Efficient Self-Evaluation for Diffusion Language Models via Sequence Regeneration

Linhao Zhong; Linyu Wu; Wen Wang; Yuling Xi; Chenchen Jing; Jiaheng Zhang; Hao Chen; Chunhua Shen

cs.CLMar 3, 2026

Efficient Self-Evaluation for Diffusion Language Models via Sequence Regeneration

Authors: Linhao Zhong, Linyu Wu, Wen Wang, Yuling Xi, Chenchen Jing, Jiaheng Zhang, Hao Chen, Chunhua Shen

Organizations: Zhejiang University · National University of Singapore · Zhejiang University of Technology

Abstract

Diffusion large language models (dLLMs) have recently attracted significant attention for their ability to enhance diversity, controllability, and parallelism. However, their non-sequential, bidirectionally masked generation makes quality assessment difficult, underscoring the need for effective self-evaluation. In this work, we propose DiSE, a simple yet effective self-evaluation confidence quantification method for dLLMs. DiSE quantifies confidence by computing the probability of regenerating the tokens in the entire generated sequence, given the full context. This method enables more efficient and reliable quality assessment by leveraging token regeneration probabilities, facilitating both likelihood estimation and robust uncertainty quantification. Building upon DiSE, we further introduce a flexible-length generation framework, which adaptively controls the sequence length based on the model's self-assessment of its own output. We analyze and validate the feasibility of DiSE from the perspective of dLLM generalization, and empirically demonstrate that DiSE is positively correlated with both semantic coherence and answer accuracy. Extensive experiments on likelihood evaluation, uncertainty quantification, and flexible-length generation further confirm the effectiveness of the proposed DiSE. Project page: https://zhongzero.github.io/DiSE/

Figures & tables

Figure 1: A simplified illustration of self-evaluation confidence quantification methods for clarity. (a) Monte Carlo simulation approach for dLLMs. A total of

N_{mc}

simulations are performed. In the

i

-th simulation, a set of masked positions

\{mask^{i}_{j}\}

is sampled. The tokens at these positions are replaced with mask tokens, and the model predicts the probability of correctly generating these tokens. The final estimation is obtained by aggregating the results across all

N_{mc}

simulations. (b) The proposed DiSE for dLLMs. The set of selected positions

\{U_{j}\}

is predefined. The model receives the entire sequence and estimates the regeneration probability of the tokens at

\{U_{j}\}

Figure 2: Generalization ability of dLLMs: Different tokens map from distinct start points to similar end points in the latent space.

Figure 3: Histogram and Cumulative distribution function (CDF) of GT token probability ranks.

Figure 4: Mean pairwise distribution distances for GT, mask, and random tokens using JS Divergence and Wasserstein Distance.

Figure 5: Differences between the DiSE scores of natural sentences and randomized sentences using the LLaDA-Instruct-8B model under four selection modes: ‘full’ (entire sentence), ‘first-10’ (first 10 tokens), ‘mid-10’ (10 tokens from the middle) and ‘last-10’ (last 10 tokens). Each subfigure contains 15 blocks, representing 15 sampled sentences. All blocks are shown in green (difference $>0$ ), indicating that natural sentences consistently achieve higher DiSE scores than randomized sentences.

Figure 6: Comparison between the DiSE scores of correct and incorrect answers across four datasets using the LLaDA-Instruct-8B model under two selection modes ‘full’ and ‘last-10’.

Figure 7: Illustration of the flexible-length dLLM generation framework with DiSE versus fixed-length generation. (a) Fixed-length generation baseline. (b) Flexible-length generation with DiSE.

	Method	ARC-Challenge	GPQA	# NFE
LLaDA-Instruct-8B	MC, $N_{mc}=1$	0.306	0.212	1
	MC, $N_{mc}=32$	0.478	0.286	32
	DiSE (ours)	0.542	0.301	1
LLaDA-1.5-8B	MC, $N_{mc}=1$	0.311	0.203	1
	MC, $N_{mc}=32$	0.488	0.275	32
	DiSE (ours)	0.567	0.299	1

Table 1: Conditional likelihood estimation results on ARC-Challenge and GPQA. The table compares the proposed DiSE against the Monte Carlo simulation baseline with varying

N_{mc}

, and also includes a comparison with the probability estimates from auto-regressive LLMs. The last column reports the average number of model forward passes per computation.

		Countdown			GSM8K			MATH500			SVAMP			Avg. ROC-AUC $\uparrow$
	Method / Gen Len	128	256	512	128	256	512	128	256	512	128	256	512	Avg. ROC-AUC $\uparrow$
LLaDA-Instruct-8B	MC, $N_{mc}=1$	0.524	0.520	0.528	0.539	0.513	0.540	0.497	0.541	0.532	0.563	0.575	0.509	0.532
	MC, $N_{mc}=32$	0.595	0.534	0.558	0.590	0.552	0.595	0.528	0.578	0.531	0.616	0.551	0.647	0.573
	DiSE (ours)	0.578	0.521	0.622	0.633	0.644	0.658	0.611	0.634	0.604	0.688	0.692	0.755	0.637
	LLaMA perplexity	0.574	0.419	0.392	0.675	0.605	0.577	0.575	0.637	0.551	0.686	0.650	0.590	0.578
LLaDA-1.5-8B	MC, $N_{mc}=1$	0.525	0.588	0.528	0.516	0.559	0.525	0.558	0.514	0.525	0.562	0.466	0.554	0.535

Table 2: ROC-AUC scores for uncertainty quantification on the Countdown, GSM8K, MATH500 and SVAMP datasets with varing generation lengths. The table compares Monte Carlo simulation baseline with varying

N_{mc}

, the proposed DiSE, and the perplexity calculation using the auto-regressive model LLaMA3-Instruct-8B. The last column reports the average ROC-AUC scores across the preceding 12 settings.

Figure 8: Qualitative example of uncertainty quantification with four generated answers using LLaDA-Instruct-8B.

Figure 9: Ablation study of different DiSE selection modes for uncertainty quantification using LLaDA-Instruct-8B with a generation length of 512.

		Countdown			GSM8K			MATH500			SVAMP			Avg. Accuracy
	Method / Base Len	128	256	512	128	256	512	128	256	512	128	256	512	Avg. Accuracy
LLaDA-Instruct-8B	Baseline	26.17	15.23	12.50	68.01	76.65	79.23	26.20	32.80	36.80	84.67	85.00	83.67	52.24
	Baseline (Max Len)	25.00	16.41	15.62	69.29	76.80	78.85	25.60	31.60	36.40	85.33	84.67	83.00	52.38
	DiSE-flexible (ours)	27.73	18.36	15.62	70.96	79.68	79.30	26.00	33.60	36.60	87.33	86.00	84.33	53.79
LLaDA-1.5-8B	Baseline	24.22	15.62	17.19	70.51	77.48	79.53	26.80	34.00	36.80	87.00	84.67	86.67	53.37
	Baseline (Max Len)	24.22	17.58	17.58	71.95	78.77	79.53	25.80	34.20	37.00	86.33	83.00	86.33	53.52

Table 3: Results of flexible-length dLLM generation with DiSE on the Countdown, GSM8K, MATH500, and SVAMP datasets with varying base lengths. The table shows two fixed-length baselines: Baseline , generating sequences of base length

L

, and Baseline (Max Len) , generating sequences of length

L+M_{max}

. These are compared with the proposed flexible-length generation with DiSE (DiSE-flexible). The final column reports the average accuracy across the preceding 12 configurations.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

		Countdown			GSM8K			MATH500			SVAMP			Avg. Accuracy
	Method / Gen Len	128	256	512	128	256	512	128	256	512	128	256	512	Avg. Accuracy
LLaDA-Instruct-8B	Baseline	26.17	15.23	12.50	68.01	76.65	79.23	26.20	32.80	36.80	84.67	85.00	83.67	52.24
	MC, $N_{mc}=1$	24.61	21.48	17.19	68.84	78.17	80.59	25.80	33.80	36.40	84.67	84.00	85.00	53.38
	MC, $N_{mc}=32$	29.69	21.88	16.41	71.11	78.70	82.79	27.60	34.80	36.20	86.33	85.67	86.67	54.82
	DiSE (ours)	30.86	24.22	27.34	73.01	82.41	83.02	29.80	34.60	38.20	88.33	87.00	90.00	57.40
	LLaMA perplexity	30.86	17.19	11.33	74.22	79.61	81.20	28.60	35.40	34.80	88.33	86.67	85.67	54.49

Appendix

Table S1: Best-of-N sampling results for uncertainty quantification on the Countdown, GSM8K, MATH500, and SVAMP datasets with varying generation lengths. The table compares the baseline without best-of-N sampling, Monte Carlo simulation with varying

N_{mc}

, the proposed DiSE, and the perplexity calculation using the auto-regressive model LLaMA3-Instruct-8B. The last column reports the average accuracy across the preceding 12 settings.

Figure S1: Additional qualitative examples of uncertainty quantification using LLaDA-Instruct-8B. DiSE assigns higher scores to correct answers and lower scores to incorrect answers, while the Monte Carlo simulation ( $N_{mc}=32$ ) produces scores that do not consistently reflect correctness.

		Countdown			GSM8K			MATH500			SVAMP			Avg. ROC-AUC $\uparrow$
	Method / Gen Len	128	256	512	128	256	512	128	256	512	128	256	512	Avg. ROC-AUC $\uparrow$
LLaDA-Instruct-8B	DiSE (full)	0.616	0.672	0.698	0.597	0.585	0.560	0.514	0.555	0.517	0.665	0.549	0.571	0.592
LLaDA-Instruct-8B	DiSE (last-10)	0.578	0.521	0.622	0.633	0.644	0.658	0.611	0.634	0.604	0.688	0.692	0.755	0.637
LLaDA-1.5-8B	DiSE (full)	0.591	0.664	0.681	0.593	0.569	0.546	0.489	0.590	0.532	0.630	0.574	0.552	0.584
LLaDA-1.5-8B	DiSE (last-10)	0.610	0.471	0.586	0.610	0.616	0.613	0.606	0.553	0.533	0.599	0.629	0.677	0.592

Appendix

Table S2: Additional ROC-AUC results for uncertainty quantification to investigate the impact of different selection modes on performance. We evaluate two selection modes ‘full’ and ‘last-10’. The table reports ROC-AUC scores across the Countdown, GSM8K, MATH500, and SVAMP datasets with varying generation lengths, as well as the average ROC-AUC scores over all settings.

		Countdown			GSM8K			MATH500			SVAMP			Avg. Accuracy
	Method / Gen Len	128	256	512	128	256	512	128	256	512	128	256	512	Avg. Accuracy
LLaDA-Instruct-8B	DiSE (full)	30.86	28.52	27.34	71.87	79.76	79.53	27.20	34.60	34.20	87.67	85.33	87.00	56.16
LLaDA-Instruct-8B	DiSE (last-10)	30.86	24.22	27.34	73.01	82.41	83.02	29.80	34.60	38.20	88.33	87.00	90.00	57.40
LLaDA-1.5-8B	DiSE (full)	27.34	25.00	32.81	72.33	79.45	80.06	24.80	37.20	38.00	88.33	86.67	85.00	56.42
LLaDA-1.5-8B	DiSE (last-10)	29.30	17.97	28.91	74.53	81.96	83.55	28.60	34.40	37.40	88.00	86.33	87.67	56.55

Appendix

Table S3: Additional best-of-N sampling results for uncertainty quantification to investigate the impact of different selection modes on performance. We evaluate two selection modes ‘full’ and ‘last-10’. The table reports accuracy across the Countdown, GSM8K, MATH500, and SVAMP datasets with varying generation lengths, as well as the average accuracy over all settings.

	Countdown			GSM8K			MATH500			SVAMP			Avg. Accuracy	Avg. # NFE	Avg. Time(s)
Method / Base Len	128	256	512	128	256	512	128	256	512	128	256	512	Avg. Accuracy	Avg. # NFE	Avg. Time(s)
Baseline	26.17	15.23	12.50	68.01	76.65	79.23	26.20	32.80	36.80	84.67	85.00	83.67	52.24	149.3	3.27
DiSE-flexible (K=2)	27.34	18.36	16.02	70.43	79.30	79.23	26.40	34.00	36.60	87.33	86.00	84.67	53.81	182.3	3.86
DiSE-flexible (K=3)	27.34	17.97	15.62	70.74	79.61	79.23	25.80	33.80	36.60	87.33	86.00	84.33	53.70	199.3	4.18
DiSE-flexible (K=4)	27.73	18.36	15.62	70.96	79.68	79.30	26.00	33.60	36.60	87.33	86.00	84.33	53.79	215.3	4.55

Appendix

Table S4: Ablation study on flexible-length dLLM generation using LLaDA-Instruct-8B, analyzing the impact of different patience values

K

on performance. The table reports accuracy on the Countdown, GSM8K, MATH500, and SVAMP datasets with varying base lengths, along with the average accuracy, the average number of model forward passes and average time for each configuration.

	Countdown			GSM8K			MATH500			SVAMP			Avg. Accuracy	Avg. # NFE	Avg. Time(s)
Method / Base Len	128	256	512	128	256	512	128	256	512	128	256	512	Avg. Accuracy	Avg. # NFE	Avg. Time(s)
Baseline	24.22	15.62	17.19	70.51	77.48	79.53	26.80	34.00	36.80	87.00	84.67	86.67	53.37	149.3	3.25
DiSE-flexible (K=2)	24.61	17.19	19.14	72.40	79.23	80.06	27.40	35.60	37.40	87.33	84.67	87.00	54.34	182.0	3.84
DiSE-flexible (K=3)	24.61	19.14	20.70	72.48	79.45	80.06	27.80	35.40	37.40	87.00	84.67	87.00	54.64	198.9	4.20
DiSE-flexible (K=4)	26.17	19.53	22.27	72.33	79.53	80.06	27.20	35.60	37.40	87.00	85.00	87.00	54.92	214.7	4.53

Appendix

Table S5: Ablation study on flexible-length dLLM generation using LLaDA-1.5-8B, analyzing the impact of different patience values

K

Figure S2: Ablation study on the Countdown dataset using LLaDA-Instruct-8B with base length $L=512$ for flexible-length dLLM generation with DiSE, examining the effects of different patience $K$ and mask sizes $D$ on performance. The figure presents both accuracy and the average number of model forward passes for each configuration.

Figure S3: Ablation study on the Countdown dataset using LLaDA-1.5-8B with base length $L=512$ for flexible-length dLLM generation with DiSE, examining the effects of different patience $K$ and mask sizes $D$ on performance. The figure presents both accuracy and the average number of model forward passes for each configuration.

Figure S4: Comparison between our method and DAEDAL on the GSM8K and MATH500 datasets using LLaDA-Instruct-8B. The numbers in parentheses on the x-axis indicate the maximum output length for each setting.

	Method	ARC-Challenge	GPQA	# NFE
Dream-Instruct-7B	MC, $N_{mc}=1$	0.290	0.223	1
	MC, $N_{mc}=32$	0.316	0.259	32
	DiSE (ours)	0.472	0.297	1

Appendix

Table S6: Conditional likelihood estimation results on ARC-Challenge and GPQA using Dream-Instruct-7B. The table compares the proposed DiSE against the Monte Carlo simulation baseline with varying

N_{mc}

. The last column reports the average number of model forward passes per computation.

		Countdown		GSM8K		MATH500		SVAMP		Avg. ROC-AUC $\uparrow$
	Method / Gen Len	128	256	128	256	128	256	128	256	Avg. ROC-AUC $\uparrow$
Dream-Instruct-7B	MC, $N_{mc}=1$	0.463	0.548	0.473	0.503	0.538	0.464	0.510	0.551	0.506
	MC, $N_{mc}=32$	0.600	0.482	0.483	0.497	0.471	0.576	0.469	0.467	0.506
	DiSE (ours)	0.651	0.505	0.591	0.530	0.543	0.445	0.604	0.482	0.544

Appendix

Table S7: Additional ROC-AUC results for uncertainty quantification using Dream-Instruct-7B. The table reports ROC-AUC scores across the Countdown, GSM8K, MATH500, and SVAMP datasets with varying generation lengths, as well as the average ROC-AUC scores over all settings.

Figure S5: Boxplot of Pairwise Distribution Distances for GT, mask and random tokens using JS Divergence and Wasserstein Distance.

Figure S6: Cumulative distribution function (CDF) of GT token probability ranks for fixed sequence and position. (Part 1)

Figure S7: Cumulative distribution function (CDF) of GT token probability ranks for fixed sequence and position. (Part 2)

Figure S8: More examples for distributions of GT, mask, and random tokens and their comparison.

Language Model Self-Assessment

Language model self-assessment studies how models judge their own confidence, capabilities, errors, and behavior.

See all 93 papers

SelFusion: Self-distillation for Diffusion Language Models
Aug 24, 2026Hyeongsoo Lim, Jinyoung Kim, Eunseo Seo +2Diffusion Language Model Training Diffusion Language Models
Department of Artificial Intelligence, Chung-Ang University
Reasoning on the Manifold: Bidirectional Consistency for Self-Verification in Diffusion Language Models
Apr 17, 2026Jiaoyang Ruan, Xin Gao, Yinda Chen +5Diffusion Model Alignment Diffusion Language Model Inference
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China · University of Science and Technology of China, Hefei, China +1
DARE: Diffusion Language Model Activation Reuse for Efficient Inference
May 1, 2026Natalia Frumkin, Bokun Wang, Hung-Yueh Chiang +3Self-Attention Diffusion Language Model Inference
Chandra Family Department of Electrical and Computer Engineering The University of Texas at Austin · Cornell University

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

SelFusion: Self-distillation for Diffusion Language Models

Hyeongsoo Lim, Jinyoung Kim, Eunseo Seo, Minho Jang+1

Department of Artificial Intelligence, Chung-Ang University

Diffusion Language Model Training Diffusion Language Models Diffusion Language Model Distillation Self-Distillation Knowledge Distillation

Open paper AskPDF arXiv

Diffusion language models (DLMs) alleviate the inherent latency bottleneck of autoregressive (AR) large language models (LLMs), but their degraded generation quality limits practical applicability. Although knowledge distillation (KD) can be a promising direction for improving performance, we empirically find that naively applying conventional KD yields only marginal gains, or even degrades generation quality. Based on these observations, we propose a novel self-distillation framework for DLMs, namely SelFusion. To enable effective KD without an external teacher model, SelFusion performs two forward passes with different masking levels, defining the hard mode with a larger masking probability and the easy mode with a smaller masking probability. However, the easy mode is not always more accurate than the hard mode and can be overconfident on incorrect tokens. Thus, we introduce bidirectional KD between the two modes, which can dynamically determine the distillation direction based on token-level correctness. Experimental results on instruction-following tasks show that the proposed self-distillation substantially outperforms other KD methods with external LLM and DLM teachers. In many configurations, the student trained with SelFusion even surpasses the performance of the LLM teacher, providing a practical path toward improving DLM generation quality. Source code can be found at https://github.com/scai-research/SelFusion_official

Abstract

Language Model Self-Assessment

Explore similar work

SelFusion: Self-distillation for Diffusion Language Models

Reasoning on the Manifold: Bidirectional Consistency for Self-Verification in Diffusion Language Models

DARE: Diffusion Language Model Activation Reuse for Efficient Inference