acceptodds
Under review as a conference paper at ICLR 2027

TARING: Training-Anchored Reweighting for Masked Radiology Report Generation

Abstract

Radiology report generators can assign high probability to corpus-common content even when it is only weakly supported by the current examination. This ambiguity is particularly important for iterative masked generators, where token predictions depend jointly on image evidence and the evolving partial report. We introduce Training-Anchored Reweighting for Iterative Non-autoregressive Generation (TARING), a training-free decoding procedure that contrasts current-study prediction with a deterministic training-mean evidence reference under a matched generative state. At each unresolved position, the same frozen model evaluates the current study and the reference while holding the partial report, refinement time, report support, queried coordinates, and valid actions fixed. TARING reweights the current-study distribution according to the predictive contrast between the two contexts. A KL-regularized objective yields this update in closed form, providing a principled trade-off between reference-relative contrast and fidelity to the original current-study prediction. On MIMIC-CXR-JPG, TARING increases three-seed CheXbert micro-F1 from 0.512 ± 0.006 without reference guidance to 0.534 ± 0.002, primarily through an increase in recall from 0.489 ± 0.006 to 0.525 ± 0.002. Alternative empirical references yield closely matched F1 values, suggesting that the observed effect is not specific to one exact reference construction. Several higher-order lexical-overlap metrics decrease modestly, indicating a trade-off between extracted clinical-label agreement and reference-text similarity. These results characterize the behavior of training-anchored reweighting for one frozen generator and dataset and do not establish clinical factuality or cross-institution generalization.

Primary area:
generative models
Keywords:
radiology report generation, vision-language models, masked generative models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jvbeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the issue of corpus prior bias by proposing a training-free token reweighting strategy based on training-mean evidence reference comparisons, followed by iterative non-autoregressive report generation.

Critical strengths
  1. The paper’s training-free reweighting strategy for addressing corpus bias is quite interesting and demonstrates a degree of innovation.
  2. The ablation experiments are relatively comprehensive, including multiple ablations and sensitivity analyses.
Critical weaknesses
  1. Using the training-mean evidence as reference to solving corpus bias lacks theoretical and experimental support. The paper’s method emphasizes addressing corpus bias and uses the average image features as a reference, but it does not provide the relationship of using the training mean as a reference and corpus bias.
  2. The experiments are insufficient regarding whether the method effectively reduces the occurrence of bias. There is a lack of a detailed comparison to demonstrate that the method truly improves model performance by reducing bias occurrence.
  3. Practical value remains unvalidated. The paper notes that its method leads to an increase in model false positives; however, this may increase the risk of misinterpretation of diagnostic reports, and whether it ultimately improves report quality remains to be verified.
  4. Experimental evaluation is insufficient: the method was evaluated using only the MIMIC-CXR-JPG test set, which does not fully demonstrate its effectiveness.
  5. Why is it necessary to use non autoregressive methods in this paper, as previous non-autoregressive methods have not been better than autoregressive methods in Table 1. This paper does not explain the motivation clearly, nor does it have experimental support.
Questions for authors
  1. Does the subtraction of the training-mean reference prediction actually cancel corpus-common predictive support? The subtraction operation in the method implies this assumption, but the paper provides no direct experimental validation. Averaged image features lack interpretability on their own; you need to add the experiments to first clarify this point, then the subsequent logic will be reasonable.
  2. Does this method truly improve evaluation metrics by effectively reducing the occurrence of prior knowledge? Your method enhances performance by effectively reducing the occurrence of prior knowledge just based on the improvement in evaluation metrics; however, you need to demonstrate that TARING's performance gains stem from reduced reliance on corpus-common content. You need to add comparative experiments with baselines to demonstrate whether prior knowledge has indeed been reduced.
  3. Could the model’s false positives lead to a misleading impression? You need to add experiments to show that the improvement in metrics outweighs the increase in misleading impressions; otherwise, the method’s contribution to report generation is too low.
2 comments
AuthorsOpenReview ↗

Response to jvbe (1/2): Reference and Corpus Effects

We thank Reviewer jvbe for the questions on the mechanism.

Why the training mean stands for corpus-common predictions (W1, Q1). At 170 decoding states of 100 test studies (steps 10, 40 and 70 of 80), the mean context's top token matches the average prediction of 256 random training studies at 96.4% of positions, with KL 0.017, against 92.0% and 0.080 for a single training study. The submission gave only the motivation; this measures how close the reference is to the average training study's prediction.

A first-order expansion motivates this approximation. Deviations of training image contexts from their mean sum to zero, so the linear terms of each token logit cancel on average. Nonlinear remainder terms remain, and Appendix A.10 retains them in the exact relation between the TARING score and the score based on averaged training-study logits. The measurements quantify predictive agreement at sampled decoding states; they do not directly measure or bound the remainder. Section 3 and Appendix A.10 present this argument and its empirical evaluation.

Whether corpus-common content decreases (W2, Q2). Corpus-common content decreases, but these analyses do not establish that this reduction causes the F1 gain. Corpus-common sentences, those that appear in at least 1% of the 172,608 training reports, make up 25.7% of the no-reference reports and 21.9% of TARING's (a significant drop, 95% paired bootstrap over patients; radiologists' reports: 11.0%), over all 2,357 test studies and three reveal seeds. Atelectasis, which the mean context writes on its own, is the label whose false positives TARING cuts most (498 to 327 in Table 8; its false negatives rise, 465 to 520).

TARING also recovers findings that the training majority would suppress: on the 3,327 label instances whose report-derived label goes against the majority training label (seed 0), recall of the 216 positives doubles, from 0.107 to 0.218 (+0.111 [0.045, 0.173]), while agreement on explicit negatives falls from 0.551 to 0.496, so this is part of a general shift toward positive statements.

At s = 1.75, image removal and a random-study reference also reduce the corpus-common sentence share (20.3% and 21.7%) without larger F1 gains. Image noising gives 25.0% with a micro-F1 gain that does not differ significantly from the training mean's. On seed 0, the association between changes in sentence share and correct-label counts is weakly positive (Spearman rho = 0.06), so studies with larger reductions in common-sentence share gain fewer correct labels. Appendix D.8 reports this analysis.

With one prespecified permutation of patients' images, the micro-F1 gain is +0.013 [0.003, 0.023], compared with +0.019 [0.010, 0.028] using the correct images. The difference between gains is +0.006 [-0.007, 0.019]. These results do not establish that the improvement is caused by more accurate use of the current study's images. Appendix D.10 reports the shuffled-image analysis and fixed-state token-choice diagnostic.

AuthorsOpenReview ↗

Response to jvbe (2/2): Clinical Errors and Limitations

False positives (W3, Q3). TARING does add false findings, and the revised abstract states them next to the gain: per report, CheXbert-derived false negatives fall by 0.10 and false positives rise by 0.06 (Table 7); on seed 0, GREEN counts 0.19 more false findings, 0.14 fewer missing findings and 0.09 more significant errors, with a significantly lower GREEN score (Table 11; significant here and below: 95% paired bootstrap over patients). Over all three seeds (2,322 studies with every report parseable), GREEN counts 0.18 more false findings, 0.15 fewer missing findings and 0.03 more severity errors per report, and the change in GREEN score, −0.003-0.003 [−0.008-0.008, 0.002], is not significant (Appendix D.10). The guidance strength sets this balance (table in our reply to dcME; Appendix D). On our seed-0 re-run of all 2,357 test studies, s = 1.50 gives a gain as large as s = 1.75 (+0.018 and +0.016; Table 5 gives +0.022 for s = 1.75 over three seeds) with 1.24 false findings per report against 1.26 (1.19 without guidance; the two strengths are not tested against each other) and no significant RadGraph or GREEN loss. s = 1.75 stays because Section 4 selected it on validation micro-F1; choosing s = 1.50 now would use test results, so the paper reports it beside s = 1.75 instead. Whether the omissions saved outweigh the false findings added needs radiologists' judgement, which we could not obtain in this period.

Evaluation beyond MIMIC-CXR (W4). One dataset is not enough, we agree, and the added test does not establish transfer: on all 590 IU X-ray test studies (the R2Gen test split) decoded without retuning, averaged over three reveal seeds, micro-F1 is 0.447 without the reference and 0.454 with TARING (+0.007, not significant; whole-report scoring gives +0.003). The claim stays scoped to MIMIC-CXR.

Why a masked generator (W5). The contrast itself does not need one: CAD, VCD and M3ID apply it in autoregressive decoding, and we agree the submission did not say why it uses a masked generator. The reason is the setting of the abstract (lines 13–14): a masked generator fills positions anywhere in the report at every step, from text on both sides, so the reference prediction has to be made on the same partial report as the current one, which TARING specifies. The revised introduction states this reason.

Presentation. The paper should have used plain terms; coined terms are now replaced, the abstract and Section 3 give the decoding rule in one line, proof steps are left to the appendix, and a new Section 4 paragraph says how the image enters LLaDA.

Reviewer_adXaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose TARING, a decoding technique that requires no retraining. At each pending position in the clinical report, the same frozen model runs twice: once on the current study, with the actual image being reported, once on a reference input, which is an averaged "evidence" representation built from the mean of additional training data. Every other setting should remain identical across the two runs. After this stage the paper becomes very hard to parse, but the general idea is that the TARING re-weights its current-study predictions to favor tokens that the current image supports more strongly than the reference does.

Critical strengths

None

Critical weaknesses

This manuscript relies on unnecessary, convoluted, and verbose explanations to obscure the key ideas in the work. Inaccurate and unnecessary definitions are found all over the paper. For example, the paper names the output space "valid actions," suggesting an output space similar to an agent exploring or interacting with an unknown environment. This paper never mentions which are the valid actions, and from the task formulation and datasets one can infer that the output space is only language tokens.

Another example, lines 168 to 170 read:

"The mean is their Euclidean barycenter: it minimizes summed squared Frobenius distance and avoids selecting or retrieving a donor"

This statement simply means that Equation 1 calculates the mean for the X_n vectors (which is already explicit in Equation 1). There is no need to express this concept in such an overly complex manner, nor there is any need to introduce the word "donor", which makes no sense in this context and is never mentioned again in the paper.

Figure 2 provides yet another example, being overloaded with equations and unnecessarily complex terms for standard components. "training evidence" (an unnecessarily complex name for features or reference features), "slot-wise mean" (simply a mean over studies), "terminal head" (a meaningless term), "queried coordinates" (the index of a 1D array), and "valid actions" (again, the only goal is to generate text tokens).

To my understanding none of the proofs and derivations outlined in the appendix are necessary, as the method simply compares the outputs of two text generators, promoting or demoting candidate generations based on their relative scores.

Finally, the main task (Radiology report generation) maps visual input to text output using only the GSAI-ML/LLaDA-8B-Instruct. That model is a language model without any visual inputs. I can't understand how the visual features are embedded or even aligned with the proposed method. Despite the verbosity and over reliance on complex explanation in this paper, not a single implementation detail hints at the visual input, its processing or its alignment to the textual data.

Although the authors state that GPT and Codex were used to draft, plan, edit, and polish, the manuscript seems to be entirely the output of automated text generation tools instructed to obscure a simple method behind complex jargon, irrelevant proofs and over-complicated math. This pattern is so evident and repetitive that the entire submission feels more like a synthetic paper injected to test the reviewers' due diligence, or to deliberately waste their time.

Questions for authors

No amount of revision can upgrade this version of the paper to an acceptable standard, it must be entirely rewritten.

2 comments
AuthorsOpenReview ↗

Response to adXa (1/2): Architecture and Terminology

We thank Reviewer adXa for the specific comments on Figure 2 and Section 3.

How the image reaches LLaDA. The original submission described the image-conditioning components, but the exact interface to LLaDA was insufficiently explicit. Each view passes through the frozen RAD-DINO encoder and a learned 128-query resampler; a set encoder over up to three views, with view-position embeddings, produces a 128 × 768 image context. A LayerNorm and a linear layer map these 128 vectors to LLaDA's 4096-dimensional input space, where they are prepended to the 128 report-token embeddings together with one embedding of the decoding step.

Rank-16 LoRA on LLaDA's query, key, value and output projections is trained with the resampler, set encoder and projection under the masked-token loss; LLaDA's own weights stay frozen. During TARING decoding, the reference branch substitutes CR\mathbf{C}_R for CX\mathbf{C}_X at the 128×768128\times768 image-context interface, before the shared LayerNorm and linear projection. The step and report embeddings remain unchanged. The training-mean reference is constructed after training, and Appendix A.10 evaluates predictions under this reference. Section 4 and the redrawn Figure 2 now show this path.

Terminology. The original submission defined A\mathcal{A} as content tokens excluding special and model-only symbols. We have renamed it "candidate vocabulary" for clarity. In the revised main text, Figure 2, and Algorithm 1, "evidence context" becomes "image context", "training evidence" becomes "training studies' image contexts", "terminal head" becomes "length head", "queried coordinates" becomes "positions decoded at a step", "refinement time" becomes "decoding step", and "slot-wise mean" becomes "mean over training studies". Appendix A.1 explains the correspondence with the terminology retained in the proofs.

The rule is now stated in the abstract and in the first paragraph of Section 3: each masked position decoded at a step takes the token that maximizes ℓR+s(ℓX−ℓR)\ell_R + s(\ell_X - \ell_R); ℓX\ell_X and ℓR\ell_R are LLaDA's logits under the current study's image context and under the mean context, and s=1s = 1 gives the unguided model.

AuthorsOpenReview ↗

Response to adXa (2/2): Reproducibility and Novelty

How the paper was written. The AI-use statement discloses the assistance received, and the authors take responsibility for the text, code, and results. We rewrote the abstract and Sections 1–4, simplified Figures 1–2 and Algorithm 1, and replaced unnecessary terminology. The implemented procedure now precedes its mathematical interpretation. The checkpoint re-runs and their differences from the original results are reported below.

The results and their reproduction. Re-running the original submission's checkpoint gives micro-F1 0.5178 without the reference and 0.5342 with TARING (reveal seed 0, all 2,357 test studies), against 0.5178 and 0.5369 for the same seed in Table 6. The TARING number is 0.0027 lower, inside the tolerance we fixed before the run (0.004, twice Table 2's seed standard deviation for TARING), and the re-run's paired gain is a significant +0.0164 (95% paired bootstrap over patients) against +0.0192 in Table 6. GREEN on the re-run reports also matches Table 11: the GREEN score changes by −0.0069 against −0.0073 there, and false findings rise by 0.1858 against 0.1939.

Seeds 1 and 2 reproduce as well: 0.5066 and 0.5316 against 0.5064 and 0.5323, and 0.5125 and 0.5305 against 0.5118 and 0.5337. The re-run used other hardware and a newer PyTorch; the paper keeps its numbers and reports the re-run in Appendix D.

Lines 168–170 and Figure 2. Agreed on both: the barycenter and "donor" sentences are deleted, and Figure 2 is redrawn with the plain terms and two equations, the mean and the combination rule above. Appendix A.7 keeps the barycenter identity; the new "Choice of the mean" argument uses one step of it, that the deviations from the mean sum to zero.

The derivations. Equations 6–8 and Algorithm 1 specify the score, token selection, report length, and decoding procedure. Appendices A and B provide the derivations and their limits. The paper should have said so; the revision now says it at the start of each appendix, and Section 3 no longer carries their proof steps. One derivation matters for reading the rule: Eq. 5 is the closed-form solution of one KL-regularized step, and Eq. 6 is the same rule on logits, which makes s−1s - 1 the inverse weight on staying close to the unguided prediction.

What is new. The combination rule follows standard contrastive decoding and guidance, as acknowledged in Appendix G.3 (lines 1430–1433). What TARING adds is the choice of second prediction: the mean of the training studies' image contexts, computed once from training data and used at the current study's masked decoding state. Over three re-run seeds it raises CheXbert micro-F1 significantly (95% paired bootstrap over patients), by 0.020, through recall with precision not significantly changed. Against the inputs a CAD- or VCD-style contrast would use in the same model at s = 1.75, the mean has significantly higher precision, with fewer false findings, than removing the image and significantly higher recall than noising it, with micro-F1 not significantly different; one random training study is significantly worse at s = 1.75. Tuned on validation by the same seed-0 grid, the controls gain +0.017 (random study), +0.019 (image removed) and +0.001 (noised image) over no reference, against +0.020 for the mean. The comparisons leave the optimal reference unresolved. We retain the training mean as a fixed reference computed from training data.

The mean context also agrees with the average prediction of 256 training studies at 96.4% of top tokens over 170 decoding states.

Reviewer_dcMEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This article introduce a method called TARING, a training-free decoding method for masked radiology report generation. It contrasts predictions conditioned on the current examination with predictions conditioned on a fixed training set average image representation, while keeping the generation state matched. On MIMIC-CXR, it improves CheXbert micro F1 primarily through higher recall.

Critical strengths
  • Simple and well specified. The method requires no additional parameter updates or external expert model and can be applied to the frozen generator.
  • Controlled evaluation. The comparison with unguided decoding holds the checkpoint and decoding settings fixed, with consistent CheXbert gains across three reveal seeds and paired uncertainty estimates.
Critical weaknesses
  • Limited methodological novelty. The core reweighting follows established contrastive decoding and guidance methods, including CAD and VCD. The incremental contribution of the training-average reference and matched-state procedure is not sufficiently established.
  • Insufficient reference baselines. Training-average, validation-average, and representative-example references perform similarly. Without null, random, or corrupted-reference comparisons, the specific value of training anchoring remains unclear.
  • Uncertain overall benefit. CheXbert improvements coexist with lower RadGraph scores and worse seed-0 GREEN results, including more false findings. The evidence supports an omission–commission trade-off rather than a general improvement in report quality.
  • Narrow validation. Experiments cover one checkpoint and one institution-derived dataset. Reveal seeds do not establish robustness across model training runs, and prior test-set inspection limits the independence of the evaluation. Also situations for longtail diagnosis, out-of-distribution were not discussed.
Questions for authors
  1. What substantive contribution distinguishes TARING from existing contrastive decoding beyond the choice of reference and adaptation to masked generation?
  2. How does the training-average reference compare with null, random-example, and corrupted-image references under the same decoding and tuning budget?
  3. Given the similar performance of the three tested references, what evidence supports the importance of training anchoring?
  4. The authors acknowledge an omission–commission trade-off. Can a validation-selected parameter setting retain the CheXbert gains without degrading RadGraph or GREEN, or is this trade-off consistent across guidance strengths?
  5. Can the findings be replicated on an independent checkpoint or an untouched external dataset?
3 comments
AuthorsOpenReview ↗

Response to dcME (1/3): Reference Controls

We thank Reviewer dcME for the specific questions.

Null, corrupted, and random references (W2, Q2, Q3). We agree that the original submission lacked these controls. At s = 1.75, the noised-image reference has a slightly higher micro-F1 point estimate than the training mean (0.533 versus 0.532), but the difference is not significant. Image removal also does not differ significantly from the mean, while the random-study reference is significantly lower.

We additionally selected each control's strength using its best seed-0 validation micro-F1 over the same strength grid, then evaluated the selected settings over three reveal seeds. After selection, image removal and the random study do not differ significantly from the mean, while the selected noised-image setting is significantly lower. The mean additionally received a multi-seed validation comparison, so the full selection procedures were not identical. The revised Table 3 reports both comparisons.

ReferenceSelected smicro-F1ΔF1 vs. No referenceΔF1 vs. Training mean
Training mean (TARING)1.750.532+0.020†
Image removed (CAD/M3ID-style)1.250.531+0.019†−0.001
Noised image (VCD-style)1.250.514+0.001−0.019†
Random training study1.500.529+0.017†−0.003

At the mean's own strength (s = 1.75), the controls trade differently:

Reference (s = 1.75)PRmicro-F1Omissions per reportFalse findings per reportRadGraphΔF1 vs. No referenceΔF1 vs. Training mean
No reference (s = 1)0.5370.4890.5121.461.200.172
Training mean (TARING)0.5430.5220.5321.361.250.167+0.020†
Image removed0.5090.5430.5251.301.490.162+0.013†−0.007
Noised image0.5660.5030.5331.421.100.168+0.021†+0.001
Random training study0.5350.5130.5241.391.270.160+0.012†−0.008†

Conditions: re-runs of the original submission's checkpoint on all 2,357 test studies. Reported values are means over reveal seeds 0–2. The image-removed reference excludes image vectors from attention in the reference branch. The noised-image reference applies forward-diffusion noise at step 500 of 1,000 to the current images. The random reference uses one training study sampled once for each test study. CheXbert labels are scored against the radiologist-written report. Delta F1 is the difference between each condition's reported mean micro-F1 and that of either no reference or the training mean, as indicated by the column. The dagger marks a significant difference using a 95% paired patient-cluster bootstrap interval.

At s = 1.75, against the mean, removing the image significantly lowers precision and RadGraph and adds false findings (1.49 against 1.25 per report); noising it significantly lowers recall and leaves more omissions (1.42 against 1.36), in exchange for higher precision; one random training study is significantly lower in micro-F1, recall and RadGraph, while the mean-nearest validation study achieved 0.533 in Table 3 of the original submission. This original comparison is summarized in Section 5 of the revised paper. In micro-F1, removing the image is not significantly different from the mean at s = 1.75 or tuned, nor is the noised image at s = 1.75. Other references also improve F1. These comparisons do not establish an optimal reference.

A single fixed study is not a stable alternative. In an additional seed-0 run at s = 1.75 with three training studies fixed in advance, micro-F1 ranges from 0.490 to 0.553 (training mean, same seed: 0.537). A reference suppresses the findings it contains: the two studies that show cardiomegaly, atelectasis, effusion and support devices lower those four findings' positive rate in other patients' reports by 0.12 [0.11, 0.13], and the normal study, the one above the mean in micro-F1, lowers "No Finding" and has lower RadGraph and GREEN and 0.39 more false findings per report than the mean (Appendix D.10).

AuthorsOpenReview ↗

Response to dcME (2/3): Novelty and Guidance

What TARING adds to CAD and VCD (W1, Q1). The scoring rule follows CAD and VCD, as acknowledged in Appendix G.3 (lines 1430–1433). TARING uses the mean of training-study image contexts as a fixed reference at the current study's masked decoding state. The tables above compare the mean with CAD-style image removal and VCD-style image noising in the same model. After seed-0 validation selection, none of the controls has a higher micro-F1 point estimate than the mean. At s = 1.75, the mean has significantly higher recall than the noised-image reference and significantly higher precision than image removal, with fewer false findings. The training mean is computed once and reused for all studies. VCD encodes a noised copy for each study. The mean also matches the average training study's prediction: its top token agrees with that of 256 training studies' average logits at 96.4% of positions over 170 decoding states, against 92.0% for one training study and 77.6% with the image removed (revised Section 3, Appendix A). Shifting the reference's decoding step by one changes only 0.48% of guided tokens (Table 10), so we do not claim step matching as a contribution of its own. The reference changes content tokens only: report length comes from the current study, and no test report changed length under any reference we tested (Appendix D.10). Our contribution is this deterministic, training-only reference and the evaluation of its error trade-offs. Other references also improve micro-F1.

The trade-off across guidance strengths (W3, Q4). It holds at every tested strength in recall and false findings, and milder strengths retain F1 gains without statistically significant decreases in RadGraph or GREEN: on all 2,357 test studies, every guided strength significantly raises micro-F1 through recall with precision not significantly changed, false findings rise with s, and RadGraph and the GREEN score fall significantly only from s = 1.75 on (GREEN is significantly higher at s = 1.25). We agree it is a trade-off rather than a general improvement; the revised abstract says so. This does not establish equivalence, and false findings remain above the unguided baseline.

sRecallmicro-F1ΔF1 vs. No referenceOmissions per reportFalse findings per reportRadGraph (complete)GREEN
1.00 (no reference)0.4960.5181.441.190.1730.284
1.25 (new)0.5160.531+0.013†1.381.220.1710.290†
1.50 (new)0.5240.536+0.018†1.361.240.1720.285
1.75 (TARING)0.5250.534+0.016†1.351.260.168†0.277†
2.00 (new)0.5320.537+0.019†1.341.280.162†0.274†

Conditions: the original submission's checkpoint re-run on all test studies, reveal seed 0, training-mean reference; omissions and false findings are CheXbert labels against the radiologist's report, GREEN as in Table 11, on the re-run reports; ΔF1 is the micro-F1 at each strength minus that without a reference; † marks a significant difference from no reference (95% paired bootstrap over patients).

Before these runs we fixed, in a dated note, a rule for an additional milder strength: on validation, the largest micro-F1 gain among strengths whose RadGraph and GREEN are not below the unguided model's. It selects none, because validation RadGraph is lower at every guided strength (0.206 to 0.192 against 0.212), so GREEN was not needed there. We keep s = 1.75, the strength chosen on validation micro-F1 (Section 4), rather than pick one from these test results; s = 1.50, which validation ranks level with it (Table 4: 0.5469 against 0.5474 over three seeds), gives a similar observed F1 gain with no statistically significant decrease in RadGraph or GREEN, and Appendix D reports it next to s = 1.75.

AuthorsOpenReview ↗

Response to dcME (3/3): Generalization and Limitations

One checkpoint and one dataset (W4, Q5). Both limit the evidence, we agree, and the external test does not lift the second: on all 590 test studies of IU X-ray (the R2Gen test split), decoded with the MIMIC-CXR training mean and nothing retuned, TARING's micro-F1 is 0.454 against 0.447 without the reference over three reveal seeds (per seed −0.008, +0.008 and +0.019), a difference within the variation, not significant (scoring the whole generated report instead: +0.003, also not significant).

A second checkpoint could not be trained within the discussion period on the compute available to us; Appendix F.1 also records that the original preprocessing cannot be fully rebuilt, so a retrained model would differ in data handling as well as seed. The claim therefore stays scoped to this checkpoint and to MIMIC-CXR.

Earlier test inspection (W4). The reported configuration was selected on validation data only: the guidance strength and the length quantile were chosen exclusively on the validation split and fixed at (qEOS,s)=(0.60,1.75)(q_{\mathrm{EOS}}, s) = (0.60, 1.75) before the test set was evaluated with this configuration (Section 4, lines 277–285; Appendix D.1). The earlier inspection disclosed in Section 7 (lines 452–454) concerned other configurations during development; we still state it as a limit, because the test set is not a fully untouched holdout. IU X-ray does not have this limit: it was used only for the test evaluation above, with the configuration selected on MIMIC-CXR validation applied unchanged and no IU data used for training or tuning.

Rare findings and distribution shift (W4). The submission examined neither, we agree; it only said external-institution evaluation was needed (lines 455–456), and IU X-ray above is that test. The three rarest labels in the training reports are pleural other (1.1%), fracture (1.8%) and lung lesion (2.5%). Table 8 (test, seed 0) shows no gain on them: false negatives go from 78 to 76, 110 to 111 and 124 to 121; false positives rise for two (13 to 25, 13 to 31) and fall for fracture (48 to 36). Grouping 12 findings into thirds by training frequency (seed 0), F1 changes by −0.006-0.006 [−0.035-0.035, 0.022] in the rarest third, where the unguided model scores 0.061, against +0.023 [0.002, 0.045] and +0.018 [0.005, 0.031] in the other two (Appendix D.10). Section 7 now states both.

Presentation. Agreed: the revision uses plain terms throughout the main text, states the decoding rule in the abstract and in Section 3's first paragraph, and leaves proof steps that are not needed to apply the rule to the appendix.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response

We thank the reviewers. Each reply below is self-contained; new runs of the original submission's checkpoint reproduce Table 6 within 0.004 in micro-F1, the tolerance fixed before running. Line numbers in our replies refer to the original submission, as in the reviews.

Whether the training-mean reference matters (dcME, jvbe). No control tuned on validation by the same seed-0 grid beats the mean in micro-F1. Image removal and a random training study are not significantly different from the mean (95% paired bootstrap over patients), image removal at lower precision; the noised image falls significantly below it, though its strength was chosen by a 0.004 validation margin on one seed. At s = 1.75 the mean has significantly higher recall than the noised image and higher precision than image removal. These comparisons show that F1 gains are not unique to the training-mean reference. We retain the mean as a deterministic, training-only choice; the results do not establish its optimality or the mechanism responsible for the gains. Its top token is that of the average training study at 96.4% of positions.

The false findings that come with the gain (dcME, jvbe). TARING trades omissions for false findings, and the revised abstract says so. The guidance strength sets the balance: on the test set (seed 0), s = 1.50 gives a gain as large as s = 1.75 without a significant RadGraph or GREEN loss; s = 1.75 stays as chosen on validation, not by test results.

Evaluation beyond MIMIC-CXR (dcME, jvbe). On all 590 IU X-ray test studies without retuning, the micro-F1 change over three reveal seeds is +0.007, not significant; the claim stays scoped to MIMIC-CXR.

How the method works and how the image reaches LLaDA (adXa; presentation, all three). The revision states the decoding rule in the abstract and Section 3, replaces coined terms, moves proof steps to the appendix, explains in Section 4 how the image enters LLaDA, and redraws Figures 1–2.

One correction: the paper said the two branches run in one batch; the code runs two forward passes. Sections 3–4 now say so; no reported number depends on it. New results are in Table 3, Sections 6–7 and Appendix D; the supplementary material has a marked version, with additions in red and deletions struck through.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.