LLMs by Hallucination Rate

Original Article ↗ Hacker News Discussion ↗

Claims that GPT‑4 and other large language models now hallucinate as little as 3% of the time in summarization tasks spark debate over how meaningful such benchmarks really are. Commenters question the methodology (including using one model to judge another), argue over whether “hallucination” or “confabulation” is the right concept, and note that performance degrades sharply outside narrow tasks like summarization. Many see hallucinations as inherent to current autoregressive architectures and stress the need for better ways to detect, contextualize, or redesign around them before relying on LLMs for high‑stakes use.

Nature of LLM “hallucinations”

  • Core view: autoregressive token-by-token prediction inherently produces guesses; you can’t fully eliminate fabrications.
  • Some argue LLMs are always doing the same process, so either they’re “always hallucinating” or the term is misleading.
  • Others stress that while the mechanism is the same, outputs differ in factual correctness; “hallucination rate” is still a useful metric.

Comparisons to Human Cognition

  • Several comments note humans also “fill in gaps,” misinterpret questions, or confidently state falsehoods; assumptions are necessary for efficient communication.
  • Disagreement over whether intelligence requires filling gaps vs. knowing when to say “I don’t know.”
  • Some liken LLM behavior to young children’s storytelling or “brain farts” and confabulations, especially when memory/context is limited.

Architectures and Mitigations

  • Proposed alternatives: models that internally work for a while and then emit a whole answer rather than autoregressive streams.
  • Techniques discussed: chain-of-thought, self-critique by feeding prior outputs back in, iterative verification until “convergence.”
  • Desire for models that more often answer “I don’t know,” but skepticism that today’s LLMs truly represent knowledge or truth internally.
  • Ideas like a “backspace token” or more aggressive RLHF to penalize fabrication are floated, but data and implementation are unclear.

Terminology Debates

  • Many argue “hallucination” is the wrong metaphor; “confabulation” or “statistical guess” is preferred.
  • Objection that “hallucination” anthropomorphizes systems that only model language, not perception or truth.

Benchmark Scope and Methodology

  • Leaderboard is summarization-only, which commenters say is a relatively low-hallucination task; numbers may not generalize.
  • Hallucination detection itself uses another model; critics note that if model-based detection were fully reliable, we’d largely solve the problem.
  • External critique of the benchmark’s methodology is referenced; some therefore place limited weight on the reported percentages.
  • Note that ~17% of documents were rejected by at least one model due to content filters.

Model Differences and Practical Reliability

  • GPT‑4’s ~3% summarization hallucination rate is widely doubted; users report far higher error rates on niche or technical questions, especially code (nonexistent libraries/functions).
  • Some observe Mistral and other models readily produce “plausible nonsense,” possibly due to RLHF choices.
  • Commenters emphasize that for critical applications, even the best current rates are still unacceptably high without human oversight.