Return or Revise? Learning When Revision Helps Retrieval-Augmented QA

Nicholas Kashani Motlagh; Tim Anderson; Jeremy Gwinnup; Grant Erdmann

cs.CLSep 24, 2026

Return or Revise? Learning When Revision Helps Retrieval-Augmented QA

Authors: Nicholas Kashani Motlagh, Tim Anderson, Jeremy Gwinnup, Grant Erdmann

Organizations: DCS Corp · Air Force Research Laboratory

Abstract

We consider the decision of whether to return an existing draft answer or revise it using retrieved evidence, as in answer-revision systems. Draft confidence estimates whether the current answer is correct, but the decision requires estimating the effect of a specified revision. For offline training and evaluation, we grade both the returned draft and its candidate revision under the same correctness judge, which makes repair, harm, and the gap to an oracle observable. We call this paired effect its recoverability, and we train policies to predict it before revision. On 25,870 held-out open-domain questions across three revision setups, a scorer trained on the paired outcome has greater area under the accuracy--revision-rate curve than a matched draft-correctness scorer in all nine Llama setup--seed fits, and gains 0.23--0.68 accuracy points on average at development-selected thresholds, a difference significant across training runs only for dense retrieval. The resulting policy improves on always revising and on average closes more than a third of the oracle gap, although it still applies 38--46% of the harmful revisions. When a draft-free standard-RAG answer is also available, however, choosing between the draft and that answer is stronger by about two points for Llama and four for OLMo, and adding candidate revision as a third option yields no significant gain. Recoverability describes one revision; its value as an available action also depends on the alternatives.

Figures & tables

Panel of Figure 1 — Figure 1: The revision decision is defined by paired outcomes. Evaluating both the draft and its candidate revision places an example in one of four cells. The diagonal cells are ties. Only the shaded off-diagonal cells distinguish returning from revising. Draft correctness alone identifies only a row.

REPAIR NQ-Open, DPR

Who wrote the song Going to Kansas City ?

Charlie Christian

\rightarrow

Jerry Leiber and Mike Stoller

[1] Kansas City (Leiber and Stoller song) …“Kansas City” is a rhythm and blues song written by Jerry Leiber and Mike Stoller in 1952.

HARM TriviaQA, DPR

Who was the second wife of Henry VIII?

Anne Boleyn

\rightarrow

Anne of Cleves

Figure 2: Revision repairs one draft and corrupts another. Two held-out cases, each giving the question, draft

\rightarrow

revision, and excerpts of the top retrieved passage. In the repair case, the passage states the correct answer and the refiner adopts it. In the harm case, the passage confirms the draft, but the refiner switches to another name from the same passage.

Panel B: Policy behavior
	Revision rate		Repairs revised		Harms revised		Macro-averaged gain
Revision setup	(% of examples)		(%) $\uparrow$		(%) $\downarrow$		(points) $\uparrow$
DPR	38.8 $\pm$	5.1	96.0 $\pm$	1.1	46.4 $\pm$	8.9	$+$ 1.16 $\pm$	0.16 †
BM25	59.5 $\pm$	13.3	94.6 $\pm$	0.4	43.0 $\pm$	5.3	$+$ 1.34 $\pm$	0.21 †
BM25 $\rightarrow$ MonoT5	35.5 $\pm$	7.7	94.0 $\pm$	0.6	38.0 $\pm$	4.2	$+$ 1.10 $\pm$	0.11 †

Table 1: Measured recoverability and learned-policy results. Panel A: repair and harm are the examples that revision changes from incorrect to correct and from correct to incorrect. Accuracy is reported for always returning the draft, always revising, the development-selected full-prompt LoRA policy, and the return-or-revise oracle. Gain is the policy’s accuracy minus always revising. Gap closed divides this gain by the oracle gap. Panel B: the fractions of all examples, of repairs, and of harms that the policy revises, and the macro-averaged gain, which is the gain over always revising when NQ-Open, TriviaQA, and PopQA are weighted equally rather than by example count. Policy cells are the mean

\pm

SD over three training seeds.

\dagger

marks the run-level test of Section 4 .

Figure 3: Paired-outcome minus draft-correctness accuracy at matched revision rates for the three revision setups. Lines are training seeds, and the gray band is a 95% bootstrap interval for the seed mean. Each area label gives the mean

\pm

SD of the integrated difference, in accuracy points.

Revision setup	Draft-correctness		Paired-outcome		(points)		Revision rate (%)		Harms revised (%) $\downarrow$
	Policy accuracy (%) $\uparrow$				Paired $-$ draft $\uparrow$		Draft-correctness policy
DPR	54.23 $\pm$	0.06	54.91 $\pm$	0.18	$+$ 0.68 $\pm$	0.18 †	57.8 $\pm$	2.0	74.1 $\pm$	3.6
BM25	55.74 $\pm$	0.03	55.96 $\pm$	0.22	$+$ 0.23 $\pm$	0.24	48.5 $\pm$	1.9	45.3 $\pm$	3.9
BM25 $\rightarrow$ MonoT5	58.51 $\pm$	0.05	58.84 $\pm$	0.10	$+$ 0.33 $\pm$	0.14	52.0 $\pm$	2.9	58.0 $\pm$	7.3

Table 2: Matched paired-outcome and draft-correctness methods. Cells are the mean

\pm

SD over three training seeds.

\dagger

marks the run-level test of Section 4 . The paired-outcome column reuses the Table 1 policies.

Panel of Figure 4 — Figure 4: Which information supports prediction. Gap closed for three inputs within each model class and revision setup (one training run). Whiskers are 95% bootstrap intervals. The ring marks the full-prompt LoRA policy of Table 1 in this run. Dotted lines separate final-state models, the token-state model, and the fine-tuned model. The paired oracle, at 100%, is off scale.

Generator/refiner	Revision setup	gain $\uparrow$		gain $\uparrow$
		Primary judge:		Second judge:
Llama 3.1 8B Instruct	DPR	$+$ 1.26 $\pm$	0.18 †	$+$ 1.37 $\pm$	0.19 †
	BM25	$+$ 1.33 $\pm$	0.22 †	$+$ 1.34 $\pm$	0.21 †
	BM25 $\rightarrow$ MonoT5	$+$ 1.10 $\pm$	0.10 †	$+$ 1.05 $\pm$	0.10 †
GPT-OSS-20B	DPR	$+$ 0.12 $\pm$	0.03 †	$+$ 0.05 $\pm$	0.03
	BM25	$+$ 0.17 $\pm$	0.03 †	$+$ 0.06 $\pm$	0.03

Table 3: Policy gains across generator/refiner families and judges. Each entry is the mean gain in points over always revising

\pm

SD across three training seeds.

\dagger

marks the run-level test of Section 4 . Appendix Table 13 gives the oracle gaps and gap closed.

Panel B: Seed-matched accuracy contrasts (points)
	{Return,RAG}		{Return,RAG}		{Return,Revise,RAG}
Revision setup	$-$ paired policy		$-$ {Return,Revise}		$-$ {Return,RAG}
DPR	$+$ 2.17 $\pm$	0.08 †	$+$ 2.05 $\pm$	0.09 †	$-$ 0.15 $\pm$	0.16
BM25	$+$ 2.37 $\pm$	0.16 †	$+$ 1.95 $\pm$	0.06 †	$+$ 0.00 $\pm$	0.08
BM25 $\rightarrow$ MonoT5	$+$ 2.30 $\pm$	0.04 †	$+$ 2.10 $\pm$	0.06 †	$-$ 0.12 $\pm$	0.11

Table 4: The available answers determine value. Test accuracy (

n{=}25{,}870

per setup). Return is returning the draft, Revise the candidate revision, and RAG a draft-free standard-RAG answer. Paired policy is the Table 1 policy. Fixed actions are deterministic. Learned cells are the mean

\pm

SD over three training seeds. Panel B uses seed-matched differences.

\dagger

marks the run-level test of Section 4 .

Llama 3.1 8B
	Unique	Captured	Fixes to the	Damaging	Net change in
Revision setup	revision wins	unique wins	binary choice	switches	correct answers
DPR	181	22.7	57.3	117.7	$-$ 37.7
BM25	222	34.3	44.0	77.7	0.7
BM25 $\rightarrow$ MonoT5	323	53.0	77.0	161.7	$-$ 31.7
OLMo 3 7B

Table 5: What changes when revision becomes a third action. Unique wins count examples where only revision is correct. Captured unique wins and fixes to the binary choice are successful switches from the same per-action model’s return-or-RAG choice; fixes to the binary choice occur when an existing action was correct but the binary selector chose the wrong one. Their sum minus damaging switches equals the net change in correct answers. Switch counts are means over three fits; unique wins are fixed. Each setup contains 25,870 examples.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Model class	Information read	Fit and score	Development selection	Trainable params.
Ridge	Final frozen state after the available input	Standardized ridge regression on $y=z_{I}-z_{d}$ ; score is predicted change	Fixed convex fit; accuracy-maximizing threshold	4,097
Linear	Same final frozen state	$4096{\rightarrow}3$ head; unweighted CE; score is $P(\mathrm{repair})-P(\mathrm{harm})$	Lowest $\Delta_{I}$ MSE checkpoint; accuracy-maximizing threshold	12,291
MLP	Same final frozen state	LayerNorm– $4096{\rightarrow}256$ –GELU–dropout– $256{\rightarrow}3$ ; unweighted CE	Same	1,057,795
Attn-pool	All frozen final-layer token states in the available prefix	Learned single-query attention pooling and the same MLP head; unweighted CE	Same	1,061,891
LoRA model	Exact available pre-revision input, at most 4,096 tokens	Rank-16 LoRA on all attention and MLP projections plus a three-way head; unweighted CE	Lowest $\Delta_{I}$ MSE checkpoint; accuracy-maximizing threshold	$\approx$ 41.96M
Matched draft correctness	Complete pre-revision prompt	Same LoRA recipe; two softmax logits; unweighted CE	Lowest draft Brier score; revise below accuracy-selected threshold	$\approx$ 41.96M

Appendix

Table 6: Decision-model recipes. Checkpoints and any policy thresholds are selected on development. Threshold-selection ties favor fewer revisions; three-action logit ties prefer return, standard RAG, then revision.

Retrieved-evidence feature	DPR	BM25	BM25 $\rightarrow$ MonoT5
No passage carries a known gold alias: harm given correct draft	439/2,666 (16.5)	548/3,309 (16.6)	421/1,992 (21.1)
$\geq 3$ passages carry a known gold alias: harm given correct draft	149/5,532 (2.7)	109/4,921 (2.2)	162/6,775 (2.4)
Revision string occurs in evidence: share of harms	676/790 (85.6)	675/862 (78.3)	681/792 (86.0)
Revision in evidence and no known gold alias: share of harms, pooled	1,138/2,439 (46.7)
PopQA lexical-mismatch pattern: harm given correct draft	221/256 (86.3)	227/276 (82.2)	223/255 (87.5)
PopQA lexical-mismatch pattern: coverage of harms	221/411 (53.8)	227/444 (51.1)	223/422 (52.8)

Appendix

Table 7: Retrieved-passage features associated with harmful revision. Entries are count/denominator (percent). The pooled row counts harms whose accepted answers have normalized aliases to check. It sums the three setups, which share the same test questions, so we give no interval for it.

Revision setup	Tian-style accuracy (%)	Tian-style revision rate (%)	(points) $\uparrow$
			LoRA $-$ Tian-style
DPR	53.70	93.8	$+$ 1.22 $\pm$	0.18 †
BM25	54.72	94.1	$+$ 1.25 $\pm$	0.22 †
BM25 $\rightarrow$ MonoT5	57.80	94.4	$+$ 1.04 $\pm$	0.10 †

Appendix

Table 8: Tian-style baseline against the Table 1 policy. The Tian-style baseline is deterministic. The difference is the mean

\pm

SD over three LoRA training seeds, and

\dagger

marks the run-level test of Section 4 .

Target	DPR		BM25		BM25 $\rightarrow$ MonoT5
Repair/harm/tie	54.93 $\pm$	0.14	56.13 $\pm$	0.06	58.92 $\pm$	0.08
Two correctness heads	54.99 $\pm$	0.17	56.29 $\pm$	0.03	59.11 $\pm$	0.06
Scalar utility regression	54.74 $\pm$	0.13	55.77 $\pm$	0.14	58.46 $\pm$	0.13
Tie-aware preference	54.79 $\pm$	0.19	55.55 $\pm$	0.79	58.64 $\pm$	0.13

Appendix

Table 9: Separate matched target-formulation experiment. Accuracy (%), mean

\pm

sample SD across three training seeds on the same 25,870 test questions. All methods include the revision outcome in supervision. These fits are separate from the Table 1 policies, so rows compare within this table.

Revision setup	Model class	Question only	Question + draft	Question + draft + evidence
DPR	Ridge	$-$ 1.0	2.4	7.7
	Linear	$-$ 1.0	$-$ 0.1	4.1
	MLP	0.3	6.8	2.4
	Attn-pool	$-$ 0.5	2.7	2.7
	LoRA	0.1	5.8	35.8
BM25	Ridge	$-$ 0.6	4.5	9.2

Appendix

Table 10: Every cell of Figure 4 . Percentage of the return-or-revise oracle gap closed over always revising for each revision setup, model class, and pre-revision input (one training run).

Input	DPR		BM25		BM25 $\rightarrow$ MonoT5
Full prompt	1.26 $\pm$	0.18	1.33 $\pm$	0.22	1.10 $\pm$	0.10
Draft removed	0.64 $\pm$	0.19 †	0.69 $\pm$	0.49	0.56 $\pm$	0.17 †
Evidence shuffled	0.18 $\pm$	0.04 †	0.25 $\pm$	0.07 †	0.23 $\pm$	0.05 †
Evidence masked	0.18 $\pm$	0.02 †	0.21 $\pm$	0.20 †	0.23 $\pm$	0.01 †

Appendix

Table 11: Gain over always revising, in accuracy points on all test examples, when the full-prompt LoRA policy is retrained with part of its input removed or replaced. Mean

\pm

SD over three training seeds.

\dagger

marks a drop from the full prompt that passes the run-level test of Section 4 .

setup	Dataset	(%) $\uparrow$	(%) $\downarrow$	Return $\uparrow$	Revise $\uparrow$	RAG $\uparrow$	Policy $\uparrow$	Oracle $\uparrow$
Revision		Repair	Harm			Standard
DPR	NQ-Open	15.04	4.18	47.40	58.25	56.65	58.86	62.44
	TriviaQA	7.32	2.85	78.22	82.68	77.24	83.89	85.54
	PopQA	8.99	2.88	30.10	36.22	31.46	37.37	39.10
BM25	NQ-Open	12.16	5.07	47.40	54.49	48.53	56.04	59.56
	TriviaQA	8.82	2.94	78.22	84.10	80.33	85.46	87.04

Appendix

Table 12: Results by dataset and revision setup (one training run). Repair and harm rates, accuracy of the three fixed actions and the policy, and the return-or-revise oracle, all in percent.

		Llama 3.3 70B judge					GPT-OSS-120B judge
Generator/refiner	Revision	Oracle gap	Gain		Gap closed		Oracle gap	Gain		Gap closed
	setup	(points) $\downarrow$	(points) $\uparrow$		(%) $\uparrow$		(points) $\downarrow$	(points) $\uparrow$		(%) $\uparrow$
Llama 3.1 8B Instruct	DPR	3.05	$+$ 1.26 $\pm$	0.18 †	41.4 $\pm$	5.9	3.41	$+$ 1.37 $\pm$	0.19 †	40.2 $\pm$	5.6
	BM25	3.33	$+$ 1.33 $\pm$	0.22 †	40.0 $\pm$	6.5	3.56	$+$ 1.34 $\pm$	0.21 †	37.6 $\pm$	6.0
	BM25 $\rightarrow$ MonoT5	3.06	$+$ 1.10 $\pm$	0.10 †	35.9 $\pm$	3.3	3.44	$+$ 1.05 $\pm$	0.10 †	30.7 $\pm$	2.8
GPT-OSS-20B	DPR	1.24	$+$ 0.12 $\pm$	0.03 †	9.9 $\pm$	2.3	1.67	$+$ 0.05 $\pm$	0.03	2.9 $\pm$	1.6

Appendix

Table 13: Complete operating points behind Table 3 . Gain and gap closed are the mean

\pm

SD over three training seeds.

\dagger

marks the run-level test of Section 4 .

	Primary	Per-answer labels		Second-judge label for each primary outcome ( $n$ )					Retained
Revision setup	outcome	Agree (%)	$\kappa$	Preserved	Repair	Harm	Unrecovered	Total	(%)
DPR	Preserved	96.46	0.929	10,602	227	160	479	11,468	92.4
	Repair	96.59	0.932	20	2,257	5	129	2,411	93.6
	Harm			13	8	684	85	790	86.6
	Unrecovered			58	29	33	11,081	11,201	98.9
BM25	Preserved	96.46	0.929	10,562	238	133	463	11,396	92.7

Appendix

Table 14: Primary and second judge on the same test examples. The per-answer label columns give binary correctness agreement and Cohen’s

\kappa

for the draft ( Preserved line) and the revision ( Repair line). Drafts are shared across setups.

Panel B: Human-labeled paired outcomes (decisive pairs)
		Return	Revise			$\Delta$ [95% CI]
Stratum	$n$	acc. (%)	acc. (%)	Harm (%)	Repair (%)	(points)
Pooled	470	47.9	50.6	5.7	8.5	$+$ 2.8	[ $-$ 0.6, $+$ 6.1]
NQ-Open	157	34.4	42.0	3.8	11.5	$+$ 7.6	[ $+$ 1.9, $+$ 13.8]
TriviaQA	156	80.1	77.6	9.0	6.4	$-$ 2.6	[ $-$ 8.9, $+$ 3.8]
PopQA	157	29.3	32.5	4.5	7.6	$+$ 3.2	[ $-$ 1.9, $+$ 8.4]

Appendix

Table 15: Single-annotator human diagnostic of paired outcomes. Outcome agreement is computed on the three-way paired outcome (Harm, Tie, Repair).

\Delta

is Repair

-

Harm in points with pair-level bootstrap 95% intervals.

Setup / dataset

Fixed

R

Fixed

G

D/R

D/G

D/G/R

D/G-D/R

D/G/R-D/G

Llama 3.1 8B ; always return: 47.38%

DPR / pooled

53.65

49.12

55.03

\pm

0.18

57.08

\pm

0.10

56.93

\pm

0.25

2.05

\pm

0.09

-

0.15

\pm

0.16

NQ-Open

59.21

\pm

0.08

63.04

\pm

0.21

62.93

\pm

0.28

3.83

\pm

0.28 †

-

0.11

\pm

0.11

TriviaQA

84.31

\pm

0.05

86.10

\pm

0.11

86.08

\pm

0.05

1.79

\pm

0.16 †

-

0.02

\pm

0.08

PopQA

37.57

\pm

0.33

39.31

\pm

0.19

39.09

\pm

0.41

1.75

\pm

0.17 †

-

0.23

\pm

0.24

BM25 / pooled

54.63

50.51

56.39

\pm

0.08

58.33

\pm

0.07

58.33

\pm

0.11

1.95

\pm

0.06

0.00

\pm

0.08

Appendix

Table 16: Action-set comparison by family, revision setup, and dataset. Accuracies are percentages; differences are points.

D

returns the draft,

R

revises it, and

G

returns the draft-free standard-RAG answer.

D/R

D/G

, and

D/G/R

are learned policies; each learned cell is mean

\pm

sample SD over three training seeds. Within each family, setup, and seed, all three action sets restrict the same per-action model; answers and labels are fixed. Differences report the mean and SD of the within-seed difference, not a difference of independent estimates. Each setup uses 3,610 NQ-Open, 7,993 TriviaQA, and 14,267 PopQA examples. No policy is retuned by dataset.

\dagger

marks the run-level test of Section 4 for the dataset rows.

Revision setup	Primary judge		Second judge		Second-judge 95% CI
Paired outcome minus draft correctness
DPR	$+$ 0.68 $\pm$	0.18	$+$ 0.75 $\pm$	0.20	[ $+$ 0.612, $+$ 0.880]
BM25	$+$ 0.23 $\pm$	0.24	$+$ 0.38 $\pm$	0.18	[ $+$ 0.215, $+$ 0.555]
BM25 $\rightarrow$ MonoT5	$+$ 0.33 $\pm$	0.14	$+$ 0.39 $\pm$	0.13	[ $+$ 0.240, $+$ 0.537]
Return-or-RAG minus return-or-revise
DPR	$+$ 2.05 $\pm$	0.09	$+$ 2.09 $\pm$	0.08	[ $+$ 1.823, $+$ 2.359]

Appendix

Table 17: Policy contrasts under both judges, in accuracy points. Mean

\pm

SD describes three policy-training seeds. The final column is a paired example-bootstrap interval for the mean of those fixed fits. Every decision, checkpoint, and threshold is unchanged. Both judges retain all 25,870 examples per setup, counting invalid and ambiguous verdicts as incorrect. The matched binary targets and the action-set comparisons are separate experiments.

Revision setup	Neutral $R$	Original oracle	Neutral oracle	Unique wins	Oracle $-$ learned $D/G$
DPR	52.74	56.70	57.22	198	$+$ 0.14	[ $-$ 0.14, $+$ 0.42]
BM25	54.06	57.96	58.38	221	$+$ 0.05	[ $-$ 0.23, $+$ 0.34]
BM25 $\rightarrow$ MonoT5	57.45	60.81	61.21	334	$+$ 0.07	[ $-$ 0.24, $+$ 0.38]

Appendix

Table 18: Prompt sensitivity on the same Llama test examples. Accuracy and oracle cells are percentages. The last column is an accuracy difference in points with a paired 95% example-bootstrap interval for the mean of three fixed policies. Neutral revision has lower standalone accuracy but a higher oracle; the neutral oracle is above the learned seed mean in every setup. No neutral-revision selector was fit.

Explore similar work

Answer Presence Drives RAG Rewriting Gains
Jun 4, 2026Yuejie Li, Yueying Hua, Ke Yang +8Rewrite Answer
Ant Group
Beyond Self-Knowledge: Propagating Uncertainty Across Reasoning and Retrieval in LLMs
Jul 28, 2026Chandan Kumar Sah, Li Zhang, Xiaoli LianConfidence Estimation Uncertainty
Beihang University, Beijing, China
ProvenAI: Provenance-Native Traces of Evidence in Generated Answers
Jun 24, 2026Mohammad Faizan, Dalal AlharthiCitations Traceability
College of Information Science University of Arizona

Jun 4, 2026, cs.AIJ/K move · Enter open · S save

Answer Presence Drives RAG Rewriting Gains

Yuejie Li, Yueying Hua, Ke Yang, Li Zhang+7

Ant Group

Rewrite Answer Random

Open paper AskPDF arXiv

Retrieval-augmented QA pipelines often route retrieved passages through an LLM \emph{rewriter} before a smaller reader, lifting F1 by tens of points on multi-hop benchmarks; this gain is typically credited to improved evidence quality. We ask whether that lift is causally driven by the gold answer string appearing in the rewritten context rather than by curation per se, using a controlled intervention audit. For each rewritten context we re-run the reader after one of four controlled edits to the compile output: removing the gold answer span, replacing a length-matched random non-answer span (placebo), or injecting the gold into rewrites where it was absent (at the prefix or at a midpoint sentence boundary). Across twelve completed (cell, baseline) intervention runs spanning three reader families (Qwen2.5-7B, Qwen3.5-35B, GLM-4.7), two datasets (HotpotQA, 2WikiMultihopQA), and three compiler arrangements (MA-only, MB-only, MA

+

verify), removing the gold answer drops reader F1 by

28

64

points beyond the length-matched placebo on paired \texttt{answer-in-compile} strata, and prepending the gold into rewrites that lacked it raises F1 by

+0.7

+9.7

points in

10

12

(cell, baseline) combinations. A companion five-sentinel audit shows the conventional single-\texttt{[MASK]} probe is itself sentinel-fragile: on 2Wiki it reports a

+4.12

~F1 ``non-leakage residual'' that flips to

-3.33

-7.81

~F1 under four alternative sentinels and fails an equivalence test for three of those four (

1/4

~pass). We do not propose a new rewriter or mitigation; we release the intervention runner and the sentinel panel so that other rewriter-gain claims can be tested against the same standard.

Abstract

Explore similar work

Answer Presence Drives RAG Rewriting Gains

Beyond Self-Knowledge: Propagating Uncertainty Across Reasoning and Retrieval in LLMs

ProvenAI: Provenance-Native Traces of Evidence in Generated Answers