ECHO: Efficient Chest X-ray Report Generation with One-step Block Diffusion Lifeng Chen1,2,∗,⋆ , Tianqi You1,3,∗,⋆ , Hao Liu1,†,‡ , Zhimin Bao1 , Jile Jiao1 , Xiao Han1 , Zhicai Ou1 , Tao Sun1 , Xiaofeng Mou1 , Xiaojie Jin2 , Yi Xu1,†
arXiv:2604.09450v1 [cs.LG] 10 Apr 2026
1
AIRC, Midea Group, 2 Beijing Jiaotong University, 3 Dalian University of Technology Equal Contribution, † Corresponding Author, ‡ Project Leader, Work completed during an internship at Midea AI Research Center. ∗
⋆
Abstract Chest X-ray report generation (CXR-RG) has the potential to substantially alleviate radiologists’ workload. However, conventional autoregressive vision–language models (VLMs) suffer from high inference latency due to sequential token decoding. Diffusion-based models offer a promising alternative through parallel generation, but they still require multiple denoising iterations. Compressing multi-step denoising to a single step could further reduce latency, but often degrades textual coherence due to the mean-field bias introduced by token-factorized denoisers. To address this challenge, we propose ECHO, an efficient diffusion-based VLM (dVLM) for chest X-ray report generation. ECHO enables stable one-step-per-block inference via a novel Direct Conditional Distillation (DCD) framework, which mitigates the mean-field limitation by constructing unfactorized supervision from on-policy diffusion trajectories to encode joint token dependencies. In addition, we introduce a Response-Asymmetric Diffusion (RAD) training strategy that further improves training efficiency while maintaining model effectiveness. Extensive experiments demonstrate that ECHO surpasses state-of-the-art autoregressive methods, improving RaTE and SemScore by 64.33% and 60.58% respectively, while achieving an 8× inference speedup without compromising clinical accuracy. Keywords: CXR Report Generation , One-step Block Diffusion , Direct Conditional Distillation Date: April 13, 2026 Correspondence: Hao Liu, Yi Xu Project Page: https://echo-midea-airc.github.io/
1
Introduction
In recent years, Vision-Language Models (VLMs) [3, 19, 24, 26, 31, 42, 46], where visual features are aligned with language instructions to enable complex cross-modal understanding, have demonstrated significant progress in many fields, notably medical image analysis [6, 21, 25, 34, 37, 39, 44, 52, 59]. Within this field, automated chest X-ray report generation (CXR-RG) [7, 23, 35, 43, 48, 54] has emerged as a critical application. As one of the most common clinical imaging exams, CXR’s high volume places a heavy diagnostic burden on radiologists, creating strong demand for high-throughput automated reporting systems to ease workloads. Despite the promising performance achieved, canonical autoregressive (AR) VLMs often 1
Vanilla Mean-field
60
∏ �� (��,� |�� ) — factorized
Multi-step Diffusion
Autoregressive/ LLaDA-based
One-step Diffusion
ECHO (Ours）
55
CXR
bilateral
pleural
focal
CXR-Reports SemScore (%)
no
✗
�� ��,� �� , �<� — conditional
DCD (Ours)
Review the findings of this chest X-ray: [M]
[M]
(TPF: 8.0, SemScore: 53.4)
SDAR*
[M]
50
T3D
CD4LM
45
SDAR*1step
d3LLM 40
dParallel ECHO (Ours) 35
SDAR*
MedGemma-27B
[M]
Distillation Variants
25
right
lower
lobe
opacity
✓
15
(a) Illustration of the mean-field bias and our DCD
External Baselines
LLaDA-MedV 0
2
4
6
TPF (Tokens per forward)
8
10
(b) TPF vs. SemScore across methods
Figure 1 Motivation and performance of ECHO. (a) Decoding all tokens simultaneously in one step produces incoherent outputs, as standard diffusion models predict each position independently. Our Direct Conditional Distillation (DCD) distills from a non-factorized target, yielding coherent one-step-per-block outputs. (b) Compared to both autoregressive and diffusion-based baselines, ECHO achieves a favorable trade-off between generation quality (SemScore) and decoding throughput (tokens per forward pass). Here, SDAR* denotes SDAR [9] implemented under the multimodal large language model setting.
suffer from high inference latency due to their sequential decoding mechanism, which becomes a primary bottleneck for producing reports rapidly and at scale. Fortunately, thanks to parallel decoding capabilities, emerging diffusion-based Vision Language Models (dVLMs) [9, 28, 57, 58] offer a highly promising route toward achieving fast report generation. Despite the promise of parallel decoding, dVLMs in practice necessitate multiple denoising steps to ensure output coherence. This requirement stems from the mean-field approximation underlying their token-factorized denoisers, which introduces a structural bias that scales monotonically with the noise ratio. Although compressing decoding into a single step theoretically maximizes throughput, it compels the model to predict all tokens simultaneously from a fully masked input, precisely the scenario where mean-field bias is most acute. As illustrated in Fig. 1, the resulting inter-token incoherence leads to substantial quality degradation, prompting the critical question: Can we achieve the upper bound of decoding speed without compromising output fidelity?
To address this challenge, distillation emerges as a natural solution, seeking to transfer the quality of multistep decoding into a student model that operates in fewer passes. Specifically, several dLLMs [8, 12, 29, 36, 60] implement this via self-distillation, which aligns token-wise predictions with teacher’s ones from less corrupted states. However, these teacher targets remain factorized across positions, ignoring dependencies between concurrently predicted tokens. For few-step inference, this factorization remains tolerable, as progressive unmasking provides the inter-token context that token-independent targets fail to capture. In contrast, onestep decoding lacks such a corrective mechanism, allowing the full mean-field bias to resurface unchecked. Therefore, enabling reliable one-step decoding requires a fundamentally different training objective, one that is unfactorized and directly encodes the joint dependencies among tokens predicted in parallel. To achieve this goal, we propose a new Direct Conditional Distillation (DCD), which constructs supervision from the teacher’s on-policy remasking trajectory by conditioning each target on committed high-confidence context. The constructed supervision can encode inter-token dependencies into the training signal and enable stable one-step-per-block parallel inference. Building upon DCD, we present ECHO (Efficient Chest X-ray Report Generation with One-step Block Diffusion), a new foundational dVLM for CXR report generation that achieves both clinical accuracy and one-step inference efficiency. More concretely, we first establish an enhanced AR-based CXR-RG VLM by training on our curated data, where more information, such as report with more finding content or symptoms, is complemented. Based on this, we propose Response-Asymmetric Diffusion (RAD) adaptation to efficiently 2
converts the AR model into a block diffusion decoding paradigm. Subsequently, we apply DCD to distill this multi-step model into a one-step counterpart. Notably, compared with state-of-the-art autoregressive methods, ECHO improves RaTE and SemScore by 64.33% and 60.58%, respectively, while achieving up to an 8× speedup. Our main contributions are summarized below: • We present ECHO, a novel dVLM for CXR report generation that delivers strong clinical accuracy through one-step-per-block parallel decoding, outperforming pioneering both autoregressive and diffusionbased models by a large margin across standard benchmarks. • We propose a novel Direct Conditional Distillation (DCD), to our best knowledge the first one-step distillation framework for discrete diffusion language models, achieving up to a 390% inference speedup with only marginal quality degradation. • We design Response-Asymmetric Diffusion (RAD) adaptation, which further reduces theoretical training FLOPs by 72.3%, translating to a 3.61× speedup in training efficiency.
2
Related Works
2.1
Multimodal Medical Foundation Models
Large-scale vision-language models (VLMs) [3, 19, 24, 26, 31, 42, 46] have established strong multimodal understanding by aligning visual representations with language instructions, achieving remarkable performance across diverse visual reasoning tasks. Motivated by this progress, researchers have adapted VLMs to the medical domain [6, 21, 25, 34, 37, 39, 44, 52, 59], where precise alignment between clinical images and textual descriptions is essential. Within medical imaging, automated CXR report generation has emerged as a central task [7, 23, 35, 43, 48, 54], where existing methods predominantly follow an autoregressive paradigm and achieve strong clinical accuracy but at the cost of sequential decoding throughput, motivating the need for faster generation paradigms.
2.2
Diffusion Language Model
Discrete diffusion language models (dLLMs) represent one such paradigm, enabling parallel token prediction in place of sequential decoding. Specifically, dLLMs [2, 4, 33, 40, 55] define a forward process that progressively masks tokens and a learned reverse process that recovers them over a discrete token vocabulary. Building on this foundation, dVLMs [9, 28, 57, 58] extend the framework to vision-language tasks through a two-stage process of visual encoder alignment followed by instruction fine-tuning, enabling image-conditioned parallel generation. Block diffusion [1, 14, 17] further refines these models by adopting a semi-autoregressive decoding scheme that generates outputs block by block in causal order, naturally supporting variable-length sequence generation. A further line of work specifically addresses the high training cost of building such models from scratch by directly adapting pretrained autoregressive models into block diffusion models [4, 5, 9, 15, 49], offering a practical and scalable alternative. Collectively, these advances have improved both the training scalability and inference throughput of dLLMs and dVLMs.
2.3
Acceleration of dLLMs
As dLLMs and dVLMs continue to mature, further accelerating their inference has attracted growing research attention. This has been pursued from two complementary directions: inference-time optimization techniques [10, 18, 27, 32, 47, 50] that reduce per-step or per-token computation overhead, and trainingtime distillation [8, 12, 29, 36, 60] that compresses the multi-step denoising process into fewer steps. Along the distillation direction, methods such as SDTT [12] and dParallel [8] progressively align predictions at higher noise levels with those at lower ones, using either cross-entropy loss on self-generated trajectories or KL divergence between predicted distributions. Most recently, T3D [60] takes a distinct angle, employing a DPO-style optimization objective to directly penalize mean-field bias during training rather than matching noise-level predictions. Despite these advances, all existing methods retain token-factorized prediction targets and therefore still require multiple denoising steps to produce coherent outputs, leaving dLLM and dVLM 3
throughput far below its theoretical upper bound. ECHO addresses this gap with Direct Conditional Distillation (DCD), a distillation framework for dLLMs and dVLMs that supports one-step denoising, applied here to block diffusion.
3
Preliminaries
To motivate our approach, we formalize the dLLM framework and analyze how its token-factorized parameterization gives rise to the parallelism bottleneck identified in Section 1. Let V denote the vocabulary and L the sequence length. A token sequence is x0 = (x0,1 , . . . , x0,L ) ∈ V L . Discrete diffusion language models (dLLMs) [38, 40, 55] define a forward process q(x1:T | x0 ) that progressively masks tokens, and a reverse process pθ (x0:T −1 | xT ) that recovers them. Mean-field parameterization. Since the joint posterior p(x0 | xt , t) over V L requires exponentially many
parameters, existing dLLMs [1, 33, 38] universally adopt a token-factorized (mean-field) approximation [2, 51, 56, 62]: L Y pθ (x0 | xt , t) = pθ (x0,i | xt , t), (1) i=1
trained by minimizing the per-token denoising objective: # " L X log pθ (x0,i | xt , t) . LMF (θ) = Ex0 , t, xt −
(2)
i=1
As Eq. (2) decomposes into L independent per-position cross-entropies, the global optimum is attained when each factor independently matches the trueQconditional marginal, i.e. p⋆θ (x0,i | xt , t) = q(x0,i | xt , t). The resulting optimal factorized joint p⋆MF = i q(x0,i | xt , t) aligns each position’s marginal with the true posterior, but cannot capture the cross-positional correlations present in q(x0 | xt , t). We measure this gap by the mean-field bias: (3) ϵMF (xt , t) ≜ KL q(x0 | xt , t) p⋆MF (x0 | xt , t) , which quantifies the irreducible joint dependence structure across token positions that no factorized distribution can represent. Why multi-step sampling mitigates the bias. The expected mean-field bias ϵ̄MF (t) ≜ Ext [ϵMF (xt , t)] is
governed by the corruption level: when xt has few masked tokens, the remaining unknowns are sparsely distributed and largely conditionally independent, so ϵ̄MF (t) is small; when all tokens are masked (t = T ), the posterior exhibits strong cross-positional dependence and ϵ̄MF (T ) reaches its maximum. In general, ϵ̄MF (t) grows monotonically with the number of masked positions. [56] Multi-step reverse sampling systematically exploits this monotonicity. At each reverse step s, tokens decoded with high confidence from pθ (· | xs , s) are committed as observed context, reducing the number of masked positions from ms to ms−1 < ms . Consequently, the model at step s − 1 operates under a strictly reduced corruption level and, accordingly, a lower expected bias ϵ̄MF (s − 1) ≤ ϵ̄MF (s). By iterating this process, the multi-step chain progressively shifts decoding from the high-bias fully-masked regime toward a low-bias, nearly-observed state, thereby recovering the inter-token dependence structure that Eq. (1) otherwise discards.
4
Methodology
4.1
Overview
As illustrated in Fig. 2, the training pipeline of ECHO comprises three successive stages. In Stage 1, we perform continued pre-training (CPT) on Lingshu-7B [52] using a curated CXR corpus. This produces ECHOAR , an autoregressive vision-language model specialized in radiology report generation. Subsequently, we in Stage 2 propose Response-Asymmetric Diffusion (RAD) adaptation, which converts ECHOAR into ECHOBase , a block diffusion decoding model that retains the domain knowledge of ECHOAR , achieving an initial inference speedup. Lastly, in Stage 3, to further push decoding speed toward its theoretical upper bound, we apply Direct Conditional Distillation (DCD) to train ECHOBase toward one-step-per-block decoding, yielding the final ECHO model. 4
Data & Foundation
Before
Training Paradigm
Noisy x VP
VP
VP
Clean 𝑥 R
R
VP
VP
VP
R
R
VP
Base Model
Stage 1 Continual Pretraining
VP VP
KL Divergence
R R
ECHO-Base
VP VP
Normalized Report (Clean)
Chest X-Ray
Unnormalized Report (Raw) Findings: \n Lungs are clear. Cardiac silhouette mildly enlarged. Slight prominence of right paratracheal soft tissue possibly due to vascular structures. \n Impression: \n Mild cardiomegaly.
Findings: \n The thorax is symmetric bilaterally. The mediastinum and trachea are midline; slight prominence of right paratracheal soft tissue noted. ... lungs are clear with no definite abnormal opacities in either lung field. Cardiac silhouette is mildly enlarged ... Impression: \n Mild cardiomegaly. No significant abnormalities in the lungs, hila ...
Medical LLM
VP
ECHO
R R
Stage 2
After (Ours)
Response-Asymmetric Diffusion Adaption
VP
𝑥
x
Context VP
VP
R
R
R R
ECHO-Base
R R
VP
condition tokens
reserved tokens
mask tokens
logits distribution
VP VP
Stage 3 Direct Conditional Distillation
R
R
(I) Phase 1: Target Trajectory Construction
R
(II) Phase 1: Student One-Step Align
R
(a)
Response-Asymmetric Diffusion Adaption
(b) Direct Conditional Distillation (DCD)
Figure 2 Overview of the ECHO training pipeline. ECHO is built in three successive stages: continued pre-training (Stage 1) produces ECHOAR ; Response-Asymmetric Diffusion adaptation (Stage 2) converts it into the block diffusion model ECHOBase ; and Direct Conditional Distillation (Stage 3) distills ECHOBase into the final one-step-per-block model ECHO. (a) RAD duplicates only the response portion of the training sequence, avoiding the redundant duplication of long vision token contexts required by prior two-stage conversion methods. (b) DCD proceeds in two phases: the teacher’s confidence-heuristic remasking trajectory is collected to form a joint, non-factorized supervision target, which is then used to align the student’s single-step prediction via KL divergence.
4.2
From ECHO-AR to ECHO
Response-Asymmetric Diffusion Adaptation. While ECHOAR achieves high-quality report generation, its token-
by-token decoding mechanism limits inference throughput. To address this, our first objective is to convert ECHOAR into ECHOBase , a block diffusion model that preserves the domain knowledge acquired during pretraining. Prior work [9] has implemented this through a two-stage adaptation, i.e., a pretraining phase followed by SFT. In these methods, the entire training sequence, covering vision, instruction, and response tokens, is duplicated across both stages to construct block-causal teacher-forcing targets. For CXR report generation, where vision token sequences are substantially long, this full-sequence duplication incurs prohibitive training costs. To overcome this inefficiency, we propose Response-Asymmetric Diffusion (RAD), which achieves this adaptation through a single SFT stage. As illustrated in Fig. 2, RAD duplicates only the response portion of the training sequence, and constructs a block attention mask such that each noisy response block attends to all vision and instruction tokens as well as all previously decoded blocks. As illustrate in Fig. 3(b), this asymmetric design eliminates the redundant duplication of long vision token contexts, substantially reducing training FLOPs while consolidating the originally two-stage conversion into one. Furthermore, we conduct a series of experiments (Sec. 5.4) to investigate the relationship between training data volume, model quality, and inference speed during the RAD conversion. We find that ECHOBase attains performance comparable to ECHOAR using only a small fraction of the original training data, demonstrating the high knowledge-transfer efficiency of RAD. Direct Conditional Distillation. With ECHOBase established, our next objective is to convert it into a model
capable of one-step-per-block decoding, fully maximizing inference throughput. However, as discussed in Sec. 3, discrete diffusion language models inherently rely on multi-step decoding to progressively reduce the mean-field bias, and our dVLM setting is no exception. This calls for a training target that is itself nonfactorized: by aligning the student to such a target, the inter-token dependencies accumulated across the multi-step denoising trajectory can be captured in a single forward pass. To this end, we propose Direct Conditional Distillation (DCD), which constructs this non-factorized target by collecting and stitching the high-confidence token distributions at each step of the teacher’s multi-step denoising process, forming a joint supervision signal that encodes cross-token dependencies.
The full self-distillation procedure is detailed in Algorithm 1 and consists of the following two phases, applied iteratively during training.
5
Algorithm 1: Direct Conditional Distillation (DCD) Input : Context xctx , teacher θ, student ϕ, N blocks, block length L, sampling confidence threshold τ Output : Updated student ϕ
// Phase 1: On-policy Teacher Trajectory Collection for n = 1, . . . , N do 2 Initialize B ← {1, . . . , L},; 3 xcurr ← (xctx , b̂1:n−1 ); 4 while B ̸= ∅ do 5 Query teacher pθ (xi | xcurr ) for all i ∈ B; 6 U ← {i ∈ B | ci ≥ τ };; 7 if U = ∅ then U ← {arg maxi∈B ci }; 1
(n)
Ptch [i] ← pθ (xi | xcurr ),; x̂i ← arg max pθ (xi | xcurr ) for all i ∈ U ; xcurr ← xcurr ∪ {x̂i }i∈U ,; B ← B \ U;
8 9 10 11
b̂n ← (x̂1 , . . . , x̂L );
12
// Phase 2: Student One-Step Alignment (cf. Sec. 4.2 and Fig. 2) 13 Construct RAD training sequence xtrain using (xctx , b̂1:N ); X (n) 14 LDCD ← DKL Ptch [i] pϕ (bn,i | xtrain ) ; n,i 15
Update ϕ ← Optimizer(ϕ, ∇ϕ LDCD );
Phase 1: On-policy Teacher Trajectory Collection. The teacher trajectory is the trajectory sampled by its multi-step denoising process. To obtain this, we first run ECHOBase through a confidence-heuristic denoising
process, where tokens are progressively unmasked in order of their prediction confidence ci ≜ maxv pθ (xi =v | xcurr ). Specifically, at each step of a block’s denoising, we record the predicted distribution and the committed token label for every position as it is unmasked. The collected token labels form the pseudo labels b̂n , which are used to construct the training sequence for the subsequent student alignment phase. The collected (n) distributions are then stitched together into the joint target Ptch for the n-th block, as the distillation target in Phase 2. Phase 2: Student One-Step Alignment. With the pseudo-labels and stitched joint distribution collected
in Phase 1, the next step is to align this distribution with the student’s one-step prediction. Concretely, we first use the pseudo-labels to construct a block teacher-forcing training sequence following the RAD scheme (Sec. 4.2), ensuring that the student makes its prediction under exactly the same conditions as the L (n) teacher. Letting bn,i denote the token at position i of block n and Qϕ ≜ pϕ (bn,i | xtrain ) i=1 the student’s one-step distribution over that block, we minimize the forward Kullback–Leibler (KL) divergence predicted (n) (n) for each block and aggregate the loss over all blocks. Intuitively, a token that is unmasked DKL Ptch ∥ Qϕ at a later step within a block requires stronger conditioning to reach a high-confidence prediction, indicating that its position is subject to more severe mean-field bias. Therefore, we introduce a token reweighting scheme that assigns each position a weight proportional to the step at which it is unmasked. Under this scheme, positions with larger mean-field bias receive stronger supervision, while tokens committed at early steps serve as a regularization signal for the alignment process. After DCD, we can generate each block with only one step, acheiving fast decoding speed without compromising clinical accuracy.
4.3
Hallucination Mitigation
Hallucination is a critical concern in CXR report generation, where inaccurate descriptions of clinical findings can directly undermine diagnostic reliability. During the development of our vision-language model, we 6
9× 0.822
80
0.880
0.8
0.6
0.4
0.2
0.0
60
5× 40
3×
20 High Resolution Image Section
Natural Image Inputs Section
Token type Content Tokens EOS Token
blk4
7× CXR (~2,870 vis. tokens) −73.0% FLOPs 3.70× speedup
0
Training Speedup (× )
0.877
Attention FLOPs Saved (%)
Confidence (mean ± std)
1.0
0.753
0
500
1,000
1,500
2,000
2,500
-
3,000
3,500
4,000
1×
Vision Token Count
blk8
(a) EOS Token Confidence: blk4 vs blk8
(b) RAD Efficiency vs. Vision Context Length
Figure 3 (a) Confidence analysis of <eos> tokens vs. content tokens under block sizes blk4 and blk8 in standard multi-step inference.(b) RAD attention FLOPs saved (%) and training speedup (×) as a function of vision token count.
identify two predominant hallucination patterns in medical dLLMs: inaccurate identification of symptoms, and degenerate repetition loops that fail to terminate. For inaccurate symptom identification, we attribute this pattern to an implicit negative bias inherent in the training data. In routine clinical practice, radiologists follow a “reporting by exception” convention, describing only abnormal findings while omitting normal structures or dismissing them with a single catchall phrase such as “no other significant abnormality.” Consequently, for the majority of anatomical regions that are clinically unremarkable, their normal status is absent from the report rather than explicitly stated as negative. This systematic omission leaves the model without explicit negative evidence during training, leaving its conditional distribution under-constrained. At inference time, this under-constraint manifests as two complementary failure modes: false-positive hallucinations, where the model fabricates findings for normal regions, and false-negative omissions, where genuine abnormalities are overlooked. To address this, we propose a data normalization paradigm tailored for CXR report generation. Specifically, we reformulate every training report so that each predefined anatomical region receives an explicit annotation, either a positive finding or a negative assertion. This ensures unambiguous supervision at every position, directly eliminating the implicit omission bias described above. Our experiments in Sec. 5.6 show that the hallucination reduction brought by data normalization consistently benefits both ECHOAR and the final ECHO model. Beyond inaccurate symptom identification, ECHO also suffers from degenerate repetition loops that fail to terminate. To understand the root cause, we examine ECHOBase ’s prediction confidence of content tokens and <eos> tokens under standard confidence-heuristic inference. As shown in Fig. 3(a), <eos> tokens exhibit systematically lower mean confidence and substantially greater variance compared to content tokens. Using such a flat and unstable distribution as the distillation target for the <eos> token makes it difficult for the one-step model to terminate generation with high confidence. This problem is further exacerbated as the block size grows: as shown in Fig. 3(a), increasing the block size leaves content token confidence nearly unaffected while causing a notable decline in <eos> confidence. To address this, we apply an additional cross-entropy loss on the <eos> token during distillation, explicitly pulling its predicted distribution toward a sharp, high-confidence one-hot target. Together, the two hallucination mitigation strategies consistently reduce generation errors in ECHO, contributing to improved clinical accuracy.
4.4
Inference Paradigm
To further maximize the inference throughput of ECHO, we optimize the inference paradigm used in dLLMs. A common technique in semi-autoregressive decoding is block KV cache [50], which caches the key-value states of previously decoded blocks to avoid redundant recomputation. As illustrated in Fig. 4(a), after all tokens of 7
Prompt
Block 0
Block 1
Prompt
Block 0
Block 1
Prompt token
Cached token
Decoded token Compute cache Decode cache
(a) Vanilla Block KV Cache
(b) Fused Block KV Cache
Figure 4 (a) Vanilla block KV cache: after all tokens of a block are committed, a dedicated forward pass is performed to update the KV cache. (b) Fused block KV cache: the KV cache update for the preceding block is fused into the current block’s denoising forward, eliminating the dedicated KV update pass.
a block are committed, a dedicated forward pass is performed solely to update the KV cache with the newly decoded block. This additional forward pass incurs acceptable inference time overhead in multi-step decoding, but for our one-step model it doubles the total number of forward passes for each block. To eliminate this overhead, we propose Fused Block KV Cache. As illustrated in Fig. 4(b), after a block is decoded, we defer its KV cache update and fuse it into the next block’s denoising forward. In this fused forward, the model simultaneously computes and caches the key-value states for the preceding decoded block while denoising the current block’s fully-masked tokens, removing the dedicated KV update pass entirely. As proven in Sec. E, Fused Block KV Cache introduces no additional FLOPs while halving the number of forward passes, directly reducing inference latency.
5
Experiments
In this section, we present experimental results to assess the effectiveness of ECHO for fast and reliable CXR report generation. We compare ECHO against state-of-the-art autoregressive baselines on report quality, and against diffusion-based distillation baselines on distillation efficiency. We then ablate the individual components of DCD, and analyze the effect of training data volume during RAD adaptation and the impact of report normalization across all training stages.
5.1
Experimental Setups
5.1.1
Dataset
We conduct experiments on a unified CXR report corpus built from multiple public datasets, including MIMIC-CXR [22], CheXpert-Plus [20], ReXGradient [61], and IU-Xray [11]. For training, we apply a standardized cleaning and normalization pipeline across all source datasets to construct a bilingual training set, with a small subset of LLaVA-ReCap-558K [26] included to mitigate catastrophic forgetting. For evaluation, we sample 2,000 English and 2,000 Chinese reports from each of normalized MIMIC-CXR, CheXpert-Plus, and ReXGradient, ensuring balanced coverage across datasets and languages. Full preprocessing details and data statistics are provided in Appendix. B.1. 5.1.2
Models
All ECHO models are initialized from Lingshu-7B [52], an autoregressive VLM pretrained on large-scale medical corpora, which provides a strong domain-specific knowledge and vision alignment for CXR report generation. 8
Table 1 Performance comparison with state-of-the-art models. We report average metrics across the MIMICCXR, CheXpert-Plus, and ReXGradient test sets. Bold values indicate the best performance. Please refer to Appendix. D for detailed results on individual datasets. CheXpert-Plus
Methods
ReXGradient
MIMIC-CXR
Speed
ROUGE-L CIDEr RateScore SemScore ROUGE-L CIDEr RateScore SemScore ROUGE-L CIDEr RateScore SemScore
TPF
TPS
Proprietary general models
Gemini3-Pro [16] Qwen3-Max [3]
26.95 27.19
1.53 1.53
40.86 41.63
29.72 29.24
32.10 30.50
1.81 1.72
33.52 39.12
LLaVA-Med [25] Lingshu-7B [52] Hulu-Med-7B [21] MedGemma-27B [39] Lingshu-32B [52] Hulu-Med-32B [21]
6.92 21.95 22.38 20.30 20.40 19.46
0.65 1.12 1.15 0.79 1.06 0.96
10.11 31.34 26.25 38.83 30.92 27.17
11.39 27.54 23.43 31.79 24.32 22.07
7.36 23.63 21.66 23.41 22.10 22.31
LLaDA-MedV [13] ECHOBase CD4LM [29] d3LLM [36] dParallel [8] T3D [60] ECHOblk8 ECHOblk4
7.69 0.22 26.72 56.90 4.25 59.12 13%↓ 20%↓ 49.21 3.38 55.905%↓ 44.0422%↓ 2.9231%↓ 55.396%↓ 42.2225%↓ 3.1725%↓ 44.6524%↓ 54.993%↓ 4.064%↓ 56.903%↓ 55.212%↓ 4.142%↓ 56.853%↓ 56.141%↓ 4.201%↓ 57.402%↓
17.76 52.64 43.8616%↓ 40.3523%↓ 36.7230%↓ 49.865%↓ 51.402%↓ 49.575%↓
9.72 0.23 18.19 66.46 5.64 63.02 11%↓ 21%↓ 58.73 4.44 58.367%↓ 18%↓ 26%↓ 54.18 4.17 57.997%↓ 48.7326%↓ 4.0228%↓ 45.5027%↓ 64.363%↓ 5.237%↓ 58.636%↓ 65.132%↓ 5.482%↓ 59.924%↓ 66.390%↓ 5.541%↓ 62.181%↓
39.02 38.13
27.08 26.86
1.55 1.51
41.42 40.78
30.45 27.32
×1.0 ×1.0
– –
17.04 34.82 23.56 37.74 31.49 26.67
6.69 22.24 22.26 20.80 20.73 17.96
0.65 1.16 1.18 0.82 1.10 0.99
10.07 32.79 22.89 38.48 32.40 24.54
9.22 26.94 22.30 30.24 24.14 18.23
×1.0 ×1.0 ×1.0 ×1.0 ×1.0 ×1.0
36.33 53.70 38.78 16.78 23.60 15.67
22.09 66.68 55.6216%↓ 53.5019%↓ 54.5318%↓ 61.657%↓ 64.004%↓ 66.280%↓
9.60 0.23 27.47 54.45 4.16 55.85 15%↓ 24%↓ 46.21 3.13 52.116%↓ 41.1324%↓ 2.6536%↓ 50.998%↓ 39.9826%↓ 2.9628%↓ 40.4727%↓ 52.383%↓ 3.857%↓ 52.386%↓ 53.072%↓ 4.003%↓ 52.975%↓ 54.120%↓ 4.052%↓ 53.953%↓
17.26 ×1.30 46.10 ×1.89 15%↓ 38.76 ×3.61 25%↓ 34.20 ×3.84 31.1532%↓×4.42 42.517%↓ ×2.00 44.832%↓ ×8.00 45.571%↓ ×4.00
12.86 48.86 98.49 101.64 110.12 60.25 274.21 129.19
Autoregressive Medical Models
0.72 1.25 1.14 0.91 1.20 1.11
10.01 20.26 22.77 26.00 20.33 23.01
Diffusion Medical Methods
5.1.3
Implementation Details
For RAD, we use the same data as the SFT stage in continued pre-training. The vision encoder and projector weights of ECHOAR are frozen throughout, and only the LLM backbone is trained. For DCD, we randomly sample 30,000 examples from the RAD training set, drawn proportionally across datasets. During teacher trajectory collection, samples that produce degenerate repetition loops are automatically discarded. 5.1.4
Evaluation Metrics
To provide a comprehensive assessment of the generated chest X-ray reports, we evaluate our model across three complementary dimensions: linguistic quality, clinical fidelity, and structural stability. These metrics enable a rigorous comparison between ECHO and existing state-of-the-art (SOTA) models while specifically addressing the unique challenges of diffusion-based generation. • Linguistic Quality Metrics (LQM). We use standard natural language generation NLG metrics, including ROUGE-L [30] and CIDEr [45], to measure lexical overlap and fluency between generated and reference reports, enabling direct comparison with prior methods. • Clinical Fidelity Metrics (CFM). Linguistic fluency does not guarantee accurate symptom identification, so we use RaTEScore [63] and SemScore [41] to assess clinical content fidelity. Since our evaluation set consists of normalized reports, a fair comparison across models requires focusing solely on symptom identification accuracy rather than reporting style. To this end, we only adopt the positive-finding score of RaTEScore, and SemScore is insensitive to negative findings by design. • Structural Stability Metrics (SSM). To assess generation stability specific to discrete diffusion models, we report Perplexity (PPL) computed with Qwen3-1.7B as the judge model. Lower PPL indicates more stable and fluent generation.
5.2
Comparison with State-of-the-art Methods
In Table 1, to evaluate the effectiveness of ECHO, we compare it against three categories of state-of-the-art models: (1) General-purpose proprietary models, including Gemini3-Pro [16] and Qwen3VL-Max [3]. (2) Autoregressive (AR) medical VLMs, such as LLaVA-Med v1.5-7B [25], Lingshu-7B/32B [52], MedGemma27B [39], and Hulu-Med-7B/32B [21]. (3) Diffusion-based medical models, including LLaDA-MedV-8B [13] and several distilled variants [8, 29, 36, 60]. These variants share the same base model ECHOBase and distillation data, but with different distillation targets. 9
Table 2 Ablation on DCD components. SW: step-wise token reweighting; CE: cross-entropy loss on the <eos> token; RKL: reverse KL divergence. Results are reported on CheXpert-Plus, ReXGradient, and MIMIC-CXR. DCD settings
CheXpert-Plus
ReXGradient
MIMIC-CXR
SW CE RKL ROUGE-L CIDEr RateScore SemScore ROUGE-L CIDEr RateScore SemScore ROUGE-L CIDEr RateScore SemScore
✓ ✓ ✓ ✓
✓ ✓
✓ ✓
52.57 52.44 56.14 52.51 53.25
3.61 3.53 4.20 3.41 3.72
54.87 56.30 57.40 55.20 57.25
46.93 46.66 49.57 45.32 47.07
63.28 62.90 66.39 61.71 63.82
5.17 5.07 5.54 4.94 5.21
61.32 61.33 62.18 61.26 61.47
58.36 59.90 66.28 59.55 63.46
51.73 51.76 54.12 50.91 51.48
3.63 3.65 4.05 3.48 3.67
51.19 53.14 53.95 53.31 53.42
43.50 43.70 45.57 42.43 43.86
PPL
23.72 21.07 18.83 20.23 21.32
Overall Performance. ECHO consistently achieves superior performance compared to both general and special-
ized models. Notably, it significantly outperforms larger medical VLMs like MedGemma-27B, ranging from 17% to 40% in CFM. Even when compared to the most powerful models like Gemini3-Pro and Qwen3VL-Max, ECHO maintains a clear advantage in all medical clinical metrics. Efficiency of Distillation. To further differentiate distillation strategies, we compare all methods under the challenging block size L = 8 configuration. As shown in Tab. 1, ECHOblk8 achieves an 8× decoding speedup with at most 2–5% quality degradation relative to ECHOBase across all metrics. In contrast, the closest qualitypreserving competitor T3D incurs a comparable quality drop yet delivers only a 2× speedup, and speedcomparable methods such as dParallel lose 18–32% in clinical fidelity metrics at a 4.4× speedup. ECHOblk4 further demonstrates that at a 4× speedup, quality degradation can be reduced to within 1.25% on average, confirming that DCD consistently achieves a more favorable quality–speed trade-off than existing distillation methods.
5.3
Ablation Study of Distillation Component
To investigate the individual contributions of the components in our proposed Direct Conditional Distillation (DCD) strategy, we conduct extensive ablation studies. Specifically, we evaluate the impact of Step-wise Weighting (SW), Cross-Entropy loss for the <eos> token (CE), and Reverse KL divergence (RKL). The results across three benchmarks are summarized in Tab. 2. Effect of Step-wise Weighting (SW). As shown in row 2 of Tab. 2, SW consistently improves generation
stability across all datasets. Upweighting tokens that are unmasked later in the denoising trajectory, which tend to carry stronger inter-token dependencies, reduces PPL from 23.72 to 21.07. RaTEScore also improves steadily, e.g., from 54.87 to 56.30 on CheXpert-Plus, indicating that the reweighting helps the model better capture the conditioning provided by earlier committed tokens. Effect of <eos> Cross-Entropy Loss (CE). Adding CE supervision on the <eos> token produces the largest
single improvement across all settings. As discussed in Sec. 4.2, one-step diffusion models tend to assign low and unstable confidence to <eos> positions, leading to degenerate repetition loops. Explicitly supervising the <eos> token with a one-hot target directly addresses this failure mode, resulting in a substantial gain in ROUGE-L from 52.44 to 56.14 on CheXpert-Plus and CIDEr from 3.65 to 4.05 on MIMIC-CXR. PPL drops further to 18.83, its lowest value across all configurations, confirming that reliable termination is essential for overall report quality. Effect of Reverse KL (RKL). We also replace the forward KL objective with reverse KL (RKL) to examine whether its mode-seeking behavior benefits distillation. As shown in rows 4 and 5 of Tab. 2, adding RKL consistently degrades performance. Compared to the SW-only baseline, CIDEr drops from 3.65 to 3.48 and SemScore from 43.70 to 42.43 on MIMIC-CXR. A likely cause is that clinical reports require the model to cover all plausible findings rather than concentrate probability mass on a single mode, which is the tendency encouraged by reverse KL. Forward KL, which preserves the full teacher distribution, is therefore better suited for this task.
5.4
Data Scale Analysis for RAD
We study how the amount of training data used in the RAD stage affects both model quality and inference throughput, by evaluating checkpoints at different training steps across all benchmarks. Fig. 5 reports five 10