CheXPO-v2: Optimizing Chest X-ray VLM Reasoning

• The paper introduces a process-supervised extension of GRPO that transforms free-form chain-of-thought reasoning into structured medical triplets for chest X-ray VLMs.
• It combines supervised fine-tuning, hard‐example mining, retrieval-based augmentation, and entity-relation matching to boost accuracy to 86.00% with only 5k preference samples.
• The framework employs a novel knowledge graph consistency reward that evaluates extracted entities and relations, ensuring verifiable and clinically reliable reasoning.

Searching arXiv for CheXPO-v2 and closely related chest X-ray VLM preference-optimization papers. CheXPO-v2 is a preference-optimization framework for chest X-ray vision-LLMs (VLMs) that is intended to improve not only final-answer accuracy but also the clinical reliability of the reasoning process. It is presented as a process-supervised extension of Group Relative Policy Optimization (GRPO) in which free-form chain-of-thought reasoning is parsed into structured medical triplets and evaluated with a knowledge graph consistency reward. In the reported formulation, CheXPO-v2 combines a supervised fine-tuning warm-up, hard-example mining, retrieval-based augmentation, and entity-relation matching, and it is reported to reach 86.00% overall accuracy on a chest X-ray VQA benchmark with 5k preference samples using Phi-4MM (Liang et al., 19 Dec 2025).

1. Clinical motivation and problem setting

The framework is motivated by a specific failure mode of medical VLMs: they can produce fluent, radiology-like outputs that are clinically unsafe because the reasoning is hallucinated, incomplete, or internally inconsistent. The paper characterizes this as a reliability problem rather than a purely predictive one. In chest radiology, a model may invent findings that are not visible, miss important abnormalities, justify a correct answer using incorrect reasoning, or generate verbose and difficult-to-verify chain-of-thought traces. The authors describe this tendency as “overthinking” and argue that it is encouraged by sparse, outcome-based rewards that privilege the final answer while ignoring intermediate reasoning quality (Liang et al., 19 Dec 2025).

A central claim is that outcome-only reinforcement learning is insufficient in medical settings because a correct answer can be paired with clinically invalid reasoning. The paper identifies two associated failures in standard RL alignment: overthinking, in which verbose reasoning is effectively rewarded because only the final answer matters, and process blindness, in which clinically invalid intermediate steps are not penalized if the output label is correct. The resulting concern is not stylistic; the authors explicitly frame hallucinated radiological justifications as a safety risk.

This position differs from narrower chest X-ray systems that focus on pathology classification alone. Earlier chest radiograph work emphasized single-model versus ensemble performance for multi-label screening (Mitra et al., 2020), and later work also addressed weak supervision from radiology reports through CheXpert-style label extraction (Wollek et al., 2023). CheXPO-v2 instead targets reasoning supervision in chest X-ray VLMs, where the model is expected to answer and justify.

2. Two-stage pipeline and training formulation

CheXPO-v2 is organized as a two-stage pipeline. The first stage is a supervised fine-tuning warm-up. The second stage applies preference optimization with GRPO (Liang et al., 19 Dec 2025).

In the supervised stage, the model is fine-tuned on a synthesized chest X-ray instruction dataset. The implementation described in the paper fine-tunes Phi-4MM with LoRA and introduces explicit output-format tokens: > ... for chain-of-thought and <answer> ... </answer> for the final answer. The training corpus is built from MIMIC-CXR, MIMIC-CXR-VQA, Medical-Diff-VQA, and Chest ImaGenome, and the paper states that 640k expert-level reasoning chains are synthesized with GPT-4o. The resulting instruction set covers 3 tasks—Basic QA, Region QA, and Comparison QA—and 10 question types: Presence, Abnormality, Anatomy, Severity, Plane, Type, Difference, Attribute, Size, and Gender (Liang et al., 19 Dec 2025).

The second stage retains the GRPO optimization structure but changes the reward design. GRPO samples multiple candidate responses for a prompt and computes a group-relative advantage,

$$A_i = \frac{R_i - \text{mean}(\{R_1, \ldots, R_G\})}{\text{std}(\{R_1, \ldots, R_G\})}.$$

The paper contrasts this with PPO by noting that PPO uses a critic or value model, whereas GRPO avoids a critic and instead uses group-relative normalized rewards. It also contrasts the method with DPO: DPO relies on preference pairs, while CheXPO-v2 uses automated verifiable rewards derived from structured reasoning comparison rather than expensive human preference annotations (Liang et al., 19 Dec 2025).

A recurrent theme is that CheXPO-v2 does not replace GRPO as an optimizer; it redefines what is rewarded.

3. Knowledge graph consistency reward

The defining innovation of CheXPO-v2 is the Knowledge Graph Consistency Reward. Rather than evaluating only answer correctness or formatting compliance, the framework parses the chain-of-thought into a structured medical graph and compares that graph to ground-truth reasoning (Liang et al., 19 Dec 2025).

The extraction pipeline has two steps. First, a radiology-oriented NER model from ReXKG extracts entities of six types: Anatomy, Disorder, Concept, Device, Procedure, and Size. Second, a relation extractor predicts directed triplets using three relation types: Located at, Suggestive of, and Modify. These are assembled as triplets of the form $(e_h, r, e_t)$ and are also described conceptually as $(\text{Disease}, \text{Relation}, \text{Anatomy})$ (Liang et al., 19 Dec 2025).

The final reward is a weighted sum of answer reward, entity consistency, and relation consistency:

• $w_1 = 1$ for answer reward
• $w_2 = 0.5$ for entity reward
• $w_3 = 0.5$ for relation reward

The paper’s interpretation is that a model should be rewarded for three things simultaneously: giving the right answer, invoking the right entities, and connecting them with the right relations. This shifts alignment from outcome supervision to process supervision. The intended effect is to produce what the paper describes as a more clinically consistent, proof-like reasoning chain.

Evaluation is similarly expanded beyond answer accuracy. CheXPO-v2 reports:

• Answer accuracy via strict string matching
• Entity-Relation Matching with Precision, Recall, and F1
• Knowledge graph metrics: KG-NSC, KG-AMS, and KG-SCS
• Hit Rate, measuring whether the final answer appears in the reasoning trace
• CoT length statistics, used to analyze verbosity and overthinking

The paper explicitly argues against the common assumption that answer-only reward is sufficient. Its ablations indicate that answer-only reward can improve accuracy while worsening reasoning fidelity and increasing verbosity, whereas entity-relation supervision improves the structural quality of the reasoning trace (Liang et al., 19 Dec 2025).

4. Hard-example mining and retrieval-based augmentation

A major practical component of CheXPO-v2 is its hard-example mining strategy. The motivation is computational and statistical: GRPO requires sampling multiple outputs per prompt, and easy examples where the SFT model is already accurate and confident provide little useful gradient. The framework therefore targets examples on which the model is wrong or uncertain (Liang et al., 19 Dec 2025).

The paper describes a stratified sampling stage over question and answer types. Conceptually, the subset ratio is less than 5%; in the implementation description, $\gamma = 2\%$ is used, producing roughly 10k samples for initial assessment. For each sample, the SFT model generates a chain-of-thought answer. A sample is treated as hard if the short answer is incorrect. It is also treated as hard if the answer is correct but the model confidence is low, using a length-normalized log-probability threshold:

• low-confidence criterion: $\mathbf{p} < \sigma$
• implementation threshold: $\sigma = -0.25$

The paper’s rationale is that low-confidence correct cases can still be valuable because they induce high reward variance under group sampling and improve GRPO learning.

After identifying hard cases, CheXPO-v2 performs retrieval-based augmentation using BioMedCLIP features over question, rationale, and image. Similarity is aggregated as

$$\mathcal{S} = \mathcal{S}_{\mathcal Q} + \mathcal{S}_{\mathcal T} + \mathcal{S}_{\mathcal V},$$

and Top-K nearest samples are retrieved from the remaining data. The purpose is to enrich the GRPO training set with semantically related, clinically relevant examples rather than abundant easy cases (Liang et al., 19 Dec 2025).

This design differs from the earlier CheXPO framework, which used confidence-similarity joint mining and counterfactual rationale within a DPO pipeline (Liang et al., 9 Jul 2025). CheXPO-v2 retains the emphasis on hard cases and semantic expansion but replaces the preference mechanism with process-aware knowledge graph supervision under GRPO.

5. Quantitative performance and ablation findings

The main reported result is that CheXPO-v2 on Phi-4MM achieves:

• 82.80% with 1k samples
• 86.00% with 5k samples

The paper presents 86.00% as the best overall result in its chest X-ray VQA benchmark table and explicitly emphasizes that this is achieved with only 5k preference samples (Liang et al., 19 Dec 2025).

The reported gains are substantial relative to both raw foundation models and SFT-only baselines. For example:

• Phi-3.5V improves from 31.30% to 73.20% with SFT-CoT
• Phi-4MM improves from 38.17% to 75.17% with SFT-CoT
• CheXPO-v2 then improves further from 75.17% to 86.00% with 5k preference samples

By question type, the paper reports especially strong CheXPO-v2 performance on:

• Abnormality: 85.93%
• Anatomy: 76.36%
• Plane: 100%
• Difference: 100%
• Gender: 100%
• Size: 92.31%

The ablation studies support two claims. First, hard-example selection matters. With 1k samples, the paper reports:

• SFT baseline: 75.17%
• Random: 77.17%
• Uniform: 80.50%
• Hard-Uniform: 82.80%

Second, reward design matters. The ablations compare answer-only, entity-only, relation-only, and combined rewards, as well as Precision, Recall, F1, and Jaccard variants for matching. The paper states that Jaccard is the best metric overall because it balances precision and recall. Its best configuration, acc + Jaccard(E+R), yields:

• Accuracy: 0.828
• KG-NSC: 0.880
• KG-AMS: 0.781
• KG-SCS: 0.904

Additional analyses reinforce the process-supervision claim. GPT-4o-based reasoning assessment indicates more correct reasoning traces when answers are correct; CoT length analysis shows that answer-only reward tends to produce longer and more verbose chain-of-thought; and relation miss-rate analysis shows lower miss rates for common relation triplets under CheXPO-v2. The paper interprets these results as evidence that process supervision reduces hallucination and pathological overthinking (Liang et al., 19 Dec 2025).

6. Relation to prior CheXPO, limitations, and significance

CheXPO-v2 is explicitly positioned against the authors’ earlier CheXPO framework. The earlier method used DPO, confidence-similarity joint mining, and counterfactual rationale, and it reported 79.74% overall accuracy on a unified chest X-ray VQA benchmark with 30k preference samples; on MIMIC-CXR-VQA, it reported 81.8 accuracy and 0.84 micro-F1 for CheX-Phi3.5V(30k) (Liang et al., 9 Jul 2025). By contrast, CheXPO-v2 is described as more process-aware, more data efficient, and less dependent on manually constructed preference signals. The paper repeatedly emphasizes the efficiency contrast: 86.0% overall accuracy with 5k preference samples in CheXPO-v2 versus 30k in the prior version (Liang et al., 19 Dec 2025).

The limitations are also explicit. The reward depends on automatically extracted entities and relations, so errors in the NER or relation-extraction pipeline can affect supervision quality. The framework is designed around chest X-ray and radiology-style reasoning, and the paper indicates that adaptation would be required for other modalities or broader clinical tasks. It also notes that the current formulation focuses on textual consistency and identifies tighter region-aware visual interpretability as future work (Liang et al., 19 Dec 2025).

In clinical terms, the significance of CheXPO-v2 lies less in replacing classification systems than in changing the alignment target for chest X-ray VLMs. The framework is intended to make explanations verifiable rather than merely fluent. It aligns with the paper’s stated model of radiological reasoning: identify anatomy, identify abnormal findings, connect findings to disease, and infer the final diagnosis. A plausible implication is that CheXPO-v2 belongs to a broader transition in chest X-ray AI from narrow pathology prediction toward models evaluated on structured reasoning, explanation quality, and generalist VQA behavior rather than final labels alone.