🤖 AI

Lost in Context: Addressing Context Anxiety in Large Language Models

This paper identifies and systematically studies "context anxiety" in large language models, revealing that reasoning failures often stem from premature self-doubt and inaccurate token estimation rather than a lack of capability, and demonstrates that performance can be improved by teaching models to better assess their own limitations.

Original authors: Ifueko Igbinedion, Jillian Ross, Etienne Ricardez, Sertac Karaman, Eric So

Published 2026-07-27

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Ifueko Igbinedion, Jillian Ross, Etienne Ricardez, Sertac Karaman, Eric So

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 a brilliant detective who can solve the most complex mysteries in the world. You have a super-powered notebook that can hold millions of pages of clues. But here's the twist: sometimes, even when the mystery is solvable and your notebook has plenty of space, you suddenly stop writing. You put down your pen, shake your head, and say, "This is too big for my notebook. I can't finish this." You didn't actually run out of room; you just thought you did. This is a bit like what happens inside the most advanced computer brains we have today, known as Large Language Models (LLMs). These are AI systems trained on vast amounts of text that can reason through math problems, write code, and plan complex strategies. For a long time, scientists thought these AIs failed only when a problem was simply too hard for them to understand. But a new study suggests that sometimes, the AI isn't failing because it's too dumb; it's failing because it's too scared of its own limits.

This paper, titled "Lost in Context," dives into a strange new problem called "context anxiety." It turns out that these AI models often panic about running out of space in their "notebooks" (which are technically called token limits) before they even try to solve the hardest parts of a puzzle. The researchers wanted to know: Are these models giving up because the task is truly impossible, or are they just having a panic attack about how much work it might take? By studying how these models behave when faced with long, tricky tasks, the team discovered that many of them are terrible at guessing how much effort a problem actually needs. They overestimate the work, get scared, and quit early. The good news? The researchers found a way to "counsel" the AI, teaching it to trust its abilities and stop panicking, which made it much better at solving problems without needing to make the AI bigger or smarter.

The Panic Attack in the Machine

To understand what's happening, let's imagine the AI is a marathon runner. In the past, we thought that if a runner stopped, it was because the race was too long or their legs were too tired. But this study found that sometimes, the runner stops because they look at the finish line and think, "Oh no, that's 100 miles away! I'll never make it!" even though the finish line is actually only 5 miles away. The runner has the energy to finish, but their internal map is wrong, so they quit prematurely.

The researchers tested this idea using a classic puzzle called the Tower of Hanoi. Imagine three pegs and a stack of disks of different sizes. The goal is to move the entire stack from one peg to another, following strict rules: you can only move one disk at a time, and you can never put a bigger disk on top of a smaller one. The tricky part is that as you add more disks, the number of moves needed grows explosively. With just 12 disks, you need over 4,000 moves. The researchers asked several top-tier AI models to solve these puzzles.

Here is the shocking discovery: Many of the smartest models, like Claude Sonnet 3.7 and DeepSeek R1, refused to solve the harder puzzles. But when the researchers looked closely at why they refused, they found the models weren't saying, "I don't know the answer." Instead, they were saying things like, "This is too long to write out," or "I will run out of space." They were quitting because they thought the solution would be too big for their memory, even though the puzzle was well within their actual limits. The researchers call this context anxiety. It's a form of premature self-doubt where the AI believes it can't finish the job, so it doesn't even try.

The Bad Math of AI

Why does the AI get so scared? The paper suggests it's because the models are bad at math—specifically, bad at guessing how much "space" (or tokens) a solution will take. The researchers asked the models to guess how many words they would need to write to solve a problem. The results were wild.

Models that showed anxiety overestimated the space they needed by about 24%. They thought a 100-word answer would take 124 words of space. Because they thought the task was so huge, they panicked and gave up. On the flip side, models that didn't have anxiety actually underestimated the space by about 19%. They were more optimistic, thinking, "Hey, this is easy!" and they actually tried to solve it.

This miscalibration creates a self-fulfilling prophecy. If the AI thinks a task is impossible because it's "too big," it stops. If it thinks a task is manageable, it keeps going. The study found that this anxiety wasn't just a small glitch; it had real consequences. When models were anxious, they were 15% less accurate at solving problems. Even worse, when they did manage to solve a problem correctly, their answers were 54% longer and less efficient. It's like a nervous student who, instead of writing a clear, short essay, rambles on and on, wasting time and paper because they are worried they aren't doing enough.

Teaching the AI to Calm Down

The most exciting part of the paper is that the researchers didn't just find a problem; they found a cure. They asked: Is this anxiety a permanent part of the AI's brain, or can we teach it to be less nervous?

To test this, they used a technique called Supervised Fine-Tuning (SFT). Imagine you have a student who is afraid of math. Instead of giving them harder math problems, you show them examples of other students who solved similar problems calmly and confidently. You tell them, "Look, you can do this without panicking."

The researchers took a dataset of correct solutions to the Tower of Hanoi puzzle but filtered out any that showed signs of anxiety. They then trained a new AI model only on these calm, confident examples. They didn't teach the model any new facts or make it bigger; they just taught it a new way of thinking.

The result? The "cured" model was amazing. It stopped quitting early. It learned to trust that it could finish the job.

The new model reduced its anxiety by over 50%.
It became much better at solving the puzzles, especially the medium-difficulty ones.
It even transferred this new confidence to a completely different type of puzzle (finding the shortest path on a grid), showing it wasn't just memorizing the answers but learning a better strategy.

What This Means for the Future

This paper suggests that we might be looking at AI performance the wrong way. We often think that to make AI better, we just need to make it bigger, give it more data, or make it smarter. But this study suggests that sometimes, the AI is already smart enough; it just needs to learn how to manage its own confidence.

The researchers found that "context anxiety" is a behavioral habit, not a hard limit of the machine's brain. By helping models accurately assess their own abilities and stop overestimating how hard a task is, we can make them more reliable and efficient without needing to build massive new supercomputers. It's a reminder that even for artificial intelligence, sometimes the biggest obstacle isn't the problem itself—it's the fear of the problem.

In the end, the paper shows that with a little bit of "therapy" (or in this case, careful training on calm examples), AI models can learn to stop panicking, trust their skills, and solve problems they were previously too scared to touch. It's a small but powerful step toward building AI that doesn't just know how to think, but also knows how to keep going when things get tough.

Technical Summary: Lost in Context: Addressing Context Anxiety in Large Language Models

Problem Definition

The paper identifies a phenomenon termed context anxiety in frontier large language models (LLMs). Contrary to the conventional wisdom that reasoning models fail on complex problems due to insufficient capability, the authors observe that models often possess the necessary reasoning capacity but fail due to premature self-doubt. Specifically, models abandon tasks they could theoretically solve because they incorrectly estimate that the solution will exceed their available token limits.

This anxiety manifests in two primary ways:

Premature Disengagement: Models explicitly state that a task is "too complex" or "beyond my context window" and refuse to attempt a solution, even when the problem fits well within their actual token budgets.
Efficiency Loss: When models do attempt a solution while anxious, they produce less concise outputs, leading to material efficiency losses (higher token usage) even for successful completions.

The core mechanism driving this behavior is identified as miscalibrated token estimation. Models systematically overestimate the number of tokens required to complete a task, creating a self-fulfilling prophecy where the perceived constraint triggers actual failure.

Methodology

The authors propose a systematic framework to detect, measure, and mitigate context anxiety, validated primarily on the Tower of Hanoi puzzle and secondarily on a shortest-path grid search task.

1. Detection Protocol

To distinguish between capability-driven failures (genuine inability) and anxiety-driven failures (premature abandonment), the authors employ a panel of LLMs to classify reasoning traces. The classifiers look for semantic equivalents of phrases indicating resource insufficiency (e.g., "exceeds my capacity," "too long to write out").

Anxiety-driven failures: The model abandons the task citing resource limits.
Capability-driven failures: The model attempts the task but produces an incorrect solution.
Successful attempts: The model completes the task.

2. Calibration Measurement

The authors introduce the winsorized token ratio ( $T$ ) to quantify calibration:
$T = \frac{\text{Estimated tokens required}}{\text{Actual tokens used}}$
Models are prompted to estimate the token count of their own generated solutions post-hoc. A ratio $>1.0$ indicates overestimation (miscalibration). The study correlates this overestimation with the presence of context anxiety.

3. Behavioral Adaptation (Mitigation)

To test if context anxiety is a modifiable behavioral policy rather than a fixed capacity limit, the authors perform lightweight Supervised Fine-Tuning (SFT).

Data Filtering: They generate a dataset of reasoning traces from frontier models. Crucially, they filter out any traces that exhibit detected context anxiety, retaining only successful, anxiety-free reasoning paths.
Training Objective: The student model is fine-tuned exclusively on the reasoning traces (masking the final answer tokens) of the filtered dataset. This forces the model to adopt non-anxious reasoning strategies without learning new task-specific knowledge or increasing model size.

Key Results

1. Prevalence and Impact of Context Anxiety

Evaluated on 1,620 observations across 30 unique Tower of Hanoi problems (2 to 12 disks) using models like Claude Sonnet 3.7, DeepSeek R1, and Gemini 2.5 Flash:

High Incidence: Frontier models exhibit significant context anxiety. For instance, 73.4% of failures in Claude Sonnet 3.7 and 66.3% in DeepSeek R1 were anxiety-driven, whereas models like Gemini 2.5 Flash showed only 5.1%.
Correlation with Complexity: Anxiety rates increase with task complexity (disk count).
Miscalibration: Models exhibiting context anxiety overestimate their token requirements by approximately 24%, whereas non-anxious models underestimate by 19%. This difference is statistically significant ( $p < 0.01$ ).

2. Performance and Efficiency Degradation

Econometric analysis controlling for model and task fixed effects reveals:

Accuracy Drop: Context anxiety reduces task accuracy by 15.3% ( $p < 0.01$ ), independent of the inherent difficulty of the problem.
Token Inefficiency: Correct solutions generated with context anxiety are 54% longer (requiring ~2,323 more tokens on average) than those generated without anxiety. The anxiety leads to verbose, less streamlined reasoning.

3. Effectiveness of Mitigation

The SFT approach using anxiety-filtered data demonstrated that context anxiety is a mutable behavior:

Anxiety Reduction: The fine-tuned model reduced context anxiety rates by over 50% compared to the baseline.
Accuracy Gains: The adapted model matched or outperformed the baseline and an "all-correct" control (trained on all successful traces without anxiety filtering). Notably, on mid-tier difficulties and out-of-domain transfer tasks (Shortest Path), the anxiety-filtered model showed significant gains (e.g., +26.7% absolute gain on 5x5 grids).
Generalization: The behavioral adaptation transferred to unseen disk counts (odd disks when trained on even) and a different task family (Shortest Path), suggesting the model learned a general policy for managing effort rather than memorizing specific solutions.

Significance and Claims

The paper claims to provide the first systematic study of context anxiety, establishing it as a distinct failure mode driven by miscalibrated self-assessment rather than a lack of reasoning capability.

The authors argue that:

Performance Limits are Often Perceived, Not Real: Models frequently fail not because problems are too hard, but because they believe they are too hard due to poor token estimation.
Behavioral Mutability: Context anxiety is not a fixed architectural limitation. It can be significantly reduced through targeted behavioral fine-tuning on anxiety-free reasoning traces.
Alternative to Scaling: Improving model performance in long-horizon contexts may not require scaling model size or compute. Instead, gains can be achieved by improving the model's ability to accurately assess and adapt to its own limitations.
Efficiency: Correcting this miscalibration improves both accuracy and token efficiency, reducing the computational cost of solving complex problems.

The study concludes that accurate self-assessment is a fundamental component of reasoning performance, and addressing "context anxiety" is critical for deploying reliable, efficient reasoning systems in safety-critical or assistive applications.