acceptodds
Withdrawn from ICLR 2027

Finite Clinical Programs: An Inspectable Discrete Interface for Radiology Report Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We study radiology report generation with a complete case-dependent conditioning message reconstructible from discrete IDs. Finite Clinical Subspace Learning (FCSL) combines named clinical states, task and prior-study fields, and quantized residual codes, using 202 bits on MIMIC-RG4 and 160 bits on MIMIC-ABN. A 0.8B language backbone accesses this message through a prefix and layer-local writes. Training uses reference-derived states; inference uses predicted IDs. Three-seed system comparisons improve clinical-label agreement but yield lower language and RadGraph scores. A supervision-matched continuous interface attains similar micro-F1 and higher complementary scores. However, matched RG4 interventions on 3,696 shared eligible requests from 245 patients yield 33.05% joint selective success for FCSL versus 18.60% for continuous conditioning: a paired difference of 14.44 percentage points (95% patient-bootstrap interval [11.87, 16.63]). Joint success requires the requested target change with all thirteen non-target labels unchanged. FCSL also exceeds prefix-only's 29.63%, without a demonstrated reduction in non-target changes. These single-seed, exploratory interventions reveal a report-quality–selective-response trade-off on the common eligible subset, rather than universal accuracy gains or clinical safety. FCSL makes complete conditioning reconstructible and its editing benefits and limits measurable at a compact backbone scale.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
radiology report generation, discrete representation learning, concept bottleneck models, multimodal learning, interpretable machine learning

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

This paper proposes Finite Clinical Subspace Learning (FCSL), a discrete conditioning framework for radiology report generation. The method represents visual evidence using 14 predefined clinical states and quantized residual codes, which are integrated into a language model through global prefixes and layer-wise updates. Experiments on two chest X-ray datasets demonstrate improved clinical-label agreement over a continuous-prefix baseline and better selective editing performance. However, these improvements come at the cost of lower language generation and RadGraph scores.

Critical strengths
  1. The paper explores an interesting direction by introducing an explicit, inspectable discrete representation for radiology report generation, with potential benefits for interpretability and controllability.

  2. The experimental evaluation is relatively comprehensive, including multiple datasets, continuous-conditioning baselines, ablation studies, and targeted clinical-state interventions.

Critical weaknesses
  1. Insufficient justification of the representation design. The proposed 202-bit representation relies on several manually specified design choices. While the 14 clinical states are derived from the CheXbert taxonomy, the selection of 8 residual queries, two quantization stages, and 256-entry codebooks lacks sufficient justification. The current ablations do not adequately establish the necessity or optimality of these configurations. A systematic analysis of the representation capacity would strengthen the methodological contribution.
  2. Limited validation across backbone architectures. The experiments are restricted to the Qwen3.5 model family. Given the potential influence of backbone pre-training and post-training, the comparisons with existing methods using different architectures are not sufficiently controlled. Evaluation with alternative language model backbones would provide stronger evidence for the generalizability of the proposed framework.
Questions for authors

Please refer to the weaknesses

2 comments

W1. How sensitive are the results to residual capacity? The two task fields, fourteen current states, and fourteen prior states account for 74 bits on RG4. The design choice concerns the residual block. Full uses eight queries, two quantization stages per query, and 256 entries per codebook. The sixteen 8-bit indices add 128 bits, giving 202 bits in total (Eq. 7). We fixed this configuration during interface design and did not tune it by grid search.

Table 31 in Appendix C.13 compares four budgets on the 6,215 RG4 test cases (seed 42, validation-selected checkpoints). R0, R1, and R3 share one training recipe and code version. Full is the original seed-42 run.

ConfigurationQueries × stages × entriesTotal bitsMicro-F1BLEU-4ROUGE-LRadGraph
R0No residual block740.63570.13110.32590.1978
R14 × 2 × 641220.63780.17370.36750.2416
Full8 × 2 × 2562020.63900.17660.36900.2421
R316 × 2 × 2563300.63660.17750.37160.2467

Adding residuals raises BLEU-4 from 0.1311 to 0.1737–0.1775 and RadGraph from 0.1978 to 0.2416–0.2467. Retraining Full in the matched-control batch gives micro-F1 0.6340 (Table 2), versus 0.6390 here. This single-seed comparison cannot resolve small capacity effects on micro-F1. All three residual configurations use two stages, and R1 changes both the query count and the codebook size relative to Full. The sweep therefore leaves the choice of stage count and the optimal budget open.

We also ran single-state edits on a fixed subset of 500 validation cases. Off-target rates were 2.95% (R0), 2.01% (R1), and 3.12% (R3). The corresponding on-target rates were 54.87%, 45.08%, and 45.82%. Off-target is the mean fraction of the thirteen non-target labels that change. On-target is the fraction of requests whose target report label flips as requested. Both measures give equal weight to each target concept. Dashes in the Full row indicate that its edit evaluations used separate protocols.

W2. Does the comparison hold on another backbone family? We trained FCSL and the continuous-prefix baseline (D+section in the submission) on the Llama-3.2-1B base model. We kept the RG4 data, seed 42, training budget, and validation checkpoint rule used for the corresponding Qwen runs. We also reused the decoding settings selected on Qwen validation. Llama FCSL and the baseline have 135.25M and 45.22M trainable parameters, respectively.

Llama-3.2-1B, RG4 testMicro-F1Macro-F1BLEU-4ROUGE-LRadGraph
FCSL0.63490.49020.17040.36670.2419
Baseline0.54390.39790.19590.39270.2726

FCSL gains 9.10 micro-F1 points over the Llama baseline, with lower BLEU-4, ROUGE-L, and RadGraph (Table 32, Appendix C.14). The same within-backbone comparison on Qwen3.5-0.8B gives a 9.73-point micro-F1 gain at seed 42. The existing Qwen3.5-4B comparison is a scale check within one family (Table 9). The Llama result is exploratory and follows prior test-set feedback during development. Evidence about supervision-matched Continuous and selective editing remains specific to Qwen3.5-0.8B.

Dear Reviewer,

Both weaknesses you raised are now answered by experiments run during the discussion period: W1 by the residual-capacity comparison (Appendix C.13, Table 31) and W2 by the Llama-3.2-1B comparison (Appendix C.14, Table 32). Three points go beyond our first response.

W2: the direction holds for three decoders. RG4, seed 42, validation-selected checkpoints. Differences are FCSL minus baseline, in points.

DecoderMicro-F1 (FCSL / baseline)Micro-F1BLEU-4RadGraph
Qwen3.5-0.8B0.6390 / 0.5417+9.73-1.32-2.17
Qwen3.5-4B0.6420 / 0.5696+7.24-1.40-1.75
Llama-3.2-1B0.6349 / 0.5439+9.10-2.55-3.07

Every difference has the same sign in all three rows. Each row is one seed, so we read the direction and not the size.

W2: the Llama result does not depend on checkpoint choice. The last-epoch FCSL checkpoint gives micro-F1 0.6360, against 0.6349 for the validation-selected one.

W1: the larger residual budgets are used. At the end of training, the residual codes carry 39.5 of 48 bits (R1) and 181.0 of 256 bits (R3) of summed marginal entropy. The codes do not collapse, and the language metrics still plateau from 122 bits upward.

If you consider another configuration necessary, for example a one-stage quantizer, please let us know.

Reviewer_jxViOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Radiology report generation needs to specify the clinical findings in written language. The report generator should be able to integrate multiple sources of input, including the current image, prior images and findings and different views. When the generator fails to report a finding, it is difficult to identify the underlying cause. The error can arise from encoding process or the report generation process. In order to inspect the source of errors more easily, the paper proposes Finite Clinical Subspace Learning (FCSL).

In contrast to other radiology report generation approaches that use continuous representation from the evidence encoder to the decoder, FCSL insets a discrete interface between these two processes. All case-specific information is encoded to a fixed size tuple of IDs that carries clinical meaning and residual codes with additional information. The finite clinical states allow editing, providing a discrete control over the report generation.

Overall, FCSL improves the classification performance as its CheXbert F1 outperforms other methods that use continuous conditioning with more parameters. However, FCSL does not substantially improve the report quality on the language aspects. In the requested changes, flipping a clinical states produces the requested change with all other labels unchanged in 33.05% of 3,696 cases, compared with 18.60% for continuous conditioning.

Critical strengths

The study demonstrates an inspectable and editable conditioning interface between the evidence encoding and report generation in contrast to the continuous conditioning. When compared with LLM-RG4, the continuous conditioning counterpart, FCSL yields better performance on micro-F1 on the clinical labels while using smaller model backbone.

Critical weaknesses

Overall, the practical utility is underspecified and the paper lacks clarity in writing. It will be helpful to articulate a clinical use case for the finite interface or for state editing. An immediate question is the practicality of editing state in order to get the correct generation versus editing the the generation directly for correction. It is unclear if the interface will override the generation result given an edit request against the factual grounding.

Several terms are first used without proper definition. It is unclear what does head metrics and manifest labels mean (Line 277). Other examples are endpoint group (Line 138), canonical residual IDs (Line 240), concept-macro (Line 309).

Line 75 to 79: The introduction should clarify the meaning of the methods when they are not yet specified. E.g., it will be hard for the reader to know upfront what the authors precisely mean by Continuous conditioning, Full, shared eligible RG4 edit requests, etc.

Questions for authors

What is the definition of joint selective success? How is it being measured?

Several acronyms are used without being defined at first use, including SN/MN, LM, VQ and EMA.

Table 3. It will be helpful to elaborate the meaning of each column name.

2 comments

What can a reader use the states for? A reader can check whether an omitted finding was recorded before generation. Among RG4 decisions with a positive reference endpoint, 4.14% have a positive encoded endpoint and a nonpositive report label. In 29.28%, both are nonpositive (§5.6, Table 24). These are agreements and disagreements with CheXbert labels of the reference. Establishing why the model missed a finding requires independent evidence. A reader can also edit the state before generating a new report. Whether clinicians prefer that workflow or obtain better corrections than by editing the text directly remains an open question.

Will an edit override the evidence? Changing one current-state ID leaves the other IDs fixed. We recompute the corresponding change relative to the prior study, then regenerate through both decoder routes. The prior and residual fields can still carry information about that finding. We measured three outcomes for Full and Continuous on the same 3,696 requests (1,596 cases, 245 patients, RG4 seed 42). The edited report reaches the requested target endpoint at rates of 60.51% and 52.21%, respectively. Requiring the intact report also to express the original endpoint gives 54.22% and 43.75%. Requiring all thirteen non-target labels also to remain unchanged gives 33.05% and 18.60%. Each rate gives equal weight to the fourteen target observations (Table 3 and Appendix Table 12). The edits reverse states on which all evaluated models agree. Their agreement with the images has not been adjudicated.

Definitions added to the revision

  • The header contains the fourteen current-state IDs and, on RG4, fourteen prior-state IDs. Header micro-F1 compares current predictions with the manifest labels, the four-state CheXbert labels of each reference report stored with its dataset row (§§3.3, 4.3). The two evaluation endpoints group positive with uncertain and blank with negative. Equation 3 selects the group, then the more probable state within that group.
  • Section 3.3 describes how codebook lookup reconstructs each residual atom from its stored indices. SN and SW use one image view, MN and MW use multiple views, and W denotes access to the prior report. Section 3.5 defines language-model (LM) loss, vector quantization (VQ), and exponential moving average (EMA).
  • Section 1 introduces Full (the complete discrete model), Continuous (state probabilities and unquantized residuals), and Prefix-only (the discrete model with the input prefix as its sole decoder route). Section 4.3 defines shared requests. For each target, the three models must agree on an explicitly positive or negative state before reversal. We generate intact and edited reports with identical decoding settings and label each with CheXbert.
  • Table 3 now defines every column. Align requires the intact target to match the original endpoint. Direction additionally requires the edited target to match the requested endpoint. Joint also requires all thirteen non-target labels to stay fixed. Unaligned requests remain in the denominator. Any off counts requests with at least one non-target change. Mean off measures the fraction of the thirteen non-target labels that change. Concept-macro means that each target contributes equally.

Table 1 identifies LLM-RG4 as a published comparison with its own preprocessing and scoring. The controlled baseline comparisons in our paper use the same decoder within each pair.

Dear Reviewer,

Your review asked what the states are for, and listed terms the submission never defined. Every listed term is now defined where it first appears, and the Table 3 caption defines each column. Two points go beyond our first response.

A concrete use, with its limit. Figure 2 now follows one finding, pneumothorax, from the image to its stored state and to the sentence in the report. It then shows the report regenerated after that single state is reversed. The regenerated sentence adds a location that no input specifies. This is why the caption and §6 describe an edit as a test of the decoder and not as a correction of the report.

Which part of the model makes edits selective. Removing the layer updates (Prefix-only) lowers joint selective success from 33.05% to 29.63% (difference +3.41 points, 95% CI [1.21, 5.33]). Replacing the discrete IDs with continuous inputs (Continuous) lowers it to 18.60%. The discrete interface accounts for most of the difference. Most edits still fail, and §1 and §6 say so.

If any term or step remains unclear, a pointer to the sentence would let us fix it in this revision.

Reviewer_3dhgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

As claimed by this paper, it researches radiology report generation and studies this task with a complete case-dependent conditioning message reconstructible from discrete IDs. It proposes Finite Clinical Subspace Learning (FCSL) that combines named clinical states, task and prior-study fields, and quantized residual codes.

However, after carefully reading this paper, the reviewer cannot figure out what this work is, even from the abstract. The whole paper is hard to follow. Although each sentence reads well but when they are put together, the reviewer feels difficult to figure out what exactly is done.

Critical strengths

The reviewer cannot identify any critical strengths because they cannot figure out what this work has done after carefully reading this paper.

Critical weaknesses

This paper is hard to follow. After carefully reading this paper, the reviewer cannot figure out what this work is, even from the abstract.

Questions for authors

The reviewer believes that this paper needs to be carefully proofread or even fully rewritten because it is really hard to figure out what has been done in this paper.

2 comments

Thank you for pointing out that the abstract made the method difficult to identify. Figure 1 in the revision shows the problem on one held-out study. The reference mentions atelectasis at both lung bases. The baseline describes the lung parenchyma as normal. FCSL records atelectasis as positive and mentions the left base in its report. We can inspect whether the finding reached the decoder even though the report still misses the right side.

FCSL separates two operations. An encoder first writes a clinical program. On RG4 its 46 IDs store the reporting task, current and prior findings, and quantized residual features, using 202 bits per study. A language model then reads this program through an input prefix and updates at eight decoder layers. Editing a named state lets us ask whether the corresponding report finding changes.

We rewrote the abstract and §1 to distinguish the two comparisons. Against the continuous-prefix baseline (D+section in the submission), FCSL improves clinical-label agreement but scores lower on BLEU-4 and RadGraph. The closer control, Continuous, shares clinical supervision, both decoder routes, and the same 79,747,648 trainable parameters as Full. Their RG4 micro-F1 scores are 0.6331 and 0.6340, respectively. For 3,696 shared edit requests from 1,596 cases, we computed success separately for each target and averaged equally over fourteen targets. The requested change occurred with the other labels preserved at rates of 33.05% for Full and 18.60% for Continuous (paired difference +14.44 percentage points, 95% patient-bootstrap CI [11.87, 16.63], seed 42). The comparison changes both the representation and its training losses.

Dear Reviewer,

Your review said the work could not be identified from the abstract and asked for the paper to be rewritten. We treated this as the main instruction for the revision. The changes, in reading order:

  • Abstract. Rewritten in the order problem, method, what the program stores, what a reader can do with it, results.
  • Figure 1. Replaced with one held-out study that shows the problem before any method is introduced.
  • §1. Rewritten as four short paragraphs: the problem, the method, the comparison with the baseline, the matched control and the edit test. Each model name and the edit metric are defined in the sentence that first uses them.
  • §3.1. A new overview describes the method in plain words before any equation.
  • Figure 2. Redrawn to follow one finding from the image to the report sentence.
  • Terminology. "D+section" is now "the baseline", and "local writes" are now "layer updates". Every acronym is expanded at first use.

The technical content and the results are unchanged. If the first page still does not make the contribution clear, we would be grateful to know which sentence fails, and we will revise it.

Comments on the paper

1 comment

Thank you for identifying where the submission lost its readers. We rewrote the abstract and introduction around a question that Figure 1 makes concrete. When a generated report omits a finding, did the model record it before writing? FCSL stores a named state for each of fourteen observations. These states, the task and prior-study fields, and learned residual codes provide all study-specific input to a 0.8B decoder. A reader can inspect a state, change it, and regenerate the report.

Across three seeds, FCSL gains 9.48 points in CheXbert micro-F1 on RG4 and 6.78 on ABN over the section-conditioned continuous-prefix baseline, with lower BLEU-4 and RadGraph. A supervision-matched continuous model reaches a similar RG4 micro-F1. The clearest difference between the two interfaces appears when we ask for a single finding to change and the other thirteen to stay fixed (Table 3).

Figure 2 now follows one finding from the image to the generated text. Section 5.6 describes how a reader can use the recorded state. Appendix C.13 adds a residual-capacity comparison (Table 31), and Appendix C.14 adds a comparison on Llama-3.2-1B (Table 32).

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

Related papers

Discussion (0)

Sign in to comment.