🤖 machine learning

Instruction-Tuned, but Not More Verifiable Instruction-Following: A Cross-Task Diagnosis for LoRA Adapters

This paper demonstrates that nominal labels like "instruction-tuned" are unreliable predictors of actual capability gains, revealing a phenomenon termed "capability drift" where LoRA adapters often fail to improve verifiable instruction-following performance despite their training objectives, thus necessitating routine cross-task evaluation before deployment.

Original authors: Junyi Zou

Published 2026-03-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Junyi Zou

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

The Big Idea: "The Resume Lie"

Imagine you hire a chef based on their resume. The resume says they are a "Pastry Specialist." You expect them to make amazing cakes.

But when they start working, something weird happens:

They don't make better cakes than before (in fact, their cake quality drops slightly).
However, they suddenly become incredible at chopping vegetables and organizing the pantry—tasks they weren't even hired for!

This paper is about a similar phenomenon in Artificial Intelligence. Researchers found that when they train AI models with a specific label (like "Instruction-Tuned"), that label often doesn't tell the whole truth about what the AI actually gets better at.

The Cast of Characters

The Base Model: A smart but untrained AI (like a raw, unskilled intern).
The LoRA Adapter: A small, cheap "add-on" or "patch" you attach to the AI to teach it a new skill without retraining the whole brain. Think of it as a specialized training manual you give the intern.
The Nominal Label: The title on the training manual. For example, "Instruction-Tuned" (meaning: I learned to follow orders).
The Reality: What the AI actually does after reading the manual.

The Experiment: The "Strict Boss" vs. The "Math Test"

The researchers took an AI and gave it a "Instruction-Tuned" manual. They expected the AI to get much better at following strict rules (like "Write a poem with exactly 5 lines and no commas").

They tested the AI on two things:

The Strict Boss (IFEval): A test where the AI must follow very specific, verifiable rules. If it misses one rule, it fails.
The Math Test (Numeric Reasoning): A test where the AI has to solve math problems.

The Shocking Result:

The Expectation: The "Instruction-Tuned" AI should be great at the Strict Boss test.
The Reality: The AI got worse at following strict rules.
The Twist: The AI got much better at the Math Test!

It's like hiring a "Pastry Specialist," and instead of making better cakes, they suddenly become a master chef at grilling steaks, but their cakes get slightly worse.

The "Capability Drift" Analogy

The authors call this "Capability Drift."

Imagine you are driving a car. You put on a new set of tires labeled "Off-Road Tires."

Goal: You expect the car to handle mud and rocks better.
Drift: Instead, the car actually drives worse on the mud, but it suddenly drives much faster on the highway.

The label on the tires (the training objective) didn't match the actual performance on the road (the real-world capability). The AI "drifted" away from what it was supposed to learn and learned something else entirely.

Why Does This Happen?

The paper doesn't know exactly why this happens (the "mechanism" is still a mystery), but they offer a guess:

Maybe when the AI learns to follow instructions, it changes its "personality" or style. It might start answering questions in a way that accidentally helps it solve math problems (because math problems often look like instructions), but this new style actually confuses it when trying to follow strict formatting rules.

The "Drift Score"

To measure this confusion, the researchers created a simple score called the Drift Score.

If the score is positive, it means the AI got better at the wrong thing (the Math Test) more than it got better at the right thing (The Strict Boss).
They found that this happens often, but not always. It depends on the specific settings, like how big the AI is or how the training was done.

The Takeaway: Don't Trust the Label!

The most important lesson from this paper is for anyone using AI:

Do not trust the name on the box.

Just because an AI model is labeled "Instruction-Tuned" or "Math-Trained," it doesn't guarantee it will actually be good at those specific tasks in the real world. It might have drifted and become good at something else entirely.

The Advice:
Before you deploy an AI model into a real business or app, you must test it on the specific tasks you care about. Don't just assume the label is true. Check the car's tires on the actual road, not just the label on the package.

Summary in One Sentence

This paper warns us that AI models often learn the wrong skills despite their labels, so we must test them ourselves rather than trusting the marketing name.

1. Problem Statement

The paper addresses a critical gap in the deployment of Large Language Model (LLM) adapters, specifically LoRA (Low-Rank Adaptation).

The Assumption: Adapters are typically selected based on nominal labels (e.g., "instruction-tuned," "numeric-reasoning-tuned"). Practitioners assume these labels reliably predict the specific capabilities that will improve after deployment.
The Reality: The authors hypothesize that nominal training objectives do not necessarily align with realized cross-task capability gains. Specifically, an adapter trained to follow instructions may not actually improve strict, automatically verifiable instruction following (compliance with hard constraints), and might instead improve unrelated tasks (e.g., numeric reasoning) or degrade performance.
Core Question: Do nominal training objectives reliably predict realized performance gains across different tasks, particularly on verifiable compliance metrics?

2. Methodology

The authors employ an empirical cross-task diagnosis framework rather than mechanistic attribution or metric innovation.

Experimental Setup

Base Models: Fixed base models (e.g., Qwen3-8B, Qwen3-14B, Llama3-8B-Instruct).
Adapters: Two primary LoRA adapters trained with different nominal objectives:
1. instr: Instruction-tuned.
2. reason: Numeric-reasoning-tuned.
Evaluation Tasks & Metrics:
1. Verifiable Instruction Following (Target): Measured using IFEval (Instruction Following Evaluation).
  - Metrics: Instruction-Level Accuracy (ILA) and Prompt-Level Accuracy (PLA). PLA is the strict, end-to-end metric used for the primary analysis.
2. Numeric Reasoning (Off-Target): Measured using Numeric Match (NM) (exact match of the final numeric answer).
3. Domain QA: Used as a secondary setting.

Diagnostic Metric: Drift Score

To quantify the mismatch between nominal goals and realized gains, the authors define a Drift Score:
$\text{DriftScore}(a \to b) = \text{OffTargetGain}(a, b) - \text{TargetGain}(a)$

Where $a$ is the adapter, $b$ is the off-target task, and $t(a)$ is the nominal target task.
Interpretation: A large positive Drift Score indicates that the adapter improved significantly on an off-target task while failing to improve (or even degrading) on its nominal target task.

Robustness Checks

The study evaluates performance across:

Multiple random seeds.
Different base models.
Various LoRA hyperparameters (ranks, attention/MLP modules, learning rates).
Additional benchmarks (FollowBench, IFBench) to test benchmark heterogeneity.

3. Key Contributions

The paper makes three restrained, evidence-based contributions:

Deployment-Relevant Diagnostic Framing: Proposes a method to test whether an adapter's nominal label aligns with realized cross-task gains, moving beyond single-task evaluation.
Robustness Evidence: Demonstrates that the "nominal-realized mismatch" is recurrent across seeds and models but configuration-sensitive. It is not a universal law but a frequent occurrence, including cases where the drift is near-zero or negative.
Controlled Counterexample: Provides a concrete case where an instruction-tuned adapter substantially improves off-target numeric reasoning performance while failing to improve (and sometimes degrading) strict, verifiable instruction following (IFEval).

4. Key Results

The "Capability Drift" Phenomenon

The authors observe a pattern they term Capability Drift: the mismatch between what an adapter is labeled to do and what it actually improves.

Strongest Case (Table 1):
- Base Model: NM = 0.133, IFEval PLA = 0.250.
- Instruction-Tuned Adapter:
  - NM (Off-Target): Improved drastically from 0.133 to 0.632.
  - IFEval PLA (Target): Decreased from 0.250 to 0.143.
  - IFEval ILA: Decreased from 0.313 to 0.271.
- Conclusion: The adapter became much better at math but worse at following strict constraints, despite being labeled "instruction-tuned."
Robustness (Table 2 & Figure 2):
- Across 5 seeds on Qwen3-8B, the Drift Score remained positive (mean ~0.511), confirming the mismatch is not a single-seed artifact.
- Across different models (Qwen, Llama) and LoRA settings, the drift was generally positive but showed configuration sensitivity. One specific model/setting yielded a slightly negative drift score (-0.040), proving the phenomenon is not universal but highly dependent on configuration.
Benchmark Heterogeneity:
- Results varied by benchmark. While IFEval and IFBench (strict/verifiable) showed no improvement for the instruction adapter, FollowBench (which operationalizes instruction following differently) showed improvement. This highlights that "instruction following" is not a monolithic capability; agreement across benchmarks cannot be assumed.
Selective Shifts (Figure 3):
- The degradation in IFEval was not uniform. Specific constraint types (e.g., language_constraints, detectable_format, keywords:existence) saw sharp declines, while others improved. This suggests the adapter shifts response priors in specific, localized ways rather than a global "reasoning" shift.

5. Significance and Practical Takeaways

Nominal Labels are Unreliable Proxies: The paper argues that training labels (e.g., "instruction-tuned") should not be treated as reliable indicators of capability. An adapter labeled for one task may inadvertently optimize for a completely different, off-target capability.
Need for Cross-Task Evaluation: Before deployment, practitioners must perform routine cross-task evaluation. Relying solely on the nominal target task evaluation can hide unintended shifts (drift) that degrade performance on critical, verifiable constraints.
Distinction Between "Helpfulness" and "Verifiability": Instruction tuning improves general helpfulness or conversational flow but does not guarantee strict, automatically verifiable compliance. The paper emphasizes that verifiable instruction following is a narrower, distinct capability that requires specific, strict evaluation.
No Simple Mechanistic Explanation: Probing experiments did not find a simple localized explanation (e.g., a single module causing the drift), suggesting the phenomenon arises from complex, distributed interactions within the model weights.

Final Conclusion:
The paper serves as a warning to the LLM community: Do not trust the label. An adapter's nominal objective is a poor predictor of its realized cross-task behavior. Rigorous, multi-task evaluation is essential to detect "capability drift" and ensure that deployed adapters actually possess the verifiable compliance they are claimed to have.