When Does Evidence-Graph Structure Predict Answer Correctness? Label-Free Abstention for Structured Retrieval

Anjan Goswami

cs.IROct 16, 2026

When Does Evidence-Graph Structure Predict Answer Correctness? Label-Free Abstention for Structured Retrieval

Authors: Anjan Goswami

Organizations: SmartInfer, Inc.

Abstract

Retrieval systems return scored answers, but a high score does not show that an answer should be trusted. The proposed training-free abstention method applies the boundary cut to the selected evidence in an evidence graph. The signal is the weighted cut between selected and unselected candidates and requires no correctness labels or confidence-model training. On MemoryAgentBench (MAB) FactConsolidation multi-hop, with unresolved queries ranked last for every selector, it raises area under the precision-coverage curve from 0.736 for the best score baseline to 0.785 (+0.049, 95% CI [+0.008, +0.092]); under the native evaluation, where score baselines keep their unresolved values, the gap is +0.108. A zero-cut subset contains 71 of 400 queries at 100% observed precision. In this pipeline, the subset corresponds to structural closure; controlled analyses attribute much of the remaining separation to resolution status, with little benefit from continuous edge weights. Across external benchmarks, usefulness varies with graph construction, edge semantics, and relation; typed structure alone does not ensure useful confidence. A supervised gradient-boosted selector achieves 0.814 versus the boundary cut's 0.785 when the retrieval stack matches. These findings provide empirical guidance for using available evidence graphs for lightweight, auditable post-retrieval abstention without additional model inference.

Figures & tables

Regime	Base Accuracy	Coherence AUPCC	Status-matched smoothed	Gap
Multi-hop	0.420–0.700	0.785	0.736	+0.049
Single-hop	0.900–0.960	0.975	—	—

Table 1: Aggregate results with the status-matched multi-hop comparison primary: unresolved queries are ranked last for every selector. Native comparisons are reported below the table; resolved-only results are in § 4.1 .

Split	Accuracy	Coherence	Smoothed	Raw	Margin
6k	0.910	0.980	0.928	0.914	0.718
32k	0.900	0.936	0.951	0.950	0.813
64k	0.960	0.977	0.940	0.928	0.875
262k	0.900	0.971	0.910	0.898	0.862
All	0.917	0.975	0.941	0.932	0.823

Table 2: Per-split single-hop AUPCC (

n{=}100

per split). Smoothed score leads on 32k.

Figure 1: Single-hop precision–coverage. The boundary cut helps even at 90–96% base accuracy.

Subset	Size	Precision	Overlap w/ zero-energy
Zero-energy ( $E{=}0$ )	71	1.000	—
Matched-coverage raw (top 71)	71	0.676	14 (Jaccard 0.110)
Top 10% raw	40	0.775	—
Top 10% smoothed	40	0.825	—

Table 3: Zero-energy overlap analysis on multi-hop MAB.

Scorer	Encoder	$d$	Coh. AUPCC	Native smoothed	Gap
$(\mathbf{q}^{\top}\mathbf{v})^{2}$	Qwen	128	0.785	0.677	+0.108
$\mathbf{q}^{\top}\mathbf{v}$	Qwen	128	0.775	0.690	+0.085
$(\mathbf{q}^{\top}\mathbf{v})^{2}$	MiniLM	384	0.788	0.680	+0.108

Table 4: Transfer across scorers and encoders (multi-hop, 400 queries each). Same 71 zero-energy queries, same 100% precision in all configurations.

AUPCC	Analysis	Relation to 0.785
0.785	Original end-to-end MAB (pooled 400, d128)	headline
0.782	Regenerated mechanism audit (400; 70 zero-energy)	different regeneration
0.765	Graph ablation (cached candidates/path; graph varied)	within-table only
0.765–0.766	Gamma sweep (separate cached harness; path re-resolved per $\gamma$ )	within-sweep only
0.750	Edge perturbation (own cached seed/protocol)	perturbation baseline

Table 5: Provenance of MAB boundary-cut AUPCC values across analyses.

Figure 2: Multi-hop precision–coverage (6k/32k/64k/262k). The boundary cut (red) attains the highest AUPCC on all four multi-hop context lengths.

Population / signal	AUPCC
Full ( $N=400$ )
Weighted cut	0.782
Unweighted cut	0.783
Resolved non-closure ( $N=246$ )
Weighted cut	0.600
Unweighted cut	0.610

Table 6: Boundary-signal AUPCC on the regenerated MAB multi-hop population: full (

N=400

) and resolved non-closure (

N=246

). Paired intervals in the accompanying analysis resample exact question texts.

Benchmark	$k$	$\gamma$	Retrieval / graph	Reproduction
LongMemEval temporal	5	0.100	session chain, continuous	reproduced
LongMemEval semantic	5	0.100	mutual-kNN(5), continuous	reproduced
MetaQA-3hop	10	0.100	typed answer-relation + type filter	reproduced
CoDEx-S	10	0.100	co-occurrence / typed mask	exploratory
LoCoMo	5	0.100	temporal chain	blocked

Table 7: External-benchmark evaluation configurations and reproduction status.

Variant	Agg.	6k	32k	64k	262k	Zero-energy precision
Typed entity–relation	0.765	0.866	0.743	0.754	0.592	1.000
Semantic kNN	0.716	0.827	0.683	0.655	0.598	0.917
Hybrid ( $\alpha=0.250$ )	0.745	0.855	0.722	0.723	0.567	0.986
Hybrid ( $\alpha=0.500$ )	0.747	0.858	0.722	0.738	0.560	0.986
Hybrid ( $\alpha=0.750$ )	0.748	0.857	0.724	0.737	0.564	0.986

Table 8: Full MAB graph ablation on cached local neighborhoods (multi-hop).

Figure 3: Graph-representation ablation on cached multi-hop MAB neighborhoods. Results are computed post hoc on cached local neighborhoods and are not directly comparable to the end-to-end AUPCC values in Tables 1 and 13 . Typed entity–relation structure outperforms semantic kNN and hybrid alternatives. The typed graph is the only variant with a perfect zero-cut subset (71 queries, 100% precision). Adding semantic edges dilutes the signal, consistent with Remark 2.1 .

Benchmark / graph	Coherence	Best comparator	Diagnosis
LongMemEval / temporal	0.736	0.887	temporal adjacency
LongMemEval / semantic kNN	0.682	0.890	generic similarity
LoCoMo / temporal	0.286	0.430	temporal adjacency
CoDEx-S / typed entity–relation	0.373	0.543	heterogeneous, App. L

Table 9: Boundary-condition experiments. The boundary cut fails when the graph does not encode support-compatible structure.

Graph family	$N$	Acc	$R_{\text{median}}$	AUROC	AUPCC	ZE prec
MAB typed	400	0.560	8.150	0.755	0.765	1.000
MAB semantic kNN	400	0.560	2.750	0.692	0.716	0.917
LongMemEval temp.	500	0.740	1.010	0.488	0.736	—
LongMemEval kNN	500	0.750	0.510	0.414	0.682	—
LoCoMo temporal	1986	0.270	1.010	0.517	0.286	—
CoDEx-S co-occ.	1828	0.330	0.670	0.332	0.215	0.083

Table 10: Support-compatibility probe across six graph families. High

R_{\text{median}}

and AUROC describe observed correctness separation; values near 1.0 or below 0.5 indicate negative transfer. This secondary diagnostic consumes correctness labels; its threshold was chosen on the original evaluated families, not independently validated as a general predictive criterion. ZE denotes zero-energy.

PID	Label	$N$	Acc	AUPCC coh	AUPCC raw	$R_{\mathrm{med}}$	ZE prec	ZE $n$
P101	field of work	20	0.800	0.766	0.896	1.400	—	0
P463	member of	270	0.580	0.700	0.624	52.000	0.764	89
P1303	instrument	71	0.560	0.577	0.719	1.470	0.600	20
P27	citizenship	105	0.600	0.485	0.795	0.690	0.000	2
P530	diplomatic relation	287	0.240	0.256	0.390	N/A	0.244	287
P106	occupation	593	0.250	0.215	0.313	0.880	1.000	1

Table 11: Per-relation boundary-cut performance on CoDEx-S with typed entity–relation mask.

R_{\mathrm{median}}

: ratio of median cut on incorrect vs. correct answers (reported as N/A when both medians are zero). Relations ordered by boundary-cut AUPCC. This secondary diagnostic consumes correctness labels; its threshold was chosen on the original evaluated families, not independently validated as a general predictive criterion.

Answer relation	$N$	Base acc	AUPCC coh	AUPCC raw	$\Delta$ (95% CI)
in_language	46	0.435	0.546	0.317	$+0.230$ $[+0.030,+0.407]$
release_year	41	0.341	0.367	0.306	$+0.062$ $[-0.120,+0.246]$
has_genre	56	0.304	0.352	0.326	$+0.026$ $[-0.128,+0.157]$
starred_actors	31	0.226	0.168	0.141	$+0.027$ $[-0.137,+0.192]$
directed_by	39	0.282	0.287	0.333	$-0.046$ $[-0.292,+0.160]$
written_by	37	0.297	0.209	0.271	$-0.062$ $[-0.256,+0.137]$

Table 12: Illustrative relation-level results on MetaQA-3hop (

n{=}250

stratified), boundary cut vs. raw score, recomputed from the persisted per-query evaluation artifact. Both directions shown; selection is illustrative, not statistically representative. Intervals are paired bootstrap 95% CIs using the per-query sampling unit (10,000 replicates, seed 42), descriptive and uncorrected for multiple comparisons; small-

N

relations have wide intervals. Only in_language has an interval excluding zero.

Signal	AUPCC	AUROC
GBT	0.814	0.889
Coherence	0.785	0.807
Smoothed score	0.677	0.645
Raw score	0.653	0.611
LR	0.648	0.620
Score margin	0.547	0.413

Table 13: Learned and training-free confidence on headline MAB multi-hop (

N=400

), using the reconstructed mapping (Appendix M ). GBT and LR use five-fold cross-validation grouped by exact question; paired uncertainty resamples exact question texts.

Labels (%)	GBT (all)	Coherence (all)	GBT (262k)	Coherence (262k)
1	0.613	0.785	0.430	0.661
5	0.665	0.785	0.539	0.661
10	0.751	0.785	0.631	0.661
20	0.790	0.785	0.660	0.661
50	0.786	0.785	0.660	0.661
100	0.814	0.785	0.656	0.661

Table 14: Label-budget comparison on headline MAB multi-hop (

N=400

overall;

N=100

at 262k). Budgets are percentages of training rows within folds grouped by exact question; paired intervals resample exact question texts. Question identity uses the reconstructed mapping (Appendix M ).

Figure 4: Label-efficiency curve on multi-hop MAB. Boundary cut (zero-fit, horizontal line) vs. GBT trained on subsampled labels. Top: aggregate AUPCC. Bottom: 262k point estimates. GBT first exceeds the boundary cut in aggregate by point estimate at 20% labels; paired intervals include zero at 20%, 50%, and 100%. Question identity uses the reconstructed mapping (Appendix M ).

Train / test	base	minilm	linear
base	0.818	0.771	0.825
minilm	0.739	0.828	0.763
linear	0.742	0.696	0.838

Table 15: GBT transfer AUPCC across retrieval configurations on headline MAB multi-hop (

N=400

per target). Rows identify training configurations and columns deployment configurations. Folds and paired resampling use exact question texts under the reconstructed mapping (Appendix M ).

Cell	GBT	Coherence	$\Delta$	95% CI	Direction
base $\to$ base	0.818	0.785	0.033	[-0.0000, +0.065]	unresolved
base $\to$ minilm	0.771	0.788	-0.017	[-0.050, +0.024]	unresolved
base $\to$ linear	0.825	0.775	0.050	[+0.016, +0.087]	positive
minilm $\to$ base	0.739	0.785	-0.046	[-0.097, -0.012]	negative
minilm $\to$ minilm	0.828	0.788	0.041	[+0.005, +0.073]	positive
minilm $\to$ linear	0.763	0.775	-0.012	[-0.059, +0.029]	unresolved

Table 16: Paired AUPCC differences between transferred GBT and the target-configuration boundary cut on headline MAB multi-hop (

N=400

per target). Intervals use 10,000 exact-question bootstrap replicates, conditional on out-of-fold predictions, under the reconstructed mapping (Appendix M ).

Train / test	base	minilm	linear
base	0.827	0.805	0.824
minilm	0.808	0.820	0.792
linear	0.808	0.771	0.831

Table 17: GBT transfer AUPCC after transductive population-rank normalization on headline MAB multi-hop (

N=400

per configuration). Exact-question folds and paired resampling use the reconstructed mapping (Appendix M ). Feature ranks include held-out rows; this diagnostic is not within-query normalization.

Method	Accuracy	$\Delta$
Baseline	0.660	—
Leaf MCTS	0.660	+0.000
Chain MCTS	0.580	$-$ 0.080
Path reranking	0.580	$-$ 0.080

Table 18: Search baselines on multi-hop 6k. Search does not improve accuracy in this setting; chain methods hurt.

	Status-matched (primary)		Native		Resolved-only
Signal	AUPCC	AUROC	AUPCC	AUROC	AUPCC	AUROC
Coherence	0.782	0.810	0.782	0.810	0.822	0.635
Raw score	0.755	0.814	0.755	0.813	0.788	0.643
Smoothed score	0.754	0.815	0.754	0.814	0.787	0.643
Smoothed full-vector margin	0.666	0.729	0.593	0.559	0.675	0.479
NQC-local	0.669	0.730	0.558	0.520	0.679	0.481

Table 19: Frozen regenerated population: minimum attempted-hop aggregation. NQC and WIG use a local-background score proxy. All scores are recomputed on the same rows.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Split	Accuracy	Coherence	Smoothed	Raw	Margin
6k	0.700	0.861	0.740	0.730	0.670
32k	0.560	0.767	0.621	0.599	0.554
64k	0.590	0.771	0.762	0.736	0.672
262k	0.420	0.661	0.495	0.468	0.355
All	0.568	0.785	0.677	0.653	0.547

Appendix

Table 20: Per-split multi-hop AUPCC (

n{=}100

per split).

Split	Qwen $d{=}128$	MiniLM $d{=}384$	Winner
6k	0.861	0.855	Qwen (+0.006)
32k	0.767	0.769	MiniLM (+0.002)
64k	0.771	0.776	MiniLM (+0.005)
262k	0.661	0.669	MiniLM (+0.008)

Appendix

Table 21: Per-split boundary-cut AUPCC by encoder (multi-hop, scorer held fixed at

(\mathbf{q}^{\top}\mathbf{v})^{2}

Figure 5: AUPCC by context length. Boundary cut vs. the smoothed-score baseline across both regimes. The boundary cut has higher AUPCC at every multi-hop context length and at three of four single-hop context lengths; smoothed score is higher at single-hop 32k.

Test scale	Purged	GBT	Coherence	$\Delta$	95% CI
6k	21	0.859	0.861	-0.001	[-0.077, +0.067]
32k	18	0.794	0.767	0.027	[-0.047, +0.095]
64k	11	0.795	0.771	0.024	[-0.057, +0.109]
262k	2	0.673	0.661	0.012	[-0.075, +0.094]

Appendix

Table 22: Question-disjoint leave-one-scale-out evaluation on headline MAB multi-hop (

N=100

test rows per scale). Training rows sharing test-question text are purged. GBT-minus-boundary-cut intervals resample exact question texts under the reconstructed mapping (Appendix M ).

Feature	Importance
Raw score	0.058
Smoothed score	0.062
Score margin	0.139
Second-highest score	0.050
Score ratio	0.691

Appendix

Table 23: Mean GBT impurity importance across five exact-question folds of headline MAB multi-hop (

N=400

), using the reconstructed mapping (Appendix M ). These are descriptive fold means, without bootstrap resampling.

Population	$N$	Accuracy	Weighted cut	Unweighted cut	Boundary-weight fraction	Boundary-edge fraction
Full	400	0.562	0.782	0.783	0.813	0.813
Resolved	316	0.712	0.822	0.819	0.861	0.861
Non-closure	330	0.470	0.584	0.590	0.657	0.662
Resolved non-closure	246	0.630	0.600	0.610	0.697	0.701

Appendix

Table 24: Structural-control AUPCC on regenerated MAB multi-hop: full (

N=400

), resolved (

N=316

), non-closure (

N=330

), and resolved non-closure (

N=246

). The second panel retains the additional full and non-closure controls. Paired intervals resample exact question texts.

Comparison	$\Delta$	95% CI
Weighted cut - unweighted cut	-0.010	[-0.027, +0.035]
Boundary-weight fraction - weighted cut	0.097	[-0.0002, +0.196]
Boundary-edge fraction - unweighted cut	0.090	[+0.007, +0.201]
Boundary-weight fraction - boundary-edge fraction	-0.004	[-0.024, +0.016]

Appendix

Table 25: Paired AUPCC differences on regenerated MAB resolved non-closure queries (RN,

N=246

). Intervals use 10,000 paired bootstrap replicates grouped by exact question text.

Benchmark / graph	$N$	Reproduced vs published	Gate
LongMemEval / temporal	500	$0.736$ vs $0.736$	PASS
LongMemEval / semantic	500	$0.682$ vs $0.682$	PASS
MetaQA-3hop / typed	250	$\Delta{=}{+}0.020$ vs ${+}0.023$	PASS
CoDEx-S / co-occurrence	1828	$0.237$ vs $0.373$	FAIL (exploratory)
MAB single-hop	—	—	BLOCKED (provenance)
LoCoMo	—	—	BLOCKED (provenance)

Appendix

Table 26: E2 reproduction gates. Confirmatory lanes reproduce the published result; CoDEx-S fails (unseeded upstream training); two lanes are provenance-blocked.

Population	Weighted cut	Unweighted cut	Boundary-weight fraction	Boundary-edge fraction
LongMemEval temporal	0.736	0.745	0.733	0.736
LongMemEval semantic	0.682	0.779	0.853	0.774
MetaQA-3hop	0.324	0.324	0.283	0.283

Appendix

Table 27: Confirmatory E2 AUPCC on external populations: LongMemEval temporal and semantic (

N=500

each) and MetaQA-3hop (

N=250

). Paired intervals resample each benchmark's external query identifier. These populations are separate from headline and regenerated MAB.

Lane	Comparison	$\Delta$	95% CI
LongMemEval temporal	boundary-edge fraction $-$ unweighted cut	-0.009	[-0.026, +0.015]
LongMemEval temporal	boundary-edge fraction $-$ weighted cut	0.001	[-0.043, +0.044]
LongMemEval temporal	boundary-weight fraction $-$ boundary-edge fraction	-0.003	[-0.045, +0.039]
LongMemEval semantic	boundary-edge fraction $-$ unweighted cut	-0.005	[-0.017, +0.007]
LongMemEval semantic	boundary-edge fraction $-$ weighted cut	0.093	[+0.041, +0.146]
LongMemEval semantic	boundary-weight fraction $-$ boundary-edge fraction	0.078	[+0.048, +0.111]

Appendix

Table 28: External E2 paired AUPCC differences: LongMemEval temporal and semantic (

N=500

each), MetaQA-3hop (

N=250

), and exploratory CoDEx-S (

N=1828

). Intervals use 10,000 replicates grouped by external query identifier. CoDEx-S failed its reproduction gate.

Signal	Aggregate AUPCC	262k point estimate
Coherence	0.775	0.624
Smoothed score	0.688	0.518
Raw score	0.667	0.490
Margin	0.612	0.435

Appendix

Table 29: Selector comparison on multi-hop MAB under the linear scorer

\mathbf{q}^{\top}\mathbf{v}

(alternative to the default

(\mathbf{q}^{\top}\mathbf{v})^{2}

Perturbation	AUPCC	Delta
Baseline	0.750	—
Drop 10% edges	0.716 $\pm$ 0.005	$-0.034$
Drop 20% edges	0.717 $\pm$ 0.005	$-0.033$
Drop 30% edges	0.720 $\pm$ 0.002	$-0.029$
Add 10% spurious	0.743 $\pm$ 0.001	$-0.007$
Add 20% spurious	0.743 $\pm$ 0.001	$-0.006$

Appendix

Table 30: Robustness of the boundary cut to graph corruption on cached multi-hop MAB neighborhoods.

$\gamma$	Coherence AUPCC	ZE queries	ZE precision
0.000	0.766	71	1.000
0.050	0.765	71	1.000
0.100	0.765	71	1.000
0.200	0.765	71	1.000

Appendix

Table 31: Gamma sweep on multi-hop MAB (400 queries). Boundary-cut AUPCC is numerically stable across the tested

\gamma

values (0.765–0.766), with the same 71 zero-energy queries at 100% precision throughout.

Signal	All	6k	32k	64k	262k
GBT	0.889	0.912	0.888	0.885	0.850
Coherence	0.807	0.753	0.793	0.792	0.845
Smoothed score	0.645	0.586	0.605	0.732	0.617
Raw score	0.611	0.573	0.584	0.683	0.578
LR	0.620	0.586	0.592	0.655	0.605
Score margin	0.413	0.458	0.447	0.508	0.351

Appendix

Table 32: AUROC for headline MAB multi-hop (

N=400

overall;

N=100

per context scale). GBT and LR use folds grouped by exact question under the reconstructed mapping (Appendix M ); paired uncertainty uses exact-question resampling.

	Probe 20		Probe 30		Probe 50
Family (truth)	Acc	Med $R_{m}$	Acc	Med $R_{m}$	Acc	Med $R_{m}$
MAB typed (useful)	0.860	8.140	0.920	8.580	0.950	7.530
MAB sem. kNN (useful)	0.550	2.250	0.610	2.390	0.680	2.500
LongMemEval temp. (not)	0.970	1.000	0.980	0.970	1.000	0.990
LongMemEval kNN (not)	0.850	0.650	0.920	0.650	0.970	0.590
LoCoMo temp. (not)	1.000	1.020	1.000	1.030	1.000	1.030

Appendix

Table 33: Prospective

R_{\text{median}}

gate (1,000 trials per cell, threshold

=2

). Decision accuracy is the fraction of trials where the gate matches the full-run ground truth. The gate is conservative: false-positive rate

<1\%

at probe size 50, meaning it rarely approves a graph that should be rejected. This secondary diagnostic consumes correctness labels; its threshold was chosen on the original evaluated families, not independently validated as a general predictive criterion.

Train / test	base	minilm	linear
base	0.891	0.841	0.857
minilm	0.761	0.909	0.753
linear	0.798	0.744	0.868

Appendix

Table 34: GBT transfer AUROC on headline MAB multi-hop (

N=400

per target). Rows identify training configurations and columns deployment configurations. Folds and paired resampling use exact question texts under the reconstructed mapping (Appendix M ).

Configuration	Mean	Std	Range	Median
base	0.008	0.002	$[0.000,0.009]$	0.008
minilm	0.003	0.001	$[0.000,0.003]$	0.003
linear	0.390	0.131	$[0.000,0.625]$	0.406

Appendix

Table 35: raw_score distribution per configuration (

n{=}400

queries).

Selective Inference

Selective inference develops valid statistical conclusions after data-dependent choices such as feature selection, clustering, or adaptive evaluation.

See all 23 papers

AbstRAG: Learning to Abstract for Retrieval Problems
Jun 8, 2026Lei Xu, Xin Quan, Daniel Pedronette +1Retrieval-Augmented Generation Adaptive Retrieval
Idiap Research Institute, Switzerland · São Paulo State University, Brazil · Department of Computer Science, University of Manchester, United Kingdom +1
Entity-Memory Graph Retrieval Improves Evidence Coverage in Long-Conversation Question Answering
Aug 28, 2026Shumao SunConversational QA Evidence Retrieval
Tsinghua University
Regime Boundary Alignment for Evidence-Gated Question Answering
Sep 25, 2026Zeyan Li, Qirong Guo, SIyuan Qiu +3Evidence-Grounded QA LLM Grounding
Shanghai Jiao Tong University · The Hong Kong University of Science and Technology (Guangzhou)

Jun 8, 2026J/K move · Enter open · S save

AbstRAG: Learning to Abstract for Retrieval Problems

Lei Xu, Xin Quan, Daniel Pedronette, André Freitas

Idiap Research Institute, Switzerland · São Paulo State University, Brazil · Department of Computer Science, University of Manchester, United Kingdom+1

Retrieval-Augmented Generation Adaptive Retrieval RAG Evaluation

Open paper AskPDF arXiv

Retrieval-augmented generation often fails when the query, the document evidence, and the user's intent are expressed at different levels of abstraction. A query may ask about a class, a relation, or an event, while the document only states specific instances, indirect framings, or scoped formulations. We define this mismatch as an abstraction gap: the minimal set of typed assumptions required to align query intent with the available evidence. To close this gap, we introduce AbstRAG, which treats abstraction as an explicit retrieval object. AbstRAG decomposes the query--evidence gap into expression, conceptual, intent--evidence, and event-type components, and scores relevance by combining match quality, a query-independent utility prior, and the cost of the required bridges. Its central mechanism is reflective refinement: a critic diagnoses retrieval failures, localizes the failed abstraction operator, proposes a minimal stage-specific patch, and accepts the patch only under sufficiency and compression controls. Across three within-document retrieval benchmarks against seven baselines, AbstRAG outperforms on nDCG@10 in 18 of 21 paired-bootstrap contrasts and improves generation accuracy by 1.9%, 5.2%, and 4.0% across the three benchmarks; ablations confirm that reflective refinement drives most of the retrieval gain and the compression control alone reduces over-expansion false positives from 73.7% to 0% on a stress slice.

Abstract

Selective Inference

Explore similar work

AbstRAG: Learning to Abstract for Retrieval Problems

Entity-Memory Graph Retrieval Improves Evidence Coverage in Long-Conversation Question Answering

Regime Boundary Alignment for Evidence-Gated Question Answering