Mechanistic Interpretability

Why Do Neural Networks Hallucinate?

LLMs predict plausible next tokens via statistical associations, not verified truth. When prompts cross sparse training regions or high sampling temperatures flatten probability distributions, autoregressive drift compounds errors into confident confabulations.

Inference Parameters

75%

Lower values simulate out-of-distribution prompts with sparse training support.

0.70

Higher softmax temperature flattens logits, boosting improbable, non-factual tokens.

6 tokens

Each hallucinated token becomes context for the next, exponentially increasing drift.

Hallucination Probability: Moderate (38%)

Model remains within plausibility bounds but minor factual deviations may emerge.

Token Probability Manifold
Factual Basin Hallucination Zone
Generated Autoregressive Trajectory

1. Statistical Plausibility ≠ Truth

Neural networks optimize for maximum likelihood estimation over token sequences, learning syntax and stylistic cadence rather than formal causal proofs or truth invariants.

2. OOD Manifold Gaps

When prompts land in unrepresented feature space, attention heads interpolate smoothly across nearest neighbors, synthesizing confident fiction from disjoint memory vectors.

3. Compounding Autoregression

Generation is autoregressive: output tokens are fed back as immutable ground truth. A single early deviation skews all subsequent attention weights down an invalid path.

Enjoy this tool? Build your own with Super