Binesh Sadanandan Dissertation
Read dissertation

The defense story

Ask the same question twice.

One small habit of distrust runs through this dissertation: take a question the model answers, say the same thing in different words, and watch whether the answer holds. The defense tells that story in five thrusts. Measure the failure, diagnose it, explain and fix it, build the safety screen, then decide what deployment can trust. Every number below links into the evidence it comes from.

Thrust 1 · Measure

Flips are common, and the obvious suspects do not explain them.

How often do equivalent questions change model answers?

Start with what an LLM calls a paraphrase: judged against a clinical equivalence rubric, only half of the candidate rewordings preserve the question. After the audit, the PadChest regeneration, and filtering, 92,856 pairs over three countries remain. On that set the six models flip between 6.4% and 54.7% of the time, and the bigger models are not the steadier ones.

Thrust 2 · Diagnose

The steadier the answers, the less the model looks at the scan.

Does consistency mean the model is using the image?

Consistency can mean the model reads the image the same way every time, or that it never reads it at all. On a 107-pair diagnostic set the answer is uncomfortable: the most consistent backend keeps answering when the image is removed, and across the three backends consistency and text-reliance rise together. The set predates the equivalence audit, so the figures are illustrative rather than estimates.

Thrust 3 · Explain & fix

A candidate circuit, a 59% fix, and a cause in the training data.

Can paraphrase flips be reduced without degrading observed accuracy?

The mechanism work points to a narrow band where the answer commits, then partly undoes itself: the layers the diagnosis named are not the best place to intervene, and single-feature ablation restores only a handful of flips. The parametric fix that works is blunt by comparison, a targeted LoRA on layers 15 to 19 that cuts flips by about 59% on a clean patient-disjoint split. A controlled probe then shows why data-side coverage is the larger lever, and a re-audit shows the adapter buys part of its consistency from the question text, so the fix is real but bounded.

Diagram of the candidate two-stage account. A chest X-ray and a reworded question enter at left; the signal passes through layer 17 feature 3818, the operator or register gate, then layer 29 feature 12139, the answer-selection gate, then the yes or no answer. When 3818 is active the flip starts at layer 17 and propagates downstream, restoring 24 percent of flips on MIMIC and 17 percent on PadChest; when 3818 is flat, 12139 is the proximal cause, restoring 58 percent and 27 percent. The link between the gates is direction-coherent on all 37 mediation pairs.
The candidate two-stage account: a register gate at layer 17 and an answer-selection gate at layer 29, one direction-coherent chain across 37 mediation pairs.
The layer-16 residual transplant. Left, the method: the same chest X-ray with a detecting phrasing answered Yes and a missing phrasing answered No, transplanting the residual stream at one layer. Right, a line chart of answer flip rate against patched layer: near zero through layer 13, 8 percent at layer 14, 35 percent at 15, 73 percent at 16, saturating by 19, with the image-token control flat at zero. The per-pair commit layer has median 16, a 95 percent confidence interval of 16 to 16, and 71 percent of pairs within layers 15 to 17.
The lens-free transplant: across 1,396 pairs, the answer commits at layer 16; the image-token control never exceeds 0.0007.

Figures use an illustrative, licence-cleared chest X-ray (Wikimedia Commons); the underlying pairs are credentialed data and are not shown.

Thrust 4 · Safety-evaluate

Consistency without grounding is the safety paradox.

Do consistency improvements produce safer, more image-dependent behavior?

A model can pass every consistency check and still not be reading the image. Crossing consistency with image reliance in a four-quadrant screen shows most consistent predictions are image-invariant, 81% on average across ten settings, and consistency-optimized adapters concentrate in exactly the quadrant you would not want. The screen flags behavior, not per-prediction error: on PadChest the text-reliant cell is often the more accurate one.

Thrust 5 · Deploy

Abstaining buys safety.

Can uncertainty and audit signals identify predictions that should be escalated?

The deployment question is what can be caught before a prediction leaves the building. Predictive entropy from a single forward pass ranks flip risk and error risk well, and it transfers across architectures. The gate built on these signals is honest about its limits: it is an offline readiness audit whose region term needs a radiologist's box, and what it admits skews toward questions the text already answers.

The story in full

The dissertation carries the full argument; the papers carry the peer-reviewed pieces; the evidence explorer carries every number with its sample and source.

Read the dissertation The papers The evidence

Cite

Permanent link

Type to search. Press Escape to close.