💬 NLP

TELLME: Test-Enhanced Learning for Language Model Enrichment

This paper introduces TELLME, a novel method that integrates Test-Enhanced Learning with continual pre-training to efficiently acquire domain-specific knowledge and improve long-term memory retention in large language models while overcoming the data and computational challenges of traditional approaches.

Original authors: Minjun Kim, Inho Won, Hyeonseok Lim, MinKyu Kim, Junghun Yuk, Wooyoung Go, Jongyoul Park, Jungyeul Park, KyungTae Lim

Published 2026-08-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Minjun Kim, Inho Won, Hyeonseok Lim, MinKyu Kim, Junghun Yuk, Wooyoung Go, Jongyoul Park, Jungyeul Park, KyungTae Lim

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

Imagine you are trying to teach a super-smart robot how to be an expert in a specific field, like finance or medicine. You can't just give it a library of textbooks and hope it memorizes everything; that's too slow and expensive. This is the world of Large Language Models (LLMs), the AI brains behind many modern chatbots. Usually, to make these robots smarter about a new topic, scientists use a method called Continual Pre-training (CPT). Think of CPT like feeding the robot a massive, endless stream of articles and news stories about that topic. The robot reads them, learns the patterns, and gets better.

However, there's a catch. Just reading a book doesn't always mean you truly understand it or remember it next week. In human psychology, there's a famous idea called Test-Enhanced Learning (TEL). It's the counter-intuitive discovery that taking a quiz while you are studying actually helps your brain lock the information in better than just re-reading the material. It's the difference between passively watching a cooking show and actually trying to cook the dish while the chef asks you, "What happens if you add salt too early?" The paper we are exploring today asks a simple but powerful question: What if we taught AI the same way we teach humans? What if, instead of just feeding it more text, we made it take quizzes while it learns?

The Paper: TELLME – Teaching AI to Take Quizzes

The researchers behind this study, a team from Korea, have built a new method they call TELLME (Test-Enhanced Learning for Language Model Enrichment). Their goal was to fix the problems of the traditional "just read more" approach. They wanted to see if mixing in "quizzes" during the training process would help AI learn faster, remember better, and actually understand the deep concepts of a subject rather than just memorizing words.

The Problem with "Just Reading"

Traditionally, when scientists want to teach an AI about finance, they feed it thousands of news articles. This is the Continual Pre-training (CPT) method. Sometimes, they follow this up with a separate phase called Instruction Tuning (IT), where they show the AI questions and answers to teach it how to chat.

The authors argue this two-step process is clunky. It's like studying for a history exam by reading the textbook for three hours, and then, only after you're done, being handed a practice test. By the time you take the test, you might have already forgotten the details you just read. Other researchers tried to mix the two steps, feeding the AI text and simple questions at the same time. But the authors noticed a flaw: those questions were often too easy, like "What is the name of the bank mentioned in this paragraph?" The AI could just copy the answer directly from the text without actually thinking. It was a "reading comprehension" trick, not real learning.

The TELLME Solution: The "Deep Dive" Quiz

TELLME changes the game by introducing a specific type of quiz based on the psychological principles of TEL. Here is how it works:

The Setup: The AI is given a piece of text (like a news article about a bank).
The Twist: Instead of asking, "What did the bank do?", the AI is asked to generate or answer questions that require outside knowledge. For example, if the text mentions "high-frequency trading," the quiz might ask, "How does high-frequency trading impact market volatility?" The answer isn't written in the article. The AI has to use what it already knows about the world to connect the dots.
The Training Loop: The AI reads the text, then immediately tries to answer these deep, open-ended questions. Crucially, the system only "grades" the AI on the text it read and the answers it gave, not on the questions themselves. This forces the AI to focus on understanding the material and retrieving its own internal knowledge to solve the puzzle.

To build this, the team used a powerful AI (GPT-4o-mini) to create 100,000 of these special quiz pairs for two difficult fields: Finance and Medicine. They made sure the questions were diverse and required reasoning, not just copying.

What They Found: Faster Learning, Better Memory

The team tested TELLME against the old methods using several different AI models (including Llama and SmolLM). The results were quite promising:

Smarter in Less Time: In the finance domain, TELLME improved the AI's performance by up to 23.6% compared to the standard methods. It learned the material more efficiently.
Better Long-Term Memory: This was the most exciting part. The researchers tested if the AI would "forget" what it learned about finance when they started teaching it about medicine.
- The old method (CPT) saw a 5.72% drop in finance performance after learning medicine.
- The TELLME method only dropped by 0.94%.
- This suggests that by "testing" itself during training, the AI built stronger, more durable memories that didn't wash away when new information arrived.
Cost Efficiency: Because TELLME learns faster, it reached the same level of performance in fewer steps. The authors noted it was about 1.4 times faster to train than the standard method, saving time and computing power.

Why It Matters

The paper suggests that the way we train AI has been too passive. By borrowing a trick from human education—testing yourself while you learn—we can build AI that understands concepts deeply and remembers them longer.

The researchers also checked if this worked for other languages. They tried it with Korean, and it worked there too, improving the AI's ability to understand Korean sentences even if it hadn't been trained on Korean text before. This hints that the "quiz" method might be a universal key to making AI smarter, regardless of the language.

However, the authors are careful to note that this is a specific improvement for continual learning (teaching AI new things after it's already been built). They also point out that they tested this on models up to 70 billion parameters, but using some shortcuts to save money, so the results might look slightly different on massive, full-scale models.

In short, TELLME suggests that if you want an AI to be a true expert, don't just make it read more. Make it take a test.

Technical Summary: TELLME (Test-Enhanced Learning for Language Model Enrichment)

Problem Statement

Continual Pre-training (CPT) is a standard method for adapting Large Language Models (LLMs) to specific domains. However, CPT faces two significant challenges: the difficulty of acquiring large-scale, domain-specific datasets and the high computational costs associated with training. While recent approaches have attempted to integrate Instruction Tuning (IT) data directly into the CPT process (e.g., INSTPT), these methods often rely on reading-comprehension style Question-Answering (QA) where answers are extracted directly from the provided text. The authors argue that such formats fail to leverage the Test-Enhanced Learning (TEL) principle effectively. Educational psychology research suggests that TEL improves long-term memory retention most when testing involves open-ended, descriptive responses that require the learner to recall and synthesize internal knowledge, rather than simple recall or multiple-choice formats constrained by the context.

Methodology

The authors propose TELLME, a framework that integrates the TEL principle into CPT to enhance both the efficiency of knowledge acquisition and the durability of learned representations.

1. Dataset Construction

TELLME utilizes a cost-efficient pipeline to generate 100,000 domain-specific samples (covering financial and medical domains) using GPT-4o-mini.

Prompt Design: The generation prompt instructs the synthesizer to create QA pairs based on general domain knowledge related to a plain text excerpt, but explicitly forbids questions that can be answered solely by extracting information from the text.
Goal: This forces the model to establish "concept mappings" between the text and broader domain knowledge, requiring the model to retrieve internal knowledge to answer.
Quality Control: The generated dataset achieved an average human-evaluation score of 4.03/5. Samples requiring context (e.g., containing "in this context") were filtered out to ensure the QA pairs are solvable independently of the text.

2. Training Framework

TELLME modifies the standard Causal Language Modeling (CLM) objective during the continual pre-training phase.

Data Structure: Each training sample $X$ consists of a sequence: Plain Text ( $t$ ), followed by Question ( $q$ ) and Answer ( $a$ ).
Loss Function Modification: To simulate the "testing" effect, the loss function is configured to exclude the question tokens from the calculation.
- The indicator function is set such that $1(x_i \in t \cup a) = 1$ and $1(x_i \in q) = 0$ .
- The model is trained to predict the plain text and the answer, but the question acts as a "test" that the model must answer using its internal knowledge, rather than a prompt to be predicted.
Comparison: This differs from methods like INSTPT, which often compute loss over all tokens (including questions) or treat plain text and QA as separate sequential steps (CPT then IT). TELLME processes mixed samples in a single pass.

Key Contributions

TELLME Framework: Introduction of a CPT framework that leverages descriptive, open-ended QA to stimulate intrinsic knowledge recall, aligning LLM training with the psychological principles of Test-Enhanced Learning.
Efficient Data Pipeline: A cost-effective method (approx. $12 for 100k samples) to generate diverse, domain-specific QA data that requires reasoning beyond the provided text.
Empirical Validation: Comprehensive evaluation across financial and medical domains using various model scales (Llama-3.2-1B/3B/8B, SmolLM2-1.7B) demonstrating significant improvements over standard CPT and CPT+IT baselines.

Experimental Results

Domain Adaptation Performance

TELLME was evaluated on financial benchmarks (FOMC, NIFTY, MMLU-F) and medical benchmarks (HeadQA, MedMCQA, MMLU-C).

Financial Domain: TELLME outperformed the CPT baseline by up to 23.6% and surpassed the INSTPT method by an average of 5.1% across both finance and medical domains.
Medical Domain: TELLME showed consistent improvements, outperforming the baseline model by an average of 2.58 points.
Efficiency: TELLME achieved the same perplexity (PPL) as standard CPT with 1.4x faster training steps, indicating higher learning efficiency.

Long-Term Retention

To test memory retention, models trained on finance data were subsequently trained on medical data (overwriting).

Retention Metrics: When evaluated on the original finance domain after medical training, the TELLME-based model showed a 0.94% performance drop, whereas the standard CPT model suffered a 5.72% drop.
Perplexity: The TELLME model maintained lower perplexity on the finance corpus compared to the CPT model, suggesting better preservation of prior domain knowledge.

Ablation Studies

Ordering: Placing QA before text (Inv-TEL) resulted in lower performance, confirming the importance of the specific ordering (Text $\to$ QA) in TELLME.
Loss Masking: Including the question in the loss calculation (TEL-Q/L) reduced performance, validating that excluding the question is crucial for the testing effect.
Synthesizer Robustness: TELLME remained effective even when using weaker synthesizers (Mistral-7B) or self-generated data, though GPT-4o-mini yielded the best results.
Multilingual Generalization: The framework was successfully adapted to Korean (TELLME-KO), improving performance on the KoBEST benchmark by 4.5% in the 0-shot setting and 8.4% in the 5-shot setting for models that previously lacked Korean capabilities.

Significance and Claims

The paper claims that TELLME offers a superior alternative to conventional CPT and sequential CPT+IT approaches by addressing the limitations of data scarcity and computational cost. By integrating the TEL principle, the method not only accelerates domain adaptation but also significantly enhances long-term memory retention, preventing catastrophic forgetting when models are exposed to new domains. The authors position TELLME as a practical, scalable solution for developing specialized LLMs that retain their knowledge over time.

The study acknowledges limitations, noting that experiments on larger models (70B parameters) utilized parameter-efficient fine-tuning (LoRA) and quantization, which may introduce variables not present in fully dense training. Additionally, the scope is currently limited to finance and medicine, with future work needed to expand to other domains and establish standardized benchmarks.