acceptodds
Under review as a conference paper at ICLR 2027

MedSteer: Clinically constrained representation steering for debiasing medical large language models

Abstract

Large language models are increasingly used for medical decision support and clinical question answering, but their responses can suffer from clinically unjustified demographic bias. Mitigating this bias is complex because attributes like age, gender, and ethnicity may be relevant in certain contexts while introducing inappropriate sensitivity in others. Existing steering methods often fail to separate sensitive attributes from diagnosis-relevant clinical information. To address this, we introduce MedSteer, a clinically constrained, inference-time representation steering framework for medical LLMs. MedSteer estimates attribute-associated representation shifts from diagnosis-conditioned matched pairs and prioritizes steering directions using semantic stability and clinical orthogonality, reducing non-decisive demographic sensitivity while preserving clinical reasoning. It further incorporates an auxiliary lexical control signal from stigma–neutral contrasts without parameter updates. Evaluated across four medical LLMs on sensitivity analysis, long-text entity retention, and medical question answering, MedSteer outperforms an unweighted MeanDiff baseline. It improves average question answering accuracy by 2.6 percentage points across 12 model–dataset combinations, better retains clinical entities, and achieves a superior fairness–utility trade-off. These findings highlight MedSteer as a practical post-deployment solution to mitigate demographic bias in medical LLMs without compromising clinical performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Medical Large Language Models, Representation Steering, Bias Mitigation, Fairness, Inference-time Intervention

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ju6fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on bias mitigation in LLMs for medical decision support. It proposes MedSteer, an inference-time representation steering method for reducing demographic bias in medical LLMs while preserving clinically relevant information. It combines diagnosis-conditioned representation differences, semantic-stability and clinical-orthogonality weighting, and an auxiliary stigma–neutral lexical direction. Evaluations across four models cover medical question answering, clinical entity retention, and demographic sensitivity. Compared with unweighted MeanDiff steering, the authors report an average QA accuracy improvement of 2.6 percentage points and higher entity retention, with fairness outcomes varying across models.

Critical strengths
  • The paper highlights the significance of separating sensitive attributes from diagnosis-relevant clinical information.
  • MedSteer intervenes on hidden representations at inference time while keeping model parameters frozen, enabling post-deployment debiasing without fine-tuning the underlying LLM.
  • The paper conducts diverse experiments in 12 model–dataset combinations.
Critical weaknesses
  • This paper does not sufficiently position MedSteer relative to existing work on mitigating demographic bias while preserving clinically relevant information. Prior studies, including [1], [2], [3], [4], and [5], have investigated related objectives and reported improvements in medical fairness and performance via diverse strategies, including model fine-tuning, reasoning steering, and prompt engineering. However, the paper states that existing steering methods “often fail to separate sensitive attributes from diagnosis-relevant clinical information” without identifying which methods exhibit this limitation, under what clinical conditions, or according to which evaluation criteria. In particular, the paper has cited and discussed [5], which is similar to MedSteer, but [5] has not been included as a baseline. Additionally, the paper has not positioned inference-time steering methods, such as MedSteer, among the medical debiasing approaches set to demonstrate the superiority of inference-time steering methods. For example, because both MedSteer and prompt engineering can mitigate bias without extra model training or tuning, the advantages of the proposed methods are not clear and need more justification.

  • The presentation of motivation can be misleading, and there is a gap between claim and evidence. The claims, such as "MedSteer strictly targets ungrounded sensitivity without disrupting valid diagnostic pathways, " and "The missing capability is therefore a selective intervention: identifying which components of an attribute-associated representation shift can be attenuated without perturbing the clinical evidence underlying the model’s response...," may mislead readers into thinking that MedSteer can distinguish whether a sensitive attribute, such as sex, age, or race, is a clinically relevant attribute or a non-clinically relevant attribute for specific medical diagnosis. For instance, some diseases, such as Systemic Lupus Erythematosus (SLE) are strongly sex-related. So, sex is a clinically relevant attribute in SLE diagnosis.

However, the evaluation of the paper does not directly distinguish clinically decisive from non-decisive demographic effects based on a specific medical question. Instead, the paper uses changes in diagnostic metrics, such as accuracy and F1, as a proxy for measurement, whose reliability needs more solid evidence. Aggregate medical QA accuracy, entity retention, and bias metrics do not establish that unjustified demographic effects are reduced while clinically necessary effects are preserved. Direct, clinically grounded evaluation of both cases is needed to substantiate this selectivity.

The definition and measurement of bias need more justification. The paper defines an attribute as “clinically non-decisive” when changing it should not alter the medically justified conclusion (Section 3.1). However, it does not specify how this judgment is established for individual cases, including the clinical criteria, annotation procedure, or expert validation used. Without this operational definition, it is unclear how the evaluation distinguishes unjustified demographic sensitivity from legitimate clinical variation, limiting its ability to assess the stated conditional-invariance objective.

  • Table 1 and Table 2 demonstrate that the results among methods are very close, and sometimes MeanDiff is better than MedSteer. Therefore, it is difficult to recognize that 2.6 percentage points is a significant improvement without statistical analysis in these experimental tasks. The paper has not reported the results of repeating the experiment and the standard deviation across different runs of the same experimental settings. Thus, we cannot determine whether the 2.6 percentage points of improvement is a benefit from the proposed approach or just uncertainty of LLMs.

  • MeanDiff seems to be an ablation of MedSteer and not an external baseline to demonstrate the superiority of MedSteer against existing methods. Therefore, diverse types of existing methods may be more suitable as baselines, such as [1], [4], and [5].

Questions for authors
  1. What is the novelty of MedSteer and its advantages over existing bias-mitigating strategies, such as prompt engineering?

  2. What direct, clinically grounded evidence shows that MedSteer suppresses unjustified demographic sensitivity while preserving demographic effects necessary for correct clinical decisions, beyond improvements in aggregate QA accuracy, entity retention, and bias metrics?

  3. Is the reported 2.6-percentage-point gain statistically reliable, and can you provide confidence intervals or paired significance tests, along with repeated-run variability where applicable?

  4. How does MedSteer compare with existing approaches such as [1], [4], and [5] under comparable evaluation settings?

2 comments

Thank you for the detailed review. We clarify the contribution and provide additional experimental evidence below.

Q1. Contribution and relationship to existing approaches.

MedSteer's contribution is diagnosis-conditioned direction construction and weighted aggregation. Each demographic contrast matches admissions on diagnosis and the other recorded demographic/social covariates. An exhaustive audit verified these matching constraints and source-field/template consistency for every pair. Semantic weighting favors pairs with similar representations, while centroid-based weighting reduces the contribution of shifts aligned with diagnosis representations:

wi=max⁡(0,cos⁡(hi,1,hi,2))(1−∣cos⁡(Δhi,vc)∣).w_i=\max(0,\cos(h_{i,1},h_{i,2}))(1-|\cos(\Delta h_i,v_c)|).

An auxiliary stigma–neutral branch adds a framing-related signal using semantic weighting. MedSteer then applies layer-wise gated interventions during inference. This introduces a clinical prior into direction construction and provides a lightweight mechanism for balancing demographic sensitivity and medical utility.

Relative to prompt engineering, MedSteer acts directly on hidden states without adding a debiasing instruction to each input. Model parameters remain frozen, and the intervention strength is selected on validation data. The distinction is the clinical information used to construct and weight the steering directions.

Q2. Direct behavioral evidence under fixed clinical evidence.

We additionally evaluated two sets of 100 cases, changing only the Black/White descriptor while keeping each case's clinical evidence fixed. The ancestry-associated set includes sickle-cell disease (60), thalassemia (20), and G6PD deficiency (20). The case-level non-decisive set includes appendicitis, cholecystitis, pancreatitis, bowel obstruction, and epilepsy (20 each).

Case setCorrect under both descriptors, Base → MedSteerDiagnosis agreement, Base → MedSteer
Ancestry-associated diseases69% → 82%76% → 89%
Race non-decisive for the specified case78% → 86%83% → 91%

Joint correctness increases by 13 and 8 percentage points, respectively, alongside improved agreement. Measuring correctness under both descriptors distinguishes successful invariance from two identical but incorrect outputs. This evaluates diagnostic behavior directly, complementing the representation projections and aggregate QA results.

The operational criterion is case-specific: given the fixed diagnostic evidence in these cases, changing only the race descriptor should preserve the diagnosis. Both cohorts assess this form of invariance, including cases involving ancestry-associated conditions; the evaluation does not treat race as a substitute for genotype or assume equal prevalence across populations.

Q3. Accuracy gains and uncertainty evidence.

The 12 model–dataset differences in Table 1 average 2.57 percentage points, reported as approximately 2.6. MedSteer improves observed QA accuracy over MeanDiff in 11 of the 12 combinations. The reported confidence intervals are entirely above zero for eight configurations, spanning all four backbones. For example, Qwen's gains are +3.00 points [+1.00, +5.00] on MedQA and +4.00 [+2.00, +6.00] on MedMCQA; Internist improves by +4.66 [+2.66, +6.66] on MMLU, and Meditron by +3.38 [+1.38, +5.38] on MMLU. These intervals provide setting-level uncertainty information alongside the descriptive average.

Entity-retention F1 also improves on all four backbones: absolute increases over MeanDiff are 0.0786 for Internist, 0.0453 for Qwen, 0.0034 for Meerkat, and 0.0104 for Meditron.

Thank you for the response, which partially addresses my concerns. Two main concerns remain.

  1. The new diagnostic evaluation is helpful, but both case groups test whether predictions remain consistent when race is changed while clinical evidence is held fixed. This demonstrates improved invariance, but does not establish whether MedSteer preserves clinically justified demographic sensitivity when changing an attribute should legitimately affect the medical decision. This distinction is central to the paper's motivation and claimed contribution.

  2. While the authors clarify how MedSteer differs from prompt engineering and other debiasing strategies, the response does not provide empirical evidence demonstrating its advantages over these alternatives. The evaluation still relies primarily on MeanDiff, making it difficult to assess the practical significance of the proposed method relative to existing approaches.

Regarding Q4, I apologize for inadvertently omitting the reference list from my original review, which made the cited works ambiguous. This was an oversight on my part, and I do not expect the authors to address that question.

Overall, the additional evidence addresses some of my concerns, but I remain unconvinced that the central claim of clinically selective debiasing has been sufficiently validated. I therefore maintain my current rating.

Reviewer_aShJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper aims to address the demographic biases that LLMs may show when used in medical domains using inference-time steering methods.

Critical strengths
  • The paper overall is easy to follow, and the gap is discussed fairly clearly.
  • Experiments span several small LLMs and two benchmarks.
Critical weaknesses
  • The paper seems to offer a modest contribution to the field, as prior work has already established contrastive mean-difference steering for bias mitigation (e.g., contrastive activation addition, FairSteer, Representation Surgery) and inference-time demographic interventions in medical LLMs (e.g., Ahsan et al., 2025), leaving MedSteer’s main addition, a heuristic per-pair weighting scheme, largely incremental and insufficiently ablated.
  • Related to the above note, the experiments seem to leave out key baselines beyond MeanDiff.
  • The method does not seem to use a clear way to preserve the justified effect of demographic variables. For example, using race for skin cancer detection and gender for breast cancer may be justified.
Questions for authors

Is there a guarantee that the LLM-based process for generating the scenarios from MIMIC does not make mistakes?

3 comments

Thank you for the constructive review. We clarify the methodological contribution and data construction and provide additional experimental evidence.

1. Contribution and relationship to existing methods.

MedSteer contributes a diagnosis-conditioned procedure for constructing and weighting demographic steering directions. It combines matched admissions sharing a diagnosis and the other observed demographic/social attributes; semantic-stability and diagnosis-alignment weights on their representation differences; and a complementary stigma-neutral lexical direction. For demographic pairs, the weight is

wi=max⁡(0,cos⁡(hi,1,hi,2))(1−∣cos⁡(Δhi,vc)∣).w_i=\max(0,\cos(h_{i,1},h_{i,2}))(1-|\cos(\Delta h_i,v_c)|).

This gives greater influence to pairs whose representations remain similar and whose differences have lower alignment with the diagnosis centroid. The lexical branch applies semantic weighting to framing-related contrasts. Layer-wise gated interventions then apply the resulting direction to a frozen model. These components adapt contrastive steering to balancing demographic sensitivity with medical utility, without parameter updates or an added debiasing prompt.

Ahsan et al. (2025) use activation patching to identify and manipulate demographic representations in LLMs on clinical tasks. FairSteer uses a learned bias detector with dynamic activation steering. MedSteer's methodological distinction is the use of diagnosis-conditioned contrasts and dual weighting during direction estimation. We will sharpen this positioning in the related-work discussion.

MeanDiff provides a controlled comparison with unweighted aggregation. Sharing the prompt format, decoding configuration, and coefficient-search protocol allows us to assess the contribution of the proposed direction construction. The component ablation below provides a more detailed comparison.

2. Additional component-level evidence.

Our gender-steering ablation yields:

MethodAccuracy
Base61.90%
MeanDiff61.19%
Semantic-stability weighting only54.75%
Clinical-orthogonality weighting only55.25%
MedSteer63.60%

The full method achieves the highest accuracy: 2.41 percentage points above MeanDiff and 1.70 points above Base. Its advantage over the individual weighting variants supports using the two criteria jointly in this experiment.

3. Clinical relevance and direct behavioral evaluation.

Our objective is to reduce sensitivity to demographic wording when the clinical evidence supports the same conclusion. Diagnosis conditioning and the alignment penalty incorporate clinical context during direction estimation, while intervention strength controls the extent of the modification.

We added a case-level evaluation that changes the race description between Black and White while holding clinical evidence fixed. It includes 100 cases where race is non-decisive for the given diagnosis and 100 cases covering ancestry-associated conditions: sickle-cell disease (60), thalassemia (20), and G6PD deficiency (20).

Case groupCorrect under both labels, Base to MedSteerDiagnostic agreement, Base to MedSteer
Race non-decisive for the case78% to 86%83% to 91%
Ancestry-associated conditions69% to 82%76% to 89%

These results directly measure robustness to race wording under fixed clinical evidence. The gains in joint correctness show that increased agreement accompanies more correct diagnoses, including in the ancestry-associated disease group.

4. MIMIC construction and exhaustive structural audit.

The single-attribute direction-estimation datasets use deterministic construction from recorded MIMIC fields. Two admissions share the same recorded admission diagnosis, differ in the target attribute, and match on the other demographic attributes. Our exhaustive audit verified these constraints, exact source-field consistency, and template consistency for every pair.

The inputs are standardized templates populated with demographic/social metadata and the recorded preliminary admission diagnosis, as described in Appendix A.1 and C.1. The audit establishes the provenance and structural validity of the matched contrasts used for direction estimation. Diagnosis contributions reflect the number of eligible matched pairs and their pair-level weights.

Dear Reviewer,

As we approach the final hours of the author-reviewer discussion period, I would greatly appreciate it if you could take a moment to review and respond to the authors’ rebuttal.

Please briefly indicate which concerns have been addressed and which remain, and consider updating your scores if needed.

Thank you again for your time and valuable contributions!

Best regards,
AC

I appreciate the authors' response. My main concerns are not fully resolved. I maintain my score.

Reviewer_kMehOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MedSteer is an inference-time representation-steering framework for reducing unjustified demographic sensitivity in medical LLMs. It estimates attribute-associated directions from diagnosis-conditioned MIMIC-III pairs and aggregates them using semantic-similarity and diagnosis-orthogonality weights. An auxiliary direction derived from stigma–neutral lexical contrasts complements the demographic intervention. The resulting direction is applied to selected layers of frozen models using a model-specific steering coefficient. Across four backbones, the paper reports improved clinical-entity retention and an average 2.6-percentage-point medical-QA improvement over unweighted MeanDiff, alongside mixed subgroup-level fairness results.

Critical strengths

An important and appropriately framed problem. The paper recognizes that demographic invariance is not always clinically appropriate and explicitly distinguishes unjustified sensitivity from medically relevant demographic information. This is a meaningful objective for medical-model debiasing.

A lightweight and interpretable intervention. MedSteer operates on frozen models without parameter updates. Its matched-pair construction and weighted direction aggregation provide a relatively simple approach that can be inspected and adapted across backbones.

Evaluation considers both fairness and utility. The experiments include four backbones, medical QA, clinical-note rewriting, and demographic-disparity assessments. The consistently improved entity-retention scores and average QA gains over MeanDiff provide encouraging evidence that weighted steering can be less disruptive than uniform direction averaging.

Critical weaknesses

The central claim of clinically conditional debiasing is not directly established. Diagnosis conditioning is used during direction estimation, while inference applies an input-independent direction and coefficient. The evaluation does not separately test cases in which demographic changes should preserve the answer and cases in which they should legitimately alter it. Consequently, aggregate QA accuracy and reduced demographic disparities do not establish that MedSteer suppresses inappropriate sensitivity while preserving appropriate demographic effects.

The proposed clinical constraint is an unvalidated geometric proxy. Clinical orthogonality is implemented through scalar weighting based on alignment with a diagnosis centroid, rather than protection of a validated clinical subspace. A direction can be orthogonal to that centroid while still affecting clinically important features. Moreover, the extraction template primarily encodes demographic metadata and a diagnosis label rather than detailed clinical evidence. The paper does not establish that these proxies identify directions that preserve diagnostic reasoning.

The fairness evidence is incomplete and sometimes reflects reduced utility. For Meerkat, the gender gap decreases while both male and female accuracy decline relative to MeanDiff. For Qwen, the gender gap increases, and all three medical-QA scores remain below the unmodified model. Table 2 also omits the unmodified baseline. The projection-based metrics require clearer behavioral validation before they can substantiate reduced harmful bias.

Limited baselines and ablations prevent attribution of the reported gains. MeanDiff is the only substantive intervention baseline, and the main ablation removes the lexical branch. There are no separate ablations of diagnosis conditioning, semantic stability, or clinical orthogonality. It therefore remains unclear whether the proposed constraints are necessary or whether comparable results could be obtained through simpler direction construction, lexical steering, or alternative intervention methods.

Insufficient evaluation details limit reproducibility and confidence in the findings. The manuscript does not adequately specify evaluation sizes and splits, patient-level separation, steering layers and injection sites, multi-category contrast orientation, normalization, or the NER evaluation procedure. Repeated use of patients across matched pairs also makes the effective number of independent observations unclear. The absence of uncertainty estimates limits interpretation of the smaller fairness and accuracy differences.

Questions for authors
  1. Can you directly evaluate clinically appropriate invariance and clinically appropriate sensitivity separately? Please construct or identify clinician-annotated cases where changing a demographic attribute should leave the answer unchanged, alongside cases where it should change the correct answer. How does MedSteer perform on each group?

  2. What validates the diagnosis centroid as a proxy for clinically relevant information? Could you measure whether the weighted directions preserve clinically meaningful features, or compare the current weighting against a learned clinical subspace and an explicit orthogonal-projection baseline? How do results change when direction estimation uses symptoms, laboratory findings, and clinical histories rather than diagnosis labels alone?

  3. Do the fairness improvements persist without reducing absolute subgroup performance? Please include Base in Table 2 and report overall accuracy, worst-group accuracy, and paired confidence intervals. How should the reduced Meerkat gender gap be interpreted when both subgroup accuracies decrease?

  4. How are the EquityMedQA projection metrics defined and independently validated? Please specify how the evaluation bias direction is constructed, whether it is independent of the steering data, and where representations are measured. Do improvements correlate with fewer unjustified changes in generated diagnoses or recommendations?

  5. Which proposed components contribute independently to the results? Could you ablate diagnosis conditioning, semantic stability, clinical orthogonality, and the lexical branch separately under matched tuning budgets? Please also compare against stronger applicable steering and debiasing baselines.

  6. Can you clarify the data splits and intervention implementation? Are direction-estimation, validation, and note-rewriting samples separated at the patient level? Please report unique patient counts, evaluation sizes, contrast orientation for multi-category attributes, the direction used for each task, exact normalization, selected layers, injection sites, and decoding settings.

  7. Does improved entity retention correspond to preserved clinical meaning? Could you supplement NER F1 with evaluations of negation, dosage, temporal relations, and unsupported additions, ideally with clinician assessment? How is semantic equivalence checked for lexical pairs that may alter meaning rather than wording alone?

  8. Could you broaden the discussion of related work on medical-model fairness and clarify how it informs your evaluation? Relevant examples include Unveiling Performance Challenges of Large Language Models in Low-Resource Healthcare: A Demographic Fairness Perspective, which examines demographic disparities across healthcare tasks; FMBench: Benchmarking Fairness in Multimodal Large Language Models on Medical Tasks, which considers demographic fairness in free-form medical QA and report generation using clinically informed evaluation; and Toward Fair Medical Advice: Addressing and Mitigating Bias in Large Language Model-Based Healthcare Applications, which studies mitigation through prompting and fine-tuning. Please consider incorporating these studies into the related-work discussion and explaining how MedSteer’s conditional-invariance objective and evaluation differ from their approaches. In particular, this would help clarify what reduced representation sensitivity and subgroup accuracy gaps establish about fairness in generated clinical responses.

2 comments

Thank you for the constructive review. We provide additional results and clarifications below.

Q1–2. Diagnosis-aware steering and direct behavioral evaluation. MedSteer constructs contrasts within the same diagnosis and matches the remaining observed demographic/social attributes. It then weights each candidate using

wi=max⁡(0,cos⁡(hi,1,hi,2))(1−∣cos⁡(Δhi,vc)∣),w_i=\max(0,\cos(h_{i,1},h_{i,2}))\bigl(1-|\cos(\Delta h_i,v_c)|\bigr),

where vcv_c is the diagnosis centroid. This reduces the contribution of contrasts with low semantic similarity or strong alignment with the shared diagnosis representation. The weighted direction and intervention strength provide a lightweight mechanism for controlling the fairness–utility trade-off.

We added a fixed-evidence diagnostic audit with two groups of 100 cases: ancestry-associated diseases (60 sickle-cell disease, 20 thalassemia, 20 G6PD deficiency) and cases in which race is non-decisive given the supplied clinical evidence (20 each of appendicitis, cholecystitis, pancreatitis, bowel obstruction, and epilepsy). Each case is evaluated with Black and White descriptions while keeping its clinical evidence fixed.

Case groupBoth descriptions correct: Base → MedSteerDiagnosis consistency: Base → MedSteer
Ancestry-associated diseases69% → 82%76% → 89%
Race-non-decisive diseases78% → 86%83% → 91%

Thus, paired diagnostic correctness improves alongside consistency under these controlled changes in race description.

Q3. Expanded subgroup comparison. We add the following Base results, reporting absolute accuracy alongside subgroup disparities:

ModelMaleFemaleGender gapEthnicity Max–MinEthnicity CV
Qwen0.672 [0.63, 0.72]0.650 [0.60, 0.70]0.022 [0, 0.07]0.035 [0.02, 0.07]0.022 [0.01, 0.04]
Meerkat0.555 [0.51, 0.60]0.535 [0.49, 0.58]0.020 [0, 0.07]0.041 [0.02, 0.09]0.025 [0.01, 0.06]
Meditron0.450 [0.40, 0.50]0.423 [0.38, 0.47]0.027 [0, 0.08]0.053 [0.03, 0.10]0.038 [0.02, 0.07]

For Meditron, Table 2 reports MedSteer male/female accuracies of 48.55%/48.78%, both above these Base values, with a gender gap of 0.22 percentage points. Meerkat's reported gender gap decreases by approximately 50% relative to MeanDiff.

Q4. Projection metrics. As defined in Fig. 2, Mean Bias Delta is the average projection difference between matched evaluation pairs along the evaluation bias direction. Mean Bias Magnitude is the average absolute projection magnitude on that direction after intervention. These representation-level measurements complement the subgroup accuracy and paired diagnostic outcomes reported above.

Q5. Contributions of the weighting components. The updated QA ablation yields:

MethodAccuracy
Base61.90%
MeanDiff61.19%
Semantic-stability weighting only54.75%
Clinical-orthogonality weighting only55.25%
MedSteer63.60%

Full MedSteer achieves the highest accuracy in this setting(QA): +2.41 percentage points over MeanDiff and +1.70 over Base. The full method also outperforms either component-only variant, supporting the combined design. Table 1 separately reports the lexical-branch ablation.

Q6. Construction and implementation. Appendix C specifies 13,264 diagnosis-conditioned pairs across five attributes. Construction is deterministic: each pair shares its recorded diagnosis and the other four snormalized demographic/social attributes. Our exhaustive structural audit verified the matching constraints and exact source-field/template consistency for all pairs. Direction extraction uses the final prompt token (Sec. 3.1), followed by layer-wise gated inference-time interventions with frozen model parameters. Appendix B uses a shared coefficient search, α∈{0.5,1.0,…,10.0}\alpha\in\{0.5,1.0,\ldots,10.0\}, guided by validation entity F1 and the gender gap. Steering methods share prompt and decoding settings.

Q7. Entity retention and lexical construction. MedSteer improves entity-retention F1 over both Base and MeanDiff on all four backbones (Table 1). F1 increases over MeanDiff are 0.0786 for Internist, 0.0453 for Qwen, 0.0034 for Meerkat, and 0.0104 for Meditron. The lexical resource comprises 150 replacement pairs instantiated in five templates, yielding 750 prompt pairs. Each contrast keeps the template and surrounding context fixed, and this branch uses semantic-stability weighting (Sec. 3.3; Appendix C.4).

Q8. Related work. We will incorporate Zhou et al.'s demographic audit, FMBench, and FairMed. Their evaluations and prompting/fine-tuning approaches provide complementary context for MedSteer's diagnosis-aware hidden-state intervention, which operates with frozen parameters and without an added debiasing prompt.

Dear Reviewer,

As we approach the final hours of the author-reviewer discussion period, I would greatly appreciate it if you could take a moment to review and respond to the authors’ rebuttal.

Please briefly indicate which concerns have been addressed and which remain, and consider updating your scores if needed.

Thank you again for your time and valuable contributions!

Best regards,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.