💬 NLP

Information Abundance Paradox: Long-Context Training Undermines Parametric Knowledge

This paper introduces the "Information Abundance Paradox," demonstrating that training large language models with excessively long contexts can undermine their ability to internalize knowledge parametrically, causing them to over-rely on contextual cues and ultimately reducing performance in tasks requiring knowledge retrieval without support.

Original authors: Arda Uzunoglu, Benjamin van Durme, Daniel Khashabi

Published 2026-08-13

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Arda Uzunoglu, Benjamin van Durme, Daniel Khashabi

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 write stories or solve riddles. For a long time, scientists believed the best way to do this was to feed the robot a massive library of books, code, and conversations, hoping that the more it read, the smarter it would become. This is the world of Large Language Models (LLMs), the kind of AI that powers chatbots and writing assistants today. The big idea everyone has been chasing is "context": giving the robot a giant window to look at. Instead of just reading one sentence at a time, the robot can look at a whole page, a whole book, or even a whole conversation history all at once. The assumption was simple: if you give the robot more information to look at while it learns, it will get better at using that information later. It's like thinking that if you study with a bigger textbook, you'll automatically become a better student.

But what if that textbook is too helpful? What if, because the answers are right there in front of your eyes while you're studying, your brain decides it doesn't need to memorize anything? You might get really good at finding the answer in the book, but if someone takes the book away during the test, you might be completely lost. This is the tricky question that a team of researchers at Johns Hopkins University decided to investigate. They wanted to see if giving AI models a "super-sized" window to read during training actually helps them learn better, or if it accidentally teaches them to rely entirely on the information right in front of them instead of building their own internal knowledge.

The Information Abundance Paradox

The researchers discovered something they call the Information Abundance Paradox. It sounds fancy, but the idea is actually quite playful and a bit counterintuitive. They found that when you train an AI with a huge amount of relevant information right in its "context window" (the text it can see while learning), the model stops trying to memorize the facts. Instead, it learns to just look at the text it's given and copy the answer.

Think of it like a student taking a math test.

The Old Way (Parametric Knowledge): The student studies hard, memorizes the formulas, and learns the rules. When they see a problem, they solve it using what's inside their head.
The New Way (Contextualization): The student is allowed to keep their textbook open on the desk while studying. They realize they don't need to memorize the formulas; they can just flip to the page and find the answer instantly. They become experts at using the book.

The paradox is that while this "open-book" student gets perfect scores when the book is there, they fail miserably if you take the book away. The researchers suggest that by flooding the AI with too much helpful context during training, we are accidentally training it to become "context addicted." It stops building a strong internal brain and starts relying entirely on the external text.

The Evidence: When Bigger Isn't Better

To prove this, the team ran a series of experiments that felt like a scientific magic show.

1. The Pretraining Experiment (Learning from Scratch)
They took small AI models and trained them on a massive collection of books (Project Gutenberg). They varied the size of the "window" the models could see at once, ranging from a tiny 512 words to a huge 32,000 words.

The Result: At first, as they made the window bigger, the models got smarter. But then, something weird happened. After the window got to a certain size (around 2,000 to 8,000 words, depending on the task), the models started getting worse.
The Analogy: Imagine a student who starts with a small notepad. They have to memorize everything to do well. Then they get a bigger notepad, and they do even better. But then they get a giant, infinite scroll of paper. Suddenly, they stop memorizing anything. They just look at the scroll. When the test comes and the scroll is gone, they can't remember anything. The "sweet spot" for learning was in the middle, not at the maximum size.

2. The Fine-Tuning Experiment (Learning Specific Tasks)
Next, they took existing AI models and taught them specific subjects like Law, Health, and Economics. They controlled exactly how much "helpful" text the models saw during training.

The Result: When the models were trained with lots of helpful documents, they became amazing at answering questions if those documents were provided again. But if the documents were removed, or if the documents contained misleading information, the models crashed.
The more helpful context they saw during training, the less robust they became. They became like a GPS that works perfectly when you have a signal, but if the signal drops, it doesn't know how to navigate by looking at a map or the sun.

Why Does This Happen? The "Efficient" Brain

The researchers dug deep into the AI's "brain" to see what was happening under the hood. They found two main reasons for this behavior:

1. The Path of Least Resistance
Learning is hard work. It takes energy to memorize a rule and store it in your brain (the model's "weights"). But if the answer is right there in the text, it's much easier to just look it up. The researchers found that when the context was rich and helpful, the AI found a "lower complexity solution." It was like finding a shortcut. The AI realized, "Why bother memorizing the rule for 'negating a bit' when I can just look at the example right here?" So, it stopped memorizing and started looking.

2. The Shift in Brain Power
They also looked at which parts of the AI's brain were doing the work.

FFNs (Feed-Forward Networks): Think of these as the model's "memory banks" where it stores facts and rules.
Attention Modules: Think of these as the model's "eyes" that scan the text to find information.
The Discovery: When the AI was trained with long, helpful contexts, the "pressure" to learn shifted. The updates stopped going into the memory banks (FFNs) and started going entirely into the "eyes" (Attention). The model was literally re-wiring itself to be better at scanning text and worse at storing facts.

The Takeaway

The paper suggests that the race to build "infinite context" models—AI that can read entire libraries in a single glance—might be hitting a hidden wall. It's not just a matter of having more data or bigger computers. If we train these models with too much information right in front of them, we might be creating models that are incredibly good at reading but terrible at thinking on their own.

The researchers aren't saying long-context models are bad. They are just warning us that it's not a simple "more is better" situation. There is a trade-off. If you want a model that can answer questions without a book in hand, you might actually need to train it with less context, forcing it to build its own internal knowledge. If you only care about models that can read a document and summarize it, then the long context is great. But if you want a truly intelligent assistant that can reason even when the information isn't provided, the "Information Abundance Paradox" suggests we need to be careful not to make our AI too dependent on the book.

Technical Summary: Information Abundance Paradox

Problem Statement

Large Language Models (LLMs) are increasingly trained and deployed with context windows spanning millions of tokens, driven by the implicit assumption that exposing models to richer, longer contexts will linearly improve performance by providing more evidence. However, empirical observations suggest that simply scaling context length does not guarantee continued progress; in some cases, long-context variants of models (e.g., Phi-3 128K, OLMo 3 65K) underperform their shorter-context counterparts on standard benchmarks. This raises a fundamental question: Is the trajectory toward near-infinite context merely a matter of supplying more long-context data, or does the context window fundamentally alter the model's learning dynamics?

The paper challenges the view of the context window as a neutral conduit for data. Instead, it posits that abundant relevant information in the training context may shift the model's learning strategy away from parametric internalization (encoding knowledge in weights) toward contextualization (relying on evidence supplied in the input). This shift can lead to "context addiction," where models perform well when context is present but degrade significantly when it is absent or misleading.

Methodology

The authors investigate this hypothesis through a combination of empirical experiments, theoretical analysis, and mechanistic probing across two training regimes: pretraining and supervised fine-tuning (SFT).

1. Empirical Experiments

Pretraining Study: The authors pretrain Llama-2 based models at four scales (20M to 750M parameters) on a fixed token budget (10B tokens) from Project Gutenberg. They systematically vary the training context window ( $W \in \{512, \dots, 32768\}$ $W \in {512, \dots, 32768}$ ) while keeping the total number of tokens and optimization steps constant.
- Evaluation: Models are evaluated on language modeling (LAMBADA, WikiSPAN, PTB), general language understanding (SuperGLUE), and closed-book multiple-choice question answering (MCQA) to measure parametric knowledge.
Supervised Fine-Tuning (SFT) Study: Using Qwen3 models (0.6B to 14B), the authors fine-tune on MMLU-Pro domains (Health, Economics, Law, Psychology). They fix the context budget (8 documents) but vary the informativeness of the context ( $k \in \{0, 4, 8\}$ $k \in {0, 4, 8}$ target-domain documents).
- Evaluation: Models are tested under three conditions: supporting context (correct evidence), conflicting context (misleading evidence), and no context.
Synthetic Pretraining: To isolate the mechanism, the authors train models on synthetic tasks (bitwise operations, string transformations, mod10 arithmetic, Caesar cipher) with varying numbers of in-context demonstrations. They measure the "supporting-conflicting gap" to quantify context reliance and use average training gradient norms as a proxy for solution complexity.

2. Theoretical Analysis

The authors formalize the tradeoff using information theory. They define the Parametric Information Frontier ( $I_k(\rho)$ ) as the minimum task information that must be stored in the weights to achieve a risk threshold $\rho$ given a context size $k$ . They prove that for nested inputs, $I_m(\rho) \leq I_k(\rho)$ for $m > k$ , demonstrating that longer contexts theoretically reduce the minimum information required in the weights to achieve the same performance.

3. Mechanistic Analysis

Gradient Allocation: The authors analyze the ratio of gradient norms between Feed-Forward Networks (FFNs) and Self-Attention (SA) modules. FFNs are associated with parametric knowledge, while SA is associated with contextual routing.
Causal Interventions: They perform module-restricted fine-tuning, updating only FFNs or only SA heads, to observe causal effects on robustness and context reliance.
Token-Level Attention: They measure the attention mass assigned to context tokens during inference to quantify the shift in computation.

Key Results

1. The Inverted-U Performance Curve

In pretraining, increasing the context window improves performance up to an intermediate optimum (e.g., ~2048 tokens for SuperGLUE/MCQA, ~8192 for language modeling) before performance consistently declines. This degradation occurs despite fixed token budgets and model capacities, indicating that longer contexts are not inherently beneficial beyond a certain point.

2. Context Addiction in SFT

In supervised fine-tuning, increasing the amount of task-relevant context during training ( $k=8$ ) significantly improves accuracy when supporting context is available at test time. However, this comes at the cost of reduced robustness:

No-Context Performance: Models trained with high-context relevance perform significantly worse when context is absent compared to models trained with low-context relevance.
Conflicting Context: These models are more vulnerable to misleading context, showing a larger drop in accuracy when provided with conflicting evidence.

3. Mechanistic Shifts

Solution Complexity: In synthetic tasks, "context addiction" (large supporting-conflicting gaps) emerges specifically in tasks where longer contexts enable lower complexity solutions (indicated by decreasing average training gradient norms). Tasks that do not benefit from simpler contextual solutions (e.g., mod10 arithmetic) do not exhibit this shift.
Gradient Reallocation: Training with informative context shifts gradient pressure away from FFNs (parametric internalization) and toward Self-Attention modules (contextualization). This shift is observed across both pretraining and SFT.
Inference Behavior: Models trained with rich context allocate significantly more attention mass to context tokens during inference, particularly in middle layers, confirming a causal reliance on supplied information.

Key Contributions

The Information Abundance Paradox: The paper introduces the hypothesis that abundant relevant information in training contexts can reduce the incentive for models to encode that information parametrically, leading to a shift from internalization to contextualization.
Empirical Evidence of Non-Monotonic Scaling: The study provides robust evidence that scaling context windows is not a free lunch; beyond an intermediate optimum, longer training contexts can degrade closed-book performance and robustness.
Mechanistic Explanation: The work identifies the specific neural mechanisms driving this phenomenon: a shift in optimization pressure from FFNs to Attention modules and the adoption of lower-complexity solutions that rely on context rather than weights.
Theoretical Formalization: The paper offers a theoretical proof that longer contexts reduce the minimum parametric information required to achieve a target risk, explaining why models may "forget" knowledge in favor of context reliance.

Significance and Claims

The paper argues that the drive toward near-infinite context windows cannot be understood solely as a data problem. Instead, the context window acts as a hyperparameter that governs the mode of learning.

Challenge to Current Paradigms: The findings challenge the assumption that "more long-context data" is always the path forward. The authors suggest that scaling of context length may trade reusable parametric knowledge for context-dependent competence.
Implications for Robustness: Models trained with highly informative long contexts may become "context addicts," failing when the context is missing or corrupted. This has critical implications for deployment in scenarios where context availability is not guaranteed.
Design Guidance: The paper suggests that optimal training strategies must balance context length to avoid shifting the model's learning mode too far toward contextualization, potentially requiring curricula or constraints that preserve parametric knowledge.

The authors conclude that while long-context processing is a powerful capability, it fundamentally alters what models internalize and how strongly they depend on inference-time information, necessitating a re-evaluation of scaling laws for context windows.