Skip to content

Data Points Reference

Fresh Updated January 2026

All key statistics and metrics from LLM technical phenomena research, organized for quick access.

KV Cache Non-Determinism

MetricValueSource
Unique completions (temp=0)11.67 of 30Chroma Research 2024
Variation causeFloating-point batchingKV Cache Implementation
ReproducibilityNot guaranteedEven with identical prompts

Key Insight

temperature=0 does not guarantee deterministic output.

Context Rot (Hidden State Drift)

MetricValueSource
Performance degradation20-60%Thinking Machines Lab 2024
Threshold>100k tokensLong-context studies
Hallucination correlationr = 0.73Hidden state analysis
Effect typeNon-monotonic decayLayer-wise measurement

RLHF Alignment Tax

MetricValueSource
NLP performance drop5-15%ACL 2024-2025
Refusal rate increaseSignificantSafety training
Knowledge suppressionObservedAlignment fine-tuning

Trade-off Spectrum

More AlignedLess Aligned
Safer responsesBroader coverage
More refusalsMore direct answers
Conservative framingComplete information

MoE Routing Variance

MetricValueSource
Unique completions30 of 30MoE architecture studies
CauseLoad balancing decisionsExpert routing
ReproducibilityVery lowToken-level routing varies

WARNING

MoE models (GPT-4, Mixtral) show higher variance than dense models.

Lost in the Middle Effect

MetricValueSource
Middle position performance30% worseLiu et al. 2023
Best positionsFirst 20%, Last 20%Attention patterns
Worst position40-60% mark"Lost in the middle"

Position Performance Map

Prompt Injection Vulnerability

MetricValueSource
Architectural separationNoneLLM design
Reliable defenseNoneSecurity research
Mitigation effectivenessPartialHeuristic-based

Rate Limiting Effects

ProviderMetricsImpact
OpenAIRPM, TPM, TPDBatch reshaping
AnthropicRPM, TPMVariance changes
GoogleQPM, TPMLoad distribution

Interpretability Gaps

MetricValueSource
Papers published (2024-2025)100+EMNLP 2025
Consensus solutions0Research status
SAE variance explained80-90%Partial decomposition
Polysemantic neuronsCommonSuperposition

Platform Citation Statistics

PlatformTop SourceDiversityFreshness Priority
ChatGPTWikipedia (47.9%)ModerateLower
PerplexityReal-time web2-3x higherCritical
ClaudeCurated sourcesLowerModerate
GeminiGoogle indexLower (1-2 sources)Moderate

Citation Overlap

Content Optimization Targets

Length Recommendations

Content TypeOptimalMaximum
Product page2k-5k tokens10k
Blog post3k-8k tokens20k
Guide/manual5k-15k tokens30k
Documentation10k-30k tokens50k

Distractor Impact

DistractorsCitation Impact
0Baseline (100%)
1-15%
4-40%

Research Sources

SourceFocus AreaYear
Thinking Machines LabContext rot, hidden state drift2024
Chroma ResearchKV cache non-determinism2024
ACL ProceedingsRLHF alignment tax2024-2025
Liu et al.Lost in the middle2023
EMNLPInterpretability2025
AnthropicSAE decomposition2024-2025

See Also

Based on research from Thinking Machines Lab, Chroma Research, and ACL 2024-2025