CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training

This paper introduces CapTrack, a capability-centric framework that redefines LLM forgetting as systematic behavioral drift rather than mere knowledge loss, revealing through a large-scale study that post-training significantly degrades robustness and default behaviors, with instruction fine-tuning causing the most pronounced effects.

The Gist

Imagine you have a brilliant, well-read librarian named LLM (Large Language Model). Before you hire them for a specific job, they have already read almost everything in the world. They know facts, can write code, speak many languages, and tell jokes. They are your "Out-of-the-Box" (OOB) librarian.

Now, you want to hire this librarian to work specifically in a Law Firm or a Hospital. To do this, you give them a crash course (called "Post-Training") to teach them the specific jargon and rules of that field.

The Problem:
Usually, when we check if the training worked, we ask: "Do they know the new legal cases?" or "Can they diagnose this disease?" If they say "Yes," we think the training was a success.

But this paper argues that this is like checking if a chef learned to make sushi by only asking, "Can you cut the fish?" You might forget to ask: "Did they forget how to bake a cake?" or "Did they suddenly become rude to customers?" or "Do they now refuse to answer simple questions?"

The authors call this "Forgetting." And they argue that in the past, we only measured forgetting as "losing facts." But in reality, the librarian might lose their personality, their politeness, or their ability to follow complex instructions, even if they still know the facts.

Enter: CapTrack (The "Capability Tracker")

To fix this, the authors built a new tool called CapTrack. Think of CapTrack as a 360-degree performance review for the librarian, divided into three categories:

1. CAN (What they can do): This is their raw brainpower. Can they still solve math problems? Do they know history? Can they write code?
   • Analogy: Can the librarian still read a complex book?
2. WILL (What they want to do): This is their default attitude. Are they helpful? Do they talk too much or too little? Do they refuse to answer safe questions?
   • Analogy: Is the librarian now grumpy? Do they give short, rude answers? Do they refuse to help with simple tasks?
3. HOW (How they do it): This is their ability to follow rules. If you ask for a list in bullet points, do they do it? If you ask for a citation, do they format it correctly?
   • Analogy: If you ask the librarian to "write the report in blue ink," do they actually use blue ink, or do they ignore you?

What They Discovered

The team took several famous librarians (models like LLaMA, Gemma, and Qwen) and gave them crash courses in Law and Medicine. Then, they used CapTrack to see what changed. Here are the big findings:

1. The "Specialist" Curse
When they trained the librarians to be lawyers, the librarians got really good at law. But they started forgetting other things.

• The Surprise: They didn't just forget facts. They forgot how to be robust. If you asked a question in a slightly different way, they got confused. They also forgot how to speak other languages well.
• The Metaphor: It's like a chef who learns to make perfect sushi but forgets how to boil water or how to be polite to a customer. They are a "specialist" but a "bad employee."

2. The Training Method Matters
They tried two different ways to train the librarians:

• Method A (Instruction Fine-Tuning / IFT): This is like a strict teacher shouting, "Do exactly what I say! Repeat after me!"
  • Result: The librarian learned the job fast, but they became rude, short, and stubborn. They forgot how to be helpful and started refusing to answer questions they used to answer easily.
• Method B (Preference Optimization / DPO): This is like a coach saying, "Here are two answers; pick the better one."
  • Result: This was much gentler. The librarian learned the job but kept their personality and politeness. In fact, if you used Method A first and then Method B, the librarian actually recovered some of their lost kindness and helpfulness.

3. Bigger Isn't Always Better
You might think a bigger librarian (a model with more "brain power") would forget less.

• Result: Nope. A giant 80-billion-parameter model forgot just as much as a smaller one. The size of the model didn't protect them from losing their "soft skills."

4. The "No Free Lunch" Rule
The authors tried to fix the forgetting by:

• Mixing in more general data (like reading a newspaper while studying law).
• Using architectural tricks (like "merging" the new model with the old one).
• Result: It's a trade-off. You can keep the librarian's old personality (Stability), but then they won't learn the new job as well (Plasticity). Or, you can have them learn the new job perfectly, but they lose their old personality. You can't have both perfectly.

The Big Takeaway

The paper tells us to stop just asking, "Is the model smart?" and start asking, "Is the model still a good, reliable, polite, and rule-following assistant?"

If you train an AI to be a doctor, you don't just want it to know medicine; you want it to still be able to listen, follow instructions, and not suddenly decide to refuse to talk to you. CapTrack is the new checklist to make sure we aren't trading our AI's soul for a little bit of extra knowledge.

Technical Summary

Here is a detailed technical summary of the paper "CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training."

1. Problem Statement

Large Language Models (LLMs) are routinely adapted to specific domains or value alignments through post-training (e.g., Instruction Fine-Tuning (IFT) and Preference Optimization like DPO). While effective for domain adaptation, this process is known to induce catastrophic forgetting.

• Current Limitation: Existing research predominantly defines forgetting as the loss of parametric knowledge (factual accuracy) or task performance on standard benchmarks.
• The Gap: This accuracy-centric view is insufficient for modern foundation models. Post-training can cause systematic model drift in behaviors, robustness, and interaction protocols that are not captured by factual recall but critically degrade user experience. For example, a model might retain legal knowledge but become less robust to multilingual inputs, refuse safe queries more often, or fail to follow formatting constraints.
• Goal: The authors argue for a broader definition of forgetting encompassing any systematic drift that negatively impacts capabilities, behaviors, or execution reliability, and propose a framework to measure it.

2. Methodology: CapTrack Framework

The authors introduce CapTrack, a capability-centric framework designed to track and quantify model drift beyond parametric knowledge. It consists of two main components:

A. The CapTrack Taxonomy

The framework decomposes LLM behavior into three complementary capability groups:

1. CAN (Latent Competence): What the model can do under ideal prompting.
   • Sub-dimensions: Parametric knowledge (C1), Reasoning (C2), Contextual comprehension (C3), Epistemic faithfulness (C4), and Robustness (C5: prompt invariance, domain shift, multilingual stability).
2. WILL (Default Behavioral Preferences): What the model will do by default (shaped by alignment/tuning).
   • Sub-dimensions: Willingness to answer (refusal/compliance), Helpfulness/Informational scope, and Style/Elaboration (verbosity, formatting).
3. HOW (Protocol Compliance & Execution): How reliably the model executes instructions.
   • Sub-dimensions: Instruction following, Output-format fidelity, Tool use, Multi-turn state commitment, Context-window operations, and Citation mechanics.

B. The Evaluation Suite

• Benchmarks: CapTrack utilizes a curated set of 30 established benchmarks (e.g., MMLU-Pro, GSM8K, MT-Bench, BFCL, RULER) and targeted adaptations (e.g., rephrased prompts, schema-wrapped outputs).
• Metrics: It employs a mix of rule-based metrics (accuracy, parse rates) and LLM-as-a-judge evaluations for qualitative aspects (reasoning quality, hallucination, style).
• Relative Evaluation: Instead of absolute scores, CapTrack measures relative deviation from the Out-of-the-Box (OOB) base model to isolate drift caused specifically by post-training.
• Scale: The study covers 7 models across three families (Qwen, LLaMA, Gemma) ranging from 4B to 80B parameters, tested in Legal and Medical domains using IFT and DPO algorithms.

3. Key Contributions

1. Redefining Forgetting: Extends the definition of forgetting from mere factual loss to include capability-level drift in behavior, robustness, and execution.
2. CapTrack Framework: Introduces a structured taxonomy (CAN, WILL, HOW) and a practical evaluation suite that enables multifaceted analysis of post-training effects.
3. Large-Scale Empirical Study: Provides the first comprehensive analysis of forgetting across diverse model families, scales, and post-training algorithms, revealing that drift patterns are highly heterogeneous.
4. Mitigation Analysis: Systematically evaluates data-centric, architectural, and regularization-based mitigation strategies, exposing fundamental trade-offs.

4. Key Results

A. Forgetting Extends Beyond Knowledge

• Robustness & Behavior are Most Affected: While parametric knowledge (C1) is relatively preserved, robustness (especially multilingual stability, C5c) and default behavioral preferences (WILL) suffer the most severe degradation.
• IFT vs. DPO:
  • Instruction Fine-Tuning (IFT) induces the strongest relative drift, causing catastrophic drops in verbosity, formatting, and multilingual robustness (up to -29% in some cases).
  • Preference Optimization (DPO) is more conservative. When applied after IFT, it can partially recover lost capabilities (particularly in the WILL group), though it does not fully reverse IFT-induced damage.

B. Model Family & Size Dependence

• Family Differences: Drift patterns vary significantly by model family.
  • LLaMA and Gemma exhibit strong drift in latent competence and behavioral preferences.
  • Qwen is generally more stable in competence but remains vulnerable in specific areas like multilingual robustness and instruction following.
• Model Size: There is no consistent dependence on model size. Larger models (80B) are not uniformly more stable than smaller ones (4B). The only clear size-dependent trend found was in citation mechanics (H6), where smaller models degraded more significantly.

C. Mitigation Strategies & The Stability-Plasticity Trade-off

The authors tested three mitigation axes:

1. Data-Centric: Switching to general data (Tulu 3) or adding replay data did not consistently reduce forgetting. Effects were mixed and capability-dependent.
2. Architectural (Model Merging): Merging post-trained models with OOB models (e.g., via TIES) reduces forgetting but linearly sacrifices in-domain performance (plasticity).
3. Regularization (LoRA): Using Low-Rank Adaptation acts as a global regularizer. Lower ranks improve stability but reduce adaptation gains.

• Conclusion: All mitigation strategies reveal a fundamental stability-plasticity trade-off. There is no "free lunch" strategy that eliminates forgetting while preserving full adaptation gains.

5. Significance

• Paradigm Shift: The paper challenges the community to move beyond "accuracy-only" evaluations for post-trained models. It demonstrates that a model can be "factually accurate" but behaviorally broken (e.g., refusing safe queries, failing to follow formatting).
• Practical Tool: CapTrack provides a diagnostic tool for practitioners to detect subtle but critical failures in deployment-ready models, such as loss of multilingual robustness or degradation in multi-turn consistency.
• Future Directions: The findings suggest that future mitigation strategies must be capability-aware and targeted (e.g., protecting specific behavioral dimensions) rather than relying on global regularization or generic data mixing. It highlights that forgetting is an inherent, multifaceted phenomenon of the post-training process itself, not just a failure of domain adaptation.

In summary, CapTrack reveals that post-training fundamentally alters the behavioral landscape of LLMs, often degrading critical real-world capabilities even when factual knowledge remains intact. Understanding and mitigating this requires a holistic, capability-centric evaluation approach.