ROZA Graphs: Self-Improving Near-Deterministic RAG through Evidence-Centric Feedback

Matthew Penaroza

cs.AIApr 8, 2026

ROZA Graphs: Self-Improving Near-Deterministic RAG through Evidence-Centric Feedback

Authors: Matthew Penaroza

Abstract

Language model agents reason from scratch on every query, discarding their chain of thought after each run. The result is lower accuracy and high run-to-run variance. We introduce reasoning graphs, which persist the per-evidence chain of thought as structured edges. Unlike prior memory that retrieves distilled strategies by query similarity, reasoning graphs enable evidence-centric feedback: for every candidate item, the system traverses all incoming evaluation edges across prior runs to surface how that specific item has been judged before. We further introduce retrieval graphs, which feed a planner that prunes consistently-rejected candidates over successive runs. Together they form a ROZA graph: a self-improving feedback loop in which accuracy gains scale with gold-passage reuse (reasoning graph) and efficiency gains scale with candidate-pool overlap (retrieval graph). The base model remains frozen; all gains come from context engineering via graph traversal. We evaluate on MuSiQue and HotpotQA, plus a high-reuse deployment subset. Four findings stand out. (1) Dose-response: accuracy improves monotonically with evidence-profile coverage, reaching +10.6pp over Vanilla RAG at 50%+ coverage on the same questions (47% error reduction, $p<0.0001$ ; per-question Spearman $ρ=+0.144$ , $p<10^{-6}$ , $n=1{,}100$ ). (2) Multi-hop scaling: 4-hop accuracy improves by +11.0pp ( $p=0.0001$ ). (3) Cross-cluster prediction: the cluster-level gain is predicted by gold-passage reuse density ( $r=0.604$ , $p=0.001$ , $n=26$ clusters). (4) High-reuse Pareto dominance: highest or tied-for-highest accuracy alongside 46% lower cost and 46% lower latency. Per-passage decision consistency across repeated runs ( $N=73$ paired probes, $K=10$ runs each, two model families, three temperatures) rises by +8 to +13pp on a fixed 20-passage context and by +12 to +21pp when the retrieval graph also prunes (all $p<0.005$ ).

Figures & tables

Panel of Figure 1 — Figure 1: Architecture. The agent loop (retrieve, reason, act) is augmented with two feedback graphs: the reasoning graph $\mathcal{G}_{R}$ feeds evidence profiles via item-inward traversal of past evaluations; the retrieval graph $\mathcal{G}_{P}$ feeds a pipeline planner that excludes consistently-rejected items. Both update atomically after each run. Evidence-centric feedback starts from the specific items in $C_{t}$ , not from query-embedding similarity.

Method	MuSiQue Acc	MuSiQue F1	HotpotQA EM	2Wiki EM
Vanilla RAG	64.5% $\,\pm\,$ 5.7	0.729 $\,\pm\,$ .043	73.6% $\,\pm\,$ 7.2	67.3% $\,\pm\,$ 13.7
Reflexion	65.8% $\,\pm\,$ 5.6	0.740 $\,\pm\,$ .047	73.6% $\,\pm\,$ 7.3	68.3% $\,\pm\,$ 12.3
ReasoningBank	66.5% $\,\pm\,$ 6.1	0.742 $\,\pm\,$ .046	74.2% $\,\pm\,$ 7.0	65.0% $\,\pm\,$ 13.8
Ours-RG	66.2% $\,\pm\,$ 5.6	0.743 $\,\pm\,$ .039	74.0% $\,\pm\,$ 6.4	67.3% $\,\pm\,$ 13.6
Ours-Full	68.1% $\,\pm\,$ 5.5 ∗∗∗	0.758 $\,\pm\,$ .037 ∗∗∗	73.8% $\,\pm\,$ 6.7	67.6% $\,\pm\,$ 13.2

Table 1: Main results: MuSiQue (

N{=}1{,}100

, 11 clusters), HotpotQA (

N{=}500

, 5), 2WikiMultiHopQA (

N{=}1{,}000

, 10). Mean

\pm

95% CI across clusters; 2Wiki uses EM. MuSiQue averages over cold-start questions (Sec. 5.2 ); differential gains explained by gold-reuse density (Sec. 5.3 ). McNemar vs. VR,

{}^{***}p<0.001

. Best bold .

Coverage	$n$	Ours-Full	VR (same Qs)	$\Delta$	Error Red.
0%	371	54.2%	54.4%	$-$ 0.3pp	0%
(0, 20)%	258	62.0%	60.1%	+1.9pp	5%
20–49%	208	75.0%	71.6%	+3.4pp	12%
50%+	263	88.2% ∗∗∗	77.6%	+10.6pp	47%

Table 2: Controlled dose-response: Ours-Full vs. Vanilla RAG (VR) on the same questions, stratified by Ours-Full coverage. Error reduction

=1{-}(\text{Ours err}/\text{VR err})

. At 50%+ coverage,

-47\%

errors. McNemar:

{}^{***}p<0.001

{}^{**}p<0.01

Config	Sonnet Acc	Haiku Acc	S–H Gap	Sonnet Lat.	Cost/q (¢, Sonnet)
Vanilla RAG	59.7%	49.1%	10.6pp	15.77s	2.96
Reflexion	63.8%	49.9%	14.0pp	23.20s	3.73
ReasoningBank	59.4%	48.8%	10.6pp	23.07s	3.65
Ours-RG	61.5%	55.3%	6.2pp	15.31s	3.23
Ours-Full	64.6% ∗∗	56.6% ∗∗	8.0pp	8.46s	1.61

Table 3: High-reuse results (387 questions). Ours-Full is Pareto-dominant on both models. Cost/q at Sonnet list pricing for cross-row comparability; native Haiku is

\approx

\times

cheaper. Significance vs. Vanilla RAG:

{}^{**}p<10^{-3}

{}^{*}p<0.05

Model	Temp	VR	Ours-RG	$\Delta_{\text{RG}}$	$p_{\text{RG}}$	Ours-Full	$\Delta_{\text{Full}}$
Haiku	0.0	0.808	0.910	$+0.101$	$1.1\!\times\!10^{-4}$	0.963	$+0.155$
Haiku	0.5	0.704	0.822	$+0.118$	$4.1\!\times\!10^{-4}$	0.916	$+0.212$
Haiku	0.7	0.685	0.814	$+0.129$	$3.7\!\times\!10^{-6}$	0.895	$+0.210$
Sonnet	0.0	0.853	0.933	$+0.080$	$2.9\!\times\!10^{-3}$	0.971	$+0.118$
Sonnet	0.5	0.775	0.888	$+0.112$	$2.0\!\times\!10^{-4}$	0.937	$+0.162$
Sonnet	0.7	0.789	0.885	$+0.096$	$3.1\!\times\!10^{-3}$	0.916	$+0.127$

Table 4: VVIR, paired

N{=}73

K{=}10

runs, cp387. Ours-RG is confound-free (identical 20-passage context as VR); Ours-Full additionally prunes. One-sided Wilcoxon for

\Delta{>}0

; all Ours-Full

p<2\!\times\!10^{-4}

. Hard-stratum in App. Tab. 13 .

Figure 2: Verdict-vector identity on a Sonnet $T{=}0$ hard probe (3-hop, $|C_{q}|{=}20$ , $K{=}10$ runs). Rows are runs, columns are passages; cells are green (used) or red (rejected) . Vanilla RAG (left): verdict pattern flickers across runs (VVIR $=0.30$ ). Ours-RG (right): all 10 runs identical (VVIR $=1.0$ ); the same collapse occurs for $16/23$ Sonnet and $20/38$ Haiku hard probes at $T{=}0$ . Aggregate statistics in Table 4 ; hard-stratum drilldown in App. Tab. 13 .

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Coverage bin	$n$	Acc (ours)	Acc (vanilla)	$\Delta$	95% CI
[0.0, 0.2)	629	0.574	0.568	+0.006	$[-0.010,+0.022]$
[0.2, 0.4)	156	0.737	0.712	+0.026	$[-0.019,+0.071]$
[0.4, 0.6)	113	0.796	0.726	+0.071	$[+0.018,+0.133]$
[0.6, 0.8)	119	0.916	0.807	+0.109	$[+0.050,+0.168]$
[0.8, 1.0]	83	0.892	0.771	+0.120	$[+0.036,+0.205]$

Appendix

Table 5: A5 controlled-overlap analysis on the paired ours_full / vanilla_rag MuSiQue pool (Sonnet 4,

n_{\text{paired}}=1100

). Bins are formed on the ours_full per-question evidence_profile_coverage ; the vanilla_rag branch has no graph and contributes only its outcome vector. CIs are 10,000-resample bootstrap. Per-question Spearman

\rho{=}+0.144

(95% CI

[+0.087,+0.201]

), one-sided

p{=}7.89e-07

Figure 3: A5 controlled-overlap re-analysis. Per-bin Acc gap is monotonic across all five bins; the per-question Spearman rank correlation $\rho{=}{+}0.144$ (95% CI $[+0.087,+0.201]$ ) is significant at one-sided $p<10^{-6}$ on $n{=}1{,}100$ paired questions. The monotonicity, paired design, and robustness to distributional outliers collectively reaffirm that the gain attributable to evidence profiles grows with prior-decision coverage.

Figure 4: Paired accuracy gain

\Delta

(in percentage points) of ours_full over vanilla_rag , stratified by reasoning depth (rows) and evidence-profile coverage bin (columns). Each cell shows the paired Acc difference and the cell sample size; the bordered region marks the triple-alignment sweet spot (hop

\geq 3

AND coverage

\geq 0.4

n{=}199

), on which ours_full reaches

89.4\%

vs. vanilla_rag

76.9\%

(

\Delta=+12.6

pp). All cells have

n\geq 25

. Source data: paired ours_full / vanilla_rag MuSiQue logs (

n{=}1{,}100

, 11 clusters).

Figure 5: Distribution of ours_full token-F1 on the $n{=}60$ paired MuSiQue questions rescued from vanilla_rag (i.e., vanilla_rag F1 $<0.8$ and ours_full F1 $\geq 0.8$ ). Mean rescued F1 = 0.99 (vs. 0.15 under vanilla_rag on the same questions); 73% of rescued questions had vanilla_rag F1 $=0.0$ exactly (complete failures). Win/loss ratio: 60:21 ( $2.86\times$ ).

Cluster	$n_{\text{flip}}$	$\text{P(correct}\|\text{dir.)}$	$\text{P(correct}\|\text{undir.)}$	OR [95% CI]	$p_{F}$	$p_{F}^{>}$
Aggregated	916	68.4% [60.7, 75.2]	3.2% [2.1, 4.6]	66.43 [39.14, 112.76]	$<10^{-90}$ ( $\chi^{2}$ )	$<10^{-90}$ (1-sided $z$ )
musique_1	63	80.0% [37.6, 96.4]	0.0% [0.0, 6.2]	$\infty$	$<10^{-6}$	$<10^{-6}$
musique_2	35	80.0% [49.0, 94.3]	12.0% [4.2, 30.0]	29.33 [4.12, 209.01]	0.0002553	0.0002553
musique_4	18	20.0% [5.7, 51.0]	37.5% [13.7, 69.4]	0.42 [0.05, 3.43]	0.6078	0.9118
musique_6	48	80.0% [49.0, 94.3]	7.9% [2.7, 20.8]	46.67 [6.66, 327.05]	$<10^{-5}$	$<10^{-5}$
musique_8	121	33.3% [9.7, 70.0]	0.0% [0.0, 3.2]	$\infty$	0.002066	0.002066

Appendix

Table 6: C8 verdict-flip analysis. The 2

\times

2 cross-tabulates whether the prior evidence profile pointed in the direction the agent eventually flipped to (‘profile-directed’) vs whether the post-flip verdict matched the gold label (‘flip correct’). Aggregated

\chi^{2}

test (top row) is two-sided; per-cluster Fisher’s exact tests are reported two-sided (

p_{F}

) and one-sided alternative ‘greater’ (

p_{F}^{>}

); odds ratios with 95% CIs (Woolf log-method, undefined when a cell is 0); conditional flip-correctness rates with Wilson 95% CIs. Independence caveat :

\chi^{2}

assumes independent observations; flipped passages within the same question share retrieval pool and reasoning context (within-question clustering), giving a design-effect of roughly 1.2–5; the headline aggregated effect is robust under all reasonable DE values (adjusted

\chi^{2}\geq 90

Figure 6: Sliding-window accuracy (window = 25 questions) on cluster musique_18 (

n{=}100

), the cluster with the largest paired Acc gap. ours_full (blue) tracks above vanilla_rag (red dashed) as the mean evidence-profile coverage in the window (green, right axis) accumulates. The visualization shows that the gain emerges as graph-derived profiles develop, rather than being uniformly present from the cluster’s first question.

Configuration	Vanilla RAG	Ours-RG	Ours-Full
GPT-5-mini, MuSiQue	69.6	73.2 ∗∗∗	72.5 ∗∗
GPT-5-mini, HotpotQA	72.0	n/a	70.8
Sonnet (seed 13), MuSiQue	66.4	n/a	69.0

Appendix

Table 7: Cross-vendor replication: mean accuracy (%) on MuSiQue (

N{=}1{,}100

, 11 clusters) and HotpotQA (

N{=}500

, 5 clusters) at ordering seed 13. The Sonnet row is at seed 13 (apples-to-apples with GPT-5-mini); the seed-13 Sonnet

\Delta

matches the per-seed value disclosed in Table 1 ’s footnote, distinct from the headline seed-42 number. Significance: paired McNemar two-sided test vs. Vanilla RAG;

{}^{**}p<0.01

{}^{***}p<0.001

Configuration	EM	$\Delta$ vs. VR	McNemar $p$
Vanilla RAG	67.3% $\,\pm\,$ 13.7	n/a	n/a
Reflexion	68.3% $\,\pm\,$ 12.3	$+1.0$ pp	$0.29$
ReasoningBank	65.0% $\,\pm\,$ 13.8	$-2.3$ pp	$0.0018$
Ours-RG	67.3% $\,\pm\,$ 13.6	$+0.0$ pp	$1.00$
Ours-Full	67.6% $\,\pm\,$ 13.2	$+0.3$ pp	$0.77$ (null)

Appendix

Table 8: 2WikiMultiHopQA results across five configurations (

N{=}1{,}000

paired, 10 clusters of 100). McNemar

p

is paired two-sided vs. Vanilla RAG. The Ours-Full vs. VR null (

+0.3

pp,

p{=}0.77

) is the headline confirmation of the gold-passage reuse mechanism (Section 5.3 ); see also the corresponding row in Table 10 .

Cluster	VR EM	Ours-Full EM	$\Delta$	Gold-reuse	Jaccard
twowiki_0	96.0%	95.0%	$-0.01$	0.000	0.372
twowiki_2	95.0%	94.0%	$-0.01$	0.043	0.497
twowiki_3	44.0%	43.0%	$-0.01$	0.050	0.659
twowiki_4	75.0%	76.0%	$+0.01$	0.025	0.567
twowiki_5	53.0%	50.0%	$-0.03$	0.025	0.571
twowiki_6	67.0%	68.0%	$+0.01$	0.010	0.432

Appendix

Table 9: Per-cluster 2WikiMultiHopQA breakdown. “Gold-reuse” is the mean fraction of answer-critical passages that appeared as gold in any prior question within the cluster; “Jaccard” is the mean pairwise candidate-pool Jaccard overlap within the cluster. Every cluster pairs high Jaccard with near-zero gold-reuse, and every

\Delta

is within

\pm 0.05

Finding	Cohen’s $d_{z}$	$p$ -value	Category
VVIR Ours-RG (Haiku, $N{=}73$ , $T{=}0$ )	$+10.1$ pp	$1.1\!\times\!10^{-4}$	Determinism
VVIR Ours-RG (Sonnet, $N{=}73$ , $T{=}0$ )	$+8.0$ pp	$2.9\!\times\!10^{-3}$	Determinism
4-hop F1	0.365 (medium)	$5.6\times 10^{-7}$	F1
Highreuse 3-hop F1	0.241 (small)	$4.5\times 10^{-4}$	F1
Highreuse Haiku F1	0.238 (small)	$3.8\times 10^{-6}$	F1
Highreuse Sonnet F1	0.105 (small)	0.039	F1

Appendix

Table 10: Effect size summary. Determinism is reported as raw VVIR improvement (Wilcoxon test); accuracy/F1 effects as paired Cohen’s

d_{z}=\mathrm{mean(diff)}/\mathrm{std(diff)}

, the appropriate effect size for paired-sample tests. F1 rows use token-level F1; the combined-main row reports raw accuracy (correctness rate) on the paired N=1,600 union of MuSiQue and HotpotQA. The 2Wiki-accuracy row is computed separately on the paired

N{=}1{,}000

union of 10 2WikiMultiHopQA clusters (Ours-Full vs. Vanilla RAG), reported as a pre-specified null prediction from the gold-passage reuse mechanism (Section 5.3 ).

Coverage	$n$	VR	+RG only	+Full	RG share	RetG share
0%	371	54.4%	$-$ 1.1pp	$-$ 0.3pp	n/a	n/a
(0, 20)%	258	60.1%	+0.0pp	+1.9pp	0%	100%
20–49%	208	71.6%	+1.9pp	+3.4pp	57%	43%
50%+	263	77.6%	+6.8pp	+10.6pp	64%	36%

Appendix

Table 11: Component decomposition by coverage. Reasoning graph provides 64% of improvement at high coverage; retrieval graph adds 36% via pruning.

Figure 7: Component decomposition by coverage level. Left : absolute accuracy showing VR baseline (gray) with reasoning graph (blue) and retrieval graph (orange) contributions stacked on top. Right : share of improvement attributable to each component. The retrieval graph dominates at low coverage (early pruning); the reasoning graph becomes the primary driver (64%) at high coverage.

Figure 8: Dose-response visualization complementing Table 2 . At zero coverage, Ours-Full and Vanilla RAG are statistically tied (

\sim

54%). The gap widens monotonically with coverage, reaching

+10.6

pp at 50%+ coverage (47% error reduction,

p<0.001

Figure 9: Decision consistency (

D_{\tau}

) tracking the mechanism in operation. Left : accuracy increases monotonically with

D_{\tau}

bin. Right :

D_{\tau}

rises across the question sequence as profiles accumulate signal. Combined with the paired coverage dose-response (Table 2 ) and the verdict-flip decomposition above, this provides a third converging line of evidence that profile-driven verdict alignment drives the accuracy gain.

Condition	Ours	VR (same)	$\Delta$	Err. Red.
4-hop + low ( $<$ 20%) cov	38.2%	36.8%	+1.5pp	n/a
4-hop + med (20–49%) cov	74.0%	58.0%	+16.0pp	38%
4-hop + high (50%+) cov	89.2%	73.5%	+15.7pp	59%
3-hop + high (50%+) cov	94.1%	83.5%	+10.6pp	64%
2-hop + high (50%+) cov	82.1%	75.8%	+6.3pp	26%

Appendix

Table 12: Coverage

\times

hop depth interaction. At high coverage, 4-hop accuracy (89.2%) surpasses the Vanilla RAG baseline on 2-hop questions (67.2%). McNemar

p=0.004

(15:2 wins) on 4-hop

+

50%+ coverage.

Model	$T$	VR VVIR	Ours-RG	$\Delta_{\text{RG}}$	Ours-Full	$\Delta_{\text{Full}}$	RG perf.	Full perf.
Haiku	0.0	0.650	0.863	+0.213	0.953	+0.303	20/38	31/38
Haiku	0.5	0.561	0.779	+0.218	0.882	+0.321	16/38	24/38
Haiku	0.7	0.558	0.758	+0.200	0.868	+0.311	13/38	24/38
Sonnet	0.0	0.557	0.883	+0.326	0.987	+0.430	16/23	22/23
Sonnet	0.5	0.561	0.830	+0.270	0.922	+0.361	14/23	15/23
Sonnet	0.7	0.583	0.826	+0.244	0.883	+0.300	12/23	13/23

Appendix

Table 13: Hard-stratum VVIR drilldown: probes with Vanilla RAG VVIR

<0.9

T=0

(Haiku

n=38/73

, Sonnet

n=23/73

). Same

K

, checkpoint, and pairing as Table 4 . Perfect columns count probes with VVIR

=1.0

(all 10 runs identical) under each condition; the corresponding Vanilla RAG perfect counts are zero or one across all six cells. All reported

\Delta_{\text{Full}}

significant at

p<10^{-4}

by one-sided Wilcoxon signed-rank test.

Method	Haiku VVIR	Headroom	Sonnet VVIR	Headroom
Vanilla RAG	0.650	–	0.557	–
Reflexion	0.766 ( $p{=}0.006$ )	33%	0.748 ( $p{=}5{\times}10^{-4}$ )	43%
ReasoningBank	0.816 ( $p{=}5{\times}10^{-4}$ )	47%	0.770 ( $p{=}8{\times}10^{-4}$ )	48%
Ours-RG	0.863	61%	0.883	74%
Ours-Full	0.953	87%	0.987	97%

Appendix

Table 14: Hard-stratum VVIR at

T{=}0

across all memory baselines on the paired determinism probe set (Haiku

n{=}38

, Sonnet

n{=}23

K{=}10

runs each). Headroom captured

=\Delta/(1{-}\text{VR}_{\text{VVIR}})

. Both query-centric baselines (Reflexion, ReasoningBank) significantly reduce VVIR, but evidence-centric memory captures roughly twice as much of the available headroom on Haiku and roughly

2\times

on Sonnet.

p

-values are paired one-sided Wilcoxon signed-rank vs. Vanilla RAG. Ours-RG and Ours-Full

p<10^{-4}

throughout (omitted).

Panel of Figure 10 — Figure 10: Reasoning graph anatomy. Forward (left): from an agent through decided/evaluated edges, reconstruct one run’s chain of thought. Backward (right): from an evidence item inward, aggregate evaluations across runs to surface cross-run patterns. Backward traversal enables evidence-centric feedback; flat strategy stores cannot support it.

Figure 11: Conditional dilution of the headline effect. Each row shows

\Delta

accuracy (Ours-Full

-

Vanilla RAG) for one stratum or dataset, with paired 95% confidence intervals. MuSiQue strata use McNemar discordant-pair CIs; cross-dataset rows use cluster-paired CIs from Table 1 . The full-sample MuSiQue effect (

+3.55

pp) is a

4\times

underestimate of the conditional effect on triple-aligned questions (

+13.9

pp). HotpotQA and 2Wiki sit at zero, consistent with the gold-passage reuse mechanism.

Hops	$n$	Ours-Full	VR	$\Delta$	McNemar
2-hop	646	68.9%	67.2%	+1.7pp	n.s.
3-hop	253	66.0%	63.6%	+2.4pp	n.s.
4-hop	201	68.2%	57.2%	+11.0pp	$\bm{p=0.0001}$

Appendix

Table 15: Accuracy by reasoning depth. Improvement scales monotonically with hop count.

Figure 12: Per-cluster gold-passage reuse vs. accuracy improvement (Ours-Full

-

Vanilla RAG);

n=26

clusters across MuSiQue, HotpotQA, 2WikiMultiHopQA. High-reuse MuSiQue clusters show substantial improvements; near-zero-reuse HotpotQA and 2Wiki clusters cluster near zero. Pearson

r=0.604

, 95% CI

[0.336,0.801]

(10,000-resample bootstrap). Cited in Section 5.3 .

Symbol	Value	Description
$T_{\max}$	50	Max evaluations per evidence item before sampling
$N$	20	Number of recent evaluations to sample when $\|E_{\mathrm{in}}\|>T_{\max}$
$B$	2000	Token budget for injected evidence profiles
$R_{\text{thresh}}$	0.7	Rejection rate threshold for exclusion list
min_support	3	Minimum observations before optimizing filter
$k$ (MuSiQue)	25	Number of clusters for query type classification

Appendix

Table 16: Hyperparameters and their values.

Seed	Ours-Full	VR @200	$\Delta$ @200	Ours-Full F1
42	56.6% ( $n{=}387$ )	51.0%	+5.6pp	0.633
43	57.4% ( $n{=}387$ )	46.5%	+10.9pp	0.632
44	58.7% ( $n{=}387$ )	53.0%	+5.7pp	0.638
Mean	57.5% (std 0.86pp)	50.2% (std 2.74pp)	+7.4pp	0.634 (std 0.003)

Appendix

Table 17: Ordering robustness: Ours-Full and Vanilla RAG accuracy across three ordering seeds on the high-reuse dataset (Haiku). Ours-Full converges to a stable accuracy regardless of question order (range 56.6–58.7%, std 0.86pp at

n{=}387

). The delta vs. Vanilla RAG at position 200 is consistently positive across all orderings.

Noise Rate	Accuracy	$\Delta$ vs. VR	$\Delta$ vs. Clean	Flips
0% (clean)	56.6%	+7.5pp	n/a	0
10% (3 seeds)	56.4% $\pm$ 3.5pp	+7.3pp	$-$ 0.2pp	42
20%	52.5%	+3.4pp	$-$ 4.1pp	71
30%	52.7%	+3.6pp	$-$ 3.9pp	118

Appendix

Table 18: Noise robustness: Ours-Full accuracy under outcome label noise on the high-reuse dataset (Haiku). The 10% row reports mean

\pm

std across 3 noise seeds. Even at 30% label noise, the system remains above the current Vanilla RAG baseline (49.1%, Table 3 ).

Explore similar work

Aug 26, 2026

LivingRAG: Augmenting Graph RAG with Experience

Graph-based RAG improves multi-hop question answering by organizing evidence as a knowledge graph. However, most existing RAG systems process each query in isolation and discard useful reasoning from the LLM's response after inference. As a result, later related queries need to retrieve evidence and reason from scratch. We propose LivingRAG, a Graph RAG framework with writable and reusable reasoning experience. LivingRAG adds a writable experience store to a graph-based retrieval backbone, enabling verified experiences to be reused during inference in two ways. Stored graph signals help retrieval find entities and passages that were useful in earlier related queries. Stored summaries provide a reference reasoning pattern for answer generation. We analyze online QA streams and find reusable signals from shared entities, graph neighborhoods, and question templates. Experiments on multi-hop QA benchmarks show that LivingRAG improves accuracy over strong RAG baselines and reduces completion-token use when relevant prior experience is reused.

Yuzhuo Cui, Zongye Zhang, Qingjie Liu

State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China · Hangzhou Innovation Institute, Beihang University, Hangzhou, Zhejiang, China

Retrieval-Augmented Generation Graph Retrieval-Augmented Generation Multi-Hop QA Knowledge Graph Question Answering

Ask PDF

Mar 14, 2026

The Reasoning Bottleneck in Graph-RAG: Structured Prompting and Context Compression for Multi-Hop QA

Graph-RAG systems achieve strong multi-hop question answering by indexing documents into knowledge graphs, but strong retrieval does not guarantee strong answers. Evaluating KET-RAG, a leading Graph-RAG system, on three multi-hop QA benchmarks (HotpotQA, MuSiQue, 2WikiMultiHopQA), we find that 77% to 91% of questions have the gold answer in the retrieved context, yet accuracy is only 35% to 78%, and 73% to 84% of errors are reasoning failures. We propose two augmentations: (i) SPARQL chain-of-thought prompting, which decomposes questions into triple-pattern queries aligned with the entity-relationship context, and (ii) graph-walk compression, which compresses the context by ~60% via knowledge-graph traversal with no LLM calls. SPARQL CoT improves accuracy by +2 to +14 pp; graph-walk compression adds +6 pp on average when paired with structured prompting on smaller models. Surprisingly, we show that, with question-type routing, a fully augmented budget open-weight Llama-8B model matches or exceeds the unaugmented Llama-70B baseline on all three benchmarks at ~12x lower cost. A replication on LightRAG confirms that our augmentations transfer across Graph-RAG systems.

Yasaman Zarrinkia, Venkatesh Srinivasan, Alex Thomo

University of Victoria · Santa Clara University

Graph Retrieval-Augmented Generation Multi-Hop QA LLM Reasoning with Graphs CoT Reasoning

Ask PDF

Sep 10, 2026

MOSAIC: Query-Aware Exploration Policy Adaptation for GraphRAG

Graph Retrieval-Augmented Generation (GraphRAG) can connect evidence distributed across a corpus graph, but most systems use largely shared exploration procedures across queries. This creates a structural mismatch: direct facts may need compact local neighborhoods, comparisons need balanced coverage of multiple targets, and mediated questions may require deeper paths through weakly related connectors. We present Mosaic, a training-free framework that formulates GraphRAG retrieval as a per-query control problem. An LLM analyzer converts query-specific evidence requirements into a bounded policy over seed selection, graph traversal, stopping, and evidence selection, while the corpus graph, indexes, scoring functions, grounding procedure, and answer generator remain shared. On GraphRAG-Bench, Mosaic achieves query-weighted Answer Correctness of 76.97 on Medical and 64.33 on Novel, improving over the strongest previously reported overall results by 5.13 and 4.43 points. On Medical, it reaches 95.1 Evidence Recall and 86.1 Context Relevancy. Controlled comparisons on an identical graph and generator show that no fixed narrow, medium, or wide policy is consistently optimal; Mosaic improves by 9.96 points over the strongest canonical fixed policy. Relative to Fixed Wide, it evaluates 81.9% fewer paths and retains 47.2% fewer evidence items. Transfer experiments on HotpotQA, MuSiQue, and 2WikiMultiHopQA further show that the policy interface can be applied without benchmark-specific retriever training.

EunKyeong Lee, Kyeong-Jin Oh, Jinwon Kim +4

KT Corporation

Multi-Hop QA Graph Retrieval-Augmented Generation Graph-Based Retrieval

Ask PDF