Binesh Sadanandan PhD Dissertation Companion
Read thesis

Demos

Six ways into the argument

Start with what the failure looks like. Then go inside the model: what the model is thinking at every layer of one question, and where the answer commits across 1,396 pairs. Then try the two signals meant to catch the failure before a reader ever sees an answer. Everything after the first section runs on the saved evaluation rows the dissertation reports. Move a threshold, switch a rule off, or drag through the layers, and the numbers recompute in your browser.

These demos replay saved offline evaluation on a research dataset. They are not a clinical tool, they do not run a model, and nothing here is medical advice.

1. What a paraphrase flip looks like

Five curated cases show the failure the later signals are meant to catch: questions that mean the same thing get contradictory answers. Within a case the image never changes; only the wording does.

GPT-5-mini and Claude Haiku 4.5 appear as general-purpose reference models for contrast; the dissertation's benchmark covers six medical vision-language models. Each verdict chip compares a model's answers with the reference read: whether the original answer was right, and whether any rewording changed it. In practice, a flipped presence answer is the difference between a finding flagged for review and one missed.

Case 01 · Curated demonstration set (MIMIC-CXR)

Pneumothorax

3 flipped answers 3 of 3 models flip
Illustrative pictogram, standing in for the study image

The radiograph is not shown

This case uses a MIMIC-CXR study. The PhysioNet licence for MIMIC-CXR prohibits redistributing images, so the radiograph is not published here. Anyone with credentialed access can obtain it from PhysioNet.

The question and the answers are below. The image never changes within a case; only the wording does.

Original question

Is there evidence of pneumothorax in this image?

Reference answer

No

Why this case mattersThe image shows no pneumothorax, yet a single more specific rewording about pleural air pushes all three models into the same false positive.
Why these questions count as equivalentEach paraphrase keeps the same clinical operator (presence of a finding) and the same finding (pneumothorax), varying only syntax, scope wording, specificity, or synonyms, which is the class of rewrite the equivalence audit rubric treats as clinically equivalent.

MedGemma-4B

Correct until rephrased

Answer to the original question: No

  1. Does this image demonstrate a pneumothorax?

    Syntactic restructuring No Same as the original Matches the reference answer

  2. Is there evidence of pneumothorax, unilateral or bilateral, on this image?

    Scope quantification No Same as the original Matches the reference answer

  3. Is there radiographic evidence of pleural air consistent with pneumothorax in this chest radiograph?

    Specificity modulation Yes Flipped from the original Conflicts with the reference answer

  4. Is there air in the pleural space suggesting pneumothorax on this image?

    Lexical substitution No Same as the original Matches the reference answer

GPT-5-mini

Correct until rephrased

Answer to the original question: No

  1. Does this image demonstrate a pneumothorax?

    Syntactic restructuring No Same as the original Matches the reference answer

  2. Is there evidence of pneumothorax, unilateral or bilateral, on this image?

    Scope quantification No Same as the original Matches the reference answer

  3. Is there radiographic evidence of pleural air consistent with pneumothorax in this chest radiograph?

    Specificity modulation Yes Flipped from the original Conflicts with the reference answer

  4. Is there air in the pleural space suggesting pneumothorax on this image?

    Lexical substitution No Same as the original Matches the reference answer

Claude Haiku 4.5

Correct until rephrased

Answer to the original question: No

  1. Does this image demonstrate a pneumothorax?

    Syntactic restructuring No Same as the original Matches the reference answer

  2. Is there evidence of pneumothorax, unilateral or bilateral, on this image?

    Scope quantification No Same as the original Matches the reference answer

  3. Is there radiographic evidence of pleural air consistent with pneumothorax in this chest radiograph?

    Specificity modulation Yes Flipped from the original Conflicts with the reference answer

  4. Is there air in the pleural space suggesting pneumothorax on this image?

    Lexical substitution No Same as the original Matches the reference answer

The claim this illustrates · Benchmark flip rates · The PSF-Med paper

2. What each layer is thinking

You have just seen answers flip. This opens one pair up and reads the model's mind on the way down. The Jacobian lens asks, at every layer and every word of the prompt: if the model had to answer from this point, what would it say? Two phrasings of one question, side by side, on one chest X-ray. The default case is a true paraphrase, where the two columns agree all the way down; switch cases to watch the answer column split. Hover any cell to see the words.

Case

Source: jlens_paraphrase/jspace_data.json, MedGemma-4B, 33 layers. Chapter 5.

Loading this case's readout (about 400 kilobytes).

Read Thrust 3 Where the answer commits

3. Where the answer commits

The section above reads one pair. This one asks whether that split is a property of the model in general or of that pair alone. Take two versions of the same question, one the model answers correctly and one it does not, and copy the model's internal state from the first into the second at a single layer. If the answer changes, that layer is carrying the decision. Drag through the 34 layers of MedGemma-4B and watch where it happens, across 1,396 pairs at once.

Controls

Layer 16
Jump to

Source: jlens_paraphrase/patch_full.json, 1,396 pairs across 91 findings, PadChest, MedGemma-4B. Chapter 5.

—

Of 1,396 pairs flip when the state is swapped here

—

—

Flip when an image token is swapped instead, at the same layer

The control: flips here would mean patching any position changes the answer.

—

What this layer is doing

—

The answer commits in a narrow band

Flip rate by transplant layer. The line is the layer you picked.

Median commit layer 16, 95% bootstrap interval [16, 16]. 71% of pairs commit within layers 15 to 17.

The same curve as a table
LayerAnswer-position transplantImage-token control

The answer is not assembled gradually. Through layer 13 the transplant does almost nothing, under 0.2%. Then it goes 8.2% at layer 14, 35.2% at 15, 73.2% at 16, and 100% by 20. The image-token control never exceeds 0.07%, one pair of 1,396: the flips are specific to the answer position. That is what makes layer 16 the point where this model has decided.

Read Thrust 3 Sample and source The layer-17 features

4. Predictive entropy as a filter

Knowing where the answer commits does not help at inference, because you cannot open the model up on a live case. So: is there a signal you could actually read? The model produces a probability for yes and no on every question, and predictive entropy measures how close that distribution is to a coin flip. The claim in Chapter 7 is that this single number, from one forward pass, ranks both errors and paraphrase flips. Drag the threshold: everything at or below it is answered automatically, everything above is escalated to a person.

Controls

0.350 nats
Jump to

Source: results/uai/{model}/padchest/softmax_entropy.jsonl, 861 questions, Chapter 7.

—

Answered automatically

—

—

Wrong, among those answered

—

—

Would flip under paraphrase, among those answered

—

—

Escalated but actually right

The cost of caution

Where the threshold falls

Stable and flipped questions by predictive entropy. The line is your threshold.

  • Answer stayed the same under paraphrase
  • Answer flipped under paraphrase
The same numbers as a table
MeasureAt this thresholdIf you answer everything

One forward pass ranks flip risk. Entropy separates flipped from stable questions well enough to score an area under the curve of 0.823 for flips and 0.862 for errors on this model, on a scale where 0.5 is a coin flip and 1.0 is perfect. Tightening the threshold buys accuracy on what remains.

5. The margin gate, and what it cannot catch

The entropy filter above is the deployed model. This is the controlled probe of Chapter 5, a small Gemma-3 decoder over a frozen MedSigLIP-448 encoder that reads a single yes or no from its own language-model head. Its only inference-time signal is the yes-minus-no margin, and a flip is a sign change of that margin across paraphrases, so it is likeliest when the margin already sits near zero. Drag the threshold: every answer with a margin at least this wide is answered automatically, everything narrower is escalated. Switch between the training hospitals and two the probe never saw to see both what one forward pass can catch and what it cannot.

Controls

1.00
Jump to

Source: results_transfer/detect_percase.json (scripts/analysis/detect.py), Chapter 5.

—

Answered automatically

—

—

Would flip under paraphrase, among those answered

—

—

Wrong, among those answered

—

—

Escalated but actually stable

The cost of caution

Where the threshold falls

Stable and flipped answers by absolute margin. The line is your threshold; everything to its right is answered automatically.

  • Answer stayed the same under paraphrase
  • Answer flipped under paraphrase
The same numbers as a table
MeasureAt this thresholdIf you answer everything

One forward pass ranks flip risk. The absolute margin separates flipped from stable answers at an area under the curve of — on this set, where 0.5 is a coin flip and 1.0 is perfect, and it holds on the two held-out hospitals (0.974 each). It beats re-asking under several paraphrases and a trained hidden-state probe, at no extra inference.

The margin catches flips, not blind answers The same margin is at chance (0.47 to 0.52) for the other failure: an answer that ignores the image. A confident answer that never read the radiograph looks exactly like a confident grounded one, so only a second pass that swaps in another patient's image catches it. A finer question, whether a stable answer quietly changes how much it uses the image from one phrasing to another, is real at the population level but not identifiable per case: zero grounded-to-unreliant transitions over 345 patient-seed evaluations, a patient-level upper bound near 3%. Per-case image reliance is a population property here, not a per-prediction label.

6. The deployment-readiness audit

Entropy cannot tell you whether the image was used. So Chapter 7 builds a rule out of several signals at once, and asks it to admit only predictions that are both stable under rephrasing and dependent on the image. Switch the conditions off and on to see what each one is doing.

This is an audit; it cannot be deployed as a gate The region condition needs a radiologist's bounding box around the finding, which by definition does not exist for a patient who has not been read yet (it is available for 637 of these 861 questions and for none of MIMIC). The swap condition needs a second image known to show the opposite label. Both are offline evaluation checks; neither could run in a clinical workflow. The rule below is the one the dissertation reports. The "text-only" condition is not part of that rule; turn it on to see why.

The rule

Admit when the paraphrases agree and (swapping the image changes the answer or the region and the background disagree)

Conditions
Jump to

Source: results/miccai/{model}/padchest.jsonl, 861 questions, Chapter 7.

—

Admitted

—

—

Accuracy on admitted

—

—

Admitted that still flip under paraphrase

Residual instability

—

Admitted that text alone already answers

The finding of Chapter 7

Where the 861 questions go

  • Answered automatically
  • Escalated to a person

The same numbers as a table
MeasureThis ruleNo gate

Six real questions, and what the rule did with them

Each is a row from the evaluation, showing the paraphrases the rule compared, what happened when the image was swapped or masked to the annotated region, and whether the question was answerable from the text alone.

The rule works on its own terms: on the targeted adapter it admits 33.0% of questions at 96.8% accuracy, against 91.5% if you answer everything, and every condition you remove admits more and is right less often. What it does not buy is grounding. Watch the fourth number: most of what the rule admits is exactly what the question text alone already answers. On the slice where the model's text-only answer disagrees with its full answer, the same targeted adapter scores 2.9% accuracy (n=241), and Qwen2-VL gets 0 of 279 right under every rule tried. High admitted accuracy here comes from high finding base rates making the text prior usually correct. The model is not reading the radiograph. The dissertation's answer to "can a gate admit image-grounded cases?" is no.

Read Thrust 5 The grounded-slice result Risk-coverage curves

Cite


  

  

     Permanent link
    
  




  

    
    
    
    
  

  

    

    
Type to search. Press Escape to close.