Group-wise Evidence-aware Alignment Reward (GEAR)

• The paper introduces GEAR as a disease-level reward component that partitions findings into true positives, false negatives, and false positives.
• GEAR leverages disease-grounded response maps from a frozen vision-language model to align generated reports with visual evidence via tailored loss functions.
• Empirical results show that integrating GEAR improves clinical metrics like GREEN and RadGraphF1 by up to 1.4% over baseline methods.

Group-wise Evidence-aware Alignment Reward (GEAR) is the reward-design component of the ESC-RL framework, introduced for reinforcement learning in radiology report generation to address a specific limitation of prior report-level rewards: they provide only a coarse scalar signal and do not indicate which disease statement is correct, which finding was omitted, which one was hallucinated, or whether a claimed finding is grounded in the image. In this formulation, GEAR operates at the finding or disease level rather than at the token, sentence, or whole-report level, and it uses disease-grounded visual evidence under weak supervision, without manual region annotations, to reinforce grounded true positives, recover false negatives, and suppress false positives (Zhou et al., 15 Apr 2026).

1. Clinical motivation and formal setting

GEAR is motivated by two recurrent errors in RL-based radiology report generation: missed findings and unsupported hallucinations. Earlier rewards based on BLEU, RadCliQ, or other clinical-efficacy metrics may correlate with report quality, but they do not provide disease-specific guidance about omission, fabrication, or visual grounding. The paper treats this as a clinical-faithfulness problem rather than a purely lexical one: in radiology, correctness depends not only on naming the right abnormality, but also on tying that abnormality to the correct visual evidence.

Within ESC-RL, the report generator is cast as a policy model $\psi_p$ with parameters $\theta$. Given an image $\mathbf{I}$, a report $R=(y_1,\ldots,y_L)$ is generated according to

$$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$

The paper also presents a generic expected-reward RL objective, but GEAR enters the actual training system as an additional clinically grounded alignment term. A central point is that the reward unit is a disease category $k$, not a word, sentence, or complete report. This is the paper’s precise meaning of group-wise.

2. Disease-group decomposition

GEAR assumes a disease taxonomy of size $K$. Ground-truth disease statuses are denoted by

$Y^* \in \{0,1,2,3\}^{K},$

where the four values correspond to blank, positive, negative, and uncertain. For a generated or refined report $R$, CheXbert is used to extract

$Y \in \{0,1,2,3\}^{K}.$

For reward construction, only the positive/absent distinction is used; the paper explicitly states that negative and uncertain statuses are excluded during group partition for simplicity.

The disease indices are partitioned into three clinically distinct groups:

$\theta$0

$\theta$1

$\theta$2

This decomposition allows GEAR to distinguish three different situations: a finding correctly mentioned, a finding present but omitted, and a finding invented by the report. The paper emphasizes that this is fundamentally different from token-level or sentence-level scoring. The grouping is not a convenience layer over a report-level reward; it is the mechanism by which the reward becomes disease-conditional.

A plausible implication is that GEAR operationalizes error correction in a clinically structured manner. True positives require spatial consistency, false negatives require evidence recovery, and false positives require suppression rather than matching.

3. Evidence representation through Disease-grounded Response Maps

The evidence used by GEAR is a set of Disease-grounded Response Maps (DRMs), one channel per disease category. These are produced by a frozen vision-language grounding model $\theta$3 pretrained on image-report pairs. The implementation uses MAVL by default, while the appendix reports that MedKLIP can be substituted.

Given a chest X-ray image $\theta$4, the model computes DRMs conditioned on the generated or refined report $\theta$5 and on the ground-truth report $\theta$6:

$\theta$7

$\theta$8

Here $\theta$9 is the DRM tensor induced by the generated report and $\mathbf{I}$0 is the DRM tensor induced by the reference report.

The paper’s notion of evidence-aware is therefore concrete: evidence is represented as spatial disease-conditioned activation maps over image regions. Alignment is not computed by directly matching words to patches. Instead, GEAR compares disease-specific response maps induced by the generated report and the ground-truth report. This provides weakly supervised grounding without manual region annotations.

For technical interpretation, the important consequence is that GEAR uses an external frozen grounding backbone as an evidence extractor. It therefore depends on the quality of the grounding model, but it also avoids requiring dense localization labels.

4. Group-specific alignment terms and optimization role

GEAR is defined through three group-specific losses. For true positives, the aim is to ensure that a correctly predicted disease is grounded in the same image region or regions as the ground-truth report. The paper describes this as an IoU-based loss, although it is written in a soft overlap or Dice-style form:

$\mathbf{I}$1

Its role is to reward grounded correct findings rather than merely correct mentions.

For false negatives, the disease is present in the ground truth but absent in the generated report. The paper uses an MSE term:

$\mathbf{I}$2

The paper itself notes an ambiguity of sign in presentation: the formula includes a leading negative sign, but the surrounding text states that the intended semantics are to encourage the predicted DRM to match the ground-truth DRM and thereby recover missed evidence. The intended role of the term is therefore clear even though the sign convention in the typesetting is not fully consistent (Zhou et al., 15 Apr 2026).

For false positives, there is no positive reference DRM to align to, so GEAR suppresses unsupported activation:

$\mathbf{I}$3

This penalizes response energy for diseases that should not be present, discouraging hallucinated findings.

The overall alignment term is

$\mathbf{I}$4

The appendix argues that the three groups should not share the same loss form. Empirically, the best configuration is exactly the proposed combination: IoU-based TP + MSE-based FN + suppression-based FP. The paper explains that MSE on TP is too background-dominated and blurs boundaries, while IoU for FN can be unstable when the predicted FN map is weak.

GEAR is not a standalone training paradigm. It is integrated into the ESC-RL objective as

$\mathbf{I}$5

where $\mathbf{I}$6 is the SPL predictor loss and $\mathbf{I}$7 is set to $\mathbf{I}$8 by default. The appendix reports that $\mathbf{I}$9 works best: too large a value overemphasizes evidence alignment and destabilizes training, while too small a value makes the guidance too weak. The paper presents the broader framework in RL language, but it does not specify a detailed optimizer such as PPO, SCST, or REINFORCE with a baseline. Mathematically, GEAR therefore behaves like an auxiliary alignment loss term inside a hybrid objective rather than like a classic scalar episodic reward fed into an explicit policy-gradient estimator (Zhou et al., 15 Apr 2026).

5. Relation to ESC-RL and empirical contribution

Within ESC-RL, GEAR and Self-correcting Preference Learning (SPL) address different weaknesses. GEAR is the evidence-aware reward alignment mechanism: it partitions disease states into TP, FN, and FP groups and compares disease-grounded response maps. SPL is the preference-based self-correction mechanism: it constructs a disease-aware preference dataset from multiple noisy observations, filters them with a learned predictor, and uses an LLM to synthesize refined reports without human supervision. In the paper’s framing, GEAR answers whether generated disease content is visually grounded and clinically supported, while SPL answers which disease descriptions from multiple noisy candidates should be trusted and retained.

The ablation study on MIMIC-CXR isolates GEAR’s contribution most clearly. Using REVTAF as baseline, the reported scores are:

• Baseline: GREEN $R=(y_1,\ldots,y_L)$0, RadGraphF1 $R=(y_1,\ldots,y_L)$1, CheXbertF1 $R=(y_1,\ldots,y_L)$2
• GEAR only: GREEN $R=(y_1,\ldots,y_L)$3, RadGraphF1 $R=(y_1,\ldots,y_L)$4, CheXbertF1 $R=(y_1,\ldots,y_L)$5

Thus, adding GEAR alone yields GREEN $R=(y_1,\ldots,y_L)$6 absolute, i.e. +1.4%, RadGraphF1 $R=(y_1,\ldots,y_L)$7, CheXbertF1 $R=(y_1,\ldots,y_L)$8, and BLEU-4 from $R=(y_1,\ldots,y_L)$9 to $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$0. The paper states: “GEAR alone boosts GREEN by 1.4\%, supporting the benefit of evidence-aware reward shaping for clinically aligned grounding.” It also compares against a generic RL signal: RL only gives GREEN $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$1, GEAR only gives GREEN $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$2, and RL + GEAR gives GREEN $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$3, suggesting that GEAR contributes beyond plain report-level RL (Zhou et al., 15 Apr 2026).

For the full ESC-RL system, the paper reports on MIMIC-CXR:

• BLEU-1 $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$4
• BLEU-4 $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$5
• ROUGE $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$6
• BERTScore $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$7
• RadCliQ $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$8
• RadGraphF1 $$\psi_p(R\mid \mathbf{I})=\prod_{l=1}^{L}\psi_p(y_l\mid y_1,\ldots,y_{l-1},\mathbf{I}).$$9
• CheXbertF1 $k$0
• GREEN $k$1

Relative to baseline, the full model improves BLEU-1, RadGraphF1, and GREEN by 2.2%, 2.5%, and 5.0%, respectively. These gains are not attributable to GEAR alone, but the paper treats GEAR as one of the major contributors, especially on clinically focused metrics. It further reports state-of-the-art results on both MIMIC-CXR and IU-Xray, with gains over the strongest comparison model on MIMIC-CXR of 1.5 points in RadGraphF1, 1.6 points in CheXbertF1, and 1.3 points in GREEN.

The qualitative analysis is also consistent with GEAR’s intended function, although it reflects full ESC-RL rather than an isolated GEAR module. The paper says ESC-RL recovers key report content, better identifies support devices and major findings such as cardiomediastinal contours, cardiomegaly, and atelectasis, and captures location-specific abnormalities such as “left retrocardiac atelectasis” and “confluent opacity in the right infrahilar region.” Competing methods are described as producing more factual errors, incomplete descriptions, and hallucinations.

Reproducibility details further clarify the operational regime. The reward grouping uses disease-status vectors in $k$2, while CE evaluation uses 14 diseases via CheXbert. The appendix reports 98.9% extraction accuracy for CheXbert on 1,000 MIMIC-CXR test samples against ground-truth labels. DRMs are extracted from 224-resized images. The policy model backbone is pre-trained REVTAF. Training uses AdamW with learning rate $k$3, weight decay $k$4, cosine schedule, 6 epochs, and batch size 18. Runs are reported on NVIDIA A800 80GB with about 20 hours per run. The framework introduces additional overhead from pretrained models; MAVL adds about 0.58M parameters and about 0.3 s/sample at inference, while CheXbert is used only during training and not at test time.

6. Interpretation, limitations, and nomenclatural ambiguity

The paper assigns GEAR a relatively precise technical role: it is a disease-level, group-conditional evidence alignment mechanism. A common misconception is to read it as a token-level reward or as a generic report-level scalarization. That reading is incorrect. GEAR works by partitioning disease categories into TP, FN, and FP sets, producing disease-grounded response maps for generated and reference reports, and applying different alignment penalties to each group.

A second point of clarification concerns the word reward. The paper itself writes GEAR as $k$5 and adds it positively to the total objective. Technically, this makes it closer to an auxiliary alignment loss than to a classic episodic reward. This is not necessarily contradictory, but it is important for interpretation: the framework is RL-cast, while the displayed optimization is hybrid and directly regularized.

The method also has explicit limitations. It depends on the reliability of the disease-status extractor CheXbert, on the quality of the grounding model’s DRMs, on a predefined disease taxonomy, and on the simplifying exclusion of negative and uncertain statuses from group partition. The experiments are limited to chest X-rays, so generalization to CT, MRI, or other anatomical regions remains unvalidated. A plausible implication is that transferring GEAR to other modalities would require re-evaluating both the disease ontology and the grounding backbone rather than merely reusing the loss definitions.

The acronym itself is ambiguous in recent literature. Other arXiv papers use GEAR for unrelated methods, including “Geometric Embedding Alignment via cuRvature matching in transfer learning” (Ko et al., 16 Jun 2025), “Graphical Event Aggregation for Rubric rewards” (Lv et al., 2 Jun 2026), and “Granularity-Adaptive Advantage Reweighting” (Li et al., 12 May 2026). In the present usage, however, GEAR refers specifically to Group-wise Evidence-aware Alignment Reward inside ESC-RL for radiology report generation (Zhou et al., 15 Apr 2026).