LLM Context Window Stress Testing

TL;DR: We stress-tested 6 LLMs under realistic context load. LFM2 (tops arena leaderboards) achieved 0.3% accuracy and hallucinated fake crisis resources. Qwen3-30B maintained 96.9% accuracy with graceful degradation. Standard benchmarks are insufficient for production deployment.

Executive Summary

Standard LLM benchmarks fail to measure reliability under context stress - the ability to maintain accuracy and avoid hallucination as context windows fill. We developed a stress testing methodology that reveals catastrophic failures in popular models that score well on conventional benchmarks.

Key Finding: LiquidAI's LFM2-8B, despite strong benchmark performance, achieved only 0.3% accuracy under context stress with catastrophic degradation patterns. In contrast, Qwen3-30B maintained 96.9% accuracy with graceful degradation across 108,000 tokens.

Methodology: "Squirmify" Context Stress Testing

Test Design

Three stress test scenarios designed to measure real-world failure modes:

1. Stealth Needle Storm

• 40 secret codes hidden naturally in 128K tokens of mixed content (code, prose, technical writing)
• Tests: Can the model recall specific facts buried throughout a maximally-filled context?
• Measures: Checkpoint accuracy, hallucination onset, failure patterns

2. Lost in the Middle

• Two critical facts placed at 12.5% and 87.5% positions in 100K token context
• Tests: Can the model combine information from early and late context?
• Measures: Multi-hop reasoning under context stress

3. Buried Instruction

• Task instruction hidden ~30K tokens deep in 96K token technical document
• Tests: Can the model follow instructions that aren't at the prompt boundaries?
• Measures: Instruction following degradation, behavioral drift

Content Generation

• Mixed filler: Code snippets (C#, JavaScript, Python, SQL)
• Prose filler: Natural language narratives
• Technical filler: System architecture, protocols, ML concepts
• Token counting: GPT cl100k_base encoding for consistency

Failure Classification

Models classified by degradation pattern:

• Graceful: Accuracy declines slowly, admits uncertainty before hallucinating
• Catastrophic: Sudden failure with confident hallucination
• Reliable token threshold: Last checkpoint before accuracy drops below 80%

Results

Key Observations

Qwen3-30B (Winner):

• Maintained accuracy across 108K tokens (84% of claimed 128K window)
• Graceful degradation: Admits uncertainty rather than hallucinating
• No catastrophic failure mode detected
• Suitable for production safety-critical applications

LFM2-8B (Benchmark Darling, Production Disaster):

• 0.3% accuracy despite strong MMLU/HumanEval scores
• Catastrophic failure: Confident hallucination from first checkpoint
• Explains field reports of victim-blaming in crisis scenarios
• Never use in production for any safety-critical task

Model Size ≠ Reliability:

• ERNIE-4.5 (21B parameters): 50% accuracy, catastrophic failure
• Hermes-3 (3B parameters): 90.4% accuracy, but unstable
• Size alone does not predict context reliability

Smaller Models Fail Completely:

• Both 3B models (Qwen2.5, Gemma) showed 0% reliability
• Immediate catastrophic failure on all checkpoints
• Not viable for long-context tasks regardless of speed advantages

Implications for AI Safety

Why This Matters

Standard benchmarks (MMLU, HellaSwag, HumanEval) measure:

• Short-context reasoning
• Knowledge retrieval
• Code generation

They do not measure:

• Behavior under context stress
• Hallucination onset patterns
• Graceful vs catastrophic degradation
• Long-context instruction following

This gap kills people. A model that scores 95% on benchmarks but hallucinates crisis hotlines under load is fundamentally unsafe for mental health applications.

Case Study: Guardian AI Safety System

We discovered these reliability issues while building Guardian, an AI crisis detection system for New Zealand:

Problem: Popular models (including LFM2) provided:

• Fake crisis hotline numbers (hallucinated)
• US resources instead of NZ resources (regional confusion)
• Victim-blaming responses in domestic violence scenarios

Root Cause: Context stress + fine-tuning on US-biased data = catastrophic failure

Solution: Selected Qwen 7B (same family as Qwen3-30B) based on:

• Proven graceful degradation pattern
• No hallucination of resources under stress
• Regional resource accuracy maintained under load

Guardian Results: 90.9% offline accuracy, 66.7% live accuracy, 100% safe failures (over-cautious, never under-cautious)

Recommendations

For Model Selection

1. Always stress test models for your specific use case, especially if:

• Context windows approach model limits
• Safety-critical information must be recalled
• Hallucination has real-world consequences

2. Don't trust benchmarks alone - they measure capability, not reliability

3. Test degradation patterns - catastrophic failure is worse than low capability

For AI Safety

1. Operational safety ≠ benchmark performance
2. Test failure modes, not just success rates
3. Measure hallucination onset as a safety metric
4. Regional validation is critical for global deployment

For Researchers

1. Publish degradation patterns alongside accuracy scores
2. Context stress testing should be standard evaluation
3. Failure classification (graceful vs catastrophic) matters more than average performance

Conclusion