- 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
-
The paper explores an interesting direction by introducing an explicit, inspectable discrete representation for radiology report generation, with potential benefits for interpretability and controllability.
-
The experimental evaluation is relatively comprehensive, including multiple datasets, continuous-conditioning baselines, ablation studies, and targeted clinical-state interventions.
Critical weaknesses
- 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.
- 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
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.
| Decoder | Micro-F1 (FCSL / baseline) | Micro-F1 | BLEU-4 | RadGraph |
|---|---|---|---|---|
| Qwen3.5-0.8B | 0.6390 / 0.5417 | +9.73 | -1.32 | -2.17 |
| Qwen3.5-4B | 0.6420 / 0.5696 | +7.24 | -1.40 | -1.75 |
| Llama-3.2-1B | 0.6349 / 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.
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.
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.
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.