The Attribution-Compression Frontier in Retrieval-Augmented Generation

Deepanshu Mody

cs.CLSep 13, 2026

The Attribution-Compression Frontier in Retrieval-Augmented Generation

Authors: Deepanshu Mody

Organizations: Center for Data Science New York University

Abstract

Context compression reduces generator input in retrieval-augmented generation, but answer quality alone does not characterize citation attribution. We measure citation attribution across compression methods and budgets, comparing reranking, extractive selection, abstractive summarization, token pruning, and an extract-cluster-rewrite construction on ASQA and QASPER under a fixed generator and primary entailment evaluator. On ASQA at a nominal 0.25 budget (achieved compression 0.08), a RECOMP-style compressor's citations score 0.86 precision against its summaries but 0.12 against source spans under our re-attributability protocol. These estimates depend on a shared NLI model for span recovery and citation scoring and lack independent human calibration. Extractive selection's observed grounded precision ranges from 0.43 to 0.49 across nominal budgets from one-half to one-tenth of the ASQA context, while answer quality declines. For the same RECOMP setting, claim verification after source recovery yields an unsupported rate of 0.88 versus 0.17 when checking summaries. This gap persists beyond structural rejection of missing provenance, but remains evaluator-dependent. A 200-question TRUE T5-XXL audit also finds emitted--grounded gaps under both fixed and recomputed source mappings, without establishing human-calibrated support rates.

Figures & tables

Figure 1: The attribution–compression frontier on 944 common ASQA questions. Compression increases rightward (

x

: achieved fraction of retrieved context). Curves show descriptive means under the shared evaluator and recovery protocol; paired difference intervals and tests appear in Appendix B . The right panel reports str-EM.

Figure 2: The frontier on 976 common QASPER questions from 271 papers, axes as in Figure 1 . The third panel scores each compressor’s declared evidence against gold paragraphs; it is undefined for no-provenance methods. Small plotted differences do not establish equivalence.

ASQA
Compressor	$\hat{\rho}$	$P_{g}$	$P_{e}$	$U_{s}$	$U_{m}$	$U_{r}$
Identity	1.00	0.51	0.51	0.44	0.44	0.44
Rerank top- $k$	0.20	0.69	0.69	0.36	0.37	0.36
Extractive	0.26	0.49	0.49	0.46	0.46	0.46
ECR	0.26	0.48	0.49	0.46	0.45	0.46
RECOMP	0.08	0.12	0.86	1.00	0.17	0.88

Table 1: Citation precision and unsupported rates at nominal budget 0.25 (identity: full context), on 944 ASQA and 976 QASPER questions. Notation follows Section 3 ; no-provenance declared-span rates are structural. Two zero-claim RECOMP answers are excluded from QASPER verification rates.

Compressor	Gen. in	Ver. span	Ver. sum.	Total
Identity	798	371	382	1169
Rerank top- $k$	210	171	174	381
Extractive	277	112	114	389
ECR	277	118	107	395
RECOMP	127	0	84	127
LLMLingua-2	253	0	240	253

Table 2: Mean input tokens on 944 ASQA questions at nominal budget 0.25 (identity: full context). Total includes generator and declared-span verification input; compression, output, and recovery are excluded.

$\tau$	$P_{g}$	reattr.	$P_{g}$	reattr.
	RECOMP		LLMLingua-2
0.3	0.15	$0.48^{n=929}$	0.16	0.54
0.4	0.13	$0.44^{n=929}$	0.13	0.47
0.5	0.12	$0.41^{n=929}$	0.10	0.40
0.6	0.11	$0.38^{n=929}$	0.08	0.34
0.7	0.09	$0.34^{n=929}$	0.05	0.27

Table 3: Recovery-threshold sensitivity on 944 ASQA questions at nominal budget 0.25; citation scoring stays at

\tau=0.5

P_{g}

: grounded precision; reattr.: re-attribution rate. Superscripts count defined values; dashes mark undefined metrics.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Compressor	$\rho$	$\hat{\rho}$	$P_{g}$	$R_{g}$	$P_{e}$	reattr.	str-EM	$U_{s}$	$U_{m}$	$n$
Identity	1	1.00	0.51	0.56	0.51	–	0.34	0.44	0.44	944
Rerank top- $k$	0.5	0.40	0.61	0.62	0.61	–	0.30	0.38	0.38	944
Rerank top- $k$	0.25	0.20	0.69	0.63	0.69	–	0.25	0.36	0.37	944
Rerank top- $k$	0.1	0.11	0.38	0.37	0.38	–	0.19	0.63	0.63	944
Rerank top- $k$	0.05	0.06	0.24	0.22	0.24	–	0.12	0.78	0.78	944
Extractive	0.5	0.51	0.43	0.52	0.43	–	0.32	0.47	0.48	944

Appendix

Table 4: All 21 primary settings on 944 common ASQA questions; question-weighted means. Notation and exclusions follow Appendix A . No-provenance declared-span rates are structural.

Compressor	$\rho$	$\hat{\rho}$	$P_{g}$	$R_{g}$	$P_{e}$	reattr.	EF1	$U_{s}$	$U_{m}$	$n$
Identity	1	1.00	0.48	0.47	0.48	–	0.08	0.53	0.53	976
Rerank top- $k$	0.5	0.50	0.35	0.37	0.35	–	0.07	0.63	0.63	976
Rerank top- $k$	0.25	0.25	0.39	0.40	0.39	–	0.06	0.59	0.60	976
Rerank top- $k$	0.1	0.10	0.42	0.40	0.42	–	0.04	0.60	0.60	976
Rerank top- $k$	0.05	0.05	0.39	0.36	0.39	–	0.03	0.64	0.64	976
Extractive	0.5	0.50	0.35	0.41	0.35	–	0.09	0.59	0.59	976

Appendix

Table 5: All 21 primary settings on 976 common QASPER questions; question-weighted means. Notation and exclusions follow Appendix A . No-provenance declared-span rates are structural.

ID	Data	Left	Right	Endpoint	$\Delta$	95% CI	$p_{H}$	$n$
C001	A	RC/.5	RC/.5	$P_{e}-P_{g}$	$+0.744$	$[+0.715,+0.771]$	0.0148	944
C002	A	RC/.25	RC/.25	$P_{e}-P_{g}$	$+0.740$	$[+0.711,+0.767]$	0.0148	944
C003	A	RC/.1	RC/.1	$P_{e}-P_{g}$	$+0.743$	$[+0.715,+0.770]$	0.0148	944
C004	A	RC/.05	RC/.05	$P_{e}-P_{g}$	$+0.621$	$[+0.588,+0.654]$	0.0148	944
C005	A	LL/.5	LL/.5	$P_{e}-P_{g}$	$+0.426$	$[+0.394,+0.458]$	0.0148	944
C006	A	LL/.25	LL/.25	$P_{e}-P_{g}$	$+0.425$	$[+0.391,+0.458]$	0.0148	944

Appendix

Table 6: Planned comparisons 1–30 of 148.

\Delta

: left-minus-right mean difference;

p_{H}

: Holm-adjusted

p

-value. Setting and metric labels follow Appendix B .

ID	Data	Left	Right	Endpoint	$\Delta$	95% CI	$p_{H}$	$n$
C031	A	RC/.25	Id	$P_{g}$	$-0.387$	$[-0.421,-0.354]$	0.0148	944
C032	A	RC/.1	Id	$P_{g}$	$-0.392$	$[-0.424,-0.359]$	0.0148	944
C033	A	RC/.05	Id	$P_{g}$	$-0.334$	$[-0.370,-0.299]$	0.0148	944
C034	A	LL/.5	Id	$P_{g}$	$-0.353$	$[-0.386,-0.320]$	0.0148	944
C035	A	LL/.25	Id	$P_{g}$	$-0.406$	$[-0.438,-0.374]$	0.0148	944
C036	A	LL/.1	Id	$P_{g}$	$-0.457$	$[-0.488,-0.427]$	0.0148	944

Appendix

Table 7: Planned comparisons 31–60 of 148.

\Delta

: left-minus-right mean difference;

p_{H}

: Holm-adjusted

p

-value. Setting and metric labels follow Appendix B .

ID	Data	Left	Right	Endpoint	$\Delta$	95% CI	$p_{H}$	$n$
C061	A	ECR/.1	X/.1	$P_{g}$	$-0.004$	$[-0.029,+0.021]$	1.0000	944
C062	A	ECR/.05	X/.05	$P_{g}$	$-0.027$	$[-0.049,-0.005]$	0.4158	944
C063	Q	ECR/.5	X/.5	$P_{g}$	$+0.063$	$[+0.033,+0.092]$	0.0148	976
C064	Q	ECR/.25	X/.25	$P_{g}$	$+0.030$	$[+0.002,+0.059]$	1.0000	976
C065	Q	ECR/.1	X/.1	$P_{g}$	$+0.025$	$[-0.001,+0.051]$	1.0000	976
C066	Q	ECR/.05	X/.05	$P_{g}$	$+0.018$	$[-0.009,+0.046]$	1.0000	976

Appendix

Table 8: Planned comparisons 61–90 of 148.

\Delta

: left-minus-right mean difference;

p_{H}

: Holm-adjusted

p

-value. Setting and metric labels follow Appendix B .

ID	Data	Left	Right	Endpoint	$\Delta$	95% CI	$p_{H}$	$n$
C091	A	RC/.05	Id	$\mathrm{EM}$	$-0.067$	$[-0.084,-0.051]$	0.0148	944
C092	A	LL/.5	Id	$\mathrm{EM}$	$-0.041$	$[-0.055,-0.026]$	0.0148	944
C093	A	LL/.25	Id	$\mathrm{EM}$	$-0.101$	$[-0.118,-0.083]$	0.0148	944
C094	A	LL/.1	Id	$\mathrm{EM}$	$-0.184$	$[-0.203,-0.165]$	0.0148	944
C095	A	LL/.05	Id	$\mathrm{EM}$	$-0.227$	$[-0.248,-0.207]$	0.0148	944
C096	A	RR/.25	RR/.5	$\mathrm{EM}$	$-0.052$	$[-0.064,-0.040]$	0.0148	944

Appendix

Table 9: Planned comparisons 91–120 of 148.

\Delta

: left-minus-right mean difference;

p_{H}

: Holm-adjusted

p

-value. Setting and metric labels follow Appendix B .

ID	Data	Left	Right	Endpoint	$\Delta$	95% CI	$p_{H}$	$n$
C121	Q	RR/.25	RR/.5	$\mathrm{EF1}$	$-0.0162$	$[-0.0225,-0.0097]$	0.0148	976
C122	Q	RR/.1	RR/.25	$\mathrm{EF1}$	$-0.0194$	$[-0.0275,-0.0112]$	0.0148	976
C123	Q	RR/.05	RR/.1	$\mathrm{EF1}$	$-0.0084$	$[-0.0171,+0.0008]$	1.0000	976
C124	A	LL/.25	LL/.5	$\mathrm{Rec}$	$-0.156$	$[-0.194,-0.117]$	0.0148	944
C125	A	LL/.1	LL/.25	$\mathrm{Rec}$	$-0.153$	$[-0.192,-0.114]$	0.0148	944
C126	A	LL/.05	LL/.1	$\mathrm{Rec}$	$-0.050$	$[-0.086,-0.015]$	0.2418	944

Appendix

Table 10: Planned comparisons 121–148 of 148.

\Delta

: left-minus-right mean difference;

p_{H}

: Holm-adjusted

p

-value. Setting and metric labels follow Appendix B .

		Prior DeBERTa		TRUE fixed		TRUE recovery
Data	Method	$P_{e}$	$P_{g}$	$P_{e}$	$P_{g}$	$P_{e}$	$P_{g}$
A	Identity	0.467	0.467	0.563	0.563	0.563	0.563
A	Extractive	0.402	0.402	0.497	0.497	0.497	0.497
A	RECOMP	0.790	0.090	0.780	0.130	0.780	0.140
A	LLMLingua-2	0.480	0.070	0.415	0.023	0.415	0.010
Q	Identity	0.487	0.487	0.498	0.498	0.498	0.498

Appendix

Table 11: Precision on the same sampled saved answers: 100 ASQA (A) and 100 QASPER (Q) questions (84 QASPER papers).

P_{e}/P_{g}

: emitted/grounded precision. Mappings and evaluation details follow Appendix C ; superscripts give nonmissing counts and dashes undefined values.

ID	Data	Method	View	Endpoint	$\Delta$	95% CI	$p_{H}$	$n$	$k$
T01	A	RC	$F$	$P_{e}-P_{g}$	$+0.650$	$[+0.550,+0.750]$	0.0024	100	100
T02	A	RC	$F$	$R_{e}-R_{g}$	$+0.640$	$[+0.540,+0.735]$	0.0024	100	100
T03	A	LL	$F$	$P_{e}-P_{g}$	$+0.392$	$[+0.290,+0.495]$	0.0024	100	100
T04	A	LL	$F$	$R_{e}-R_{g}$	$+0.367$	$[+0.273,+0.463]$	0.0024	100	100
T05	Q	RC	$F$	$P_{e}-P_{g}$	$+0.430$	$[+0.326,+0.530]$	0.0024	100	84
T06	Q	RC	$F$	$R_{e}-R_{g}$	$+0.425$	$[+0.320,+0.525]$	0.0024	100	84

Appendix

Table 12: All 24 planned TRUE robustness contrasts, separate from the original 148-test analysis.

D/F/T

: prior DeBERTa/TRUE fixed/TRUE recovery;

n/k

: questions/clusters.

\Delta

: left-minus-right mean difference;

p_{H}

: Holm-adjusted within this 24-test family. 10,000 replicates, seed 13. Other notation follows Appendix B .

RAG Citation Faithfulness

See all 19 papers

CAGE: Cognitive Attribution Graphs for Faithful Inline Citation Generation in Long-Form Question Answering
Jul 27, 2026Zhichao Yan, Shizhao Li, Jiapu Wang +5Retrieval-Augmented Generation Question Answering
Shanxi University, Taiyuan, China · Nanjing University of Science and Technology, Nanjing, China · Nanyang Technological University, Singapore +3
ProvenAI: Provenance-Native Traces of Evidence in Generated Answers
Jun 24, 2026Mohammad Faizan, Dalal AlharthiMulti-Hop QA Data Provenance
College of Information Science University of Arizona
Compression Beyond the Uncompressed: A Two-Stage Training Recipe for Soft Context Compression in RAG
Sep 4, 2026Shuyu Guo, Shuo Zhang, Zhaochun RenRetrieval-Augmented Generation LLM Inference Acceleration
Shandong University Qingdao, China · Bloomberg London, United Kingdom · Leiden University Leiden, The Netherlands

Jul 27, 2026J/K move · Enter open · S save

CAGE: Cognitive Attribution Graphs for Faithful Inline Citation Generation in Long-Form Question Answering

Zhichao Yan, Shizhao Li, Jiapu Wang, Haoran Luo+4

Shanxi University, Taiyuan, China · Nanjing University of Science and Technology, Nanjing, China · Nanyang Technological University, Singapore+3

Retrieval-Augmented Generation Question Answering Evidence-Grounded Generation RAG Citation Faithfulness

Open paper AskPDF arXiv

Long-form question answering increasingly relies on retrieved evidence to make LLM outputs verifiable, with inline citations tracing claims to source documents. However, existing systems often attach citations that are topically related but insufficient to support their claims. We identify attribution ambiguity as a structural challenge: end-to-end generation must implicitly resolve combinatorial claim--document assignments, obscuring evidential boundaries and increasing the risk of evidence-boundary overrun, where claims exceed cited support. To address this challenge, we propose CAGE (Cognitive Attribution Graphs for Citation Generation), a two-stage framework that introduces an explicit cognitive attribution map before answer generation. CAGE first trains a plug-and-play Cognitive Map Induction Model to construct answer-centered support subgraphs, aligning each semantic answer unit with supporting documents through explicit relations. A Structured Citation Reasoning Model then realizes these units as sentence-level claims with map-aligned citations. Experiments on ASQA, ELI5, and ExpertQA show that CAGE achieves state-of-the-art performance, demonstrating the effectiveness of attribution-space contraction and map-guided citation generation.

Abstract

RAG Citation Faithfulness

Explore similar work

CAGE: Cognitive Attribution Graphs for Faithful Inline Citation Generation in Long-Form Question Answering

ProvenAI: Provenance-Native Traces of Evidence in Generated Answers

Compression Beyond the Uncompressed: A Two-Stage Training Recipe for Soft Context Compression in RAG