Grounded Revision vs. Prior Injection: Probing Retrieval-Augmented Patent Claim Amendment

Josepha Michiko Leo; Hyun-seok Min; Yehoon Jang; Irvan Zidny; Jin-Woo Chung; Sungchul Choi

cs.CLSep 29, 2026

Grounded Revision vs. Prior Injection: Probing Retrieval-Augmented Patent Claim Amendment

Authors: Josepha Michiko Leo, Hyun-seok Min, Yehoon Jang, Irvan Zidny, Jin-Woo Chung, Sungchul Choi

Organizations: Major in Industrial Data Science & Engineering, Department of Industrial and Data Engineering, Pukyong National University · Tomocube Inc. · Connectionary · Teamreboott Inc.

Abstract

Retrieval-augmented generation is widely used in professional writing, yet whether retrieval grounds revision or merely injects templates is rarely tested where "correct" has a definable meaning. Patent claim amendment supplies that signal: the examiner names the attacked limitation and cites prior art, providing per-case ground truth. We release three artifacts: (i) a corpus of 7,385 USPTO prosecution cases with XML-aligned pre/post claims, rejection, and cited prior art; (ii) a seven-probe battery comparing random and structural-match retrieval as two policies under a fixed prompt scaffold; (iii) a deterministic five-channel metric (C1-C3 and C5 in main, C4 supplementary) requiring no LLM evaluation. Across 9,600 pre-registered calls on four frontier LLMs (Claude Sonnet 4, Claude Haiku 4.5, GPT-5.4, GPT-4o-mini), no tested model exhibits detectable classical prior-injection behavior; retrieval effects are small and direction-inconsistent between random and structural retrieval, and the null is unchanged under a dense (semantic) retriever, across retrieval depths k in {1,3,5,10}, and under a paraphrase-sensitive grounding metric. Revision locality reveals a model-specific difference that the template channel misses. The four-cell taxonomy, which we treat as exploratory, leaves the prior-injector cell unoccupied.

Figures & tables

Figure 1: Does the model address the examiner-attacked limitation (top, grounded revision) or recycle generic patent phrasing (bottom, prior injection)? Our probe battery (§ 4.1 ) and metric (§ 4.2 ) distinguish the two without an LLM judge.

Model	base	$\Delta$ F	$\Delta$ G	H1	H2	H3
Claude Sonnet 4	3.99	$-0.16$	$-0.13$	N	Y	N
Claude Haiku 4.5	3.40	$+0.19$	$+0.03$	N	N	Y
GPT-5.4	3.73	$-0.12$	$+0.06$	N	N	N
GPT-4o-mini	3.14	$+0.20$	$-0.27$	$\sim$	N	N

Table 1: Pre-registered hypothesis verdicts on C5 (canonical-phrase rate per 1,000 characters,

N{=}300

per cell = 100 cases

\times

3 replicates). Y: supported; N: not supported;

\sim

: at-boundary, see § 6.1 for the boundary-handling rule.

Model	base C2	$\Delta$ F	$\Delta$ G
Claude Sonnet 4	0.617	$-0.074$	$-0.013$
Claude Haiku 4.5	1.147	$-0.174$	$-0.189$
GPT-5.4	0.789	$+0.016$	$+0.058$
GPT-4o-mini	1.459	$-0.044$	$-0.076$

Table 2: Revision-locality (C2) effects under random (F) and structural (G) retrieval, per model.

Figure 2: Four retrieval-mechanism cells in

(\Delta_{F},\Delta_{G})

space on C5. Coordinates from Table 1 : Sonnet 4

(-0.16,-0.13)

anchor-grounder; GPT-5.4

(-0.12,+0.06)

and GPT-4o-mini

(+0.20,-0.27)

structural-matcher (opposite-sign F vs G); Haiku 4.5

(+0.19,+0.03)

inert / mildly prior-injecting. The prior-injector cell (both

\Delta>+0.2

) is unoccupied across all four models tested in this four-model, one-corpus,

k{=}3

, deterministic-structural-matcher configuration.

	C1	C2	C3	C5
C1	—	$+0.04$	$-0.21$	$-0.05$
C2	$+0.07$	—	$-0.14$	$-0.04$
C3	$-0.17$	$-0.29$	—	$+0.06$
C5	$-0.01$	$+0.03$	$+0.04$	—

Table 3: Channel correlations pooled across 3,062 (case, condition, model) triples. Pearson above the diagonal, Spearman below.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Sonnet 4 (base: C1 .009, C2 .617, C3 .887, C5 3.99)
cond	$\Delta$ C1	$\Delta$ C2	$\Delta$ C3	$\Delta$ C5
A	$+0.004$	$-0.083$	$-0.005$	$+0.01$
B	$+0.001$	$-0.038$	$-0.018$	$+0.08$
C	$-0.001$	$+0.001$	$-0.006$	$-0.09$
D	$+0.001$	$+0.047$	$-0.026$	$+0.03$
E	$+0.001$	$-0.088$	$-0.002$	$-0.30$

Appendix

Table 4: Per-model probe

\times

channel deltas. Conditions A–E are observational; F and G are retrieval interventions. Leading zeros omitted in baselines for compactness.

(probe, channel)	predicted	hit-rate	$n$
B, C1	FLAT	$0.95$	$382$
B, C3	FLAT	$0.64$	$387$
F, C5	UP	$0.50$	$383$
G, C5	UP	$0.47$	$384$
D, C5	UP	$0.42$	$392$
$C_{\text{mechtrue}}$ , C1	UP	$0.40$	$178$

Appendix

Table 5: C4 directional hit-rate per (probe, channel) pair, pooled across four models and three replicates on the 100-case cohort. FLAT predictions use a

|\Delta|<0.05

tolerance.

Model	med $\Delta$ F	$p_{F}$	med $\Delta$ G	$p_{G}$
Claude Sonnet 4	$+0.00$	0.52	$+0.00$	0.78
Claude Haiku 4.5	$+0.07$	0.11	$+0.01$	0.33
GPT-5.4	$+0.00$	0.14	$+0.01$	0.30
GPT-4o-mini	$+0.00$	0.87	$+0.00$	0.78

Appendix

Table 6: C5 baseline-versus-retrieval tests. No comparison is significant uncorrected (0 of 8), and none survives Holm correction. Median-based TOST establishes equivalence to zero within the

\pm 0.2

margin for every model under both probes.

Condition	mean cosine	struct. (0–3)	statute match
random-F	0.587	0.746	31.7%
structural-G	0.607	2.342	90.0%
dense-G	0.871	1.173	55.3%

Appendix

Table 7: Retrieval quality, F vs G vs dense-G, over each condition’s actual picks (300 per condition).

Model	$k{=}1$	$k{=}3$	$k{=}5$	$k{=}10$
Sonnet 4	4.90/4.94	4.94/4.94	4.94/4.92	5.00/4.94
Haiku 4.5	4.15/4.29	4.29/4.27	4.11/4.37	4.32/4.41
GPT-5.4	4.93/4.85	4.94/4.83	4.93/4.88	4.85/4.92
GPT-4o-mini	4.12/3.95	4.27/4.06	4.05/4.10	4.04/3.94

Appendix

Table 8: C5 median by retrieval depth, random-F

|

structural-G.

Retrieval-Augmented Generation

Retrieval-augmented generation connects language models to external knowledge so answers can use current, domain-specific evidence.

See all 796 papers

Return or Revise? Learning When Revision Helps Retrieval-Augmented QA
Sep 24, 2026Nicholas Kashani Motlagh, Tim Anderson, Jeremy Gwinnup +1Retrieval-Augmented Generation LLM Answer Verification
DCS Corp · Air Force Research Laboratory
PAVE: Premise-Aware Validation and Editing for Retrieval-Augmented LLMs
Mar 21, 2026Tianyi Huang, Caden Yang, Emily Yin +2Evidence-Grounded QA LLM Answer Verification
Ryquo · App-In Club
Not All Queries Need Rewriting: When Prompt-Only LLM Refinement Helps and Hurts Dense Retrieval
Mar 2, 2026Varun KotteQuery Reformulation Dense Retrieval
Adobe

Sep 24, 2026J/K move · Enter open · S save

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

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

DCS Corp · Air Force Research Laboratory

Retrieval-Augmented Generation LLM Answer Verification Adaptive RAG RAG Evaluation Knowledge-Intensive QA

Open paper AskPDF arXiv

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.

Abstract

Retrieval-Augmented Generation

Explore similar work

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

PAVE: Premise-Aware Validation and Editing for Retrieval-Augmented LLMs

Not All Queries Need Rewriting: When Prompt-Only LLM Refinement Helps and Hurts Dense Retrieval