🤖 AI

Beyond Endpoint Gains: A Weight-Delta Audit of Medical Specialization

This paper proposes a weight-delta path audit to analyze medical specialization in language models, revealing that while decoder-side updates strongly correlate with benchmark performance gains, the underlying component-level mechanisms remain diffuse and cannot be uniquely attributed to specific architectural families like MLPs.

Original authors: Praphul Singh, Shanu Kumar, Akshat Agarwal

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Praphul Singh, Shanu Kumar, Akshat Agarwal

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of artificial intelligence, researchers have spent years building massive, general-purpose computer brains that can read, write, and reason about almost anything. These generalist models are like broad-shouldered students who know a little bit about history, science, and literature. To make them useful for specific jobs, scientists often take these general models and give them extra training focused on one subject, such as medicine. The result is a specialist model, designed to answer questions about diseases, treatments, and medical exams better than its generalist ancestor. For years, the standard way to judge if this specialization worked was simple: compare the final test scores. If the specialist got a higher score on a medical exam than the generalist, the training was considered a success. This approach, however, leaves a crucial mystery unsolved. It tells us the destination was reached, but it reveals nothing about the journey or the specific changes made to the model's internal structure to get there. It is like knowing a car is faster after a mechanic worked on it, without ever looking at the engine to see which part was actually tightened or replaced.

A new study by researchers at IIT Kanpur, Oracle Health AI, and MBZUAI decides to look under the hood. Instead of just comparing the final test scores of two medical models, the team performed a detailed audit of the actual changes made to the computer's memory during the training process. They examined two specific pairs of models: one starting from a general model called Gemma and evolving into a medical version called MedGemma, and another starting from Qwen and becoming HuatuoGPT. By treating the difference between the general and specialist versions as a physical path through the model's internal weights, the researchers could trace exactly how the model changed as it learned medicine. They wanted to know if the improvement in medical knowledge was a clean, localized fix, or if the changes were scattered and messy throughout the system.

The researchers began by mapping the entire landscape of changes. They found that the update was not a tiny, precise edit to a single component. Instead, the changes were broad and structured, spreading across many different layers of the model's brain. The most significant shifts occurred in the parts of the model responsible for processing information and making decisions, known as the decoder. When the team simulated the full journey from the general model to the specialist, the model's performance on medical tests rose in perfect step with the changes in its internal weights. This confirmed that the observed updates were indeed the source of the medical improvement. However, the journey was not a straight line of pure medical gain. As the model moved toward specialization, it also picked up changes in other areas. Some non-medical tasks improved, while others got slightly worse, yet the overall ability to handle general knowledge remained surprisingly stable. The update was a complex package, carrying medical expertise alongside a mix of other gains and losses.

The most surprising finding came when the researchers tried to pinpoint exactly which part of the model was responsible for the medical expertise. A common theory in the field suggests that the "feed-forward" layers, which act like the model's long-term memory banks, are the primary place where new knowledge is stored. The researchers tested this by isolating these layers and seeing if they could reproduce the medical gains on their own. They found that these layers did indeed capture a large portion of the improvement, performing better than other parts of the model. But when they ran a strict control test, the story changed. They created random groups of the same size and energy from the model's updates and found that these random groups performed just as well, or sometimes even better, at reproducing the medical gains. This suggests that the apparent success of the memory layers was not because they were uniquely designed for medicine, but simply because they contained a large volume of the total changes.

To further test this, the researchers tried to reverse the process. They took the fully trained specialist model and attempted to "roll back" or remove specific parts of the update to see if they could fix the non-medical losses without hurting the medical gains. They hoped to find a clean switch: a specific component that, if removed, would restore general knowledge while keeping medical expertise intact. No such switch existed. When they removed the memory layers to fix the non-medical losses, the model's medical performance crashed. When they removed other parts, the medical performance dropped even more. The trade-off was inescapable; the medical improvement and the side effects were deeply intertwined, woven into the same fabric of changes. There was no single, isolated circuit that held the medical knowledge.

The study concludes that the path to specialization is far more complex than a simple upgrade to a specific module. The medical expertise is not stored in a single, neatly labeled compartment that can be swapped in or out. Instead, it emerges from a broad, distributed set of changes that affect the model's entire reasoning process. While the memory layers play a significant role, they are not the sole explanation, and their apparent dominance is largely a result of their size and the sheer amount of change they contain. The researchers emphasize that this work is an audit of how the model changed, not a guarantee of its safety or reliability in a real hospital. The findings suggest that when we see a model get better at a specific task, we should not assume the improvement is a clean, localized fix. It is likely a broad restructuring of the system, where gains and losses are inextricably linked, and where simple explanations about which part of the brain is responsible may be misleading.

Technical Summary: Beyond Endpoint Gains: A Weight-Delta Audit of Medical Specialization

Problem Statement
Specialist language models are typically evaluated through "endpoint gains": comparing the score of a generalist model against a domain-specialized model on a benchmark. While this demonstrates that specialization occurred, it leaves the released update itself unexamined. The standard endpoint view fails to answer critical questions: What does the specific weight update carry? Is the specialization localized to specific components, or is it a broad, distributed change? Does the update introduce off-domain regressions or gains? Current literature often treats the score difference as evidence of specialization without auditing the internal structure of the update or testing whether simple component explanations (e.g., "the MLPs contain the medical knowledge") survive rigorous controls.

Methodology
The authors propose a paired weight-delta path audit to analyze the internal structure of released specialist updates. The methodology relies on two public, tensor-aligned checkpoint pairs:

Gemma-3-4B-IT $\to$ MedGemma-4B-IT (Primary audit).
Qwen2.5-7B-Instruct $\to$ HuatuoGPT-o1-7B (Replication across architecture and lineage).

The audit treats the difference between the generalist and specialist checkpoints ( $\Delta W$ ) as an auditable path through weight space. The protocol involves four stages:

Descriptive $\Delta W$ Screen: Analyzing the Frobenius norm, relative change, and singular-value spectrum of the weight update to determine if the specialization is a localized edit or a broad, structured update.
Full Decoder-Path Reconstruction: Constructing a linear additive path $W(t) = W_{base} + t\Delta W_{dec}$ (where $t \in [0,1]$ ) to test if the full decoder-side update reconstructs the measured medical benchmark movement.
Scoped Component-Family Paths: Testing "partial sufficiency" by applying only subsets of the update (e.g., Attention layers, MLP layers, or mixed buckets like Attention + $down\_proj$ ) to see if they can reproduce the benchmark movement.
Controls and Rollbacks:
- Matched Random Controls: Comparing structured buckets against random subsets matched by parameter count, matrix count, or $\Delta W$ energy to test if performance is due to component identity or simply update mass.
- Endpoint-Anchored Rollbacks: Starting from the full specialist update and rolling back specific families (setting $\alpha \in [0,1]$ ) to observe the trade-off between recovering selected non-clinical regressions and retaining medical performance.

Evaluation Metrics
The audit uses a medical composite (1,810 examples from MedQA and medical MMLU subsets) as the target readout. It also employs:

A full non-clinical suite (7,325 examples) to check for broad forgetting.
Selected regression slices (2,469 examples) and selected gain slices (2,606 examples) defined as endpoint-conditioned diagnostics where the specialist model dropped or gained at least 0.02 accuracy compared to the base.
Metrics include Medical Retention (normalized gain), Regression Recovery, and Gain Acquisition.

Key Results

Broad, Structured Updates: The $\Delta W$ screen reveals that the specialization update is not a tiny, localized edit. It is broad and structured across decoder layers and projection families. In the Gemma pair, MLP projections account for a large share of update energy; in the Qwen pair, MLPs account for ~87% of parameters and update energy.
Full Path Reconstruction: The full decoder-side update strongly reconstructs the measured medical benchmark movement.
- Gemma $\to$ MedGemma: The full path achieves 0.974 normalized medical retention.
- Qwen $\to$ HuatuoGPT: The full path achieves 1.183 normalized medical retention (slightly exceeding the finite-sample endpoint).
- This confirms the decoder delta is the correct substrate for the audit.
Heterogeneous Off-Domain Movement: The update does not represent a simple trade-off between medical ability and general ability. While aggregate non-clinical accuracy remains nearly flat, the update carries heterogeneous movements: specific non-clinical slices regress, while others gain.
Negative Localization (The Core Finding):
- Partial Sufficiency: The MLP family is the strongest scoped component, reaching 0.797 medical retention at its best point (scale $t=0.80$ ) in Gemma and 1.024 in Qwen. However, no single bucket matches the full decoder path.
- Failure of Controls: Matched random controls (matched by parameter count or energy) often approach or exceed the performance of the structured buckets. For instance, in Gemma, parameter-matched random controls achieved 0.814 retention at $t=1$ , exceeding the structured MLP's value of 0.699 at $t=1$ (though the structured MLP's best point was 0.797). This suggests that "sufficiency" is often a function of update mass rather than unique semantic localization.
- Rollback Trade-offs: Rolling back the MLP family from the specialist endpoint recovers more selected regressions than rolling back Attention, but at a severe cost to medical retention (e.g., dropping to 0.359 retention in Gemma). No single family acts as a "clean knob" that preserves medical gains while fixing regressions.

Significance and Claims
The paper claims to provide a reusable audit protocol that sits between endpoint evaluation and mechanistic circuit discovery. Its primary contributions are:

Reconstruction: Demonstrating that measured medical benchmark movement is carried by the released decoder updates in two distinct model families.
Negative Localization: Establishing that at the tested granularity (coarse component families), the medical benchmark movement is not uniquely localized to a specific component family (such as MLPs). The apparent sufficiency of MLPs is largely explained by their size and the magnitude of the update they contain.
Methodological Discipline: Arguing that component-level claims must survive size/update-mass controls and endpoint-anchored rollbacks before being accepted as explanations.

The authors explicitly state that these results are not clinical validation, nor do they provide a circuit-level mechanism (e.g., specific neurons or circuits). The audit is limited to text-only multiple-choice benchmarks and does not measure free-form generation, calibration, or clinical utility. The conclusion is that while endpoint gains motivate the study of updates, simple component narratives (e.g., "specialization is in the MLPs") are insufficient without rigorous negative controls.

Technical Summary: Beyond Endpoint Gains: A Weight-Delta Audit of Medical Specialization

More like this