PaperlayerView exact revision on arXiv

Source-linked AI summary

CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models

Chunzheng Zhu, Jiaqi Zeng, Hongbo Zhao, Yihang Chen, Yijun Wang, Jianxin Lin

arXiv:2609.38810v1cs.CVcs.AI

TL;DR

Medical VLMs must resolve competing image, text, and learned medical signals, but prior mechanistic analyses have not explained clinical multimodal failures or commitment under insufficient visual evidence. CRAFT localizes causal heads for arbitration and brake failures; targeted excision produces distinct corrections with limited clean-input damage.

  • Problem

    Prior mechanistic analyses focus on text-only models, leaving unexplained how medical VLMs resolve competing visual, textual, and learned medical signals or commit under insufficient visual evidence.

    Source: §1 · p. 1, §3.2 · p. 4

  • Method

    CRAFT uses causal deconfounding and layer-wise conflict scoring to identify minimal failure-specific head sets for intervention.

    Source: §1 · p. 2, §7 · p. 9

  • Results

    Arbitration-head excision reduces Conflict Flip Rate (CFR) by up to 41 absolute points, while brake-head excision raises Uncertainty Rate (UR) under insufficient evidence by 75%, with minimal clean-accuracy damage.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The failures occupy structurally disjoint circuitry and respond to distinct, training-free interventions affecting fewer than one percent of total heads.

    Source: §7 · p. 9

    The analysis covers open-weight models up to 14B parameters, English prompts, and benchmarks predominantly featuring chest X-rays and fundoscopy; it is not validated for non-English inputs or CT and MRI.

    Source: §20 · p. 31

Abstract

from arXiv · show

As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at GitHub repository.

1 Introduction

Medical VLMs must arbitrate among visual evidence, textual context, and learned medical knowledge, yet prior mechanistic work does not explain clinical multimodal failures. CRAFT uses causal head interventions to distinguish textual override from commitment under insufficient visual evidence and target each failure with limited clean-input damage.

3.2 Two Failure Modes and Problem Formulation

  • Arbitration failure hijacks a correct answer with misleading text, whereas brake failure commits to a concrete diagnosis when degraded visual evidence is insufficient.The two failures represent distinct regimes: the model has the correct answer but is overridden in arbitration failure, while it lacks sufficient evidence yet commits in brake failure.Source: §1 · p. 2, §3.2 · p. 4

1 Introduction

  • CRAFT identifies failure-specific causal head sets through layer-wise conflict scoring and minimal-head selection that balances targeted failure reduction against clean-input perturbation.Linear probes and Tuned Lens trajectories provide temporal evidence about when conflict signals become decodable and answer preferences shift.Source: §1 · p. 2
  • CRAFT interventions reduce Conflict Flip Rate (CFR) by up to 41 absolute points and raise Uncertainty Rate (UR) under insufficient evidence by 75%, with minimal collateral damage on clean diagnostic accuracy.Arbitration-head excision suppresses textual override, while brake-head excision restores abstention under degraded evidence.Source: §1 · pp. 2–3

2 Related Work

  • Prior mechanistic head analyses focus on text-only models, while medical VLM work has not traced multimodal clinical failures to internal circuitry.Existing selective-prediction approaches rely on output calibration or prompt heuristics rather than mechanistic intervention.Source: §1 · p. 1, §2 · p. 3

3.1 Transformer Residual Structure

  • Each decoder layer adds parallel attention-head contributions and a feed-forward network to the residual stream, enabling per-head activation interventions.The additive structure permits modifying a single head output while leaving the rest of the forward pass unchanged.Source: §3.1 · p. 3

3.3 Interventional Tracing Setup

  • Observational head rankings can confound causal influence with correlation because latent factors open back-door paths between head activity and model output.CRAFT pairs each conflict input with a conflict-free counterfactual sharing its image and query, then fixes clean activations to block those paths.Source: §3.3 · p. 4
  • CRAFT selects sparse intervention targets empirically from intervention profiles, balancing improved targeted behavior against collateral damage rather than prespecifying target counts or boundaries.The setup defines CFR as conflict samples following misleading text, UR as correct abstention on degraded samples, and C2W/U2O as clean-input collateral damage measures.Source: §3.3 · p. 5

4 Causal Tracing of Failure Circuitry

CRAFT traces each failure from layer-level effects to sparse, failure-specific causal heads while limiting disruption to clean reasoning. Arbitration heads form a broad mid-to-deep override band, whereas brake heads form a compact, late-layer circuit for abstention.

4 Causal Tracing of Failure Circuitry

  • CRAFT first scores layers to find a critical band, then identifies heads whose causal intervention affects the targeted failure while preserving clean-input reasoning.For arbitration, the layer score measures how much patching a layer reduces the conflict effect toward the clean baseline; heads are selected using conflict effect and clean-input disruption criteria.Source: §4 · p. 5, §4.1 · p. 5

4.1 Arbitration Failure / 4.2 Brake Failure localization

  • Arbitration heads are selected for reducing textual override with limited clean-input disruption, while brake heads are selected for increasing abstention under degraded visual evidence.The brake procedure first finds layers where attenuation shifts probability toward abstention, then localizes heads using hallucination-relief and collateral-disruption criteria.Source: §4.1 · p. 5, §4.2 · pp. 6–7
  • Arbitration override effects span layers 12–17 in Hulu-Med and 14–20 in InternVL, while brake relief concentrates in layers 31–35 and 25–35, respectively.The arbitration band spans layers 12–22 at head level; the deepest Hulu-Med layers show negligible per-head conflict circuitry. Brake layers account for over 78% of relief despite spanning fewer than 15% of model depth.Source: §4.1 · p. 6, §4.2 · p. 7

4.2 Brake Failure localization

  • Only 4–7 brake heads per model carry statistically significant signal; excision yields HR gains of 34.2 for Hulu-Med and 29.7 for InternVL, with BCP below 2.1.Brake-head excision restores appropriate uncertainty without degrading correct answers, and the selected brake layers are disjoint from the arbitration band.Source: §4.2 · p. 7

5 Functional Roles and Decoupling Analysis

Targeted excision produces distinct corrections for the two failure modes, while random ablation is less reliable. Comparing selection criteria shows why CRAFT balances target suppression against collateral damage.

5.1 Decoupling of Arbitration and Brake Interventions

  • The comparison uses equal head counts across four medical VLMs, with arbitration tested on VQA-RAD and SLAKE and brake failure on SLAKE and HealMed-VQA.The study evaluates arbitration under textual conflict and brake failure under visual degradation.Source: §5.1 · p. 7
  • Arbitration-head excision reduces conflict following while leaving clean accuracy virtually unchanged, whereas brake-head excision raises unknown rates across four models with zero or near-zero U2O collateral.For brake failure, the gains come from redirecting non-abstaining predictions toward unknown; random ablation produces smaller or unstable gains and can introduce U2O collateral.Source: §5.1 · p. 8

5.2 Contribution of Individual Selection Objectives

  • The dual selection criterion balances conflict reduction and clean-input preservation: on Hulu-Med 4B, it yields C2W of 0.5, versus 1.8 for CER-only selection.On InternVL3.5 4B, CER-only selection raises C2W to 4.0, while the dual criterion matches BCP-only conflict reduction with C2W of 0.0.Source: §5.2 · p. 8
  • For brake failure on InternVL3.5 4B, HR-only selection causes U2O degradation of 21.6%, while the dual criterion keeps U2O at 0.0 with cleaner UR gain.HR-only and dual selection coincide on Hulu-Med 4B, but the InternVL3.5 comparison shows the value of including the collateral criterion.Source: §5.2 · p. 8

6 Causal Verification via Internal Probes

Temporal probes place conflict detection before failure commitment, and Tuned Lens trajectories show preference reversal during conflict. Interventions selectively weaken commitment and suppress arbitration’s reversal.

6 Causal Verification via Internal Probes

  • The verification combines per-layer probes of conflict detection and failure prediction with Tuned Lens projections that track answer-preference shifts across layers.The probes are trained on answer-position hidden states, and the analysis is conducted on Hulu-Med 4B for arbitration and brake failure.Source: §6 · p. 8, §6.1 · p. 9, §6.2 · p. 9

6.1 Temporal Probe Separation

  • For arbitration, conflict detection saturates around layers 17–21 with AUROC > 0.97, before failure prediction peaks 3–5 layers later with Macro F1 ≈0.85–0.87.After intervention, the detection probe remains intact while the failure-prediction probe drops substantially. For brake failure, degraded evidence is predictive from shallow layers onward, while failure prediction peaks at layers 22–27 and behaves non-monotonically after intervention.Source: §6.1 · p. 9, Figure 6 · p. 9

6.2 Tuned Lens Trajectory Analysis

  • Under conflict, the Tuned Lens preference margin changes from negative to positive between layers 18 and 25; excising arbitration heads suppresses this reversal and keeps the margin negative.Without conflict, the margin remains negative throughout; its multi-layer reversal under conflict is consistent with a distributed override circuit.Source: §6.2 · p. 9, Figure 7 · p. 9

7 Conclusion

CRAFT uses causal deconfounding to localize sparse, failure-critical attention heads in medical VLMs. The conclusion presents distinct, training-free interventions for textual override and degraded-evidence commitment as a lightweight post-hoc safety mechanism.

7 Conclusion

  • CRAFT localizes structurally disjoint head sets: excising a mid-layer override wideband suppresses textual hijacking, while excising sparse late-layer gating heads recovers abstention under degraded visual evidence.Both interventions affect fewer than one percent of total heads and are described as training-free and modular.Source: §7 · p. 9

8.1 Models, conflict construction, and data splits

  • The evaluation uses four medical VLMs and retains examples that are correct on clean inputs before applying textual conflict or masking diagnostically relevant image regions.Retained examples are split 70/30 into training for tracing, localization, and threshold selection, and validation for reported metrics; the mask-scale sweep keeps the unknown rate below 50% even at 3.0.Source: §8.1 · p. 13, Figure 8 · p. 13

8.2 Metric Formal Definitions and Text-Only Results

  • On ConflictMedQA and PubMedQA, CRAFT lowers conflict-following rate and improves resistance over baseline and random ablation, with especially strong effects on ConflictMedQA.On PubMedQA, the strongest reported case reduces CFR from 53.70 to 3.70 and increases RESIST from 46.30 to 96.30.Source: §8.2 · pp. 13–14

8.3 Key Parameter Selection and Sensitivity

  • Threshold sensitivity shows that looser settings improve correction at the cost of more collateral damage, while overly strict settings leave failure unresolved.The selected operating points balance target-effect retention against preservation of competence.Source: §8.3 · p. 14

9 Theoretical Analysis of Deconfounding

The analysis argues that observational head attribution can be confounded, while clean-state interchange isolates head-level causal effects and supports CER as a causal metric.

9.1 Confounding in Observational Attribution

  • Latent confounding can make observational activation rankings misidentify causal importance because Uc affects both head activation and model output.The analysis attributes Uc to factors including lexical overlap, visual ambiguity, and cross-modal coupling.Source: §9.1 · p. 15

9.2 D Separation under Clean State Interchange

  • Fixing a head to its clean-state value blocks the back-door path from Uc while preserving downstream computation through other heads and FFN components.Excising a head is treated as a fixed-value interchange intervention, so the same deconfounding guarantee applies.Source: §9.2 · p. 15
  • Clean-state interchange isolates a head’s controlled direct effect, and CER estimates the effect of excising that head on the preference margin under matched counterfactual inputs.The CER guarantee requires the paired counterfactual to share (Xv, q) with the conflict sample.Source: §9.2 · p. 15

10 Comparison with Existing Safety Methods

CRAFT’s targeted head excision outperforms the compared inference-time safety baselines on failure mitigation and preserves clean accuracy better. Against fine-tuning, it offers complementary gains, especially when combined with conflict-aware training.

10 Comparison with Existing Safety Methods

  • CRAFT achieves the lowest CFR and highest Resist with minimal C2W for arbitration, while improving brake UR more than HALP with zero U2O.The brake comparison is on Hulu-Med 4B, SLAKE image conflict; Table 8 reports brake failure results.Source: §10 · p. 16, Table 8 · p. 17

10.1 Clean Accuracy Preservation under Baseline Methods

  • Across both failure modes, CRAFT preserves clean accuracy better than compared baselines, with the sparsest arbitration head set and the highest post-intervention accuracy on all three brake backbones.The comparison uses matched or near-matched intervention budgets.Source: §10.1 · pp. 16–17, Table 9 · p. 17

10.2 Comparison with Fine Tuning Based Safety Adaptation

  • CRAFT lowers textual-conflict CFR without parameter updates, and combining it with CONFLICT SFT yields CFR 10.77 and Resist 89.23.For brake failure, CRAFT raises UR more than fine-tuning while keeping U2O at zero; the combination achieves the largest net gain, 31.82, with better clean-accuracy preservation from fine-tuning.Source: Table 10 · p. 17, Table 11 · p. 17, §10.2 · p. 18

11 Computational Cost

The section reports CRAFT’s offline localization and online intervention costs, then describes the temporal probes and Tuned-Lens analyses used to track conflict and answer preferences across layers and heads.

11 Computational Cost

  • Localization requires 227 to 346 forward passes per sample once, while hooked masking adds less than 2% latency on Hulu-Med 4B and about 14% on InternVL3.5 4B.After localization, selected heads are reused; sparse head skipping yields no meaningful FLOP savings without fused kernels.Source: §11 · p. 18, Table 15 · p. 19, Table 14 · p. 19

12.1 Temporal Probe Architecture

  • Two independent layer-specific linear probes detect injected textual conflict and predict whether the model follows the misleading text or visually grounded answer.Probe A detects conflict; Probe B predicts answer arbitration on conflict instances with unambiguous predictions.Source: §12.1 · p. 19

12.2 Tuned Lens Configuration

  • The Tuned Lens uses a learned affine transform per layer to track preference dynamics through the logit difference between the gold and misleading answers.The transforms are trained on held-out medical text by minimizing KL divergence to the model’s final output distribution.Source: §12.2 · p. 19

12.3 Head-Level Tuned-Lens Preference Distributions

  • Head-level distributions show a gold-to-wrong preference shift under textual conflict, while degraded images often fail to produce the expected shift toward unknown.For image-conflict examples, a larger abstention margin means stronger preference for unknown over the best competing concrete answer.Source: §12.3 · p. 20

13 Patching Strategy Analysis

The patching analysis selects a compact answer-centered window to keep layer-wise conflict tracing stable and locally informative.

13 Patching Strategy Analysis

  • Clean-state replacement patches only answer-neighboring states aligned by relative offset, avoiding full-sequence token-alignment artifacts while retaining most of the conflict prompt.The method tests whether answer-neighboring residual states carry the textual override signal.Source: §13 · p. 20
  • Kpatch = 8 achieves the highest Robustness across evaluated models, indicating a compact answer-centered window recovers a stable layer-wise trace.Robustness measures agreement with the local-window consensus; Smoothness measures adjacent-layer variation.Source: §13 · p. 20, Table 16 · p. 21
  • Larger windows and full-sequence patching reduce Robustness, and flatter curves can reflect over-patching rather than improved localization.Robustness is the primary window-selection criterion; Smoothness is reported only as a diagnostic.Source: §13 · pp. 20–21

14 Downstream Propagation after Arbitration Head Ablation

The experiments examine how arbitration ablation changes downstream attention, how brake localization depends on attenuation strength, and whether brake heads also regulate abstention under local visual corruption.

14 Downstream Propagation after Arbitration Head Ablation

  • After arbitration-head excision, downstream attention to injected conflict text decreases and spreads more diffusely over the remaining context.The input is held fixed before and after ablation.Source: §14 · p. 21

15 Layer-wise Attenuation Sensitivity for Brake localization

  • Brake localization requires attenuation: HR is positive for β < 1 and trends toward zero as β approaches 1.0.Although β = 0.0 gives the largest HR, β = 0.2 yields the largest unknown-rate gain while retaining a strong hallucination-relief signal, so it is used for the main experiments.Source: §15 · p. 21, Table 17 · p. 22

16 Robustness to Local Visual Noise

  • Across local-noise strengths, excising the same six brake heads raises unknown rate by roughly 32–34 points while clean-image accuracy stays unchanged.The pre-intervention unknown rate rises from 1.82 to 11.59 as noise increases.Source: §16 · p. 22, Table 18 · p. 22

17 More Experiment Results

Across benchmarks and models, the selected interventions target distinct failure behaviors: arbitration-head ablation reduces conflict following, while brake-head ablation increases abstention under visual degradation. The results also show that arbitration targets transfer more consistently than brake targets, and that arbitration localization depends on clean competence.

17.1 Head-Set Disjointness between Arbitration and Brake localization

  • Arbitration and brake head sets have zero intersection on SLAKE for both Hulu-Med 4B and InternVL3.5 4B.The reported Jaccard and overlap coefficients are both zero.Source: §17.1 · p. 22, Table 19 · p. 23

17.2 Arbitration Head Distribution in InternVL3.5 and Qwen3-VL

  • InternVL3.5 4B and Qwen3-VL-8B show broad mid-layer arbitration bands, while backbone heads are more dispersed and extend into late layers.The arbitration bands span roughly layers 13–22 in InternVL3.5 and 14–24 in Qwen3-VL-8B; their strongest CER is around layers 21 and 22, respectively.Source: §17.2 · pp. 22–23

17.3 Cross-Dataset and Cross-Model Transfer

  • Arbitration-head transfer remains positive across dataset and model shifts, whereas brake-head transfer is weaker, less directional, and more tied to the target model and degradation regime.Reported arbitration transfer gains include 30.77 and 12.44 points for Hulu-Med 4B, and 13.71 and 1.14 points for InternVL3.5 4B.Source: §17.3 · pp. 23–24, Table 20 · p. 23, Table 21 · p. 23

17.4 Clean-Competence-Stratified Head localization

  • At 80% clean correctness, selected arbitration sets remain close to the fully clean-correct set; at 50% and 30%, they broaden, and the all-wrong stratum separates.Relaxing clean competence can mix conflict arbitration with difficulty, uncertainty handling, and wrong-answer commitment.Source: Figure 13 · p. 24, §17.4 · pp. 24–25, Table 23 · p. 24

17.5 Full Results: ConflictMedQA

  • On ConflictMedQA, CRAFT lowers validation CFR and raises resist rate across models and injection positions, with larger W2C transitions and low C2W than random ablation.The results indicate generalization beyond the localization set, with the largest improvements in the Before A setting.Source: §17.5 · p. 25, Table 24 · p. 25

17.6 Full Results: PubMedQA

  • On PubMedQA, Qwen3-4B validation CFR falls from 53.70 to 3.70 and resist rate rises from 46.30 to 96.30 under Before Q.CRAFT also reduces CFR and improves resist rate across the reported models and injection positions.Source: §17.6 · pp. 25–26, Table 25 · p. 26

17.7 Full Results on Multimodal Conflict Benchmarks

  • Across VQA-RAD and SLAKE, arbitration-head ablation lowers validation conflict following, while brake-head ablation raises held-out unknown rates on SLAKE and HealMed-VQA.Random ablation is generally less reliable for arbitration; the two interventions correct distinct behaviors.Source: §17.7 · pp. 26–27, Table 26 · p. 27, Table 27 · p. 28

17.8 Validation Confusion Matrix Diagnostics

  • On Hulu-Med 4B VQA-RAD, arbitration ablation raises diagonal accuracy from 63.4% to 88.4% for gold yes samples and from 46.8% to 94.4% for gold no samples.The bidirectional correction restores the yes/no decision boundary rather than favoring one answer label.Source: §17.8 · pp. 28–29, Figure 14 · p. 28

18 Failure Case Studies

The case studies show partial but incomplete corrections after sparse head excision: conflict preference weakens in an arbitration example, and unknown gains rank in a brake example, but neither case is fully resolved.

18.1 Arbitration Failure

  • In the arbitration case, the gold-versus-wrong margin improves from −4.0417 to −0.8913, but the wrong answer remains top-ranked.The remaining error requires distinguishing medical process from physical injury, an etiology-level visual-grounding task.Source: §18.1 · pp. 29–30, Figure 15 · p. 29

18.2 Brake Failure

  • In the brake case, unknown rises above the gold answer after masking, but the wrong concrete answer remains top-ranked.The image no longer supports a reliable concrete answer, so the desired response is unknown.Source: Figure 15 · p. 29, §18.2 · p. 30

18 Failure Case Studies

  • The residual cases suggest that hard samples may require additional mechanisms beyond sparse head excision.The authors point to stronger image-grounded reasoning for difficult etiology attribution and adaptive abstention thresholds or localization-aware decoding for difficult brake cases.Source: §18 · p. 29, §18.2 · p. 30

19 Computational Cost

CRAFT’s tested head interventions are safety-oriented rather than acceleration methods: hooked masking leaves attention computation intact, while head skipping offers limited FLOP savings without fused kernels. The analysis is restricted to specific model, language, imaging, and intervention settings.

19 Computational Cost

  • Hooked head masking does not reduce profiled FLOPs, and query-head skipping offers only a small theoretical reduction without fused kernels.The implementations were benchmarked on one VQA-RAD validation sample per model.Source: §19 · p. 30, Table 29 · p. 30

20 Limitations and Future Directions

  • The mechanistic analysis covers open-weight models up to 14B, English prompts, predominantly chest X-ray and fundoscopy, and static head suppression.A fixed scan window may miss circuits in larger or closed systems; CT and MRI multi-slice inputs may involve different visual-token aggregation.Source: §20 · p. 31
  • The authors propose input-conditioned gating and tracing arbitration across multi-step agents as future directions.A preliminary single-layer classifier on the residual stream suggests gating feasibility with less than 0.5% additional latency.Source: §20 · p. 31

Related papers

Loading 2609.38810v1…