How Should Diffusion Language Models Edit Code?

Xijia Tao; Ziru Liu; Shansan Gong; Jiacheng Ye; Kecheng Chen; Zirui Wu; Lin Zheng; Xinyu Fu

cs.SESep 29, 2026

How Should Diffusion Language Models Edit Code?

Authors: Xijia Tao, Ziru Liu, Shansan Gong, Jiacheng Ye, Kecheng Chen, Zirui Wu, Lin Zheng, Xinyu Fu, +2 more

Organizations: The University of Hong Kong · Huawei Technologies Ltd. · City University of Hong Kong · Peking University

Abstract

Code editing requires a model to decide where to make changes, generate the new content, and preserve everything else. We study how masked diffusion language models divide these responsibilities across four editing interfaces: whole-file rewriting, search-and-replace, locate-then-infill, and token-level editing. Experiments on CanItEdit reveal a composition gap: diffusion models can generate coordinated changes when the correct edit locations are supplied, but much of this capability is lost when those locations must be predicted. Access to the intact original code helps the model fill multiple edit regions, yet does not resolve the difficulty of selecting those regions. By varying the editable regions while holding the generation model and decoding procedure fixed, we identify two distinct requirements for successful editing: covering every required change and placing precise boundaries around it. Missing a required region prevents the corresponding change, while widening regions to ensure coverage can sharply reduce success by requiring unchanged code to be regenerated. A sentence-level Wiki editing probe shows the same qualitative gap between supplied and predicted locations beyond code. These findings show why strong infilling capability alone does not ensure reliable editing: the interface must expose all required changes while limiting regeneration of unchanged code.

Figures & tables

Figure 1: Editing interfaces and the multi-span composition gap. (A) Four interfaces assign different structural decisions to the model and preservation to the runtime. (B) On 142 multi-span CanItEdit instances, gold-region old-visible infilling reaches 25.4% pass@1, compared with 11.3% for direct str_replace and 10.6% when predicted regions are composed with the infiller (Hybrid-multi). Matched oracle repairs of a separate predicted-hole baseline (P) expose complementary coverage and boundary errors. These repairs are diagnostics, not autonomous editors; routing and repair details appear in Appendix L .

Autonomous DiffuCoder-family editors
Interface / structure	Overall $\uparrow$	Single $\uparrow$	Multi $\uparrow$	SynErr $\downarrow$	Exc. $\downarrow$
Search-and-replace ( str_replace )	23.3	48.5	11.3	5.7	68.6
Locate-then-infill	12.9	27.9	5.6	15.2	70.0
Rewrite	11.4	27.9	3.5	23.8	63.3
Token-level editing	5.2	—	—	50.0	—
Structure and conditioning diagnostics

Table 1: CanItEdit results (%). Overall, Single, and Multi report pass@1 (

\uparrow

); SynErr and Exc. are the percentages of evaluations whose harness status is SyntaxError and Exception (

\downarrow

). Exception includes assertion failures and other runtime errors, and excludes SyntaxError and Timeout, so the columns do not sum to 100. Autonomous DiffuCoder-family recipes differ, so their scores are operating points rather than a causal interface ranking. Oracle rows supply edit positions; matched empty and old-visible fillers differ in source conditioning. Dashes indicate unavailable split or exception results.

Figure 3: Known-site generation controls (pass@1, %). Left: original-source context helps the matched filler on multi-span CanItEdit. Right: after matched SFT, middle placement helps on harder, single-gap HumanEval-Infilling tasks with oracle gap length.

Figure 4: Multi-span pass@1 of the fixed old-visible filler under controlled scaffold corruptions. Complete coverage alone is insufficient: expanding boundaries by one line performs worse than direct str_replace (dashed line), whereas a single tight extra hole retains much of the gold-scaffold advantage.

Scaffold	Recall (%)	pass@1 (%)
Clean gold	100	25.4
Drop one required span	66.0	6.3
Expand each range $\pm 1$	100	7.7
Add one false-positive hole	100	20.4
Add two false-positive holes	100	8.5

Table 2: Matched scaffold interventions on the old-visible filler. Recall is mean gold-span coverage on 142 multi-span CanItEdit examples; pass@1 is multi-span.

Interface / locations	Exact	SARI	ROUGE-L	BLEU	Edit F1	Coverage	Preserved
str_replace	0.0	0.411	0.882	0.795	0.061	6.8	97.5
Predicted locations $+$ infill	3.1	0.437	0.807	0.728	0.376	45.3	82.3
Gold locations $+$ infill	9.4	0.597	0.808	0.755	0.927	89.3	95.0

Table 3: FineEdit Wiki targeted editing on 64 grouped-split test examples. SARI, ROUGE-L, BLEU, and edit F1 use the

[0,1]

scale; exact match, coverage, and preservation are percentages. Predicted and gold locations use the same old-visible filler and decoder. Coverage measures gold edits realized in the output; preservation measures source retained outside gold edit regions.

Interface	Text generated	Structure predicted
Rewrite	$U+V$	Implicit in full output
Search-and-replace	$A+V$	Exact source anchors
Locate-then-infill	$H+V$	Regions and per-hole length
Token-level editing	Changed content	Operations across source tokens

Table 4: Generation responsibilities for a fixed edit.

U

is unchanged output text,

V

is replacement/insertion text,

A

is the copied old-anchor text, and

H

is the location-plan text, measured in tokens. Entries describe output content, excluding serialization/control overhead and repeated denoising passes; they are not runtime measurements.

Appendix figures & tables50 assets

Supplementary material from the paper’s appendix.

Appendix

Statistic	Mean	Median	Max
Source tokens	174.6	174	546
Target tokens	204.1	204	730
Instruction tokens	28.0	27	118
Edit distance (tokens)	49.5	26	722
Edit ratio ( $\frac{\text{edit dist}}{\text{source len}}$ )	0.23	0.16	1.00

Appendix

Table 5: EditPackFT Python training set (

N{=}20{,}602

). Token counts use the model tokenizer.

Field	Mean	Std	Min	P25	Med	P75	P90	P95	Max
Source tokens	174.6	84.3	4	107	174	242	288	307	546
Target tokens	204.1	92.0	6	133	204	273	320	346	730
Instruction tokens	28.0	3.9	23	25	27	29	32	35	118
Prompt + Source	202.6	84.5	28	135	202	270	316	336	577
Prompt + Target	232.1	92.2	32	161	232	301	348	374	755

Appendix

Table 6: Token-level sequence length distributions (EditPackFT Python,

N{=}20{,}602

Field	Mean	Std	Min	Med	P75	P95	Max
Source chars	748.0	383.1	10	733	1033	1384	2937
Target chars	874.0	414.3	21	863	1175	1544	2947
Source lines	27.2	13.0	1	27	36	49	94
Target lines	31.1	14.0	1	31	41	55	164
Instruction chars	44.9	17.7	16	41	51	74	444

Appendix

Table 7: Character and line count distributions for source and target code.

Operation	Mean	Std	Min	Med	P75	P95	Max
keep tokens	161.5	81.7	1	160	227	293	379
insert tokens	36.4	57.9	0	15	44	148	689
delete tokens	6.9	22.6	0	0	4	35	525
edit tokens	6.3	14.6	0	2	6	28	279
Edit distance	49.5	63.1	1	26	61	175	722
Edit ratio	0.23	0.21	0.003	0.16	0.33	0.70	1.00

Appendix

Table 8: Token-level edit operation distributions per example.

Change Type	Count	%
Mixed (insert+delete+edit)	14,362	69.7
Insert only	4,984	24.2
Delete only	993	4.8
Edit only (substitution)	263	1.3
Has insertions	17,426	84.6
Has deletions	7,603	36.9

Appendix

Table 9: Distribution of examples by change type.

Statistic	Mean	Std	Min	Med	P75	P95	Max
Groups per example	2.3	2.5	0	2	3	7	31
Tokens per group	15.6	30.2	1	6	15	65	611

Appendix

Table 10: Insert group statistics. An insert group is a contiguous run of insert operations.

#	Family	Supervised state
1	Fill pair	Fully expanded [MASK] slots; parallel fill; keep/edit/delete on source.
2	Expand trigger	Pre-mask source; <\|expand\|> on the last source token before each insert group.
3	Intermediate expand	Random partition of each group; every slot expand vs. fill; $O(\log N)$ doubling.
4	Over-expand recovery	$N{+}1$ slots; extra slot labelled EOS .
5	Partial-fill	Some insert slots already committed; remaining masks filled.
6	Spurious-expand	Bogus masks on non-insert positions, labelled EOS .

Appendix

Table 11: Token-level Levenshtein training families (compressed). Families 1–4 were the original four-type bank; 5–9 add partial-fill, recovery, residual, and no-op coverage.

Interface	Length unit	Other
Stage-1 plans	block-decision, size 32	confidence $0.1$ ; numbered old code in the best runs
Stage-2 fill	block-decision, size 16	confidence $0.3$ ; joint denoising of all holes
Rewrite SFT	fixed 256 or 512 steps	full-file generation; no Stage-1 plan
str_replace (headline)	block-append, step 1685	serialized $(s_{i}\to t_{i})$ ; runtime apply-in-order
str_replace (budget)	256-slot, step 1515	same interface, smaller generation budget
str_replace (pad)	random-pad, step 1685	same recipe as headline except padding

Appendix

Table 12: Decode settings for headline CanItEdit cells.

Role	Step	Init / path
Rewrite SFT	640	DiffuCoder-Instruct; ml-diffucoder/checkpoints/sft_editpackft_baseline_rewrite_8gpu
str_replace SFT	1685	DiffuCoder-Instruct; checkpoints/str_replace_sft_block (block-append)
Shared Stage-1/2 multitask	1935	oracle Stage-2 specialist 1610; checkpoints/stage23_multitask_sft_from_stage3_step1610
Oracle Stage-2 specialist	1610	DiffuCoder-Instruct; checkpoints/stage3_oracle_scaffold_diffucoder_block_decision_ep5

Appendix

Table 13: Checkpoint identities for the roles named in the main text. Decode protocol is in Table 12 ; recipe confounds are in Table 27 .

Attempt	Result	Lesson
Token-level Levenshtein editing	Oracle masks $\sim$ 30–32.5% exact match; model-chosen insert points $\sim$ 0%.	Bottleneck is where to insert, not fill given where.
Header reweight, skeleton weight, header-only targets, multi-hunk oversampling	All failed location gates (exact line-op $>$ 25%; mean edits toward gold 2.78). Header-only raised count but produced out-of-bounds locations.	Upweighting format tokens is not a substitute for learning the discrete structure those tokens encode.
Frozen-backbone dense ScaffoldHead	CanItEdit exact line-op 0.48%; in-domain exact event-tape 6.6%. Stopped: no fill, no unfreeze.	Coarse metrics (nearby anchors, non-zero edits) can look calibrated while the scaffold remains unusable.
OpenCoder 50/50 Stage-2 mix; hole markers; per-token DreamOn expand	Gold-loc / Qwen 38.1% vs. ep5 39.5%. Markers swapped EOS-error types; DreamOn SyntaxError $\sim$ 38%.	More generic code did not raise the fill ceiling. Block-decision was the production length harness because it avoided the worst syntax collapse.
Round-consistent / block-state-balanced Stage-2 sampler	Exact realized length $16.4\%\to 21.5\%$ ; under-generation $71.1\%\to 65.9\%$ ; pass@1 $27.1\%\to 23.8\%$ ; multi-span $25.4\%\to 19.0\%$ . Long 1-hole ( $33$ – $64$ / $65+$ ) unchanged.	Matching gold trajectory states more closely improved length statistics, not functional editing (Table 15 ).

Appendix

Table 14: Failed finer-grained interventions. None became the method of the paper.

Stage-2 training	pass@1	SynErr	1-span	Multi	Exact len	Under-gen
Aligned old-visible	27.1	27.1	30.9	25.4	16.4	71.1
Round-consistent	23.8	29.0	33.8	19.0	21.5	65.9

Appendix

Table 15: Round-consistent / block-state-balanced Stage-2 training versus the aligned old-visible baseline (CanItEdit, gold locations, original-source context, unlocked block-decision). Exact-length and under-generation rates are among aligned completions. Rates are percentages. Diagnostic, not autonomous performance.

Setting	pass@1	SynErr	1-hole	multi	2	3	$4+$
Unlocked	27.1	27.1	30.9 / 35.3	25.4 / 23.2	44.4 / 11.1	7.1 / 28.6	16.7 / 31.7
$\delta{=}0$ forced occupancy	28.1	42.4	47.1 / 32.4	19.0 / 47.2	33.3 / 35.2	21.4 / 60.7	5.0 / 51.7
$\delta{=}+1$	27.6	46.2	42.6 / 29.4	20.4 / 54.2	40.7 / 37.0	17.9 / 60.7	3.3 / 66.7
$\delta{=}+2$	24.8	50.5	41.2 / 38.2	16.9 / 56.3	35.2 / 38.9	14.3 / 60.7	1.7 / 70.0
$\delta{=}+4$	14.8	56.7	26.5 / 44.1	9.2 / 62.7	20.4 / 48.1	7.1 / 60.7	0.0 / 76.7
$\delta{=}-1$	4.3	71.0	1.5 / 69.1	5.6 / 71.8	11.1 / 70.4	3.6 / 67.9	1.7 / 75.0

Appendix

Table 16: Same-checkpoint forced-reference occupancy on the aligned old-visible filler (CanItEdit). Hole buckets are gold insert-span counts (

n{=}68/54/28/60

). 1-hole, multi, and per-count columns are pass@1 / SyntaxError. Rates are percentages. Diagnostic, not autonomous performance.

unlocked $\downarrow$ / $\delta{=}0$ $\rightarrow$	Pass	SynErr	Other
Pass	38	15	4
SynErr	12	30	15
Other	9	44	43

Appendix

Table 17: Paired unlocked

\to

\delta{=}0

outcome churn on the same evaluation instances. Pass is harness OK; SynErr is harness SyntaxError (IndentationError remains Other).

99/210

examples change outcome.

Holes	$n$	changed	Pass $\to$ SynErr	SynErr $\to$ Pass
1	68	28	0	9
2	54	21	9	3
3	28	17	0	0
$4+$	60	33	6	0

Appendix

Table 18: Unlocked

\to

\delta{=}0

churn by gold hole count. The 1-hole gain is a SyntaxError rescue; 2-hole and

4+

examples donate successful unlocked trajectories into SyntaxError.

Slice	Max gold hole	$n$	pass@1	SynErr
1-hole	1–16	14	57.1	14.3
1-hole	17–32	22	54.5	18.2
1-hole	33–64	12	8.3	33.3
1-hole	$65+$	20	0.0	70.0
All examples	$65+$	74	6.8	47.3
Total gold tokens $129+$	—	60	8.3	45.0

Appendix

Table 19: Unlocked old-visible pass@1 and SyntaxError by gold hole length. The

65+

1-hole collapse (

0\%

70\%

) is unchanged under

\delta{=}0

. Bucket boundaries are empirical for this checkpoint/protocol, not a universal threshold.

Split	$n$	cont. acc.	stop acc.	all holes	mean $P(\text{cont.})$
Overall	210	57.3	73.3	50	0.281
1-hole	68	32.4	76.5	25	0.247
2-hole	54	69.9	71.3	16	0.308
3-hole	28	51.5	70.2	5	0.312
$4+$	60	65.0	74.0	4	0.270
Multi SynErr	67	74.0	63.3	5	0.375

Appendix

Table 20: Teacher-forced control-slot agreement on the

\delta{=}0

trajectory (observer-only: gold stop/continue still forced). Continue accuracy is agreement on non-final reference blocks; stop accuracy is agreement on the gold final block. These probabilities are not unconditional length-prediction accuracy.

Placement	Layout	exact-match	pass@1
Middle	prefix [MASK] suffix	79.2	93.6
End-prompted	chat(PREFIX, SUFFIX) then [MASK]	31.9	17.7

Appendix

Table 21: HumanEval-Infilling single_line placement probe (DiffuCoder-Instruct, gold-length masks,

n{=}1033

). The end-prompted row is the off-the-shelf protocol; it is not the matched-SFT end arm in Table 22 . pass@1 is the functional metric. Rates are percentages.

Evaluation	Middle	End	$\Delta$
Off-the-shelf single_line pass@1	93.6	17.7	$+75.9$
Matched-SFT single_line pass@1	94.6	94.3	$+0.3$
exact match	84.2	82.5	$+1.7$
Matched-SFT multi_line pass@1	61.9	55.2	$+6.7$
exact match	37.3	33.1	$+4.2$

Appendix

Table 22: HumanEval-Infilling with oracle span length. Off-the-shelf end uses the end-prompted chat protocol; the matched-SFT arms share one serialization. Rates are percentages.

Evaluation	Metric	Middle	End
EditPackFT	exact match	19.2	19.8
	syntax-ok	97.0	96.8
	mean token distance	12.2	12.2
CanItEdit single-span	pass@1	42.4	47.0
	string exact match	24.2	27.3
	SyntaxError	28.8	25.8

Appendix

Table 23: Matched-SFT middle versus end on held-out EditPackFT (

n{=}400

) and gold-location CanItEdit canonical single-span fills (

n{=}66

), both with oracle span length.

Task	Output tokens	M2T	Replace	Delete	Insert	Structural share
Mathematics	121	137	4.4	1.6	1.2	5.0%
Prose	107	124	9.0	2.8	1.6	9.7%
Code writing	128	136	1.4	0.2	0.2	1.3%
Code editing	104	117	3.4	1.2	0.4	4.1%

Appendix

Table 24: LLaDA2.2 operation counts per greedy trace (

n{=}5

per task). Structural share is replacement, deletion, and insertion divided by all writes.

Interface	Explicit	Exact	Edit	Length	Multi-span	Source copied
	location	boundary	count	allocation	serialization	by runtime
Whole-file rewrite	—	—	—	—	—	—
Search-and-replace	✓	✓	✓	—	✓	✓
Locate-then-infill	✓	✓	✓	✓	✓	✓
Token-level in-place	✓	✓	✓	✓	✓	✓

Appendix

Table 25: Structural commitments consumed explicitly by each runtime. A dash means that the decision remains latent in the generated output.

Locator	Parse	Exact	Pred./gold	pass@1	Multi
DiffuCoder	91.0	2.4	1.89 / 2.78	13.3	12.0
iLLaDA-8B	94.3	4.3	1.93 / 2.78	13.8	11.3
LLaDA-8B	86.7	3.3	1.81 / 2.78	11.9	10.6

Appendix

Table 26: Matched-layout Stage 1 locator outputs and CanItEdit pass@1 with the fixed DiffuCoder old-visible Stage 2 filler (

N{=}210

). Parse and exact line-operation agreement are percentages; predicted/gold edits are means.

Knob	Rewrite	str_replace	Locate
Init	Instruct	Instruct	Stage-2 $1610$
LR	$2{\times}10^{-5}$	$1{\times}10^{-5}$	$1{\times}10^{-5}$
Batch / epochs	128 / 4	64 / 5	64 / 3
Decode temp	$1.0$	$0.0$	two-stage

Appendix

Table 27: Recipe confounds in Class A. Locate Stage 2 curriculum is part of the two-stage interface. Because these knobs are unmatched, Class A cells are within-family operating points rather than a causal ranking.

Interface	Setup	pass@1	SynErr	1-span	multi
Rewrite	SFT, 256 steps	9.0	36.2	20.6	3.5
Rewrite	SFT, 768 steps	10.0	13.8	20.6	4.9
str_replace	256-slot, step 1515	9.0	5.7	23.5	2.1
str_replace	random-pad, step 1685	4.3	11.9	7.4	2.8
Unified diff	DiffuCoder FT	4.3	1.0	13.2	0.0

Appendix

Table 28: Class A length-harness and serialization controls omitted from Table 1 (CanItEdit). Rates are percentages.

Bucket	$n$	str_replace	Locate	Rewrite	Gold loc $+$ Qwen
1–5 changed lines	54	46.3	29.6	29.6	77.8
6–20 changed lines	112	20.5	9.8	7.1	39.3
${>}20$ changed lines	44	2.3	0.0	0.0	9.1
Changed-line ratio ${<}10\%$	28	21.4	21.4	10.7	89.3
Length similar	108	34.3	20.4	21.3	62.0
Length grow	100	12.0	5.0	1.0	22.0

Appendix

Table 29: Local vs. large-edit slices of CanItEdit. Class A cells except the last column (gold-location

+

Qwen, Class C). Length similar is

|\Delta|{}\leq{}5

source lines; length grow is

\Delta\geq 6

; the remaining 2/210 shrink (

\Delta\leq-6

). Changed-line ratio

{<}10\%

is not the same slice as 1–5 changed lines: a long file with a small ratio can still be a large absolute edit.

Interface	Checkpoints	pass@1 (%)
Locate same-ckpt	645 / 1000 / 1290 / 1935	6.7 / 11.4 / 13.3 / 12.9
Rewrite 640 decode	256 / 512 / 768 steps	9.0 / 11.4 / 10.0
Rewrite sibling 0522	428 / 856 (512 decode)	10.5 / 11.4
str_replace 256-slot	303 / 606 / 909 / 1212 / 1515	11.9 / 8.1 / 8.1 / 7.6 / 9.0

Appendix

Table 30: Existing-checkpoint train and decode curves (CanItEdit). Rewrite 428/856 are a sibling batch-96 run; completions were chatty and fenced, then stripped before scoring. Block-append str_replace has only step 1685. Do not plot 428–640–856 as one curve.

Model	pass@1	1-span	multi
GPT-4o	72.4	79.4	69.0
Qwen3-8B	51.4	67.6	43.7
Qwen2.5-Coder-7B-Instruct	48.6	58.8	43.7

Appendix

Table 31: Autoregressive rewrite references on CanItEdit, 105 descriptive instructions only . These cells are a different protocol from the full-evaluation tables and are not headline parity with dLLM results. Rates are percentages.

Interface	OK	SynErr	Assert.	Other	Timeout
Stable-DiffCoder rewrite	127	0	56	26	1
str_replace block-append	49	12	71	73	5
Locate+infill same ckpt	27	32	66	81	4
Rewrite SFT, 512 steps	24	50	37	96	3
Oracle loc $+$ empty plan	52	57	45	56	0
Oracle loc $+$ Qwen CoT	90	45	15	59	1

Appendix

Table 32: Coarse failure mix on CanItEdit (counts). AssertionError is among scored programs, not a proof of “right location, wrong body.” str_replace parse/apply failures are not SyntaxError.

Figure 5: Primary error class by interface (first-match cascade, percentages). wrong_localization is defined only for predicted-plan locate-then-infill. IndentationError is a harness Exception and is folded into syntax.

Interface	Fmt	Apply	No-op	Syn	Loc	Sem	Pass
Rewrite SFT, 512 steps	0	0	11	59	0	116	24
str_replace block-append	23	29	6	14	0	89	49
Locate+infill same ckpt	9	0	9	42	112	11	27
Oracle loc $+$ empty plan	0	0	12	64	0	82	52
Oracle loc $+$ Qwen CoT	0	0	1	58	0	61	90

Appendix

Table 33: Primary error cascade on CanItEdit (counts). First match; not a manual diagnosis. Locate wrong loc. is undefined on gold-scaffold oracles.

str_replace <old> spans, 142 multi-span cases
Mean gold-span recall	42.7%
All gold spans covered	21.8% (31/142)
Predicted $K{=}1$	45/142
Enter hybrid Stage 2 branch	53/142

Appendix

Table 34: Localization audits for autonomous structure predictors.

Metric	Value
Parseable plans	95.7%
Zero predicted edits	4.3%
Exact line-op match	13.3%
Nearest-anchor exact	57.9%
Mean predicted / gold edits	1.55 / 2.78

Appendix

Table 35: Location quality of predicted Stage 1 plans on CanItEdit.

Figure 6: Gold-scaffold Stage-2 pass@1 by perturbation family (shared multitask filler, gold-aware Qwen plans). Local one-boundary errors are often survivable; expanding each range by kept lines hurts more; translating the whole scaffold is catastrophic. Dashed/dotted lines: exact-gold privileged fill 42.9% and autonomous same-checkpoint 12.9% (references, not a ranking). Diagnostic, not autonomous performance.

Local typed $\pm 1$
Scaffold	Dist.	pass@1	SynErr	1-span	Multi
Exact gold ranges	0.00	42.9%	21.4%	57.4%	35.9%
Shift insert-gap $-1$	0.69	38.6%	21.0%	50.0%	33.1%
Shift insert-gap $+1$	0.71	37.1%	22.9%	50.0%	31.0%
Shift start $+1$ (where valid)	0.56	36.7%	21.4%	54.4%	28.2%
Shift end $+1$	1.66	33.3%	31.4%	55.9%	22.5%

Appendix

Table 36: Gold-scaffold Stage-2 fill under boundary perturbations (multitask step 1935 filler, gold-aware Qwen plan). Diagnostic, not autonomous performance.

Plan / conditioning	pass@1	1-span	Multi	SynErr
Gold loc. + Qwen CoT	42.9%	57.4%	35.9%	21.4%
Gold loc. + location-free gold-aware Qwen	42.9%	57.4%	35.9%	22.4%
Gold loc. + Qwen CoT (ep5 filler)	39.5%	48.5%	35.2%	27.1%
Gold loc. + empty plan (multitask 1935)	24.8%	42.6%	16.2%	27.1%
Gold loc. + input-only Qwen3-8B	24.8%	41.2%	16.9%	21.4%
Gold loc. + Qwen headers only	22.4%	42.6%	12.7%	16.7%

Appendix

Table 37: Gold-location Stage-2 fill under different plan texts. Rows above the last midrule use the multitask step 1935 filler unless noted (ep5). The last two rows are separately trained aligned Stage-2 diagnostics from DiffuCoder-Instruct (block 16, five epochs): empty_plan and original_source at global_step_1610 , not the plan-conditioned ep5 specialist. Only that pair is a matched test of old-source visibility; 20.5/27.1 vs. 24.8 is not. The integrated-diff pre-glue

20.5\%

cell in Table 43 is a third, different

20.5\%

. All rows are diagnostics, not autonomous performance.

Interface	1 hole	2	3	$4+$
$n$	68	54	28	60
Oracle loc. $+$ Qwen CoT (C)	57.4%	50.0%	50.0%	16.7%
Oracle loc. $+$ empty (C)	42.6%	22.2%	21.4%	8.3%
str_replace block-append (A)	48.5%	16.7%	17.9%	3.3%
Locate-and-infill same ckpt (A)	27.9%	9.3%	7.1%	1.7%
Rewrite SFT 512 (A)	27.9%	7.4%	3.6%	0.0%

Appendix

Table 38: Pass@1 by gold insert-span count (CanItEdit). Gold-location rows give Stage 2 every gold hole (empty-plan row: multitask 1935 filler). Class A rows are autonomous.

4+

pools 4–10 holes (

n{=}60

Interface	1 hole	2	3	$4+$
$n$	68	54	28	60
Oracle loc. $+$ Qwen CoT (1935)	57.4 / 20.6	50.0 / 22.2	50.0 / 21.4	16.7 / 21.7
Oracle loc. $+$ empty (1935)	42.6 / 16.2	22.2 / 25.9	21.4 / 28.6	8.3 / 40.0
Aligned empty 1610	33.8 / 27.9	27.8 / 20.4	3.6 / 17.9	6.7 / 35.0
Aligned old-visible 1610	30.9 / 35.3	44.4 / 11.1	7.1 / 28.6	16.7 / 31.7

Appendix

Table 39: Pass@1 / SyntaxError by gold insert-span count (CanItEdit). SyntaxError is not monotonic in hole count. The old-visible 1-hole cell has the highest SyntaxError among 1–2 holes; privileged Qwen CoT SyntaxError is nearly flat through

4+

. Diagnostic, not autonomous performance.

Intervention	pass@1	1-span	Multi
Exact gold spans	42.9%	57.4%	35.9%
Extra false-positive span	39.5%	48.5%	35.2%
Shift first span $+1$ line	19.5%	29.4%	14.8%
Drop last gold span	5.7%	0.0%	8.5%

Appendix

Table 40: One-span interventions on gold scaffolds with gold-aware Qwen plans (filler 1935). Single-span drop is 0% by construction. Diagnostic, not autonomous performance.

Planner	Parse	Exact op	Mean pred	pass@1
Unconstrained Stage 1	95.7%	13.3%	1.55	12.9%
Gold- $K$ prompt hint	89.0%	14.3%	1.80	11.9%

Appendix

Table 41: Gold edit-count prompt hint on Stage 1 (multitask 1935, confidence 0.1) filled by the same checkpoint. Raises exact-count agreement, not pass@1. Diagnostic, not a production decoder.

Arm	Full	Unaffected	Affected
D pre-glue	20.5	34.7	0.0
D ′ glue-fix	26.2	34.7	14.0
D ′′ canonicalize	28.1	34.7	18.6
Glue-fix $+$ oracle length	27.1	30.6	22.1

Appendix

Table 42: EOF/no-newline glue split on the integrated-diff step-1610 filler (CanItEdit). Affected:

86

evaluations /

43

sources. Unaffected:

124

evaluations /

62

sources. Oracle-length locks gold occupancy on the glue-fixed scaffold (train–test mismatch). Rates are percentages. Diagnostic, not autonomous performance.

Arm	pass@1	1-span	Multi	SynErr
D pre-glue integrated	20.5	30.9	15.5	20.5
D ′ glue-fixed integrated	26.2	33.8	22.5	26.7
D ′′ canonicalized integrated	28.1	33.8	25.4	24.8
Local $\pm 2$ prefix control	28.1	38.2	23.2	21.9

Appendix

Table 43: Gold-scaffold construction with oracle locations. D, D ′ , and D ′′ share one integrated-diff step-1610 filler; only scaffold construction changes.

Scaffold	Passed	pass@1 (%)
Predicted baseline $P$	15/142	10.6
$P+C$ (coverage)	19/142	13.4
$P+B$ (boundaries)	20/142	14.1
$P+C+B$ (both)	31/142	21.8
Gold	36/142	25.4

Appendix

Table 44: Paired oracle repairs of the separate predicted-hole baseline

P

on 142 multi-span CanItEdit examples. The gold row is the existing clean-scaffold result. Repairs change scaffold geometry, not the filler or decoder; none is an autonomous editor.

Interface	Setup	pass@1	SynErr	1-span	Multi
Direct str_replace	block-append 1685	23.3	5.7	48.5	11.3
Hybrid-multi	predicted <old> $\to$ old-visible	22.9	9.0	48.5	10.6
Hybrid-all	all unique <old> $\to$ old-visible	13.3	23.8	22.1	9.2
Gold loc. $+$ old-visible	aligned 1610	27.1	27.1	30.9	25.4
Gold loc. $+$ aligned empty	aligned 1610	20.5	26.7	33.8	14.1

Appendix

Table 45: Detailed composition cells (CanItEdit). Hybrid-multi overall 22.9% is not a better editor: single-span 48.5% is inherited from direct str_replace . Rates are percentages.

Method	All 31	Routed 19	Fallback 12
Direct str_replace	10/31	7/19	3/12
Predicted scaffold $\to$ filler	8/31	5/19	3/12
Gold scaffold $\to$ same filler	13/31	7/19	6/12

Appendix

Table 46: Paired 31-case control on multi-span examples whose predicted <old> spans cover every gold hunk. Filler is aligned old-visible Stage 2 at step 1610. Routed

=

Hybrid-multi sent the example to Stage 2; fallback

=

predicted

K{=}1

kept the str_replace reconstruction.

Relation	$n$	Direct	Pred. scaf.	Gold scaf.
Merged multiple gold spans	12	3/12	3/12	6/12
Strict superset	10	5/10	2/10	3/10
Exact boundaries	3	2/3	1/3	1/3
Strict subset	3	0/3	2/3	3/3
Overlap / fragmented	3	0/3	0/3	0/3
Predicted $K{=}$ gold $K$	17	7/17	5/17	5/17

Appendix

Table 47: Boundary-relation and

K

-match slices of the 31 fully-covered examples. Relation rows use first-match precedence (merge

>

fragment

>

exact

>

superset

>

subset

>

residual overlap). The

K

rows are not exclusive of the relation rows; bins with

n{\leq}3

are not interpreted.

Figure 7: Overall CanItEdit pass@1 under gold-scaffold corruptions of the aligned old-visible filler. Missing required spans collapse performance; one extra hole is comparatively cheap; expanding gold boundaries is not. Drop-one / drop-25 / drop-50 overall rates mix emptied 1-span scaffolds; the locator-quality curve in Figure 4 uses multi-span pass@1. Diagnostic, not autonomous performance.

Scaffold	pass@1	Multi	$\Delta$	SynErr	Appl./imposs.
Clean gold	27.1	25.4	—	27.1	210/0
Drop one gold span	4.3	6.3	$-22.9$	12.9	210/0
Drop 25% of gold spans	3.3	4.9	$-23.8$	11.0	210/0
Drop 50% of gold spans	0.5	0.7	$-26.7$	10.0	210/0
Expand $\pm 1$	15.2	7.7	$-11.9$	39.0	210/0
Expand $\pm 2$	13.8	7.0	$-13.3$	42.9	210/0

Appendix

Table 48: Full gold-scaffold corruption grid on the aligned old-visible filler (CanItEdit). Relative

\Delta

is against clean gold

57/210

(

27.1\%

). Applied / impossible counts how often the operator changed the scaffold. Shift

\pm 1

also changes coverage (mean recall 29.6% / 32.9% on the 142) and extra-span count.

Family	Mode	Recall	Extra	Multi	Overall	SynErr
gold	clean scaffold	100.0	0.00	25.4	27.1	27.1
missing-span	drop-one	66.0	0.00	6.3	4.3	12.9
missing-span	drop-25%	62.2	0.00	4.9	3.3	11.0
missing-span	drop-50%	45.5	0.00	0.7	0.5	10.0
shift	$-1$ line	29.6	2.62	9.9	11.9	15.2
shift	$+1$ line	32.9	2.46	4.9	8.1	32.4

Appendix

Table 49: Locator quality

\to

multi-span pass@1 on the aligned old-visible filler. Recall and extra-span count are means on the 142 multi-span examples.

\Delta

vs. gold and vs. 11.3 are multi-span points. Corruption rows are diagnostics; str_replace rows are autonomous references.

Method	Overall	1-span	Multi-span
Direct one-shot str_replace	23.3	48.5	11.3
One-edit prompt, max 5 rounds	14.8	35.3	4.9
First-valid executor, max 1 round	19.5	48.5	5.6
First-valid executor, max 2 rounds	18.1	38.2	8.5
First-valid executor, max 3 rounds	17.1	30.9	10.6
First-valid executor, max 4 rounds	19.0	36.8	10.6

Appendix

Table 50: Sequential str_replace on CanItEdit. One-edit prompt and first-valid executor use the same checkpoint as direct one-shot. Rates are percentages. Gold-location old-visible Stage 2 is a diagnostic, not a sequential baseline.

Stop reason	$n$	Rate
noop_edit	59	28.1%
Malformed edit structure	56	26.7%
repeated_edit	29	13.8%
max_rounds	27	12.9%
code_state_loop	22	10.5%
ambiguous_old	13	6.2%

Appendix

Table 51: First-valid sequential stop reasons (max 5 rounds). too_many_edits is the one-edit-prompt death mode (52.9%) and does not apply here.

Explore similar work

Knowledge Editing for Masked Diffusion Language Models
Jun 2, 2026Haewon Park, Yohan JoMultimodal Knowledge Editing Masked Diffusion Language Models
Graduate School of Data Science, Seoul National University
CROCODIL: Cross-Model Code Editing with LLMs
Sep 3, 2026Linghan Zhong, Aditya Thimmaiah, Jayanth Srinivasa +2Strong Large Language Model
1The University of Texas at Austin · 2Cisco Research
When Models Edit Too Much: On the Fidelity of Minimal Code Edits
Sep 3, 2026Tongyao Zhu, Wei Hern Lim, Min-Yen KanCode Quality Edit
National University of Singapore

Jun 2, 2026, cs.CLJ/K move · Enter open · S save

Knowledge Editing for Masked Diffusion Language Models

Haewon Park, Yohan Jo

Graduate School of Data Science, Seoul National University

Multimodal Knowledge Editing Masked Diffusion Language Models Language Modeling Diffusion Models Next-Token Prediction

Open paper AskPDF arXiv

Knowledge editing aims to update or correct factual knowledge in a language model. A widely used approach, locate-then-edit, first localizes a fact within the model and then edits the weights there. To date, such methods have been developed exclusively for autoregressive models (ARMs). Whether they work for masked diffusion models (MDMs), which model text bidirectionally and generate by iterative denoising rather than next-token prediction, remains an open question. We address it by transferring locate-then-edit to MDMs and comparing multiple MDMs with their matched ARMs. Our central finding has two parts. First, where an edit should be applied transfers between them: the same early-to-mid-layer MLP at the last subject token is most effective for both. Second, this shared location does not guarantee a shared outcome. Single-token edits succeed in both, but as targets grow longer, editing degrades far more sharply in the MDMs than in the ARMs. The failure stems from how the edited fact is generated: producing a multi-token target passes through intermediate states in which the target is partially unmasked, for which the edit was never optimized. Guided by this diagnosis, we introduce a simple correction that optimizes the edit including such states, substantially restoring multi-token performance. Our code is available at https://github.com/holi-lab/MDM-KE.

Abstract

Explore similar work

Knowledge Editing for Masked Diffusion Language Models

CROCODIL: Cross-Model Code Editing with LLMs

When Models Edit Too Much: On the Fidelity of Minimal Code Edits