Clinical Cognition Alignment for Gastrointestinal Diagnosis with Multimodal LLMs¶

Conference: ECCV2026
Paper: ECCV Open Access
Area: Multimodal VLM
Keywords: gastrointestinal endoscopy diagnosis / multimodal large language model / clinical cognition alignment / counterfactual reinforcement learning / GRPO

TL;DR¶

Targeting two failures of multimodal large language models (MLLMs) in gastrointestinal endoscopy — reasoning that ignores the expert cognitive pathway and diagnoses driven by spurious background correlations — this paper curates a dataset whose reasoning chains follow an explicit "anatomical localization → morphological evaluation → micro-detail analysis" hierarchy, internalizes it with SFT, formally shows that standard supervised tuning necessarily converges to background shortcuts, and corrects the bias with a counterfactual-driven GRPO that synthesizes lesion-masked normal samples and optimizes clinical-cognition-centric rewards, reaching 67.67% average accuracy across five endoscopy benchmarks and surpassing Qwen3-VL-Plus by roughly 26 points.

Background & Motivation¶

Gastrointestinal malignancies account for a substantial share of the global cancer burden, and endoscopic screening is the gold standard for early detection and intervention. Because examination quality depends heavily on operator experience and inter-observer variability, computer-aided diagnosis is regarded as a key support tool for reducing miss rates. The past decade of this field was dominated by discriminative deep learning: CNNs and ViTs already reach expert-level accuracy on individual tasks such as polyp detection, anatomical landmark recognition, and ulcer classification. The problem is that these models are closed-set classifiers — they emit a category label and nothing else. The absence of a diagnostic rationale means clinicians cannot audit the output, which directly undermines trust in high-stakes medical settings.

MLLMs combine the perceptual capacity of a visual encoder with the knowledge and inference ability of an LLM, can produce a visual description and a conclusion at the same time, and in principle can mimic the workflow of an endoscopist. Deploying a general MLLM on endoscopy, however, runs into two concrete obstacles. The first is clinical cognition misalignment: an endoscopist follows a strict hierarchical flow — localize the anatomical segment, assess macroscopic morphology (shape, elevation, size, color, boundary), analyze micro-details (villous or pit patterns and microvascular architecture), and only then conclude — whereas general MLLMs reason in a scattered manner, skip critical analytical steps, and even hallucinate features that are not in the image. The paper gives a telling failure case: shown a submucosal tumor (SMT), Gemini 3 Pro is led astray by the "this is a capsule endoscopy" context and grounds its diagnosis in the modality cue rather than in the lesion itself. The second is the absence of a causal link between visual features and the diagnostic conclusion: the model is sensitive to background confounders (instrument artifacts, bubbles, bile residue) and latches onto the easiest environmental cues instead of the pathology. Together these two deficiencies make the model brittle in real clinical environments.

Existing medical multimodal work — general medical VQA models and endoscopy report generation — largely treats endoscopic analysis as a standard image captioning problem: the only supervision is image-text pairs, so the model learns shortcut features rather than clinical logic. This paper closes the gap in two steps. It first turns the expert diagnostic logic itself into data and supervision, so the model "reads" the image in the right order; it then shows that this supervision is statistically bound to slide toward background shortcuts, and pulls it back to the lesion with a constructible counterfactual penalty. Core idea: internalize the expert's hierarchical diagnostic flow as a generation paradigm through a three-stage clinical cognition dataset and SFT, then use GRPO over lesion-masked "normal" counterfactual samples with format, cognition, and diagnostic rewards to force the conclusion to hang on lesion features alone.

Method¶

Overall Architecture¶

The input is an endoscopic image plus a diagnostic instruction; the output must contain both the correct diagnosis and a reasoning chain that conforms to clinical standards. CogAlign achieves this dual alignment in two serial stages: aligning the reasoning process first, then the causal grounding. In stage one, a hierarchical clinical cognition dataset is built — every image is paired with a chain of "anatomical localization → morphological evaluation → micro-detail analysis → diagnosis" — and SFT writes this generation paradigm into the model; at that point the model talks in the right order but may still be reading the background. Stage two opens with a formal analysis of why the gradients of standard SFT preferentially absorb spurious background features, and designs a counterfactual-driven GRPO accordingly: lesion regions are masked and Gaussian-blurred to synthesize counterfactual samples labelled "normal", and format, clinical cognition, and diagnostic consistency rewards are optimized with group-relative advantages, so that any generation that answers from the background is penalized.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Endoscopic image + query"] --> B["Hierarchical clinical<br/>cognition dataset"]
    B --> C["SFT internalizing<br/>the cognition chain"]
    C --> D["Formalization of visual<br/>shortcuts + CF penalty"]
    D -->|"λ→∞ zeroes spurious sensitivity"| E["Counterfactual-driven GRPO<br/>masked normal samples<br/>three cognition rewards"]
    E --> F["Structured chain<br/>+ final diagnosis"]

Key Designs¶

1. Hierarchical clinical cognition dataset: freeze the expert's three-stage flow into a superviseable reasoning chain

Public endoscopy datasets are essentially image-label pairs without the intermediate reasoning steps a diagnosis requires; training on them pushes the model to fit shortcut features instead of clinical logic. The paper first defines an annotation protocol derived from expert workflows that enforces a strict coarse-to-fine path with three fixed stages: anatomical localization (identify the organ segment and record imaging conditions to supply spatial context), morphological evaluation (shape, elevation, size, color, and boundary of the lesion), and micro-detail analysis (fine-grained surface patterns such as villous structures, together with vascular configuration), followed by the final diagnosis. Manual annotation of reasoning chains at scale is prohibitively expensive, so curation is semi-automatic: images are aggregated from five public repositories and filtered for quality and diversity; Gemini 3 Pro serves as the teacher model and is prompted with the explicit three-stage hierarchy to generate a structured reasoning description per image; finally, a human-in-the-loop refinement pass has experts review every annotation, auto-saving the ones that pass and manually revising those containing factual errors. The corpus contains 24,515 samples (CrohnIPI, GastroVision, HyperKvasir, Kvasir-Capsule, and The SEE-AI Project), split stratifically into 19,736 training and 4,779 test samples and covering 23 single-label categories plus 49 multi-label pathology combinations, written as \(\mathcal{D}=\{(x_i,q_i,r_i,l_i)\}_{i=1}^{N}\). The value of this dataset is not its scale but the fact that it turns "how an expert looks at an image" from an implicit prior into a token-level supervision target.

2. SFT that internalizes the cognition chain: making the conclusion a consequence of the preceding analysis

With reasoning chains available, the paper constrains the generation manifold in the most direct way: the chain \(r\) and the label \(l\) are concatenated into a target sequence \(y=r\oplus l\), and the whole sequence is optimized with an autoregressive negative log-likelihood, with the image projected into the LLM embedding space by the visual encoder so that the loss covers every token of the reasoning steps and the conclusion. The effect is not merely "an extra explanation in the output": the diagnosis sits at the end of the sequence, so its conditional probability already contains the preceding morphological and micro-detail descriptions — the label becomes a conditional consequence of the earlier analysis rather than an opaque classification straight from visual features. This is precisely the cognitive misalignment of general MLLMs: they are not incapable of describing, they simply have no causal chain between description and conclusion and can jump straight to an answer at any time. SFT binds intermediate steps and conclusion into one supervised sequence, so skipping steps is suppressed in the training distribution.

3. Formalizing visual shortcuts and the counterfactual penalty: why SFT alone is not enough

SFT makes the model talk in the right order, but not necessarily for the right reason. The paper offers a simplified yet non-vacuous formalization: an image is generated from a causal latent \(Z_c\) (the lesion) and a spurious latent \(Z_e\) (background, instruments, bubbles, specular reflections), the model output is a logistic function of features, and effective feature sensitivity is defined as the Jacobian norm of the output with respect to a factor,

\[ S_i=\left\|\nabla_{Z_i} f_\theta(Z_c,Z_e)\right\|_2, \]

which measures how much the model actually relies on that factor. Within this framework, Theorem 1 states that when the spurious features are "simpler" than the causal ones, gradient descent yields \(\|w_e\|>\|w_c\|\), i.e. \(S_e>S_c\): by the simplicity bias, low-complexity feature mappings have a larger spectral norm and converge faster early in training, and once the prediction error goes to zero the optimization stagnates while the causal weights are still under-learned. In other words, supervised tuning on image-label pairs is mathematically bound to learn the background first. Theorem 2 then shows that adding a counterfactual penalty \(R_{cf}=\mathbb{E}\left[f(\Psi(0,Z_e))^2\right]\) — the prediction still produced when the causal factor is zeroed out and only the background remains — and taking \(\lambda\to\infty\) in the total objective \(J=L+\lambda R_{cf}\) forces the optimum to satisfy \(f(x_{cf})\to 0\), hence \(S_e\to 0\); to keep reducing the loss on the original samples, the gradient can only turn toward the causal weights \(w_c\). The value of this analysis is that it converts "the model looks at the background" from an empirical complaint into an actionable prescription: more data is not needed, only an input with the same background but no lesion, and a penalty on its prediction.

4. Counterfactual-driven GRPO: masked normal samples plus three clinical-cognition-centric rewards

Stage two operationalizes that prescription. For counterfactual normal sample synthesis, the MLLM first produces an initial lesion box, experts refine it into a precise mask \(M\), and high-intensity Gaussian smoothing is applied inside \(M\) to obtain \(x_{cf}\), which is then paired with a "normal" label and a negative reasoning chain. The construction is the key: the background is untouched and only the lesion is erased, so the sample forms a natural control — if the model still reports a pathology from the background, its conclusion clearly does not rest on the lesion and should be penalized heavily. Gaussian blur rather than a solid white patch is a deliberate choice: white masks introduce sharp artificial boundaries and out-of-distribution signals that destabilize the RL optimization, whereas blurring neutralizes the diagnostic evidence while keeping the visual texture natural, letting the model isolate the lesion representation cleanly. Three rewards then define correct behavior. The format reward \(R_{fmt}\) is an all-or-nothing constraint: the output must cover, in order, (1) location and imaging environment, (2) mucosal morphology and focal lesions, and (3) surface texture and microvascular architecture; missing any section yields 0 and a complete structure yields 1, turning structural integrity into a hard constraint. The clinical cognition reward \(R_{cog}\) targets semantic precision: for each ground-truth reasoning chain an LLM pre-extracts three key features per cognitive section, nine in total, and training simply checks whether those keywords occur in the generated text,

\[ R_{cog}(\mathbf{y},K_{gt})=\frac{1}{9}\sum_{k\in K_{gt}}\mathbb{I}(k\in\mathbf{y}), \]

which turns the intermediate reasoning from an optional narrative into supervision that must name concrete semiological findings — without it the model degrades into vague descriptions lacking visual-pathological grounding. The diagnostic consistency reward \(R_{diag}\) parses the conclusion at the end of the response and gives 1 if it matches the ground-truth label and 0 otherwise, ensuring the chain lands on the right judgement; the ablation shows that removing it makes the model obsess over describing and occasionally forget to conclude from its own analysis. The three rewards are summed with weights 1.0 / 1.0 / 2.0 into a total reward and optimized by GRPO with group-normalized advantages. Note that the counterfactual samples are not an extra loss term here but ordinary training samples inside the same reward system: on \(x_{cf}\) the correct behavior is a negative chain and a "normal" verdict, so a model that answers from the background receives zero diagnostic reward — formally equivalent to \(R_{cf}\) in the theory, but implemented as one more batch of synthesized data.

A Worked Example: reward computation inside one GRPO sampling round¶

Take a gastric antrum white-light image (ground-truth label Erosion, heavily obscured by frothy bile-stained mucus and bubbles). For the same query the policy samples \(G=8\) outputs; consider three of them:

Output A: all three section headers present → \(R_{fmt}=1\); 5 of 9 key features hit → \(R_{cog}=5/9\approx0.56\); conclusion Erosion correct → \(R_{diag}=1\). Total reward \(r=1+0.56+2\times1=3.56\).
Output B: only two sections written (surface texture and microvascular architecture missing) → \(R_{fmt}=0\); 4 keywords hit → \(0.44\); conclusion correct → 1. Total reward \(r=0+0.44+2=2.44\); both structure and semantics are penalized.
Output C: all three sections present, but the model reads only the bubbles and concludes "smooth mucosa, no abnormality" → \(R_{fmt}=1\); 2 keywords hit → \(0.22\); conclusion Normal → 0. Total reward \(r=1+0.22+0=1.22\).

After the mean and standard deviation of the group are computed, the normalized advantage \(\hat{A}_i\) pushes C — the "well-formatted but background-driven" sample — to the bottom. Meanwhile, the counterfactual sample of the same image (lesion masked and Gaussian-blurred, labelled normal) is trained in the same batch; if the model still reports Erosion on it, its diagnostic reward is again zero. That is exactly the boundary between the two stages: the first teaches the model to speak in the clinical order, the second forces it to speak about the lesion.

Loss & Training¶

Stage one is a standard autoregressive negative log-likelihood, with \(y_i=r_i\oplus l_i\) and \(L_i\) the sequence length:

\[ \mathcal{L}_{SFT}(\theta)=-\frac{1}{N}\sum_{i=1}^{N}\sum_{t=1}^{L_i}\log P(y_{i,t}\mid x_i,q_i,y_{i,<t};\theta). \]

Stage two samples \(G\) outputs \(\{y_1,\dots,y_G\}\) per query, computes the total reward \(r_i=R_{fmt}(y_i)+\lambda_1 R_{cog}(y_i)+\lambda_2 R_{diag}(y_i)\), standardizes it within the group into the advantage \(\hat{A}_i=(r_i-\mu_r)/(\sigma_r+\epsilon)\), and optimizes the clipped surrogate objective with a KL penalty (\(\rho_i=\pi_\theta(y_i|q)/\pi_{\theta_{old}}(y_i|q)\), with \(\beta\) controlling the cost of deviating from the reference model \(\pi_{ref}\)):

\[ \mathcal{J}_{GRPO}(\theta)=\mathbb{E}_{q\sim D}\left[\frac{1}{G}\sum_{i=1}^{G}\min\big(\rho_i\hat{A}_i,\ \mathrm{clip}(\rho_i,1-\epsilon,1+\epsilon)\hat{A}_i\big)-\beta D_{KL}(\pi_\theta\|\pi_{ref})\right]. \]

⚠️ Both formulas above are reconstructed in their standard form from the paper's prose (Equations (2) and (15) are garbled in the cached text); refer to the original paper for the exact notation.

Implementation uses the SWIFT framework on eight NVIDIA L20 GPUs. Stage-one SFT runs 400 steps with AdamW, a learning rate of \(1\times10^{-4}\), a cosine scheduler, and a global batch size of 128; the vision encoder and aligner are frozen and LoRA (rank 16, \(\alpha=32\)) is applied to all linear modules, with sequences capped at 2048 tokens. Stage-two GRPO runs 200 steps, continuing the LoRA optimization with a global batch size of 256, a reduced learning rate of \(1\times10^{-6}\) and a KL penalty \(\beta=0.04\); each query samples \(G=8\) generations, and the format, clinical cognition, and diagnostic consistency rewards are weighted 1.0, 1.0, and 2.0. Both stages are lightweight adaptations — only the LoRA parameters are trainable, which is what allows the full pipeline to run at the 2B, 4B, and 8B scales.

Key Experimental Results¶

Main Results¶

The test suite contains 4,779 endoscopic images from five benchmarks — CrohnIPI (CI.), GastroVision (GV.), HyperKvasir (HK.), Kvasir-Capsule (KC.), and The SEE-AI Project (SA.) — evaluated by accuracy. SEE-AI is markedly harder, containing 235 multi-label samples whose predictions must match the complete set of ground-truth pathologies exactly.

Model	CI.	GV.	HK.	KC.	SA.	Average
Gemini 3 Pro	30.58%	44.73%	44.40%	21.83%	19.20%	24.82%
GPT-5.2	6.80%	18.80%	33.20%	5.32%	8.32%	11.13%
Qwen3-VL-Plus	52.91%	64.10%	72.78%	34.72%	33.63%	41.16%
Hulu-Med-7B (medical expert)	19.42%	13.39%	9.46%	10.86%	6.22%	8.58%
Qwen3-VL-8B	39.32%	47.01%	67.57%	30.14%	29.22%	35.30%
Qwen3-VL-8B (SFT)	62.14%	76.92%	89.38%	72.74%	58.77%	66.31%
CogAlign-2B	50.00%	73.79%	89.77%	53.99%	50.96%	57.40%
CogAlign-4B	59.22%	76.35%	89.19%	66.77%	57.22%	64.05%
CogAlign-8B	63.11%	77.21%	91.51%	74.01%	60.18%	67.67%

Ablation Study¶

The reward ablation is given only as curves in Fig. 6 of the paper (no readable numbers in the cached text), so the table below combines the two gain levels that can be checked directly against Tab. 1 and Tab. 2 with the qualitative conclusions of the reward removals:

Config	Average accuracy	Note
Qwen3-VL-8B (untrained)	35.30%	general backbone, no endoscopic cognition
+ clinical cognition alignment SFT	66.31%	structure and chain supervision only, +31.01
+ counterfactual-driven GRPO (CogAlign-8B)	67.67%	full model, a further +1.36
w/o clinical cognition reward \(R_{cog}\)	drops clearly, see Fig. 6 (⚠️ no numbers available)	intermediate reasoning degrades into vague descriptions without visual-pathological grounding
w/o diagnostic consistency reward \(R_{diag}\)	drops significantly, see Fig. 6 (⚠️ no numbers available)	the model focuses on describing and occasionally fails to conclude

Single-label versus multi-label breakdown on the same test set:

Model	Single-Label	Multi-Label	Average
Qwen3-VL-Plus	42.76%	10.21%	41.16%
Qwen3-VL-8B	36.77%	6.81%	35.30%
Qwen3-VL-8B (SFT)	69.19%	10.64%	66.31%
Hulu-Med-7B	9.02%	0.00%	8.58%
CogAlign-2B	59.93%	8.09%	57.38%
CogAlign-4B	66.81%	10.64%	64.05%
CogAlign-8B	70.47%	13.62%	67.67%

Key Findings¶

The bulk of the gain comes from the dataset and SFT, while the counterfactual RL is a correction rather than the main driver: at 8B, the entire +31.01 jump from 35.30% to 66.31% comes from cognition-chain supervision, and GRPO adds only another 1.36. Viewed the other way, the RL stage pays off more for smaller models — 2B gains 2.91 (54.49→57.40), 4B gains 2.07 (61.98→64.05), 8B gains 1.36 — consistent with smaller models being more prone to background shortcuts.
Multi-label diagnosis is the weak spot of every model and the place with the largest headroom: general models almost fail on this subset (Gemini 3 Pro 0.85%, GPT-5 series 0.43%–2.55%), the medical expert model Hulu-Med-7B scores 0.00%, and Qwen3-VL-8B reaches 6.81%; CogAlign-8B reaches 13.62%. The authors attribute this to the reasoning chain forcing a systematic scan of the mucosal surface instead of fixation on one salient feature. Still, the ~57-point gap to the 70.47% single-label figure shows concurrent pathology recognition is far from solved.
Medical expert models have no inherent advantage: Hulu-Med-4B/7B average only 7.72%/8.58%, below general models of comparable size. The authors attribute this to training objectives centered on general medical visual question answering rather than the fine-grained recognition endoscopy demands — "medical" and "endoscopic" are not the same specialty.
Dataset difficulty separates cleanly: Kvasir-Capsule and SEE-AI are the low-scoring benchmarks for every model (even CogAlign-8B reaches only 74.01% and 60.18%), matching capsule-endoscopy image quality and the higher multi-label ratio in SEE-AI.
Robustness and the masking strategy are each verified separately: under simulated spot interference, SFT-only models degrade severely while the full CogAlign stays stable at every scale, directly supporting the "SFT overfits spurious background correlations" claim; for masking, Gaussian blur consistently beats solid white patches, which the authors attribute to the sharp artificial boundaries and out-of-distribution signals introduced by white patches that destabilize the RL optimization.
One numeric discrepancy: the average accuracy of CogAlign-2B is 57.40% in Tab. 1 and 57.38% in Tab. 2, both on the same test set (⚠️ refer to the original paper; each figure is cited here from its own table).

Highlights & Insights¶

Treating cognitive misalignment as a data problem rather than a prompting problem: instead of repeatedly prompting the model at inference time to "locate first, then describe morphology", the paper makes that trajectory a supervised target sequence so that step-skipping is disallowed in the training distribution. The idea transfers directly to any diagnostic or review task with a mandated professional workflow (pathology, radiology, industrial inspection).
The counterfactual samples are trivial to construct yet map directly onto the theoretical term: no second model and no manual redrawing are needed — one lesion mask plus Gaussian blur produces a control sample where "background unchanged, lesion gone", turning the abstract counterfactual penalty \(R_{cf}\) into a usable batch of training data. Low cost is the most elegant part of the design.
The keyword reward turns "semantic precision" into an automatically computable discrete signal: nine key features extracted offline by an LLM, matched as strings during training, yield a dense reward far cheaper than training a reward model and without human scoring. This trick transfers to any generation task where both structure and content must be right, such as report generation or chain-of-thought quality assessment.
The theory serves only one penalty term, but it delivers an actionable prescription: Theorem 1 shows SFT alone inevitably learns spurious features, and Theorem 2 shows that zeroing the causal factor and penalizing the prediction on it removes spurious sensitivity. It does not claim to characterize real MLLMs, yet it upgrades "why counterfactuals are needed" from intuition to a directionally clear derivation.
Clean division of labor among the rewards: format locks structure, keywords lock semantics, diagnosis locks the conclusion — a closed loop of structure, process, and outcome, with the 1:1:2 weighting itself signalling the priority of the diagnostic conclusion in the system.

Limitations & Future Work¶

Counterfactual samples depend on lesion masks, and masks depend on experts: the initial box comes from the MLLM but must be refined by experts to fix \(M\). In real deployment there is no expert in the loop, so mask quality degrades and weakens the control effect of the counterfactual samples — a bad mask can even erase the wrong region and inject wrong supervision.
The diagnostic reward only checks the final label (0/1) and does not constrain chain faithfulness: the keyword reward verifies that words appear, not that they are used in the right place — a model can list a pile of terminology in the morphology section while describing something other than the lesion in the image. A visual consistency check between description and image is missing.
The theoretical analysis rests on a simplified logistic feature model: Definition 1 splits the image into two latents and writes the model as \(\sigma(w^\top\phi(X))\); real MLLM representations hardly satisfy this, and \(K(Z_e)<K(Z_c)\) together with \(\lambda\to\infty\) are idealized conditions, so the conclusions should be read as directional guidance rather than quantitative predictions.
Limited data and evaluation scale: the test set has only 4,779 images, of which SEE-AI contributes just 235 multi-label samples, so multi-label percentages are volatile at this sample size; the dataset itself is class-imbalanced (23 single-label categories plus 49 multi-label combinations), and the generally low multi-label accuracy may reflect incomplete annotation coverage as much as a model capacity ceiling.
Incomplete ablation evidence: the reward-removal experiments are reported only as Fig. 6 curves without checkable numbers in the text, and there is no sensitivity analysis of the LoRA rank, the reward weights, or \(G\); the rationale for the 1.0/1.0/2.0 weighting is not stated.
Directions for improvement: replace expert refinement with iteratively self-generated masks (train on a coarse mask first, then refine it with the model's own localization ability) to cut cost; add a visual consistency term to the reward (e.g. scoring image-text agreement of the morphological description); and extend the counterfactual intervention in the other direction — change the background while keeping the lesion — to test whether the model is genuinely causal in both directions.

vs general large models (Gemini 3 Pro / GPT-5 series / Qwen3-VL-Plus): they treat endoscopy as general visual question answering and average 24.82% / 11.13% / 41.16%, with a failure mode of being led by environmental cues (the capsule-endoscopy hallucination in Fig. 1(c)). This paper differs by not relying on general visual ability but injecting the domain cognitive flow and then enforcing causal grounding.
vs medical expert models (Hulu-Med and similar): they have more scale and more data for general medical VQA but average only 7.72%–8.58% on endoscopy, with multi-label accuracy near zero. The difference is the training objective — broad medical coverage versus fine-grained endoscopic recognition — which the domain-specific hierarchical reasoning chain supplies.
vs endoscopy report generation / vision-language description methods: these treat endoscopic analysis as image captioning (VILA-M3, Fleming-VL and similar), producing descriptive text without enforcing that the description serves the diagnosis. The three rewards here make the description a causal precondition of the diagnosis rather than an optional add-on.
vs this paper's own Base-SFT (ablation view): plain SFT already reaches 66.31%, but in the noisy example of Fig. 4 (bottom row) it is misled by frothy bile and bubbles and reports Normal for an Erosion, and it degrades badly under spot interference. This contrast delimits what SFT can do — it aligns structure and language, not causality.
vs GRPO itself: GRPO's group-relative advantage was validated on mathematical and reasoning tasks; this paper replaces the reward wholesale with clinical-cognition-centric format, keyword, and diagnosis terms and folds counterfactual samples into the same reward system — a clean example of transferring a general RL algorithm to domain-specific causal correction.

Rating¶

Novelty: ⭐⭐⭐⭐ The pairing of the two problems (cognitive misalignment and missing causality) is well aimed and the combination of a theoretical account with counterfactual GRPO has ideas, though the individual techniques are mostly domain adaptations of existing components (SFT, GRPO, counterfactual intervention).
Experimental Thoroughness: ⭐⭐⭐⭐ Five datasets, three model scales, a multi-label split, robustness, and masking studies are all covered, but the reward ablation has curves without numbers and there is no sensitivity analysis of reward weights or LoRA.
Writing Quality: ⭐⭐⭐ The storyline is clear and the failure case is persuasive, but the formula typesetting is chaotic (the gradient equations of the theorems and Eqs. (11)/(12)/(14) are unreadable in the cached text), the derivations are rough (they rest on an unproven spectral-norm assumption), and some ablation conclusions are qualitative only.
Value: ⭐⭐⭐⭐ Trustworthiness of endoscopic MLLMs is a real pain point, and the two-stage paradigm — data plus SFT for cognition, counterfactual RL for causality — along with the cheap keyword reward, is directly reusable; the dataset of 24,515 reasoning chains has independent value.