acceptodds
Under review as a conference paper at ICLR 2027

TELOS: FROM SEMANTIC ACCESS TO REALIZATION VIA EXPRESSION MEMORY AND SUFFIX CONTRAST

Abstract

Radiology report generation requires accurate expression of clinical concepts, including radiographic findings, anatomical sites, negations, and diagnostic uncertainties. While vision–language models (VLMs) generate fluent narratives, overall language fluency does not guarantee the reliability of specialized terminology, particularly when clinical expressions span multiple subword tokens. Standard fine-tuning updates continuous parameters but leaves domain concepts implicit, forcing the decoder to coordinate multi-token realization through unguided subword steps. We investigate whether specialized terminology can be better supported without altering native autoregressive decoding by coupling addressable expression memory with targeted suffix supervision. To this end, we introduce TELOS (Term-level Expression Latent memory with Onset-to-Suffix contrast, an expression-level adaptation framework motivated by classical teleology (τ ´ϵλoς, realization of whole concepts from initial potential). TELOS supports native subword decoding of a frozen VLM through explicit memory and targeted supervision. It maintains an addressable Clinical Term Dictionary indexed by complete expressions. At sparse decoder depths, a Bilateral Evidence Router (BER) combines contextual language cues with visual re-attention to regulate weighted memory access, using an explicit abstention state to suppress current memory reads during non-specialized phrasing. Causal depthwise writeback propagates expression-relevant support across neighboring steps to assist multi-subword realization. To support local realization decisions, Onset-to-Suffix Contrast (OSC) provides targeted supervision at suffix positions against high-scoring non-reference competitors. The resulting interface leaves the pretrained backbone, tokenizer, and vocabulary unchanged. Extensive evaluations on MIMIC-CXR across LLaVA-1.5, Qwen2.5-VL, and LLaVA-Rad demonstrate consistent improvements in report quality and terminology realization: on LLaVA-1.5, TELOS improves RadGraph-F1 from 0.158 to 0.240 and BLEU-4 from 0.080 to 0.158 over decoder LoRA, requiring only 25% trainable parameters compared with matched LoRA.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language Models, Multimodal Learning, Representation Learning, Medical Report Generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1ns9OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

TELOS is designed to address errors that occur during the generation of multi-word clinical terms in radiology reports. Such expressions are common in medical language, but virtual language models (VLMs) may complete these expressions with frequently occurring substitutes, thereby changing their clinical meaning. The framework combines linguistic context and visual information and exploits explicit memory for medical terms while promoting correct completion through targeted suffix supervision. The evaluation covers three VLM backbone networks and four radiology datasets, employing quantitative metrics and qualitative examples to examine term generation and clinical accuracy for different subword length groups. The results show that the improvement is particularly significant for longer clinical expressions and requires far fewer training parameters than the corresponding LoRA configuration.

Critical strengths

This work focuses on a specific limitation of medical virtual language models (VLMs) and proposes a coherent mechanism to address this limitation. Error rate estimation (BER) leverages textual context and image retrieval terms, while causal rewriting supports continuity between individual lexical units in clinical expressions. Suffix-level contrastive supervision further addresses erroneous term completion. Although some unresolved issues remain in experimental evidence, these components are closely related to the proposed problem.

Due to the limited number of trainable parameters, TELOS is an ideal choice for modifying existing VLMs with limited parameter overhead.

Critical weaknesses
  1. The primary assessment did not directly demonstrate that TELOS generates more reliable multi-word medical terms. Figure 3 measures the consistency between the generated sub-word length distribution and the reference distribution, as its caption states. Matching these distributions does not prove that the generated expressions convey the correct clinical meaning, which is precisely the shortcoming shown in Figure 1. Therefore, I believe this analysis is supplementary evidence rather than a core validation of the method. The report-level clinical scores in Tables 1 and 2 provide useful background information, but they do not assess correctness at the multi-word terminology level independently. Figures 4 and 5 in the appendix seem to better align with this objective and would be more helpful in the main text.

A proper assessment should also consider the limitations of word matching. Expressions such as “vascular lesion” and “avascular lesion” may have similar word length patterns despite significant differences in meaning. Conversely, “avascular lesion” and “lesion without internal vessels” may use different phrasing to describe similar findings. These examples suggest that consistency in length distribution or exact matching alone cannot adequately measure clinical correctness. The appendix contains some useful qualitative comparisons, but broader evidence should assess whether the generated terms retain the expected concepts, negations, and uncertainties. Concept-aware assessments or radiologist assessments can compensate for this deficiency, and representative examples should be provided in the text to support this. For now, the results in the text are insufficient to fully support the core reliability claims.

  1. Length-based analyses contain significantly fewer terms in five- and six-word groups, so their improvements in estimation may not be as stable as those of shorter expressions. Figures 4 and 5 demonstrate how confidence intervals or error bars can be obtained through term-level bootstrapping resampling, explicitly specifying the resampling unit. Reporting the number of terms in each group also helps readers assess the strength of evidence for longer terms.
Questions for authors

Please address my major concerns.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1ns9

We thank Reviewer 1ns9 for the valuable recommendations regarding evidence placement, assertion-aware metrics, and statistical validation.

1. Core evidence and figure allocation: We agree that matching subword length distributions represents supplementary consistency evidence rather than direct proof of semantic correctness. In contrast, evaluating exact controlled completion rates, term-level precision, recall, and F1 across token lengths 2 to 6+ directly benchmarks terminology realization fidelity. In the revised manuscript (highlighted in blue), we have promoted both core terminology realization evaluations to the main text as Figure 3 (MIMIC-CXR) and Figure 4 (CheXpert Plus under institutional shift) in Section 4.4, fully equipped with 95% bootstrap confidence intervals, and relocated the subword length distribution analysis to Appendix A.6.

2. Clinical concept assessment: Polarity, negation, and uncertainty
Surface word matching cannot capture clinical polarity. Our evaluation explicitly addresses this through assertion-aware clinical metrics:

  • RadGraph F1: Parses reports into entities and relations, explicitly scoring observation assertion labels (definitely present, definitely absent / negated, uncertain). Mismatching opposing assertions (e.g., "vascular lesion" vs. "avascular lesion") penalizes both assertion and relation matching.
  • CheXbert Diagnostic F1: Evaluates categorical classification across 14 conditions, distinguishing positive from negative findings.
  • Radiologist evaluation: As reported in response to Reviewer WrLo (Question 2), board-certified radiologists specifically evaluated clinical consistency and polarity discrepancies across 50 cases, confirming that TELOS reduces important diagnostic errors by 18.9%--24.3% over LoRA.

3. Bootstrap confidence intervals and sample sizes for length-stratified groups: We added 95% confidence intervals to all five metrics using 1,000 paired word-level bootstrap resamples within each length group. For controlled completion, metrics average the 30 contexts per word, then average across sampled words; for free generation, metrics are recomputed from word-level occurrences across held-out test reports using identical resampled word identities across methods. The term counts across token-length groups 2 to 6+ are detailed below (and now incorporated into Table 4 in Appendix A.5 of the revised manuscript, highlighted in blue):

DatasetBackbone23456+
MIMIC-CXRLLaVA-1.53030301810
Qwen2.5-VL30302956
LLaVA-Rad3030301810
CheXpert PlusLLaVA-1.5303027166
Qwen2.5-VL30302134
LLaVA-Rad303027166

As visualized in the updated plots with 95% bootstrap confidence intervals (revised main-text Figure 3 and Figure 4), TELOS exhibits non-overlapping error bars with LoRA on completion rate and F1 across the 5- and 6+ groups, confirming that suffix drift mitigation on longer expressions is statistically reliable.

Reviewer_1ns9OpenReview ↗

Official comment of Reviewer 1ns9

The reviewer appreciates the author's feedback. However, I decided to maintain my original score for the following reasons:

  1. The statement that confidence intervals "do not overlap" does not match the graph.

  2. Clinical semantics and uncertainty validation are insufficient.

  3. The completion test uses the training context, and long-phrase samples are scarce.

AuthorsOpenReview ↗

Response to reviewer 1ns9

If there is anything further that you feel we should clarify, revise, or supplement, please do not hesitate to let us know. We would be more than happy to address any additional concerns or make further improvements to the manuscript. Please feel free to point out even the smallest details. Thank you for your patience and generous support.

AuthorsOpenReview ↗

Supplementary to question 2

We agree that surface exact matching cannot capture clinical polarity ("vascular" vs. "avascular") nor reward valid clinical paraphrases ("avascular lesion" vs. "lesion without internal vessels"). We validate that TELOS resolves both challenges across four hierarchical levels:

  • Backbone embedding discrimination on antonym pairs: To test whether frozen subword embeddings conflate similarly spelled words with contrasting meanings, we examined six word pairs from training reports, averaging subword vectors from the frozen LLaVA-Rad input embedding layer. Six pairs exhibited cosine similarities below 0.70:

    Isolated word pairInput embedding cosine
    hyperinflated / hypoinflated0.623
    calcified / noncalcified0.602
    symmetrical / asymmetrical0.592
    vascular / avascular0.514
    distended / nondistended0.427
    dilated / nondilated0.380

    Even among words sharing the same negating prefix, vector similarity varied substantially (0.602 for calcified/noncalcified down to 0.380 for dilated/nondilated, where only the final subword is shared). This demonstrates that the frozen backbone inherently represents contrasting concepts in distinct vector spaces rather than collapsing them.

  • Dictionary retrieval attribute discrimination: In the 1,265-entry LLaVA-Rad clinical dictionary, we evaluated attribute discrimination across candidate entries differing only in the queried attribute. By querying both directions and ranking entries by cosine similarity:

    • Similar-looking antonyms (displacement): 4 matched pairs (displaced vs. nondisplaced) →\to 8/8 (100%) matching terms ranked higher;
    • Same contrast, including hyphenated spellings: 5 matched pairs →\to 9/10 (90%) matching terms ranked higher;
    • Directional opposites (laterality): 64 matched pairs (left vs. right) →\to 122/128 (95.3%) matching terms ranked higher. These high retrieval discrimination rates confirm that the Bilateral Evidence Router reliably selects matching attributes and polarities rather than false-positive antonyms.
  • Semantic equivalence across varied phrasings via BERTScore: To evaluate whether reports preserve true clinical meaning under flexible phrasing (e.g., "avascular lesion" vs. "lesion without internal vessels"), we employ BERTScore, which computes greedy maximal cosine similarity over deep contextual token embeddings rather than exact string matches. Given candidate tokens x=⟨x1,…,xn⟩x = \langle x_1, \dots, x_n \rangle and reference tokens y=⟨y1,…,ym⟩y = \langle y_1, \dots, y_m \rangle with normalized contextual representations xi,yj\mathbf{x}_i, \mathbf{y}_j:

    • Recall: R=1∣y∣∑jmax⁡i(xi⊤yj)R = \frac{1}{|y|} \sum_{j} \max_{i} (\mathbf{x}_i^\top \mathbf{y}_j)
    • Precision: P=1∣x∣∑imax⁡j(xi⊤yj)P = \frac{1}{|x|} \sum_{i} \max_{j} (\mathbf{x}_i^\top \mathbf{y}_j)
    • F1 (BERTScore): F1=2⋅P⋅R/(P+R)F_1 = 2 \cdot P \cdot R / (P + R)

    Because contextual embeddings project synonymous clinical expressions ("avascular" and "without internal vessels") into proximate latent neighborhoods, BERTScore rewards semantic equivalence despite divergent surface vocabulary. Conversely, contradictory tokens ("vascular" vs. "avascular") receive low cosine similarity, heavily penalizing polarity errors. In Table 1, TELOS achieves the highest BERTScore across all backbones: 0.860 on LLaVA-1.5 (vs. 0.818 for LoRA), 0.853 on Qwen2.5-VL (vs. 0.821 for LoRA), and 0.874 on LLaVA-Rad (vs. 0.863 for base), directly confirming that TELOS maintains genuine semantic equivalence across diverse phrasing.

  • Assertion-aware clinical graphs and expert radiologist audit:

    1. RadGraph-F1: Parses reports into clinical entities and relations, explicitly scoring observation assertion labels (definitely present, definitely absent / negated, uncertain). Mismatching opposing assertions ("vascular" vs. "avascular") penalizes both assertion and relation matching. In Table 1, TELOS achieves 0.240 vs. 0.158 for LoRA on LLaVA-1.5 (+51.9% relative gain) and 0.222 vs. 0.206 on Qwen2.5-VL.
    2. CheXbert Diagnostic F1: Evaluates condition-level multi-label classification (grouping positive and uncertain findings vs. negative), confirming categorical diagnostic correctness.
    3. Double-blind radiologist evaluation: In our 50-case evaluation by board-certified radiologists (detailed in response to Reviewer WrLo, Question 2), TELOS reduced clinically meaningful diagnostic and negation errors by 18.9%--24.3% relative to LoRA.
Reviewer_MRu4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Although visual language models can generate relatively fluent expressions when producing radiology reports, clinical professional terms are often split into multiple subwords, and standard fine-tuning cannot ensure reliable generation of these terms. To address this issue, the article proposes TELOS. This framework introduces an addressable clinical terminology dictionary and a start-to-suffix comparison mechanism (OSC), while also utilizing a bilateral evidence router (BER) and a causal write-back mechanism. It improves the issue of suffix drift in the generation of multi-subword medical terms without changing the pretrained VLM backbone, tokenizer, or vocabulary. The authors conducted experiments on three models and four datasets and reported report-level metrics, clinical metrics, and term-level diagnostic results.

Critical strengths
  1. The paper focuses on the suffix drift issue in the generation of multi-subword clinical terms. This problem is clearly defined, and errors in terminology can directly affect the clinical meaning of reports, so it has practical significance.
  2. TELOS freezes the backbone model and retains the original tokenizer and autoregressive decoding, allowing it to be applied as a module to different VLMs. Experiments cover three models and four datasets, and the results show that this method can improve report quality and professional term generation with fewer trainable parameters.
Critical weaknesses

The core conclusions of the paper regarding generalization ability are insufficiently supported. Both the CheXpert Plus and COVCTR experiments were retrained on the target datasets, which is closer to supervised domain adaptation rather than strict cross-hospital or cross-modality generalization. Additionally, the paper does not evaluate the model's ability to generate out-of-vocabulary clinical terms. Therefore, the current experiments are not enough to demonstrate that TELOS can generalize to unseen clinical expressions and open-vocabulary scenarios, which limits the method's applicability in real clinical settings.

Questions for authors
  1. The paper compares several basic medical or radiological models and general adaptation baselines, but does not include representative task-specific radiology report generation methods discussed in related work. Could the authors explain the reason for not conducting these comparisons?
  2. The paper's argument for lightweight design mainly focuses on the number of trainable parameters, while TELOS introduces terminology matching, visual re-attention, and causal convolution across multiple decoder layers. Could the authors report its actual inference cost?
  3. The paper describes the CheXpert Plus experiment as cross-hospital generalization and the COVCTR experiment as cross-modal transfer, but both experiments were trained on the target datasets. Could the authors clearly distinguish between zero-shot generalization and supervised adaptation to the target domain, and indicate whether the related statements need adjustment?
  4. The paper lacks an evaluation of expressions for out-of-dictionary clinical terms. How does TELOS perform when encountering rare terms or expressions not covered by the dictionary? Does explicit memory bias the model toward in-dictionary terms, thereby affecting the generation of rare or unlisted expressions?
2 comments
AuthorsOpenReview ↗

Response to Reviewer MRu4

We thank Reviewer MRu4 for the constructive observations regarding baseline choices, runtime overhead, adaptation terminology, and out-of-dictionary terms.

1. Comparison with task-specific radiology report generation methods: Our study investigates parameter-efficient adaptation of frozen general VLMs. Comparing LoRA and Engram on identical backbones (LLaVA-1.5, Qwen2.5-VL, LLaVA-Rad) isolates the causal effect of the decoding interface while controlling for model architecture and pretraining data. Table 1 also benchmarks against full-parameter domain foundation models (LLaVA-Med, MedGemma1.5); with 5.82M parameters, TELOS on a general VLM matches or exceeds their clinical accuracy. Furthermore, task-specific methods (e.g., PromptMRG) focus on prompt retrieval or custom architectures, whereas TELOS is an orthogonal, plug-and-play decoding interface that can be integrated into them. We will make this distinction clear.

2. Actual inference cost beyond parameter count: Below we report peak training memory, per-token decode latency, and peak inference memory on an NVIDIA H200 (BS=1, BF16, 1,024 tokens):

BackboneMethodTraining peak (GiB)Decode latency (ms/token)Inference peak (GiB)
LLaVA-1.5Original—13.8913.72
LoRA (merged)117.2914.0614.03
LoRA (unmerged)117.2929.4614.10
Engram74.7416.5513.69
TELOS75.9116.8514.12
Qwen2.5-VLOriginal—15.2515.61
LoRA (merged)97.4615.3215.61
LoRA (unmerged)97.4629.4515.68
Engram72.5818.3415.69
TELOS77.0518.1715.65

While merging LoRA removes adapter computation entirely, TELOS retains dynamic memory retrieval and causal writeback, incurring a modest runtime cost of ~2.8 ms/token and negligible extra inference VRAM (<0.4 GiB). In exchange, TELOS cuts peak training memory by 35.3% (75.91 vs. 117.29 GiB) on LLaVA-1.5 and 20.9% (77.05 vs. 97.46 GiB) on Qwen2.5-VL to enable training within standard 80 GiB GPUs, while substantially improving clinical diagnostic accuracy (e.g., RadGraph-F1 increases from 0.158 to 0.240 on LLaVA-1.5, a +51.9% relative gain).

3. Clarification of domain adaptation terminology: We agree with this distinction. Because CheXpert Plus and COVCTR utilize target-domain training data, they evaluate TELOS's capability in supervised cross-hospital and cross-modal adaptation (method-level adaptability). In contrast, our IU-Xray evaluation (Appendix A.8, Table 8) tests the MIMIC-CXR-trained model directly without target fine-tuning, assessing the checkpoint's zero-shot cross-hospital generalization (model-level transfer). In the revised manuscript (highlighted in blue throughout Section 1, Section 4.1, Section 4.2, Section 5, and the Appendix), we have updated these descriptions to strictly distinguish supervised cross-hospital and cross-modal adaptation from zero-shot transfer.

4. Performance on out-of-dictionary (OOD) clinical terms: We evaluated RadGraph-XL entity micro-F1 across all 22 model/dataset combinations on MIMIC-CXR (2,463 reports) and CheXpert Plus (11,562 reports), partitioned into In-Dictionary (ID) and Out-of-Dictionary (OOD) sets:

DatasetBackboneMethodID F1 (%)OOD F1 (%)
MIMIC-CXRLLaVA-1.5Original5.136.52
LoRA33.5729.14
Engram28.5724.84
TELOS34.1929.62
Qwen2.5-VLOriginal15.6315.17
LoRA22.9722.16
Engram30.3425.93
TELOS31.1328.35
LLaVA-RadOriginal51.2431.95
Engram54.1235.39
TELOS54.8635.97
CheXpert PlusLLaVA-1.5Original0.767.46
LoRA25.5622.83
Engram20.8918.57
TELOS28.1123.09
Qwen2.5-VLOriginal11.3812.96
LoRA24.8324.07
Engram17.6820.17
TELOS24.3024.21
LLaVA-RadOriginal29.3517.66
Engram38.6028.03
TELOS42.0429.56

TELOS achieves the highest OOD F1 across all groups, outperforming LoRA and Engram. Because the native vocabulary head is unchanged and the router provides a bypass via the abstention anchor (i=0i=0), TELOS does not bias generation against unlisted terms.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer MRu4,

As the author-reviewer discussion period is drawing to a close, we would like to gently follow up on our response to your comments.

To directly address your feedback, we have updated the manuscript to strictly clarify domain adaptation terminology throughout the text (supervised adaptation vs. zero-shot transfer), provided comprehensive compute benchmarks on an NVIDIA H200, and reported out-of-dictionary (OOD) RadGraph-XL results across all 22 evaluation groups.

We hope our clarifications have resolved your concerns. Please let us know if there are any remaining questions before the discussion ends—we would be glad to provide any further clarification.

Many thanks for your time and valuable suggestions!

Reviewer_WrLoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses a problem in chest imaging report generation: a medical term may span several tokens, and even if the model gets the beginning right, it can continue toward the wrong term. It proposes TELOS to generate medical terminology more accurately while keeping the original model frozen. The approach adds clinical expression memory and visual re-attention to the LLM decoder layers. Routing balances contextual and visual evidence, while causal writeback carries memory information across nearby token positions. Onset-to-Suffix Contrast (OSC) adds supervision to help the correct continuation tokens win over competing tokens. Experiments on several chest X-ray and CT report generation benchmarks show that TELOS improves report quality with fewer trainable parameters than LoRA.

Critical strengths
  1. The paper introduces several technical ideas to improve the clarity of generated chest imaging reports.
  2. The paper is well written and easy to follow.
Critical weaknesses
  1. Some improvements over the stronger baselines are quite small. Without confidence intervals or results across multiple runs, it is hard to tell whether these gains are reliable.
  2. The framework introduces several components, so it is important to understand what each contributes. Although the paper includes some ablations, the benefits of the abstention state and causal writeback are not separately evaluated. More targeted ablations would help show whether each component is needed and whether the added complexity is justified.
Questions for authors
  1. Beyond parameter counts and theoretical complexity, could the authors report training time, inference latency, and peak GPU memory usage compared with LoRA and Engram under the same hardware and settings? This would help assess the practical cost of the added modules.
  2. If any radiologist evaluation of the generated reports is already available, could the authors share the results to help assess the method's usefulness in clinical practice? I am not requesting a new study.
  3. Does enabling TELOS affect the base VLM's general visual question-answering abilities? If any evaluation is available, could the authors share whether the method preserves performance beyond report generation?
  4. Could the authors discuss TELOS's own failure cases, particularly whether incorrect memory retrieval can lead to fluent but clinically incorrect terms? Examples would help clarify the method's limitations.
3 comments
AuthorsOpenReview ↗

Response to Reviewer WrLo (1/2)

We thank Reviewer WrLo for the constructive feedback on computational overhead, clinical utility, general VQA preservation, and failure case analysis.

1. Computational Overhead: We evaluated decoding latency and peak memory on an NVIDIA H200 (BS=1, BF16, greedy decoding up to 1,024 tokens), and estimated 3-epoch training metrics using per-GPU BS=16 across two H200s.

VLMsMethodTraining time(h)Training GPU (GiB)Decode latency (ms/token)Inference peak (GiB)
LLaVA-1.5Original——13.8913.72
LoRA (unmerged)5.61117.2929.4614.10
LoRA (merged)5.61117.2914.0614.03
Engram4.7574.7416.5513.69
TELOS4.9475.9116.8514.12
Qwen2.5-VLOriginal——15.2515.61
LoRA (unmerged)7.7597.4629.4515.68
LoRA (merged)7.7597.4615.3215.61
Engram7.2272.5818.3415.69
TELOS7.2977.0518.1715.65
  • Inference trade-off: Merging LoRA folds updates into the backbone, achieving baseline latency (14.06 / 15.32 ms/token). TELOS retains dynamic expression retrieval, visual re-attention, and causal writeback, which incur a modest runtime overhead of ~2.8 ms/token and minimal extra VRAM (<0.4 GiB). In dynamic multi-adapter serving without weight merging, TELOS is ~40% faster than unmerged LoRA.
  • Training efficiency: Restricting updates to sparse memory modules cuts training peak memory by 35.3% (75.91 vs. 117.29 GiB) on LLaVA-1.5 and 20.9% (77.05 vs. 97.46 GiB) on Qwen2.5-VL over LoRA. This allows training within a standard 80 GiB GPU envelope without activation offloading, with lower training time (4.94h vs. 5.61h on LLaVA; 7.29h vs. 7.75h on Qwen). We will include this breakdown in the revised Appendix.

2. Human Evaluation: To assess clinical utility, two certified radiologists conducted a blinded assessment on 50 randomly sampled LLaVA-1.5 cases (25 MIMIC-CXR, 25 CheXpert Plus). Reports from all methods were anonymized (coded A–D) and independently evaluated against authenticated radiologist references using two clinical criteria:

  • Important clinical errors: Defined as distinct discrepancies judged likely to materially affect clinical interpretation or patient care, including missed critical findings, false-positive abnormalities, or device positioning errors. Repeated mentions within a report were counted once. We report both the mean error count per report and the percentage of reports containing ≥1\ge 1 such error.
  • Reference consistency (0–4 Likert scale): Evaluates overall clinical fidelity according to explicit grading rubrics: 0 = largely inconsistent or uninformative; 1 = major clinical deficiencies or contradictions; 2 = partial agreement with substantial clinical gaps; 3 = mostly consistent with only minor non-critical omissions; and 4 = complete clinical consistency with no material disagreement.
MethodMean important errors ↓\downarrowReports with ≥1\ge 1 error (%) ↓\downarrowReference consistency (0–4) ↑\uparrow
Original1.60 / 1.3490 / 800.42 / 0.60
LoRA1.48 / 1.4092 / 881.14 / 1.36
Engram1.52 / 1.4288 / 821.00 / 1.20
TELOS1.20 / 1.0680 / 721.76 / 1.84

Values of table reported as MIMIC-CXR / CheXpert Plus. Relative to LoRA, TELOS reduces mean important clinical errors by 18.9% on MIMIC-CXR and 24.3% on CheXpert Plus, lowers the proportion of error-compromised reports by 12 and 16 percentage points, and increases clinical consistency by +0.62 / +0.48 points. This confirms that gains in automated metrics like RadGraph-F1 translate directly into verifiable reductions in actionable medical errors.

AuthorsOpenReview ↗

Response to Reviewer WrLo (2/2)

3. General VQA capability preservation: In the revised manuscript (Appendix A.11, Table 11, highlighted in blue), we evaluate general VQA capability preservation on ScienceQA TEST (4,241 questions) and TextVQA VAL (5,000 questions):

BackboneMethodScienceQA ALL (%)ScienceQA IMG (%)TextVQA (%)
LLaVA-1.5Original70.4169.4158.25
LoRA69.0967.3856.56
Engram45.8165.0047.71
TELOS69.9668.4757.27
Qwen2.5-VLOriginal88.6688.8985.15
LoRA87.3987.6184.74
Engram82.8679.0383.68
TELOS88.5988.7585.53

Unconditional static nn-gram retrieval (Engram) degrades non-domain reasoning (ScienceQA ALL drops by 24.60 points on LLaVA-1.5). In contrast, TELOS preserves general VLM capabilities almost losslessly across both backbones (ScienceQA drops by only 0.45 on LLaVA and 0.07 on Qwen), outperforming LoRA on all benchmarks. This validates that the router's explicit abstention state (i=0i=0) successfully suppresses active memory reads (ut→0u_t \to 0) during non-domain text generation.

4. Failure cases analysis: In the revised manuscript (Appendix A.12, Figure 9, highlighted in blue), we examine a representative failure case on MIMIC-CXR where the authentic reference reports a right chest tube and no pneumothorax. TELOS generates “The chest tube has been removed” and asserts “There is a small apical pneumothorax.” LoRA preserves the negative finding, while Engram also predicts pneumothorax. Counterfactual probing by masking the pneumothorax memory entry during inference did not remove the false-positive finding. This confirms that the error originated from perceptual ambiguity in the base vision encoder rather than hallucination forced by memory retrieval. TELOS ensures accurate multi-token realization along the decoder's chosen trajectory, but does not override upstream visual errors.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer WrLo,

With the discussion phase closing shortly, we wanted to kindly check in regarding our rebuttal to your helpful review.

In our response and revised manuscript, we provided dedicated H200 compute benchmarks (peak training memory, decode latency, and inference VRAM), a 50-case double-blind radiologist evaluation, general VQA preservation results (Appendix A.11), and a failure-case analysis with counterfactual probing (Appendix A.12).

We would be deeply grateful to know if these empirical results address your questions on computational overhead, clinical utility, and failure boundaries, or if any points require further discussion.

Thank you again for your constructive guidance and time!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.