Intent-Triggered Anchoring for Multi-Token Format Units in Parallel Decoding of Diffusion Language Models

Junhyeok Lee; Kyu Sung Choi

cs.CLOct 12, 2026

Intent-Triggered Anchoring for Multi-Token Format Units in Parallel Decoding of Diffusion Language Models

Authors: Junhyeok Lee, Kyu Sung Choi

Organizations: Interdisciplinary Program in Cancer Biology, Seoul National University College of Medicine · Department of Radiology, Seoul National University Hospital · Department of Radiology, Seoul National University College of Medicine · Healthcare AI Research Institute, Seoul National University Hospital

Abstract

In parallel decoding, diffusion language models commit several tokens per forward pass, so co-committed tokens cannot see each other, breaking multi-token format units that parsers need. On ECHO, an open-weight one-step chest X-ray (CXR) report generator, 32.7-36.7% of evaluation reports lack a parseable Impression heading and score as empty. We propose intent-triggered anchoring, a training-free decoding rule that completes a format unit once the model commits its first token, inserting only units with a unique, content-free completion; fallbacks write the unit when the model stops or loops without starting it. On three CXR evaluation sets, the rule raises readable rates from 63.3-67.3% to 100.0% with no regressions and a 95th-percentile latency ratio of 0.68-1.28 against official decoding. Broken headings state findings absent from the reference; adding the heading at the end leaves them in 15.5-31.0% of reports, our rule in at most 0.6%. For math answer markers, the rule raises accuracy in the most parallel setting of five diffusion language models by 5.8-20.8 points on GSM8K and 5.1-17.1 on GSM-Symbolic, held out from tuning. With lenient answer extraction, gains are 3.3-12.5 and 2.7-6.3 points; the rule adds 1.4-8.6 points over a stop-only fallback for four of the five models.

Figures & tables

Panel of Figure 1 — Figure 1: The failure and two repairs on one 8-token block; $\hookleftarrow$ is a newline token. ✗: no parseable heading; ✂: Findings cut mid-sentence; ✓: readable. Thick border: tokens the decoder writes; dashed: slots re-predicted with the anchor in view.

Panel of Figure 2 — Figure 2: Anchoring on the fused block key–value (KV) cache, on the last four slots of block $n$ . Gray: blocks already cached; green: block $n$ cached with its anchor. (i), (ii): the two cache invariants.

Development
	Readable (%)			p95 (s)
Split	Off.	Full	Reg.	Off.	Full	+Fwd
MIMIC valid (277)	65.7	100.0	0/0	2.71	1.68	0.18
ReX valid (3,613)	63.1	100.0	0/0	0.72	0.69	0.39
Evaluation
MIMIC test (517)	65.2	100.0	0/0	2.31	1.57	0.18

Table 1: Full pipeline against official one-step decoding (Off.); counts are studies. Bold: better of Off. and Full. Reg.: readable/complete regressions; p95: 95th-percentile end-to-end time per study in seconds; +Fwd: mean extra decode forwards per study of the full pipeline. ReXGradient (ReX); CheXpert-Plus (CXP).

Structure
	MIMIC valid			ReX valid			MIMIC test			ReX test			CXP valid
	Fixed	EOS	Full	Fixed	EOS	Full	Fixed	EOS	Full	Fixed	EOS	Full	Fixed	EOS	Full
Readable (%)	100.0	98.9	100.0	100.0	100.0	100.0	100.0	99.6	100.0	100.0	100.0	100.0	100.0	96.2	100.0
Complete (%)	100.0	94.2	100.0	99.9	95.5	100.0	100.0	88.4	100.0	99.8	95.4	100.0	100.0	90.4	100.0
Duplicate headings	9	10	0	464	16	0	10	16	0	471	20	0	4	3	0
Entities

Table 2: Placement comparison on all five splits. Fixed: fixed-position anchoring at

b=100

; EOS: anchoring at the end-of-sequence (EOS) token; Full: full pipeline; Off.: official one-step decoding. Spurious findings: reports whose broken heading fragment states a positive entity absent from the reference; retention is measured against Off. entities, and supported retention against reference-supported Off. entities. Bold: best per split, except for retention. RaTE+: positive-entity RaTEScore. ReXGradient (ReX); CheXpert-Plus (CXP).

	MIMIC-CXR valid (277 studies)						ReXGradient valid (3,613 studies)
Decoder	Chg.	Unr./Inc.	MTR	Sem	R-L	RaTE+	Chg.	Unr./Inc.	MTR	Sem	R-L	RaTE+
Full pipeline	–	0/0	45.31	48.28	55.65	58.34	–	0/0	55.75	57.37	64.46	65.93
$-$ intent completion	104	0/0	45.28	47.41	55.33	57.70	1,877	0/0	55.43	51.33	63.89	63.89
$-$ boundary rewrite	11	11/11	45.31	48.25	55.66	58.44	1	1/1	55.75	57.37	64.46	65.93
$-$ real-word guard	12	0/0	45.33	48.38	55.67	58.27	2	0/0	55.75	57.37	64.46	65.92
$-$ EOS relocation	9	9/9	43.84	46.80	53.92	56.80	82	82/82	54.48	56.50	63.11	64.78

Table 3: Leave-one-out ablation on the two development splits. Rows give the full pipeline and each variant with one component removed. Chg.: reports differing from the full pipeline; Unr./Inc.: unreadable/incomplete reports. Bold: best per score column. METEOR (MTR); Sem: SemScore; R-L: ROUGE-L; RaTE+: positive-entity RaTEScore; EOS: end of sequence.

		GSM8K test ( $n{=}1{,}319$ )								GSM-Symbolic main ( $n{=}5{,}000$ )
		Parse		Accuracy			Lenient		No-loop	Parse		Accuracy			Lenient		No-loop
Model	$\kappa$	Base	Anch.	Base	Stop	Anch.	Base	Anch.	Anch.	Base	Anch.	Base	Stop	Anch.	Base	Anch.	Anch.
SDAR, $B{=}32$	8	49.2	93.7	31.9	41.2	47.8	39.0	48.5	40.5	42.2	92.9	23.5	28.9	33.4	28.2	33.8	28.6
SDAR, $B{=}4$	4	38.0	91.9	29.3	41.8	50.1	38.1	50.6	39.7	32.5	89.9	23.1	32.0	36.0	30.3	36.6	31.1
Fast-dLLM v2	8	62.0	90.2	29.9	35.4	37.8	34.7	38.1	35.5	58.8	90.6	23.5	28.1	29.5	27.1	29.8	28.3
LLaDA	8	77.6	93.6	50.7	53.1	56.5	54.7	58.8	51.7	73.6	92.4	41.8	43.2	46.9	45.0	48.8	42.3

Table 4: Answer-marker anchoring in the most parallel setting of each model and in dParallel’s native setting. Base and Anch.: percentages without and with anchoring; Stop: accuracy with only the stop fallback, from a separate run; Parse: a boxed answer is found; Lenient: accuracy when the last number is taken if no boxed answer parses; No-loop: anchored accuracy counting answers written after a loop fallback as wrong;

\kappa

: nominal tokens per forward; nat.: dParallel with no per-forward floor. Bold: best per row within Parse, Accuracy, and Lenient.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

(a) Correct continuation after Imp (%)
Slot of Imp	$B{=}8$	$B{=}4$
Last slot	59.56 (596)	99.49 (197)
Earlier slot	50.91 (4,528)	99.34 (3,503)
(b) Official one-step block-8 outputs
	MIMIC valid ( $n{=}277$ )	ReX valid ( $n{=}3{,}613$ )
Findings: present	277/277	3,613/3,613

Appendix

Table 5: Heading-prefix continuation by slot and output categories of official one-step block-8 outputs on the two development sets. Boundary counts pool 3,890 studies; parentheses give heading-prefix occurrences. Completion ambiguity and vocabulary analyses use each development set separately; OOV: out-of-vocabulary error candidates. ReXGradient (ReX).

Panel of Figure 3 — Figure 3: Readable rate and exact entity retention on the two development sets. Numbers label fixed-position token budgets; diamonds mark the full pipeline. Orange: MIMIC-CXR; blue: ReXGradient.

	MIMIC test ( $n{=}517$ )					ReX test ( $n{=}3{,}651$ )					CXP valid ( $n{=}52$ )
Condition	R-L	MTR	CIDEr	RaTE+	Sem	R-L	MTR	CIDEr	RaTE+	Sem	R-L	MTR	CIDEr	RaTE+	Sem
Constant, canonical	50.30	37.71	3.67	60.83	40.57	64.64	54.70	5.41	69.70	62.26	50.41	35.97	3.56	60.53	38.76
Constant, oracle ref.	51.50	40.88	3.51	54.93	43.31	66.37	53.83	5.62	60.96	61.31	56.86	47.20	3.12	55.62	30.06
Teacher B4, 4 steps	54.87	46.73	3.68	60.08	49.19	64.04	55.22	5.31	68.11	58.45	55.33	43.25	3.72	61.57	51.42
Teacher B8, 8 steps	55.04	46.95	3.61	59.84	49.54	63.36	55.39	4.92	68.75	54.77	55.92	45.35	3.76	62.07	53.43
Student B4, 1 step	51.24	43.73	3.34	56.10	44.65	59.62	51.75	4.17	62.57	43.61	49.73	40.40	3.02	54.97	46.18

Appendix

Table 6: Two constant reports and the four released checkpoints on the three evaluation sets; n denotes studies. Oracle reference: approximate medoid selected from evaluation references; Teacher B

B

and Student B

B

: ECHO-Base and ECHO with block size

B

. Bold: best per metric and set. R-L: ROUGE-L; METEOR (MTR); Sem: SemScore; RaTE+: positive-entity RaTEScore. ReXGradient (ReX); CheXpert-Plus (CXP).

Data	Model	$\kappa$	TPF	Base	Rep.	Anch.	Base	Rep.	Anch.	No-loop	Reg.	Fires	Base	Anch.
				Parse			Accuracy						Forwards
GSM8K	SDAR, $B{=}32$	4	5.3	90.8	91.1	99.2	73.4	73.7	77.3	76.7	0	20/55/48	34.9	34.2
GSM8K	SDAR, $B{=}32$	8	10.2	49.2	50.6	93.7	31.9	33.0	47.8	40.5	2	54/145/449	28.0	20.4
GSM8K	SDAR, $B{=}4$	2	2.4	86.3	86.3	97.0	73.6	73.6	78.8	78.2	1	12/92/59	83.8	79.2
GSM8K	SDAR, $B{=}4$	4	4.0	38.0	38.0	91.9	29.3	29.3	50.1	39.7	0	15/215/539	78.0	36.1
GSM8K	Fast-dLLM v2	4	5.5	93.0	93.4	97.7	64.7	64.9	66.7	66.6	0	18/41/11	55.8	55.9

Appendix

Table 7: Answer-marker anchoring in every setting. Percentages are for the baseline (Base), the baseline with only the post-decoding marker collapse (Rep.), and anchoring (Anch.). No-loop: anchored accuracy counting answers written after a loop fallback as wrong; Reg.: problems that parse only without anchoring; Fires: intent/stop/loop writes; Forwards: mean forward passes per problem; TPF: measured baseline tokens per forward;

\kappa

: nominal tokens per forward; nat.: dParallel with no per-forward floor; GSM-Sym.: GSM-Symbolic.

Teacher B8 (8 steps)
View	MIMIC test	ReX test	CXP valid
Strict	55.04	63.36	55.92
Recovered	55.41	63.36	55.92
Raw	56.80	65.01	57.31
Unparseable (%)	0.8	0.0	0.0
Student B8 (1 step)

Appendix

Table 8: ROUGE-L views of identical generations on the three evaluation sets. Strict: unchanged official extractor; Recovered: all-content fallback; Raw: no section extraction; Unparseable: official evaluation-code failure rate; Published: ECHO’s reported value. Studies: 517, 3,651, and 52. Teacher B8 and Student B8: ECHO-Base and ECHO with block size 8. ReXGradient (ReX); CheXpert-Plus (CXP).

Parallel Decoding for Diffusion Language Models

See all 37 papers

Supportive Token Revealing for Fast Diffusion Language Model Decoding
Jun 2, 2026Giries Abu Ayoub, Mario Barbara, Lluís Pastor-Pérez +4Masked Diffusion Models Diffusion Language Model Decoding
Department of Computer Science, University of Haifa · Institute for AI, University of Stuttgart · IMPRS-IS
When Confidence Misleads: Suffix Anchoring and Anchor-Proximity Confidence Modulation for Diffusion Language Models
May 27, 2026Jungwon Park, Jimyeong Kim, Jungmin Ko +2Non-Autoregressive Text Generation Language Model Decoding
1RICS · 2AIIS · 3IPAI +2
Neither Parallel Nor Sequential: How DiffusionGemma Actually Commits Tokens
Jun 12, 2026Ali Asaria, Tony Salomone, Deep GandhiMasked Diffusion Models Diffusion Sampling
Transformer Lab

Jun 2, 2026J/K move · Enter open · S save

Supportive Token Revealing for Fast Diffusion Language Model Decoding

Giries Abu Ayoub, Mario Barbara, Lluís Pastor-Pérez, Tanja Bien+3

Department of Computer Science, University of Haifa · Institute for AI, University of Stuttgart · IMPRS-IS

Masked Diffusion Models Diffusion Language Model Decoding Diffusion Model Inference Acceleration Parallel Decoding for Diffusion Language Models

Open paper AskPDF arXiv

Discrete diffusion language models can generate text efficiently by updating multiple masked positions in parallel, but this parallelism introduces a quality-latency trade-off. Aggressive decoding may commit mutually dependent tokens too early, while conservative decoding requires many denoising steps. Existing methods address this tension by deciding which tokens are safe to reveal using confidence or dependency criteria. However, avoiding unsafe commits does not necessarily make the remaining masked sequence easy to decode, since uncertain tokens may depend on masked tokens, creating a bottleneck for denoising steps. We propose AXON, a training-free module that can be added on top of existing parallel decoding strategies for diffusion language models. Rather than replacing the base decoder, AXON monitors the remaining uncertain masked tokens and intervenes only when their current state suggests that additional context is needed. It then shifts the criterion from which tokens are safest to reveal to which confident reveals would best support later denoising. AXON selects anchors, confident masked tokens that uncertain positions attend to, using attention, uncertainty, and confidence signals. Experiments on reasoning and code-generation benchmarks across multiple diffusion language models show that AXON improves the quality-latency trade-off of existing parallel decoders, often reducing the number of function evaluations while maintaining or improving accuracy.

Abstract

Parallel Decoding for Diffusion Language Models

Explore similar work

Supportive Token Revealing for Fast Diffusion Language Model Decoding

When Confidence Misleads: Suffix Anchoring and Anchor-Proximity Confidence Modulation for Diffusion Language Models

Neither Parallel Nor Sequential: How DiffusionGemma Actually Commits Tokens