutation may impact family members; Enhanced surveillance is recommended; Annual mammograms are recommended at an earlier age ; Breast MRI is recommended at an earlier age ; Preventive strategies include risk-reducing medication ; Preventive strategies include prophylactic surgery; Bilateral mastectomy may be considered ; Oophorectomy may be considered ; Discussion with oncology and surgical teams is recommended; Family members may consider genetic testing; Ongoing follow-up and support resources are available.
States that a germline BRCA1 mutation increases breast and ovarian cancer risk, discusses genetic counseling, family testing, and general surveillance, and emphasizes coordination with a multidisciplinary care team without specifying concrete screening schedules or preventive interventions.
Novel Case: How does the narrative link Mayab to the origins of the Mayas and their presence in Babylon?
Required Evidences
Generation (Detailed Summary)
Mayab is a country of the Mayas; The Mayas are said to have originated from Mayab; The Mayas established themselves in Babylon.
Identifies Mayab as the Maya homeland, compares Mayas with Chaldeans through cultural and architectural parallels, and introduces an Egypt-related migration hypothesis instead of explicitly stating Maya establishment in Babylon.
Appendix
Table 8. Error analysis case studies summarizing queries, abstracted generated answers, and ground-truth evidences required for each query. Evidences missing from the context are highlighted in bold .
Dataset
Medical
Novel
False Elimination Rate (%)
1.42
0.85
Appendix
Table 9. Analysis of false deletions introduced by Interference Elimination.
Retrieval Performance ↑
Generation Performance ↑
Method
Fact
Reason
Summary
Creation
Fact
Reason
Summary
Creation
REL
REC
REL
REC
REL
REC
REL
REC
RGL
ACC
RGL
ACC
COV
ACC
COV
FTH
ACC
Stage I Indexing: Identifying the Dots by Knowledge Consolidation
Base
85.99
69.22
84.82
66.00
85.46
61.30
81.94
32.36
38.02
54.64
25.98
43.54
49.87
43.07
35.28
71.89
44.37
Ours
89.56
75.01
88.19
72.77
85.97
64.93
80.83
35.50
38.61
56.34
24.50
45.61
51.53
46.92
37.12
77.90
40.50
Stage II Retrieval: Connecting the Dots by Associative Navigation
Appendix
Table 10. Ablation analysis of core components on the Novel dataset. We conduct ablation studies by comparing our full framework with five variants: (1) Base (the original LightRAG ( Guo et al., 2024 ) baseline); (2) w/o SemA, w/o ConA, and w/o FunA, stripping the Semantic, Contextualized, and Functional Association modules; and (3) w/o Elim, excluding the Interference Elimination module.
Retrieval Performance ↑
Generation Performance ↑
Method
Fact
Reason
Summary
Creation
Fact
Reason
Summary
Creation
REL
REC
REL
REC
REL
REC
REL
REC
RGL
ACC
RGL
ACC
COV
ACC
COV
FTH
ACC
LightRAG
92.00
91.57
86.00
89.49
90.00
91.84
77.50
84.41
3.85
20.87
5.12
21.08
84.26
22.02
59.60
-
22.61
HippoRAG 2
84.00
90.64
79.50
89.33
89.50
89.37
68.00
83.75
33.52
20.69
32.91
21.85
80.76
22.90
62.51
-
22.60
Ours
91.00
92.13
87.00
90.10
87.50
92.65
72.50
85.39
25.13
21.59
17.66
23.77
85.33
23.16
63.69
-
22.73
Note: Evaluation Metric definitions are detailed in Sec. 4 . The symbol ↑ indicates that higher scores represent better performance.
Appendix
Table 11. Retrieval and generation performance on the Medical dataset with Mistral-medium-3.1 ( Mistral AI, 2025 ) . All stages use the same model. Faithfulness for Creative Generation is marked as ‘-’ due to invalid judge outputs.
Category
Metric
Formula
Description & Key Intuition
Retrieval
Context Relevance
2r1+r2
Measures how relevant the retrieved context is to the query based on annotator scores ri∈{0,0.5,1} .
Evidence Recall
∣R∣1∑c∈RI(S(c,C))
Evaluates the proportion of reference claims supported by the retrieved context.
Generation
Rouge-L
RLCS+β2PLCS(1+β2)PLCSRLCS
Captures lexical overlap between the generated and reference answers using the longest common subsequence (LCS).
Answer Accuracy
αFC+(1−α)SS
Combines factual correctness (FC) and semantic similarity (SS) with α=0.5 .
Faithfulness
∣A∣∣{c∈A∣S(c,C)=1}∣
Reflects the proportion of generated claims supported by the retrieved context, indicating hallucination avoidance.
Evidence Coverage
∣K∣∣{k∈K∣M(k,G)=1}∣
Assesses how completely the required evidence is reflected in the generated answer.
Appendix
Table 12. Summary of Evaluation Metrics for Retrieval, Generation, and Graph Structure.
Method
Hyperparameter
Value
Description
LightRAG
top_k
10
-
chunk_top_k
5
-
HippoRAG 2
retrieval_top_k
10
-
linking_top_k
20
-
qa_top_k
5
-
Ours
top_k
10
Number of initial entity and relation entries used to initiate associative retrieval
Appendix
Table 13. Hyperparameters used in the main experiments.
Method
Hyperparameter
Value
Description
Ours (Ablation)
top_k
3
Number of initial entity and relation entries used to initiate associative retrieval
chunk_top_k
0
Number of text chunks provided to the generator
top_neighbors
8
Maximum neighbors explored per node during Semantic Association
top_ppr_nodes
15
Nodes selected by PPR in Contextualized Association
top_fastrp_nodes
15
Structurally analogous nodes selected in Functional Association
Appendix
Table 14. Hyperparameters used in the ablation studies.
Retrieval-Augmented Generation (RAG) significantly enhances Large Language Models (LLMs) by providing access to external knowledge. However, current research primarily focuses on retrieval quality, often overlooking the critical ''integration bottleneck'': even when relevant documents are retrieved, LLMs frequently fail to utilize them effectively due to conflicts with their internal parametric knowledge. In this paper, we argue that implicitly resolving this conflict in a single generation pass is suboptimal. We introduce GuarantRAG, a framework that explicitly decouples reasoning from evidence integration. First, we generate an ''Inner-Answer'' based solely on parametric knowledge to capture the model's reasoning flow. Second, to guarantee faithful evidence extraction, we generate a ''Refer-Answer'' using a novel Contrastive DPO objective. This objective treats the parametric Inner-Answer as a negative constraint and the retrieved documents as positive ground truth, forcing the model to suppress internal hallucinations in favor of external evidence during this phase. Finally, rather than naive concatenation or using the DPO trained model directly, we propose a joint decoding mechanism that dynamically fuses the logical coherence of the Inner-Answer with the factual precision of the Refer-Answer at the token level. Experiments on five QA benchmarks demonstrate that GuarantRAG improves accuracy by up to 12.1% and reduces hallucinations by 16.3% compared to standard and dynamic RAG baselines.
Zhengyi Zhao, Shubo Zhang, Zezhong Wang +7
The Chinese University of Hong Kong · Ministry of Education Key Laboratory of High Confidence Software Technologies, CUHK · University of International Relations +2
Retrieval-augmented generation (RAG) has become a standard approach for knowledge-intensive LLM tasks. Complex reasoning remains challenging because queries often require composing dispersed evidence through multiple dependent reasoning steps. Yet existing RAG systems process queries independently, repeatedly retrieving overlapping evidence and reconstructing established reasoning. Preserving such reasoning makes previously retrieved evidence and intermediate results reusable for subsequent queries. However, inference-time reuse becomes increasingly costly as the history grows, motivating a compact representation that preserves useful information without substantial loss in quality. To address these challenges, we present AutoPrunedRetriever (APR), a graph-based RAG system that encodes corpus facts and query-derived reasoning in a shared symbolic codebook. The codebook serves both as the retrieval index and as persistent memory, representing facts, questions, and answers as compact edge sequences, so that the previously established structure can be directly retrieved and reused. At inference time, APR retrieves query-specific reasoning paths and constructs compact prompts, while incremental consolidation keeps the growing memory compact. For deployments where the retriever cannot be replaced, we introduce AutoPruned Layer (APL), which leaves host retrieval unchanged but replaces the retrieved context with the corresponding entries in the shared symbolic codebook, writing each question and answer back as a compact structure. Across five complex-reasoning benchmarks, APR improves accuracy over the strongest baseline by 1.8-10.8 points while using up to 96% fewer input tokens. APL reduces generator input by 60-96% across host systems and preserves answer quality on document-style contexts.
Ning Wang, Kuanyan Zhu, Daniel Yuehwoon Yee +4
Cornell University · University of Cambridge · The University of Hong Kong +2
Large language models (LLMs) are widely used in retrieval-augmented generation (RAG) to incorporate external knowledge at inference time. However, when retrieved contexts are noisy, incomplete, or heterogeneous, a single generation process often struggles to reconcile evidence effectively. We propose MASS-RAG, a multi-agent synthesis approach to retrieval-augmented generation that structures evidence processing into multiple role-specialized agents. MASS-RAG applies distinct agents for evidence summarization, evidence extraction, and reasoning over retrieved documents, and combines their outputs through a dedicated synthesis stage to produce the final answer. This design exposes multiple intermediate evidence views, allowing the model to compare and integrate complementary information before answer generation. Experiments on four benchmarks show that MASS-RAG consistently improves performance over strong RAG baselines, particularly in settings where relevant evidence is distributed across retrieved contexts.
Xingchen Xiao, Heyan Huang, Runheng Liu +1
School of Computer Science and Technology, Beijing Institute of Technology · Department of Mathematical Sciences, Tsinghua University