Research map / Large language models / Reasoning and chain of thought

Lost in Context: Adressing Context Anxiety in Large Language Models

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Studies "context anxiety" in reasoning LLMs on Tower of Hanoi, linking premature self-doubt and token-estimate miscalibration to accuracy and efficiency losses, and reduces it via SFT on anxiety-filtered reasoning traces of GPT-OSS-20B.

Problem

Reasoning models sometimes abandon solvable long-horizon problems, claiming their output would exceed token limits; it is unclear whether such failures reflect true capability limits or miscalibrated self-assessment of token needs.

Approach

Defines context anxiety operationally via LLM-judge detection of resource-insufficiency statements in reasoning traces, links it to overestimated token requirements (winsorized estimated/actual token ratio), and mitigates it by SFT on reasoning traces filtered to exclude anxious ones.

Key results

• Context anxiety reduces accuracy by 15.3% and increases completion tokens on correct answers by 2,323 tokens (fixed-effects regression, p<0.01).
• Models with context anxiety overestimate token requirements by about 24%, versus underestimation of 19% for those without anxiety.
• SFT on anxiety-filtered traces reduced detected anxiety rates (e.g., 0.28 to 0.12 at 5 disks) and raised 5-disk accuracy from 76.7% to 83.3% versus 46.7% for all-correct SFT.

Surprising findings

• Claude Sonnet 3.7 and DeepSeek R1 refuse nearly 100% of the hardest problems, with most failures anxiety-driven.
• Fine-tuning on all correct traces (without anxiety filtering) substantially degraded accuracy at intermediate difficulties.

Limitations

• Primary evaluation uses a single symbolic task (Tower of Hanoi), with a second task only in the appendix.
• Detection relies on explicit self-reported statements and may miss implicit disengagement.
• Token estimation is elicited from the model itself and may not reflect true computational cost.

Experimental setup

Datasets: Tower of Hanoi generated problems (270 eval problems, 2-12 disks) Shortest-path grid search instances

Benchmarks: Tower of Hanoi (2-12 disks) Shortest-path grid search (5x5 to 12x12)

Metrics: accuracy anxiety rate winsorized token ratio completion tokens on correct solutions

Baselines: Claude Sonnet 3.7 Claude Sonnet 4.5 DeepSeek R1 Gemini 2.5 Flash Kimi K2 Thinking o4 Mini All-Correct SFT

Builds on: chain-of-thought prompting distillation from reasoning teachers

Closest papers by meaning

• The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs (ICLR 2026 · similarity 0.89)
• Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning (NeurIPS 2025 · similarity 0.88)
• Rethinking External Slow-Thinking: From Snowball Errors to Probability of Correct Reasoning (ICML 2025 · similarity 0.87)
• Training on the Test Task Confounds Evaluation and Emergence (ICLR 2025 · similarity 0.87)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.87)
• CodeCrash: Exposing LLM Fragility to Misleading Natural Language in Code Reasoning (NeurIPS 2025 · similarity 0.86)