SMASH 2026 · Paper #2 · Result card
Gated Knowledge-Guided Prototype Transfer for Zero-Support Medical Vision Language Model
Ungated transfer made a missing grade worse in 46–50% of support draws. The gate cut that to 16% and added +2.37 points of missing-class recall.
classes No DR (3), Mild (1), Moderate (1), Severe (0), Proliferative (0)zero-support Severe, Proliferativeβ 1.0k 3temperature automodality gap 1.0separability 1.0gap shrink / grade 0.5support draws 50data synthetic ordinal embedding world
Negative-transfer rate
46–50% → 16%
share of 50 support draws where missing-class recall fell; ungated (VLM / ontology) → gated
Missing-class recall (gated)
+2.37 pts
95% CI [+1.72, +2.95] vs SS-Text+ (73.4% → 75.8%)
Supported-class recall (gated)
+0.97 pts
ΔSCR (paper: −1.20); Δtop-1 +1.12, ΔACA +1.53, ΔH +1.56
Evidence gate
0.51 / 0.22
Severe 0.514 · Proliferative 0.215 (paper: 0.582 severe, 0.214 proliferative)
From the abstract
We introduce gated Knowledge-Guided Prototype Transfer (KGPT), a pre-repulsion module that transfers supported-class visual residuals to missing classes using VLM similarity and an ontology-based clinical distance. A validation-free evidence gate scales transfer based on semantic relevance and residual agreement, thereby limiting unsupported extrapolation.
Evidence
Paper vs. your run
| Metric | Paper | Your run | |
|---|---|---|---|
| SS-Text+ Top-1 / ACA / MCR / SCR / H | 57.45 / 54.64 / 54.90 / 54.46 / 52.87 | 72.9 / 62.7 / 73.4 / 55.5 / 63.0 | absolute levels depend on the synthetic world |
| Ungated NTR (VLM / Ont), severe NTR | 46% / 48%; severe 38% / 34% | 50% / 46%; severe 16% / 14% | NTR = share of draws with ΔMCR < 0; severe = ΔMCR < −5 |
| Ungated ΔMCR (VLM / Ont) | −2.16 [−4.73, 0.32] / −2.43 [−4.68, −0.21] | -0.28 [-1.54, +1.05] / +0.32 [-0.93, +1.61] | |
| Gated ΔMCR / ΔACA / ΔH | +5.23 [3.84, 6.69] / +1.37 [0.78, 1.99] / +1.83 [1.06, 2.64] | +2.37 [+1.72, +2.95] / +1.53 / +1.56 | |
| Gated ΔSCR / ΔTop-1 | −1.20 [−1.52, −0.88] / −0.45 [−0.71, −0.19] | +0.97 / +1.12 | the price of transfer |
| Gated NTR; draws improved / unchanged | 2%; 36 improved, 13 unchanged of 50 | 16%; 42 improved, 0 unchanged of 50 | |
| Mean gate (severe / proliferative) | 0.582±0.010 / 0.214±0.003 | 0.514±0.019 / 0.215±0.008 | paper relevances 0.381 / 0.102 imply agreement a ≈ 0.94–0.95 |
Try next
- Zero gap shrink — borrowed correction is always exactly right — With gap_shrink=0 every class has the same modality gap, so the borrowed visual correction needs no scaling. Does the gate, by reducing the effective step, actually hurt missing-class recall compared to ungated transfer? This is the cleanest test of whether the gate adds value beyond overshoot-prevention.
{"gap_shrink": 0} - Middle grade with zero support — does a central missing class reveal a gate-closed configuration? — Setting support_counts to 3,1,0,1,0 leaves Moderate unsupported and flanked by two supported neighbours (Mild and Severe). Because the two neighbours may be nearly equidistant in VLM space, the NN margin above the candidate-row median (q_vlm) can be near zero, triggering a gate-closed class and showing how the gate refuses to act when evidence is ambiguous.
{"support_counts": "3,1,0,1,0"} - Lower fixed beta 0.5 — is a smaller step as safe as the gate at beta=1? — The beta sweep shows ungated NTR is around 18–24% at β=0.45–0.6. Comparing ungated β=0.5 to gated β=1 tests whether evidence-adaptive scaling adds value beyond simply turning down the step size. If gated β=1 still beats ungated β=0.5 on ΔMCR with lower NTR, the gate earns its keep.
{"beta": 0.5}
Read this before sharing
- Synthetic embeddings stand in for MESSIDOR with FLAIR features.
- SS-Text+ is approximated with a shrinkage-plus-repulsion prototype.