From Sampled Evidence to Radiology Reports: Objectives, Constraints, and Realization

Kun Zhao; Haoteng Tang; Chenghua Lin; Liang Zhan

cs.CLOct 12, 2026

From Sampled Evidence to Radiology Reports: Objectives, Constraints, and Realization

Authors: Kun Zhao, Haoteng Tang, Chenghua Lin, Liang Zhan

Organizations: University of Pittsburgh · University of Texas Rio Grande Valley · The University of Manchester

Abstract

Sampling multiple reports creates an inference problem: how should sampled evidence be converted into a final decision, and does a better structured decision yield a better report? We study these questions through a framework that separates the decision signal, scoring objective, and feasible decision set, followed by text realization. On MIMIC-CXR and IU-Xray, controlled comparisons across 13 generator configurations per benchmark under grouped development/test evaluation reveal three findings. First, changing the objective under fixed evidence and rule capacity changes predictions and their held-out metric trade-offs. Second, relaxing candidate constraints has objective-dependent effects under shared signals and frozen scores. Diagnostics connect the IU macro-oriented assertion increase to near-zero label thresholds and show that limited candidate coverage coexists with selection headroom. Third, a controlled realization study with SFT-Qwen3-VL-8B shows that decision-level gains can be attenuated, amplified, or reversed. Change decomposition reveals that realization can remove incorrect assertions while losing correct findings, so target fidelity and reference accuracy need not improve together. Finding gains also do not consistently improve reference-based report quality. These results motivate separate evaluation of decision rules, candidate constraints, and realized reports.

Figures & tables

Figure 1: Overview of the proposed framework. Multiple reports are sampled to extract finding-level evidence. The framework separates decision signals, scoring objectives, and feasible decision sets to produce structured predictions. RQ1 studies objective choice, RQ2 examines candidate constraints (constrained, partial, and free), and RQ3 evaluates report realization through retrieval and generation. Decision accuracy, target fidelity, and report quality are assessed separately.

	MIMIC-CXR				IU-Xray
Contrast	Ma-14	Mi-14	Ma-5	Mi-5	Ma-14	Mi-14	Ma-5	Mi-5
Ma–Maj [-1pt] (0/0/0)	+0.0644 [-1pt] (13/0/0)	-0.0033 [-1pt] (7/0/6)	+0.0590 [-1pt] (13/0/0)	+0.0238 [-1pt] (11/0/2)	+0.0553 [-1pt] (13/0/0)	-0.1423 [-1pt] (1/0/12)	+0.0767 [-1pt] (12/1/0)	+0.1234 [-1pt] (12/1/0)
Mi–Maj [-1pt] (0/0/0)	+0.0471 [-1pt] (12/0/1)	+0.0245 [-1pt] (11/0/2)	+0.0563 [-1pt] (12/0/1)	+0.0376 [-1pt] (12/0/1)	+0.0125 [-1pt] (6/4/3)	-0.0008 [-1pt] (4/1/8)	+0.0055 [-1pt] (5/6/2)	+0.0160 [-1pt] (6/6/1)
Ma–Mi [-1pt] (0/0/0)	+0.0172 [-1pt] (12/1/0)	-0.0278 [-1pt] (1/1/11)	+0.0028 [-1pt] (7/1/5)	-0.0138 [-1pt] (4/1/8)	+0.0428 [-1pt] (13/0/0)	-0.1415 [-1pt] (0/0/13)	+0.0712 [-1pt] (12/1/0)	+0.1074 [-1pt] (12/1/0)

Table 1: Shared-threshold comparisons: mean within-generator paired differences across 13 configurations per benchmark, with positive/zero/negative counts in parentheses. Contrasts denote first minus second. Ma/Mi denote macro-/micro-F1-14-selected shared thresholds; Maj denotes binary majority. Metric columns report macro/micro-F1 over 14/five labels.

Data	Contrast	Ma-14	Mi-14	Ma-5	Mi-5	RG	RL	GREEN	RaTE	B4
MIMIC	RG–CX [-1pt] (0/0/0)	-0.0223 [-1pt] (0/0/13)	-0.0282 [-1pt] (1/0/12)	-0.0289 [-1pt] (0/0/13)	-0.0332 [-1pt] (1/0/12)	+0.0428 [-1pt] (13/0/0)	+0.0227 [-1pt] (12/0/1)	+0.0069 [-1pt] (11/0/2)	+0.0139 [-1pt] (12/0/1)	+0.0079 [-1pt] (10/0/3)
	RL–CX [-1pt] (0/0/0)	-0.0243 [-1pt] (0/0/13)	-0.0355 [-1pt] (1/0/12)	-0.0333 [-1pt] (0/0/13)	-0.0436 [-1pt] (0/0/13)	+0.0183 [-1pt] (13/0/0)	+0.0369 [-1pt] (12/0/1)	+0.0035 [-1pt] (9/0/4)	+0.0086 [-1pt] (11/0/2)	+0.0146 [-1pt] (13/0/0)
	RG–RL [-1pt] (0/0/0)	+0.0020 [-1pt] (7/0/6)	+0.0073 [-1pt] (10/0/3)	+0.0044 [-1pt] (10/0/3)	+0.0104 [-1pt] (12/0/1)	+0.0246 [-1pt] (13/0/0)	-0.0142 [-1pt] (0/0/13)	+0.0033 [-1pt] (10/0/3)	+0.0053 [-1pt] (13/0/0)	-0.0067 [-1pt] (0/0/13)
IU	RG–CX [-1pt] (0/0/0)	+0.0012 [-1pt] (6/3/4)	+0.0108 [-1pt] (5/3/5)	-0.0015 [-1pt] (2/9/2)	+0.0021 [-1pt] (2/9/2)	+0.0715 [-1pt] (13/0/0)	+0.0188 [-1pt] (13/0/0)	+0.0085 [-1pt] (8/0/5)	+0.0337 [-1pt] (13/0/0)	+0.0153 [-1pt] (12/0/1)
	RL–CX [-1pt] (0/0/0)	+0.0028 [-1pt] (5/5/3)	+0.0052 [-1pt] (5/5/3)	-0.0008 [-1pt] (1/9/3)	-0.0019 [-1pt] (2/9/2)	+0.0364 [-1pt] (11/0/2)	+0.0316 [-1pt] (13/0/0)	+0.0030 [-1pt] (8/0/5)	+0.0158 [-1pt] (12/0/1)	+0.0187 [-1pt] (13/0/0)
	RG–RL [-1pt] (0/0/0)	-0.0017 [-1pt] (4/4/5)	+0.0056 [-1pt] (5/4/4)	-0.0007 [-1pt] (2/9/2)	+0.0040 [-1pt] (3/9/1)	+0.0351 [-1pt] (13/0/0)	-0.0129 [-1pt] (1/0/12)	+0.0055 [-1pt] (7/0/6)	+0.0179 [-1pt] (12/0/1)	-0.0034 [-1pt] (5/0/8)

Table 2: Candidate-utility comparisons across 13 configurations per benchmark. Each cell reports the mean paired difference and positive/zero/negative counts. Contrasts denote first minus second; CX/RG/RL denote CheXbert-14/RadGraph/ROUGE-L utilities. Ma/Mi denote macro/micro-F1 over 14/five labels; RaTE/B4 denote RaTEScore/BLEU-4.

	MIMIC-CXR				IU-Xray
Effect	Ma-14	Mi-14	Ma-5	Mi-5	Ma-14	Mi-14	Ma-5	Mi-5
$\Delta^{\mathrm{ma,P}}$ [-1pt] (0/0/0)	+0.0069 [-1pt] (10/0/3)	-0.0322 [-1pt] (0/0/13)	-0.0068 [-1pt] (1/0/12)	-0.0086 [-1pt] (2/0/11)	-0.0010 [-1pt] (9/0/4)	-0.2208 [-1pt] (0/0/13)	-0.0038 [-1pt] (11/0/2)	-0.1023 [-1pt] (2/0/11)
$\Delta^{\mathrm{ma,F}}$ [-1pt] (0/0/0)	+0.0181 [-1pt] (11/0/2)	-0.0269 [-1pt] (0/0/13)	+0.0138 [-1pt] (10/0/3)	+0.0141 [-1pt] (10/0/3)	+0.0439 [-1pt] (11/0/2)	-0.2206 [-1pt] (1/0/12)	+0.0526 [-1pt] (10/0/3)	-0.0411 [-1pt] (5/0/8)
$\Delta^{\mathrm{mi,P}}$ [-1pt] (0/0/0)	+0.0196 [-1pt] (13/0/0)	+0.0002 [-1pt] (6/0/7)	+0.0206 [-1pt] (13/0/0)	+0.0008 [-1pt] (8/0/5)	-0.0085 [-1pt] (0/11/2)	+0.0001 [-1pt] (4/7/2)	-0.0205 [-1pt] (0/11/2)	-0.0009 [-1pt] (2/9/2)
$\Delta^{\mathrm{mi,F}}$ [-1pt] (0/0/0)	+0.0289 [-1pt] (13/0/0)	+0.0075 [-1pt] (9/0/4)	+0.0328 [-1pt] (13/0/0)	+0.0135 [-1pt] (13/0/0)	+0.0254 [-1pt] (11/0/2)	+0.0035 [-1pt] (10/0/3)	+0.0628 [-1pt] (11/0/2)	+0.2085 [-1pt] (12/0/1)
$I^{\mathrm{P}}$ [-1pt] (0/0/0)	-0.0127 [-1pt] (0/0/13)	-0.0323 [-1pt] (0/0/13)	-0.0274 [-1pt] (0/0/13)	-0.0093 [-1pt] (0/0/13)	+0.0076 [-1pt] (10/0/3)	-0.2209 [-1pt] (0/0/13)	+0.0167 [-1pt] (12/0/1)	-0.1015 [-1pt] (2/0/11)

Table 3: RQ2: mean paired differences across 13 configurations, with positive/zero/negative counts in parentheses.

\Delta^{o,t}

denotes Partial (P) or Free (F) minus Candidate under macro (ma) or micro (mi) scoring;

I^{t}=\Delta^{\mathrm{ma},t}-\Delta^{\mathrm{mi},t}

uses the same metric. Ma/Mi-14/5 denote macro/micro-F1 over 14/five labels. Partial replaces five training-defined low-prevalence labels, distinct from the five-label evaluation subset.

		Ma-14			Mi-14			Ma-5			Mi-5
Obj.	Contrast	$\Delta D$	$A_{R}$	$A_{G}$	$\Delta D$	$A_{R}$	$A_{G}$	$\Delta D$	$A_{R}$	$A_{G}$	$\Delta D$	$A_{R}$	$A_{G}$
MIMIC-CXR
Macro	P–C	+.0215	-.0149	+.0010	-.0162	-.0037	+.0012	-.0127	+.0025	-.0025	-.0133	+.0004	+.0013
	F–C	+.0307	-.0153	+.0062	-.0154	+.0006	+.0115	+.0056	-.0084	-.0030	+.0062	-.0039	-.0013
Micro	P–C	+.0220	-.0079	-.0052	-.0010	-.0023	+.0007	+.0160	-.0018	-.0048	-.0024	+.0002	+.0012
	F–C	+.0243	+.0034	-.0066	-.0022	-.0011	+.0022	+.0215	+.0041	-.0081	+.0020	+.0013	-.0021

Table 4: Decision-level contrasts and their changes after report realization with SFT-Qwen3-VL-8B. P–C and F–C denote partial-minus-constrained and free-minus-constrained, respectively. For each metric,

\Delta D

is the decision-level contrast, and

A_{r}=\Delta M_{r}-\Delta D

measures its change after retrieval (

A_{R}

) or generation (

A_{G}

). The report-level contrast is

\Delta M_{r}=\Delta D+A_{r}

. Ma/Mi denote macro/micro F1.

	MIMIC-CXR						IU-Xray
Method	Ma-14	Mi-14	Ma-5	Mi-5	GREEN	RG	Ma-14	Mi-14	Ma-5	Mi-5	GREEN	RG
Greedy	.3531	.5807	.4712	.5729	.4866	.3146	.1020	.5521	.0467	.0727	.5731	.3122
MBR-1	.3725	.5903	.5111	.5775	.4977	.3382	.0544	.5825	.0000	.0000	.5993	.3960
MBR-2	.3742	.5833	.5145	.5762	.4942	.3347	.0629	.5843	.0000	.0000	.5692	.3585
CCS-Emb	.3881	.6092	.5178	.6059	.4952	.3313	.0775	.5842	.0000	.0000	.6022	.3740
CCS-Dev	.3789	.5898	.5145	.5824	.4970	.3310	.1022	.5787	.0707	.1565	.5553	.2997

Table 5: Report-level performance with SFT-Qwen3-VL-8B on the held-out test subset. C/P/F denote constrained, partial, and free targets; R/G denote retrieval/generation. MBR-1/2 use unigram/bigram F1 utilities. CCS-Emb uses a fixed embedding utility; CCS-Dev and SDR-Dev select their utility or variant on development Ma-14.

Appendix figures & tables59 assets

Supplementary material from the paper’s appendix.

Appendix

Generator	Rule	$\tau$	Ma-14	Mi-14	Ma-5	Mi-5	P	R
Qwen2.5-VL-8B	Maj	17	.3637	.6136	.4877	.5946	.5282	.3335
	Ma	3	.4212	.5940	.5529	.6073	.3311	.5893
	Mi	10	.4036	.6436	.5527	.6474	.4308	.4262
Q25-1	Maj	17	.2885	.5441	.4163	.5251	.3941	.2610
	Ma	8	.3651	.6023	.5148	.5998	.3943	.4002
	Mi	9	.3614	.6018	.5077	.5988	.4208	.3849

Appendix

Table 6: MIMIC: absolute shared-threshold test results. Maj/Ma/Mi denote majority/macro-selected/micro-selected rules; P/R are 14-label macro precision/recall.

Generator	Rule	$\tau$	Ma-14	Mi-14	Ma-5	Mi-5	P	R
Q35	Maj	17	.2106	.6235	.1822	.1833	.4147	.2491
	Ma	1	.2479	.5491	.3080	.3841	.2578	.4113
	Mi	9	.2102	.6054	.1811	.2222	.3142	.2853
IV35	Maj	17	.2072	.6322	.2339	.3421	.2719	.2668
	Ma	2	.2332	.5522	.2879	.3692	.1839	.4372
	Mi	18	.2025	.6319	.2151	.3087	.3106	.2611

Appendix

Table 7: IU: absolute shared-threshold test results. Maj/Ma/Mi denote majority/macro-selected/micro-selected rules; P/R are 14-label macro precision/recall.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	+0.0176 [-0.0024, +0.0364]	-0.0495 [-0.0611, -0.0380]	+0.0002 [-0.0352, +0.0301]	-0.0401 [-0.0582, -0.0208]
Q25-1	+0.0037 [-0.0022, +0.0113]	+0.0004 [-0.0037, +0.0048]	+0.0072 [-0.0053, +0.0259]	+0.0010 [-0.0059, +0.0082]
Q35	+0.0611 [+0.0360, +0.0892]	-0.0076 [-0.0253, +0.0106]	+0.0822 [+0.0409, +0.1269]	+0.0870 [+0.0459, +0.1331]
IV35	+0.0257 [-0.0008, +0.0509]	-0.0422 [-0.0552, -0.0293]	+0.0066 [-0.0167, +0.0313]	+0.0035 [-0.0201, +0.0293]
ESC-RL	+0.0120 [-0.0072, +0.0311]	-0.0362 [-0.0458, -0.0264]	-0.0121 [-0.0411, +0.0151]	-0.0377 [-0.0518, -0.0242]
MPO	+0.0150 [-0.0126, +0.0403]	-0.0632 [-0.0760, -0.0507]	-0.0373 [-0.0806, +0.0004]	-0.0674 [-0.0870, -0.0481]

Appendix

Table 8: MIMIC: Macro minus Micro.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	+0.0575 [+0.0333, +0.0803]	-0.0196 [-0.0364, -0.0036]	+0.0652 [+0.0238, +0.1057]	+0.0128 [-0.0163, +0.0449]
Q25-1	+0.0766 [+0.0558, +0.0986]	+0.0581 [+0.0431, +0.0734]	+0.0986 [+0.0655, +0.1395]	+0.0748 [+0.0465, +0.1092]
Q35	+0.0393 [+0.0183, +0.0610]	-0.0141 [-0.0284, +0.0015]	+0.0381 [+0.0141, +0.0655]	+0.0438 [+0.0135, +0.0795]
IV35	+0.0345 [+0.0070, +0.0624]	-0.0454 [-0.0602, -0.0313]	+0.0239 [-0.0024, +0.0527]	+0.0157 [-0.0105, +0.0449]
ESC-RL	+0.0761 [+0.0499, +0.1020]	+0.0016 [-0.0145, +0.0183]	+0.0650 [+0.0268, +0.0997]	+0.0178 [-0.0092, +0.0494]
MPO	+0.0547 [+0.0303, +0.0781]	-0.0329 [-0.0482, -0.0170]	+0.0127 [-0.0274, +0.0505]	-0.0279 [-0.0528, +0.0014]

Appendix

Table 9: MIMIC: Macro minus Majority.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	+0.0376 [-0.0048, +0.0829]	-0.0563 [-0.0798, -0.0318]	+0.1269 [+0.0543, +0.2034]	+0.1618 [+0.0710, +0.2584]
IV35	+0.0307 [-0.0124, +0.0770]	-0.0797 [-0.1077, -0.0515]	+0.0728 [-0.0135, +0.1632]	+0.0605 [-0.0307, +0.1541]
Q25-3	+0.0815 [+0.0455, +0.1144]	-0.1812 [-0.2218, -0.1405]	+0.1375 [+0.0976, +0.1750]	+0.2195 [+0.1486, +0.2843]
Q25-1	+0.0359 [+0.0187, +0.0565]	-0.2039 [-0.2356, -0.1722]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]
ESC-RL	+0.0769 [+0.0474, +0.1080]	-0.1753 [-0.2099, -0.1408]	+0.1154 [+0.0748, +0.1570]	+0.1628 [+0.0812, +0.2420]
MPO	+0.0775 [+0.0415, +0.1138]	-0.1806 [-0.2195, -0.1418]	+0.1663 [+0.0953, +0.2376]	+0.1753 [+0.1040, +0.2410]

Appendix

Table 10: IU: Macro minus Micro.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	+0.0373 [-0.0209, +0.1073]	-0.0744 [-0.1027, -0.0453]	+0.1258 [+0.0258, +0.2337]	+0.2007 [+0.0939, +0.3109]
IV35	+0.0260 [-0.0161, +0.0710]	-0.0800 [-0.1076, -0.0531]	+0.0540 [-0.0292, +0.1414]	+0.0271 [-0.0570, +0.1136]
Q25-3	+0.1157 [+0.0888, +0.1411]	-0.1785 [-0.2205, -0.1368]	+0.1452 [+0.1076, +0.1805]	+0.2393 [+0.1794, +0.2965]
Q25-1	+0.0359 [+0.0187, +0.0565]	-0.2046 [-0.2365, -0.1730]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]
ESC-RL	+0.1178 [+0.0883, +0.1456]	-0.1894 [-0.2304, -0.1489]	+0.1535 [+0.1150, +0.1913]	+0.2654 [+0.2089, +0.3202]
MPO	+0.1060 [+0.0728, +0.1387]	-0.1780 [-0.2188, -0.1375]	+0.1744 [+0.1045, +0.2439]	+0.1957 [+0.1371, +0.2540]

Appendix

Table 11: IU: Macro minus Majority.

Generator	Maj	Shared Ma	Shared Mi	Label Ma	NC	Const/AP
Q25-3	0.3637	0.4212	0.4036	0.4206	0.4206	0.2172
Q25-1	0.2885	0.3651	0.3614	0.4002	0.4002	0.2172
Q35	0.3547	0.3940	0.3329	0.4310	0.4310	0.2172
IV35	0.3770	0.4115	0.3858	0.4269	0.4269	0.2172
ESC-RL	0.3682	0.4442	0.4322	0.4612	0.4612	0.2172
MPO	0.3601	0.4148	0.3998	0.4444	0.4444	0.2172

Appendix

Table 12: MIMIC: 14-label macro-F1 under the common reference-only mask. Label Ma is label-specific calibration; NC is the exported nonconstant control. Const/AP denotes the coincident constant/always-positive baseline; Deg. is the exported degenerate-label count. Always-negative scores are zero.

Generator	Maj	Shared Ma	Shared Mi	Label Ma	NC	Const/AP	Deg.
Q35	0.2106	0.2479	0.2102	0.2692	0.2454	0.1210	1
IV35	0.2072	0.2332	0.2025	0.2693	0.2525	0.1210	1
Q25-3	0.0556	0.1713	0.0899	0.1884	0.1884	0.1210	0
Q25-1	0.0550	0.0909	0.0550	0.0909	0.0884	0.1210	5
ESC-RL	0.0556	0.1735	0.0965	0.1665	0.1665	0.1210	0
MPO	0.0552	0.1612	0.0837	0.1516	0.1516	0.1210	1

Appendix

Table 13: IU: 14-label macro-F1 under the common reference-only mask. Label Ma is label-specific calibration; NC is the exported nonconstant control. Const/AP denotes the coincident constant/always-positive baseline; Deg. is the exported degenerate-label count. Always-negative scores are zero.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	-0.0212 [-0.0359, -0.0069]	-0.0374 [-0.0471, -0.0277]	-0.0183 [-0.0412, +0.0083]	-0.0420 [-0.0594, -0.0259]
Q25-1	-0.0198 [-0.0316, -0.0077]	-0.0390 [-0.0505, -0.0273]	-0.0187 [-0.0364, +0.0011]	-0.0385 [-0.0574, -0.0195]
Q35	-0.0104 [-0.0290, +0.0085]	+0.0260 [+0.0135, +0.0397]	-0.0224 [-0.0510, +0.0044]	+0.0011 [-0.0190, +0.0202]
IV35	-0.0285 [-0.0459, -0.0126]	-0.0286 [-0.0432, -0.0149]	-0.0386 [-0.0638, -0.0130]	-0.0324 [-0.0504, -0.0155]
ESC-RL	-0.0253 [-0.0382, -0.0137]	-0.0377 [-0.0469, -0.0284]	-0.0341 [-0.0542, -0.0147]	-0.0431 [-0.0592, -0.0276]
MPO	-0.0320 [-0.0445, -0.0211]	-0.0385 [-0.0483, -0.0288]	-0.0507 [-0.0702, -0.0329]	-0.0457 [-0.0608, -0.0314]

Appendix

Table 14: MIMIC: RG-CX finding contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	-0.0194 [-0.0382, -0.0016]	-0.0446 [-0.0568, -0.0327]	-0.0159 [-0.0450, +0.0110]	-0.0491 [-0.0726, -0.0290]
Q25-1	-0.0260 [-0.0357, -0.0157]	-0.0456 [-0.0564, -0.0348]	-0.0251 [-0.0442, -0.0042]	-0.0450 [-0.0645, -0.0262]
Q35	-0.0017 [-0.0208, +0.0167]	+0.0278 [+0.0148, +0.0415]	-0.0115 [-0.0417, +0.0150]	-0.0051 [-0.0244, +0.0131]
IV35	-0.0346 [-0.0509, -0.0192]	-0.0395 [-0.0536, -0.0265]	-0.0568 [-0.0854, -0.0291]	-0.0518 [-0.0715, -0.0336]
ESC-RL	-0.0315 [-0.0468, -0.0167]	-0.0460 [-0.0584, -0.0337]	-0.0426 [-0.0704, -0.0165]	-0.0566 [-0.0782, -0.0364]
MPO	-0.0213 [-0.0358, -0.0069]	-0.0434 [-0.0539, -0.0331]	-0.0442 [-0.0639, -0.0254]	-0.0507 [-0.0683, -0.0344]

Appendix

Table 15: MIMIC: RL-CX finding contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	-0.0018 [-0.0188, +0.0151]	+0.0072 [-0.0035, +0.0178]	-0.0024 [-0.0280, +0.0250]	+0.0071 [-0.0115, +0.0267]
Q25-1	+0.0062 [-0.0069, +0.0198]	+0.0066 [-0.0030, +0.0164]	+0.0064 [-0.0156, +0.0286]	+0.0065 [-0.0100, +0.0233]
Q35	-0.0087 [-0.0208, +0.0027]	-0.0018 [-0.0093, +0.0058]	-0.0108 [-0.0320, +0.0085]	+0.0062 [-0.0092, +0.0217]
IV35	+0.0061 [-0.0058, +0.0183]	+0.0109 [-0.0014, +0.0234]	+0.0182 [-0.0066, +0.0438]	+0.0195 [+0.0024, +0.0371]
ESC-RL	+0.0062 [-0.0072, +0.0194]	+0.0083 [-0.0018, +0.0184]	+0.0085 [-0.0126, +0.0292]	+0.0135 [-0.0036, +0.0309]
MPO	-0.0108 [-0.0276, +0.0046]	+0.0050 [-0.0038, +0.0135]	-0.0065 [-0.0287, +0.0150]	+0.0050 [-0.0106, +0.0207]

Appendix

Table 16: MIMIC: RG-RL finding contrasts.

Table 18

Generator	RG	RL	GREEN	RaTE	B4
Q25-3	+.0257	-.0144	+.0005	+.0028	-.0064
Q25-1	+.0298	-.0197	+.0025	+.0034	-.0061
Q35	+.0103	-.0009	-.0023	+.0034	-.0000
IV35	+.0092	-.0007	+.0017	+.0016	-.0014
ESC-RL	+.0252	-.0197	+.0030	+.0046	-.0103
MPO	+.0255	-.0150	-.0023	+.0076	-.0064

Appendix

Table 19: MIMIC: RG-RL report-metric differences. RaTE/B4 denote RaTEScore/BLEU-4. All columns evaluate the narrative selected by the row utility; no report-metric intervals are available.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	+0.0128 [-0.0181, +0.0441]	-0.0131 [-0.0399, +0.0135]	+0.0049 [-0.0598, +0.0755]	-0.0308 [-0.1000, +0.0377]
IV35	+0.0072 [-0.0362, +0.0447]	+0.1695 [+0.1272, +0.2106]	-0.0239 [-0.0860, +0.0331]	+0.0589 [-0.0077, +0.1221]
Q25-3	-0.0095 [-0.0219, -0.0002]	-0.0038 [-0.0127, +0.0050]	-0.0056 [-0.0179, +0.0000]	-0.0134 [-0.0438, +0.0000]
Q25-1	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]
ESC-RL	-0.0029 [-0.0123, +0.0044]	-0.0089 [-0.0182, -0.0004]	+0.0051 [+0.0000, +0.0164]	+0.0128 [+0.0000, +0.0428]
MPO	+0.0081 [+0.0002, +0.0231]	+0.0131 [+0.0050, +0.0224]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]

Appendix

Table 20: IU: RG-CX finding contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	+0.0370 [-0.0301, +0.0879]	-0.0522 [-0.0820, -0.0217]	-0.0037 [-0.0747, +0.0803]	-0.0434 [-0.1083, +0.0235]
IV35	+0.0094 [-0.0292, +0.0489]	+0.1165 [+0.0750, +0.1578]	-0.0061 [-0.0577, +0.0440]	+0.0194 [-0.0474, +0.0857]
Q25-3 (3 epoch)	-0.0081 [-0.0208, +0.0019]	-0.0060 [-0.0159, +0.0036]	-0.0056 [-0.0179, +0.0000]	-0.0134 [-0.0438, +0.0000]
Q25-1	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]
ESC-RL	-0.0041 [-0.0129, +0.0006]	+0.0004 [-0.0054, +0.0054]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]
MPO	+0.0006 [+0.0001, +0.0012]	+0.0120 [+0.0044, +0.0210]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]

Appendix

Table 21: IU: RL-CX finding contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	-0.0242 [-0.0731, +0.0351]	+0.0391 [+0.0125, +0.0667]	+0.0086 [-0.0541, +0.0724]	+0.0126 [-0.0378, +0.0654]
IV35	-0.0022 [-0.0593, +0.0496]	+0.0530 [+0.0199, +0.0864]	-0.0178 [-0.0873, +0.0439]	+0.0396 [-0.0164, +0.0989]
Q25-3	-0.0013 [-0.0050, +0.0008]	+0.0021 [-0.0072, +0.0113]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]
Q25-1	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]
ESC-RL	+0.0012 [-0.0012, +0.0055]	-0.0093 [-0.0189, -0.0002]	+0.0051 [+0.0000, +0.0164]	+0.0128 [+0.0000, +0.0428]
MPO	+0.0075 [-0.0002, +0.0226]	+0.0011 [-0.0012, +0.0044]	+0.0000 [+0.0000, +0.0000]	+0.0000 [+0.0000, +0.0000]

Appendix

Table 22: IU: RG-RL finding contrasts.

Table 23

Generator	RG	RL	GREEN	RaTE	B4
Q35	+.0094	-.0058	+.0052	+.0054	-.0003
IV35	+.0181	-.0005	-.0020	-.0033	-.0005
Q25-3	+.0662	-.0116	+.0176	+.0357	-.0039
Q25-1	+.0105	-.0055	-.0214	+.0058	-.0151
ESC-RL	+.0500	-.0206	+.0014	+.0305	-.0148
MPO	+.0693	+.0006	-.0267	+.0307	+.0082

Appendix

Table 25: IU: RG-RL report-metric differences. RaTE/B4 denote RaTEScore/BLEU-4. All columns evaluate the narrative selected by the row utility; no report-metric intervals are available.

Utility	Ma-14	Mi-14	Ma-5	Mi-5	RG	RL	GREEN	RaTE	B4
chexbert14	0.3814	0.5928	0.4955	0.5854	0.2942	0.3133	0.4988	0.6197	0.0864
chexbert5	0.3710	0.5747	0.4937	0.5827	0.2927	0.3156	0.4964	0.6186	0.0873
rougeL	0.3571	0.5573	0.4622	0.5417	0.3124	0.3502	0.5023	0.6282	0.1011
bleu4	0.3617	0.5601	0.4659	0.5437	0.3109	0.3471	0.5031	0.6284	0.1044
embed	0.3735	0.5716	0.4846	0.5648	0.3095	0.3323	0.5059	0.6293	0.0979
bertscore	0.3627	0.5640	0.4665	0.5458	0.3152	0.3429	0.5041	0.6291	0.0988

Appendix

Table 26: MIMIC: extended utility matrix. Entries are arithmetic means over 13 configurations. Every metric scores the narrative selected by the corresponding utility.

Utility	Ma-14	Mi-14	Ma-5	Mi-5	RG	RL	GREEN	RaTE	B4
chexbert14	0.0732	0.5468	0.0348	0.0366	0.3117	0.3099	0.6007	0.6220	0.0561
chexbert5	0.0759	0.5317	0.0343	0.0331	0.3079	0.3089	0.5970	0.6197	0.0561
rougeL	0.0760	0.5519	0.0340	0.0347	0.3481	0.3415	0.6037	0.6378	0.0747
bleu4	0.0748	0.5540	0.0340	0.0355	0.3424	0.3382	0.6020	0.6328	0.0788
embed	0.0743	0.5549	0.0309	0.0387	0.3547	0.3282	0.6163	0.6416	0.0706
bertscore	0.0756	0.5544	0.0341	0.0395	0.3553	0.3311	0.5921	0.6420	0.0658

Appendix

Table 27: IU: extended utility matrix. Entries are arithmetic means over 13 configurations. Every metric scores the narrative selected by the corresponding utility.

Dataset / contrast	Index	Text	Findings
MIMIC / CX vs RG	0.9492	0.9492	0.3572
MIMIC / CX vs RL	0.9678	0.9678	0.3868
MIMIC / RG vs RL	0.8259	0.8259	0.3062
IU / CX vs RG	0.9784	0.9784	0.0844
IU / CX vs RL	0.9737	0.9737	0.0846
IU / RG vs RL	0.8837	0.8837	0.0631

Appendix

Table 28: Mean utility-selection disagreement rates over 13 configurations. Index/text disagreements are distinct from finding-vector disagreements.

Dataset	Replacement label	Train prev.
MIMIC	Pleural Other	0.0169
MIMIC	Pneumothorax	0.0334
MIMIC	Consolidation	0.0383
MIMIC	Fracture	0.0387
MIMIC	Lung Lesion	0.0395
IU	Pneumothorax	0.0043

Appendix

Table 29: Training-defined labels used in partial replacement.

Table 29

Table 30

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	+.0215 [+.0046, +.0384]	-.0162 [-.0217, -.0101]	-.0127 [-.0291, +.0043]	-.0133 [-.0200, -.0064]
Q25-1	+.0060 [-.0126, +.0253]	-.0502 [-.0585, -.0419]	-.0079 [-.0278, +.0116]	-.0152 [-.0229, -.0078]
Q35	-.0025 [-.0184, +.0130]	-.1237 [-.1328, -.1150]	-.0016 [-.0114, +.0068]	+.0029 [-.0004, +.0062]
IV35	+.0188 [+.0004, +.0392]	-.0410 [-.0492, -.0331]	+.0058 [-.0052, +.0160]	+.0068 [+.0011, +.0124]
ESC-RL	+.0131 [+.0002, +.0263]	-.0123 [-.0165, -.0078]	-.0084 [-.0234, +.0068]	-.0095 [-.0150, -.0040]
MPO	+.0097 [-.0028, +.0228]	-.0171 [-.0216, -.0125]	-.0114 [-.0246, +.0021]	-.0125 [-.0185, -.0067]

Appendix

Table 34: MIMIC: macro-oriented partial-minus-constrained target contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	+.0307 [+.0120, +.0504]	-.0154 [-.0255, -.0051]	+.0056 [-.0144, +.0257]	+.0062 [-.0068, +.0193]
Q25-1	+.0301 [+.0087, +.0519]	-.0407 [-.0554, -.0269]	+.0254 [+.0005, +.0512]	+.0211 [+.0027, +.0406]
Q35	+.0186 [+.0004, +.0370]	-.1037 [-.1172, -.0902]	+.0266 [+.0123, +.0420]	+.0408 [+.0239, +.0602]
IV35	+.0257 [+.0056, +.0490]	-.0350 [-.0474, -.0227]	+.0301 [+.0147, +.0473]	+.0413 [+.0261, +.0587]
ESC-RL	+.0181 [+.0032, +.0340]	-.0109 [-.0198, -.0017]	+.0109 [-.0090, +.0309]	+.0102 [-.0029, +.0240]
MPO	+.0209 [+.0044, +.0377]	-.0153 [-.0248, -.0058]	+.0053 [-.0124, +.0243]	+.0046 [-.0090, +.0180]

Appendix

Table 35: MIMIC: macro-oriented free-minus-constrained target contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	+.0220 [+.0064, +.0396]	-.0010 [-.0031, +.0013]	+.0160 [-.0083, +.0429]	-.0024 [-.0065, +.0017]
Q25-1	+.0317 [+.0102, +.0520]	-.0002 [-.0027, +.0023]	+.0222 [-.0068, +.0476]	-.0017 [-.0058, +.0019]
Q35	+.0137 [-.0079, +.0331]	+.0014 [-.0017, +.0039]	+.0067 [-.0098, +.0199]	+.0066 [+.0024, +.0102]
IV35	+.0383 [+.0155, +.0589]	+.0045 [+.0018, +.0074]	+.0060 [-.0155, +.0234]	+.0070 [+.0019, +.0118]
ESC-RL	+.0156 [-.0029, +.0327]	-.0014 [-.0036, +.0007]	+.0111 [-.0142, +.0361]	-.0018 [-.0054, +.0016]
MPO	+.0240 [+.0095, +.0389]	.0000 [-.0018, +.0019]	+.0283 [+.0085, +.0475]	+.0006 [-.0024, +.0035]

Appendix

Table 36: MIMIC: micro-oriented partial-minus-constrained target contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	+.0243 [+.0065, +.0439]	-.0022 [-.0079, +.0034]	+.0215 [-.0069, +.0504]	+.0020 [-.0075, +.0112]
Q25-1	+.0586 [+.0335, +.0818]	+.0121 [+.0045, +.0199]	+.0404 [+.0086, +.0685]	+.0153 [+.0028, +.0284]
Q35	+.0371 [+.0114, +.0598]	+.0487 [+.0372, +.0603]	+.0388 [+.0191, +.0586]	+.0461 [+.0297, +.0646]
IV35	+.0422 [+.0188, +.0646]	+.0146 [+.0068, +.0223]	+.0263 [+.0020, +.0482]	+.0330 [+.0184, +.0477]
ESC-RL	+.0222 [+.0019, +.0414]	-.0003 [-.0054, +.0050]	+.0152 [-.0114, +.0415]	+.0010 [-.0077, +.0093]
MPO	+.0319 [+.0146, +.0485]	+.0035 [-.0029, +.0102]	+.0390 [+.0155, +.0621]	+.0116 [+.0023, +.0216]

Appendix

Table 37: MIMIC: micro-oriented free-minus-constrained target contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	-.0005 [-.0279, +.0241]	-.0152 [-.0209, -.0091]	-.0287 [-.0634, +.0033]	-.0109 [-.0185, -.0029]
Q25-1	-.0258 [-.0573, +.0072]	-.0500 [-.0592, -.0409]	-.0301 [-.0533, -.0039]	-.0135 [-.0206, -.0063]
Q35	-.0161 [-.0379, +.0078]	-.1250 [-.1339, -.1165]	-.0082 [-.0209, +.0075]	-.0036 [-.0073, +.0007]
IV35	-.0195 [-.0433, +.0073]	-.0456 [-.0543, -.0369]	-.0002 [-.0197, +.0220]	-.0002 [-.0069, +.0066]
ESC-RL	-.0026 [-.0211, +.0180]	-.0109 [-.0153, -.0062]	-.0195 [-.0466, +.0097]	-.0077 [-.0136, -.0010]
MPO	-.0143 [-.0338, +.0050]	-.0171 [-.0216, -.0126]	-.0397 [-.0608, -.0192]	-.0131 [-.0189, -.0073]

Appendix

Table 38: MIMIC: same-metric objective-by-partial-relaxation interactions.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	+.0065 [-.0223, +.0331]	-.0131 [-.0231, -.0030]	-.0159 [-.0519, +.0182]	+.0043 [-.0080, +.0177]
Q25-1	-.0285 [-.0650, +.0100]	-.0528 [-.0679, -.0387]	-.0149 [-.0403, +.0141]	+.0058 [-.0088, +.0218]
Q35	-.0185 [-.0430, +.0093]	-.1524 [-.1634, -.1417]	-.0122 [-.0287, +.0057]	-.0054 [-.0167, +.0062]
IV35	-.0164 [-.0426, +.0130]	-.0495 [-.0617, -.0377]	+.0038 [-.0186, +.0293]	+.0083 [-.0038, +.0212]
ESC-RL	-.0041 [-.0244, +.0188]	-.0106 [-.0192, -.0018]	-.0044 [-.0309, +.0258]	+.0092 [-.0035, +.0227]
MPO	-.0110 [-.0340, +.0132]	-.0188 [-.0279, -.0099]	-.0336 [-.0586, -.0083]	-.0070 [-.0205, +.0062]

Appendix

Table 39: MIMIC: same-metric objective-by-free-relaxation interactions.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	+.0007 [-.0312, +.0332]	-.1956 [-.2169, -.1750]	+.0045 [-.0604, +.0651]	-.2368 [-.2921, -.1815]
IV35	-.0408 [-.0810, +.0004]	-.1776 [-.1942, -.1612]	-.0598 [-.1202, +.0029]	-.2525 [-.3035, -.1984]
Q25-3	-.0106 [-.0369, +.0097]	-.1975 [-.2243, -.1705]	+.0055 [+.0018, +.0101]	-.1414 [-.2178, -.0669]
Q25-1	+.0158 [+.0057, +.0276]	-.0933 [-.1158, -.0716]	+.0123 [+.0018, +.0294]	+.0198 [+.0066, +.0364]
ESC-RL	-.0015 [-.0304, +.0222]	-.1469 [-.1683, -.1252]	+.0083 [+.0018, +.0175]	-.1078 [-.1674, -.0470]
MPO	-.0249 [-.0533, +.0058]	-.2826 [-.3138, -.2516]	-.0608 [-.1275, +.0059]	-.1913 [-.2647, -.1139]

Appendix

Table 40: IU: macro-oriented partial-minus-constrained target contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	-.0357 [-.0795, +.0117]	-.3287 [-.3676, -.2918]	-.0310 [-.1101, +.0451]	-.2194 [-.2887, -.1504]
IV35	-.0484 [-.0942, +.0007]	-.3221 [-.3537, -.2911]	-.0558 [-.1221, +.0167]	-.2471 [-.3013, -.1887]
Q25-3	+.0198 [-.0158, +.0558]	-.1659 [-.2052, -.1266]	+.0495 [-.0024, +.1010]	-.0870 [-.1742, -.0038]
Q25-1	+.0931 [+.0652, +.1195]	-.0061 [-.0423, +.0287]	+.1191 [+.0850, +.1515]	+.1173 [+.0924, +.1431]
ESC-RL	+.0515 [+.0176, +.0841]	-.1020 [-.1404, -.0642]	+.0632 [+.0152, +.1112]	-.0419 [-.1157, +.0312]
MPO	+.0009 [-.0355, +.0405]	-.2802 [-.3236, -.2370]	-.0428 [-.1230, +.0475]	-.1362 [-.2215, -.0487]

Appendix

Table 41: IU: macro-oriented free-minus-constrained target contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	-.0635 [-.1071, .0000]	-.0021 [-.0082, +.0019]	-.1333 [-.2000, .0000]	-.0095 [-.0363, +.0060]
IV35	-.0476 [-.0714, .0000]	.0000 [-.0034, +.0028]	-.1333 [-.2000, .0000]	-.0045 [-.0279, +.0105]
Q25-3	.0000 [.0000, .0000]	+.0013 [.0000, +.0033]	.0000 [.0000, .0000]	+.0014 [.0000, +.0051]
Q25-1	.0000 [.0000, .0000]	.0000 [.0000, .0000]	.0000 [.0000, .0000]	.0000 [.0000, .0000]
ESC-RL	.0000 [.0000, .0000]	+.0006 [.0000, +.0021]	.0000 [.0000, .0000]	.0000 [.0000, .0000]
MPO	.0000 [.0000, .0000]	+.0007 [.0000, +.0021]	.0000 [.0000, .0000]	+.0012 [.0000, +.0045]

Appendix

Table 42: IU: micro-oriented partial-minus-constrained target contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	-.0788 [-.1350, +.0113]	+.0042 [-.0136, +.0216]	-.1549 [-.2564, +.0192]	+.0343 [-.0548, +.1214]
IV35	-.0748 [-.1161, -.0078]	-.0033 [-.0179, +.0121]	-.1781 [-.2805, -.0057]	-.0314 [-.1010, +.0331]
Q25-3	+.0179 [-.0156, +.0467]	+.0024 [-.0153, +.0201]	+.0698 [+.0203, +.1206]	+.1789 [+.0861, +.2733]
Q25-1	+.0705 [+.0523, +.0872]	+.0116 [-.0118, +.0356]	+.1520 [+.1084, +.1914]	+.3464 [+.2584, +.4246]
ESC-RL	+.0137 [-.0045, +.0336]	-.0003 [-.0178, +.0172]	+.0257 [-.0193, +.0758]	+.0927 [+.0162, +.1741]
MPO	+.0173 [-.0167, +.0509]	-.0026 [-.0218, +.0169]	+.0778 [+.0172, +.1388]	+.2226 [+.1126, +.3282]

Appendix

Table 43: IU: micro-oriented free-minus-constrained target contrasts.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	+.0642 [-.0304, +.1348]	-.1935 [-.2167, -.1698]	+.1378 [-.0595, +.2558]	-.2273 [-.2907, -.1571]
IV35	+.0068 [-.0456, +.0458]	-.1775 [-.1937, -.1614]	+.0736 [-.0961, +.1855]	-.2480 [-.3010, -.1907]
Q25-3	-.0106 [-.0369, +.0097]	-.1988 [-.2258, -.1719]	+.0055 [+.0018, +.0101]	-.1428 [-.2195, -.0684]
Q25-1	+.0158 [+.0057, +.0276]	-.0933 [-.1158, -.0716]	+.0123 [+.0018, +.0294]	+.0198 [+.0066, +.0364]
ESC-RL	-.0015 [-.0304, +.0222]	-.1476 [-.1690, -.1257]	+.0083 [+.0018, +.0175]	-.1078 [-.1674, -.0470]
MPO	-.0249 [-.0533, +.0058]	-.2833 [-.3145, -.2521]	-.0608 [-.1275, +.0059]	-.1926 [-.2664, -.1150]

Appendix

Table 44: IU: same-metric objective-by-partial-relaxation interactions.

Generator	Ma-14	Mi-14	Ma-5	Mi-5
Q35	+.0431 [-.0648, +.1237]	-.3329 [-.3720, -.2943]	+.1239 [-.0850, +.2545]	-.2537 [-.3556, -.1518]
IV35	+.0263 [-.0396, +.0773]	-.3189 [-.3540, -.2843]	+.1223 [-.0680, +.2460]	-.2156 [-.2991, -.1271]
Q25-3	+.0019 [-.0360, +.0401]	-.1683 [-.2090, -.1270]	-.0204 [-.0891, +.0426]	-.2659 [-.3959, -.1389]
Q25-1	+.0226 [-.0019, +.0479]	-.0177 [-.0558, +.0188]	-.0329 [-.0674, +.0040]	-.2290 [-.3040, -.1488]
ESC-RL	+.0378 [+.0030, +.0706]	-.1017 [-.1417, -.0622]	+.0374 [-.0112, +.0841]	-.1346 [-.2326, -.0382]
MPO	-.0164 [-.0579, +.0270]	-.2776 [-.3205, -.2350]	-.1206 [-.2157, -.0115]	-.3589 [-.4876, -.2293]

Appendix

Table 45: IU: same-metric objective-by-free-relaxation interactions.

Table 43

Table 44

Table 45

Table 46

Generator	Ma assertions	Ma added	Ma add prec.	Mi assertions	Mi added	Mi add prec.
Q25-3	2.6851	794	0.2607	2.1522	218	0.3028
Q25-1	3.0677	1476	0.2188	2.1620	376	0.4122
Q35	3.9679	2535	0.1014	2.2849	477	0.4088
IV35	2.9958	1173	0.1995	2.2004	376	0.4043
ESC-RL	2.6110	685	0.2686	2.1613	226	0.3319
MPO	2.6578	772	0.2552	2.1508	236	0.3644

Appendix

Table 54: MIMIC: frozen-threshold diagnostics. Near-zero counts use

\tau<10^{-8}

. Assertions sum unmasked per-label positive rates; added counts and precision pool reference-scored cells. Ma/Mi refer to scoring objectives.

Generator	Ma near-zero	Ma assertions	Ma added	Ma add prec.	Mi assertions	Mi added	Mi add prec.
Q35	1	5.6121	1874	0.0454	1.1996	59	0.3559
IV35	1	4.9798	1590	0.0314	1.1525	29	0.3448
Q25-3	4	9.7601	3760	0.0662	1.2534	94	0.3191
Q25-1	6	10.5448	4177	0.0864	1.3117	157	0.3376
ESC-RL	2	8.3879	3166	0.0853	1.2466	90	0.3000
MPO	2	10.1323	3927	0.0507	1.2735	119	0.3109

Appendix

Table 55: IU: frozen-threshold diagnostics. Near-zero counts use

\tau<10^{-8}

. Assertions sum unmasked per-label positive rates; added counts and precision pool reference-scored cells. Ma/Mi refer to scoring objectives.

Label	Train prev.	Ma threshold	Ma pos. rate	Ma add prec.	Mi threshold	Mi pos. rate	Mi add prec.
No Finding	0.2375	0.3540	0.4118	0.3614	0.3172	0.4375	0.2077
Enlarged Cardiomediastinum	0.0931	0.0802	0.2312	0.1115	0.3172	0.0045	0.8824
Cardiomegaly	0.2576	0.3132	0.2978	0.4626	0.3172	0.2948	0.4140
Lung Lesion	0.0395	0.1100	0.0488	0.1896	0.3172	0.0063	0.2857
Lung Opacity	0.3045	0.2864	0.4273	0.3912	0.3172	0.3990	0.3393
Edema	0.0955	0.2771	0.1332	0.5294	0.3172	0.1161	0.5587

Appendix

Table 56: MIMIC: label-wise threshold diagnostics. Thresholds and positive rates are equal-weight means over thirteen configurations; added precision pools added scored cells across configurations. These aggregate associations are descriptive, not causal.

Label	Train prev.	Ma threshold	Ma pos. rate	Ma add prec.	Mi threshold	Mi pos. rate	Mi add prec.
No Finding	0.6270	0.3676	0.8360	0.6433	0.2710	0.9283	0.2550
Enlarged Cardiomediastinum	0.0268	0.0044	0.8420	0.0195	0.2710	0.0000	–
Cardiomegaly	0.0875	0.1132	0.4250	0.1722	0.2710	0.1261	0.4353
Lung Lesion	0.0446	0.0107	0.9719	0.0402	0.2710	0.0000	–
Lung Opacity	0.1258	0.1537	0.4310	0.1754	0.2710	0.1578	0.2811
Edema	0.0126	0.0037	0.9922	0.0107	0.2710	0.0000	–

Appendix

Table 57: IU: label-wise threshold diagnostics. Thresholds and positive rates are equal-weight means over thirteen configurations; added precision pools added scored cells across configurations. These aggregate associations are descriptive, not causal.

Generator	Unique	Zero error (%)	Oracle errors	Ma errors	Mi errors	Ma missed (%)	Mi missed (%)
Q25-3	8.4525	67.5978	0.4043	1.3436	1.2325	56.8436	53.7011
Q25-1	8.6711	62.9888	0.4839	1.3561	1.2626	52.6536	51.0475
Q35	4.7968	38.0587	1.1746	1.6948	1.5656	30.2374	24.8603
IV35	6.0810	50.2095	0.8080	1.4560	1.3226	38.1983	33.1704
ESC-RL	8.9246	69.4134	0.3757	1.3547	1.2318	56.2151	54.0503
MPO	9.3149	67.2486	0.4351	1.3715	1.2584	53.8408	51.0475

Appendix

Table 58: MIMIC: candidate-oracle diagnostics. Errors are mismatched scored labels per input, not normalized Hamming rates. Missed denotes a strictly better available candidate under masked Hamming error.

Generator	Cond.	Ma-14	Mi-14	Ma-5	Mi-5
Q25-3	Oracle	0.6822	0.8797	0.8344	0.9016
Q25-3	Macro C	0.4722	0.6465	0.6158	0.6602
Q25-3	Micro C	0.4382	0.6683	0.5992	0.6752
Q25-1	Oracle	0.5996	0.8520	0.7825	0.8675
Q25-1	Macro C	0.4607	0.6318	0.5938	0.6443
Q25-1	Micro C	0.4073	0.6516	0.5808	0.6620

Appendix

Table 59: MIMIC: F1 of the Hamming-minimizing oracle and constrained selections. Oracle F1 is diagnostic and is not an F1 upper bound.

Generator	Unique	Zero error (%)	Oracle errors	Ma errors	Mi errors	Ma missed (%)	Mi missed (%)
Q35	2.0673	67.0404	0.6009	0.9709	0.8072	16.1435	13.4529
IV35	3.2601	71.3004	0.4753	0.9664	0.7691	23.0942	18.1614
Q25-3	3.9776	73.9910	0.3498	1.5179	0.8274	61.6592	33.6323
Q25-1	2.9641	68.6099	0.5022	1.7892	0.8274	73.0942	27.3543
ESC-RL	4.1143	76.9058	0.3251	1.5785	0.8117	66.1435	33.1839
MPO	3.5000	72.1973	0.3969	1.2578	0.8072	48.2063	29.3722

Appendix

Table 60: IU: candidate-oracle diagnostics. Errors are mismatched scored labels per input, not normalized Hamming rates. Missed denotes a strictly better available candidate under masked Hamming error.

Generator	Cond.	Ma-14	Mi-14	Ma-5	Mi-5
Q35	Oracle	0.3715	0.7248	0.5009	0.5793
Q35	Macro C	0.2530	0.6110	0.2581	0.3779
Q35	Micro C	0.2499	0.6386	0.3146	0.3151
IV35	Oracle	0.4754	0.7850	0.5556	0.7200
IV35	Macro C	0.2656	0.6114	0.3044	0.4071
IV35	Micro C	0.2479	0.6567	0.3641	0.4217

Appendix

Table 61: IU: F1 of the Hamming-minimizing oracle and constrained selections. Oracle F1 is diagnostic and is not an F1 upper bound.

Table 55

Objective	Target	Route	Ret.	Ins.	Ham.	GREEN	RG	RaTE	RL	B4
macro	C	R	.8946	.0360	.0211	.4375	.2722	.6010	.2886	.0799
macro	C	G	.8919	.0309	.0208	.4890	.3200	.6453	.3575	.1161
macro	P	R	.8702	.0315	.0262	.4274	.2687	.5992	.2864	.0777
macro	P	G	.8645	.0271	.0264	.4817	.3147	.6434	.3542	.1167
macro	F	R	.8286	.0284	.0375	.4243	.2647	.5973	.2825	.0723
macro	F	G	.7958	.0221	.0426	.4813	.3156	.6442	.3529	.1184

Appendix

Table 64: MIMIC: fidelity and all five report metrics for the twelve main realization configurations.

Objective	Target	Route	Ret.	Ins.	Ham.	GREEN	RG	RaTE	RL	B4
macro	C	R	.9634	.0019	.0034	.4867	.2617	.5680	.2707	.0494
macro	C	G	.8626	.0673	.0175	.4963	.3154	.5953	.3329	.0961
macro	P	R	.4443	.0031	.1954	.4367	.2182	.5460	.2303	.0277
macro	P	G	.4589	.1876	.2269	.2920	.2195	.5162	.2706	.0612
macro	F	R	.3639	.0038	.4444	.0774	.1049	.4286	.2083	.0198
macro	F	G	.3469	.1332	.4925	.1981	.1975	.4991	.2402	.0454

Appendix

Table 65: IU: fidelity and all five report metrics for the twelve main realization configurations.

Table 58

Table 59

Data	Obj.	T/R	$D_{\mathrm{TP}}$	$D_{\mathrm{FP}}$	$A_{+}$	$A_{-}$
MIMIC	macro	C/R	121	167	31	69
MIMIC	macro	C/G	117	173	26	59
MIMIC	macro	P/R	158	235	23	69
MIMIC	macro	P/G	149	265	21	57
MIMIC	macro	F/R	206	381	19	70
MIMIC	macro	F/G	215	488	20	47

Appendix

Table 70: Pooled reference-scored changes for Qwen2.5-VL-7B (3 epoch): 1,432 MIMIC and 446 IU test inputs. C/P/F denote targets, R/G denote routes. Counts are descriptive.

Method	Ma-14	Mi-14	Ma-5	Mi-5	GREEN	RG	RaTE	RL	B4
Plurality	.3900	.6214	.5229	.6209	.4947	.3227	.6307	.3518	.1124
Most-typical	.3900	.6214	.5229	.6209	.4947	.3227	.6307	.3518	.1124
Consensus-CB	.4005	.6331	.5338	.6361	.4876	.3158	.6425	.3282	.1016
CCS-Emb	.3881	.6092	.5178	.6059	.4952	.3313	.6465	.3617	.1203
Classifier probe	.4400	.6163	.5625	.6348	.4783	.3106	.6385	.3234	.1063
CLIPScore	.4037	.6046	.5295	.6101	.4812	.3118	.6266	.3346	.1131

Appendix

Table 71: MIMIC: extended benchmark, including all five reference-based report metrics and additional baselines.

Configuration	Ma-14	Mi-14	Ma-5	Mi-5	GREEN	RG	RaTE	RL	B4
Macro, C $\to$ G	.4334	.6260	.5901	.6556	.4818	.3428	.6454	.3613	.1195
Macro, P $\to$ G	.4669	.6198	.5843	.6444	.4796	.3419	.6452	.3586	.1208
Macro, F $\to$ G	.4777	.6338	.6031	.6622	.4797	.3417	.6469	.3585	.1231
Micro, C $\to$ G	.4130	.6415	.5627	.6705	.4818	.3436	.6450	.3622	.1194
Micro, P $\to$ G	.4315	.6402	.5811	.6684	.4818	.3424	.6440	.3619	.1200
Micro, F $\to$ G	.4287	.6408	.5803	.6721	.4843	.3431	.6448	.3615	.1207

Appendix

Table 72: MIMIC: alternate-generation-scale results. These are supplementary and excluded from development configuration selection; the implementation notes identify a 1.5B image-label realization checkpoint (see Appendix C.4 ).

Method	Ma-14	Mi-14	Ma-5	Mi-5	GREEN	RG	RaTE	RL	B4
Plurality	.0586	.5940	.0082	.0202	.5678	.3129	.6373	.3009	.0523
Most-typical	.0875	.5966	.0082	.0202	.6028	.3810	.6510	.3397	.0751
Consensus-CB	.0663	.5924	.0000	.0000	.5668	.3122	.6370	.3003	.0511
CCS-Emb	.0775	.5842	.0000	.0000	.6022	.3740	.6454	.3383	.0785
Classifier probe	.1266	.5156	.1046	.1795	.5259	.2977	.6131	.2930	.0578
CLIPScore	.1762	.5635	.0987	.1653	.5689	.3257	.6222	.3262	.0796

Appendix

Table 73: IU: extended benchmark, including all five reference-based report metrics and additional baselines.

Configuration	Ma-14	Mi-14	Ma-5	Mi-5	GREEN	RG	RaTE	RL	B4
Macro, C $\to$ G	.1565	.5549	.1663	.3425	.5822	.3772	.6436	.3551	.0947
Macro, P $\to$ G	.0747	.1387	.1050	.1714	.2947	.2459	.5138	.2674	.0581
Macro, F $\to$ G	.0854	.1258	.1416	.2183	.2197	.2327	.5092	.2520	.0514
Micro, C $\to$ G	.1275	.6262	.1510	.3239	.6033	.4002	.6667	.3670	.0934
Micro, P $\to$ G	.1275	.6275	.1510	.3262	.6072	.4020	.6679	.3675	.0938
Micro, F $\to$ G	.1402	.6203	.2131	.3676	.5983	.3992	.6643	.3667	.0948

Appendix

Table 74: IU: alternate-generation-scale results. These are supplementary and excluded from development configuration selection; the implementation notes identify a 1.5B image-label realization checkpoint (see Appendix C.4 ).

Table 65

Radiology Report Generation

Radiology report generation uses medical images and clinical context to produce structured, clinically accurate findings and impressions.

See all 159 papers

RadFusion: Towards Threshold-Controllable Radiology Report Generation
Aug 11, 2026Ying Jin, Noel C. F. Codella, John Corring +3Radiology Report Generation
Microsoft
ReportQA: QA-Based Radiology Report Evaluation
Jun 13, 2026Yiming Shi, Shaoshuai Yang, Xi Chen +10Radiology Report Generation Medical VLMs
Department of Electronic Engineering, Tsinghua University · Beijing Electronic Digital & Intelligence · College of AI, Tsinghua University +1
Can Jev Judge Radiology Reports? Evaluating a System One Model for Clinical Factuality
Sep 23, 2026Jiaju Huang, Hao Yang, Xinyu Ma +6Radiology Report Generation LLM-as-a-Judge
Intelligent Medical Computing Laboratory, Faculty of Applied Sciences, Macao Polytechnic University · Department of Radiology, Netherlands Cancer Institute, Amsterdam, the Netherlands · Department of Radiology and Nuclear Medicine, Radboud University Medical Center, Nijmegen, the Netherlands

Aug 11, 2026J/K move · Enter open · S save

RadFusion: Towards Threshold-Controllable Radiology Report Generation

Ying Jin, Noel C. F. Codella, John Corring, Mu Wei+2

Microsoft

Radiology Report Generation

Open paper AskPDF arXiv

Automated radiology report generation is advancing rapidly in response to the shortage of radiologists, yet unlike a perception model, existing generation models offer no control over the sensitivity-specificity trade-off of their diagnostic content. Such control is essential because clinical scenarios diverge: emergency triage prioritizes sensitivity to reduce missed findings, whereas confirmatory interpretation emphasizes specificity to limit unnecessary interventions. A single fixed report can neither adapt to these scenarios nor support the ROC-based validation widely expected for regulatory clearance. We introduce RadFusion, a framework that equips report generation with threshold controllability. Our method fuses a multi-label classifier, which provides per-disease confidence scores, with a VQA-based report generator, which describes medical findings in detail; an LLM then rewrites the report so that its stated diagnoses follow the classifier's decisions at the selected threshold while staying grounded in the generator's descriptions. On MIMIC-CXR, the performance of RadFusion conforms to the classifier's ROC curve: sweeping the threshold and mapping the reports back to class labels reproduces the classifier's validated ROC performance. This conformance makes generated reports quantitatively evaluable through ROC analysis, strengthening the case for regulatory clearance, and enables operating-point selection that matches report behavior to clinical context. Moreover, combining the two model types improves diagnostic accuracy over uncontrolled generation: sensitivity increases by 6.9% at matched specificity, and specificity by 20.7% at matched sensitivity. These results show that RadFusion makes report generation clinically adaptable, quantitatively verifiable, and diagnostically more reliable.

Abstract

Radiology Report Generation

Explore similar work

RadFusion: Towards Threshold-Controllable Radiology Report Generation

ReportQA: QA-Based Radiology Report Evaluation

Can Jev Judge Radiology Reports? Evaluating a System One Model for Clinical Factuality