Zenaique

Stacking two unrelated LoRA adapters: most likely outcome

MCQ·Hard·4.0 · 0·~1 min·Asked atInflection AiSharechatTata Digital·Relevant atCohereDatabricksMetaMicrosoft
Attempt it
TL;DR

Merging two independently-trained LoRA adapters often hurts both tasks. Their deltas collide in shared weight directions. Route between adapters or train one joint adapter instead.

Memory aid
Sign in to see the mnemonic that makes this stick.
Easy to grasp

Imagine two editors revising the same manuscript, working in separate rooms, neither seeing the other's marks. One rewrites for a legal audience; the other rewrites for children. Each edit makes sense alone. Now you stack both sets of changes onto one page. Sentences get cut twice, tone whipsaws, and the result reads worse than either editor's version on its own. That is what happens when you add two LoRA adapters trained apart: each assumed it owned the weights, so their changes fight over the same words. The fix is not to merge the pages. You keep both edited versions and pick the right one per request, or you give both editors the full brief up front so they revise together.

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

This question probes a tempting but wrong mental model: that LoRA adapters are modular skills you can snap together like building blocks. The intuition feels airtight. Each adapter is small, low rank, and trained for one job, so surely adding two gives you a model good at both jobs for free. The correct answer, C, says the opposite usually happens. Merging two independently trained adapters frequently makes the model WORSE on each task than the matching adapter was alone.

The reason sits in how fine-tuning actually works. An adapter is not a self-contained module bolted onto the side of the network. It is a delta added directly to the base model's weights. Two deltas trained in isolation never coordinated, and when you sum them onto shared parameters, they collide. Understanding why, and knowing the safer alternatives, is exactly the senior judgment this question rewards. Notice how the question is constructed: three of the four options are designed to catch a specific misconception, and only one names the real failure mode and the real fixes.

This matters in practice because the merge button is cheap and the failure is silent. Nothing crashes when you sum two adapters. The model still produces fluent text. You only discover the regression when you run both eval sets and watch each score drop below the standalone adapter it was supposed to replace. Teams that treat adapters as additive ship that regression to production and spend days debugging a problem that was baked in at merge time.

The rest of this deep dive covers the merge mechanism, why each distractor is a real-world misconception, when merging is in fact safe, and the production patterns that sidestep the problem: per-request routing, joint training, and interference-aware merge recipes.

The LoRA update and why merging is linear

LoRA freezes the base weight matrix and learns a low-rank update. The effective weight at inference is the base plus a scaled product of two small matrices.

W=W0+αrBAW = W_0 + \frac{\alpha}{r} \, B A

Here B and A are the trainable low-rank factors, r is the rank, and α/r is the scaling factor. The factor B has shape d×r and A has shape r×k, so their product has the full shape of the base weight while costing only r·(d+k) trainable parameters. Because the update is just an added term, you can fold it into the base by computing the product once. That is what "merging" means in practice: you collapse the adapter into the weights and serve a single matrix with no inference-time overhead.

Merging two adapters is therefore the sum of two such deltas onto the same base. You compute each product, scale it, and add both to W_0. The operation is always well-defined, which is precisely why option D is wrong. The α/r scalar controls the magnitude of each update. It does not block coexistence. Two adapters can always be summed onto one base; the linear algebra never forbids it.

This is the crux that trips up candidates. Feasibility was never the issue. The LoRA equation is linear in the adapter product, so addition is trivially valid. Quality is the open question, and quality depends entirely on whether the two deltas were trained to be compatible. They almost never were.

Why independent deltas interfere destructively
Why the two optimistic distractors are wrong
When merging is actually safe
Production patterns that avoid the trap
Sign in to unlock the full deep dive.

Situations where this technique stops working.

Sign in to see when this approach fails.

2–4 min · Everything important, quickly.

Sign in to see the quick scan of the deep dive.
StrategyWhat happens to qualityWhen to use
Naive merge (sum deltas)Often regresses BOTH tasks for unrelated adaptersOnly when adapters are near-identical or same warm start
Interference-aware merge (TIES, DARE)Reduces but does not eliminate conflictYou need one fused checkpoint and can tolerate some loss
Joint training on unionBest quality, tasks share gradientsYou control both datasets and can retrain
Route per request (multi-LoRA)Each task stays at full adapter qualityMulti-tenant serving, many adapters, one base

Real products, models, and research that use this idea.

  • vLLM and S-LoRA serve thousands of distinct LoRA adapters on one base model by routing per request, deliberately avoiding the merge and regress trap.
  • Predibase and Together.ai multi-tenant fine-tuning platforms keep customer adapters separate at inference rather than merging them into a shared base.
Sign in to see more production examples.

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy does low rank fail to prevent interference between two merged adapters?
A

Reason about the column spaces of the two B matrices. Even rank-r subspaces intersect, so summed deltas can land in a direction neither optimiser visited, degrading both tasks.

3 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Assuming more adapters equals more capability. Two independent fine-tunes never shared a gradient, so adding their deltas frequently degrades both source tasks rather than combining them.

Sign in to see all red flags and common mistakes.

60 second bullets to scan on the way to the call.

  • Why two independently trained adapters interfere when merged

  • How the linear LoRA update makes merging feasible but not safe

Sign in to unlock the revision sheet.

Primary sources. Browse if you want the original framing.

Similar questions

Same topic, related formats. Practice these next.

4 curated
Next question
What is RLHF, and why is it used after pretraining?
MCQ·Easy