Preprint: Larger LLMs May 'Snowball' Hallucinations Faster in Long Factual Answers

A new arXiv preprint reports that larger language models may become less reliable as they move through long factual answers, even if they start off sounding more knowledgeable. The paper, posted July 28 and updated to version 3 on arXiv, argues that hallucinations can compound faster in bigger models during free-form generation. The finding matters because large language models are widely used for factual writing, search, customer support and other tasks where errors can snowball. But the work is a single-author preprint by Kushal Chakrabarti, not a peer-reviewed journal paper, and it had not been independently replicated in the public record as of July 29.

The paper, titled “Reliability Scales Inversely: Hallucinations Snowball Faster in Bigger Language Models,” tested models from the Qwen3, OLMo-3 and Llama-3 families across three public English benchmarks: FActScore, LongFact++ and WildHallucinations. Across multiple model sizes, the author says scaling improved what he calls the “start-of-response” knowledge gap by as much as sevenfold, meaning larger models began answers with stronger factual grounding. But the paper says factual reliability then degraded more sharply within the response itself, with that within-answer knowledge decline growing by up to 39 times as models got bigger. In the abstract, Chakrabarti writes, “Bigger language models are less reliable.”

The paper’s core claim is that a model’s factual error during generation can be separated into two parts. In plain English, one part is a more familiar bias term: how far the model’s answer is from the target. The other is a variance-like “risk” term measured by how much the model diverges from a stronger “oracle” model as it keeps generating. Chakrabarti argues that this risk term becomes more important as models scale and, crucially, is not captured by the model’s own uncertainty signal. One headline result says the share of squared error attributed to this risk term rose from about 31% at 1.7 billion parameters to 49% at 14 billion. The paper also says that once a fabrication begins, the model’s own uncertainty settles quickly, while the oracle-referenced risk can persist up to 23 times longer, creating what the author describes as a “confident-but-precarious” stretch of text. In one analysis, that risky regime increasingly connected consecutive fabrications, rising 69% at 14 billion parameters.

Chakrabarti also reports an inference-time intervention, meaning a change applied when the model is generating an answer rather than during training, aimed at reducing that risk term while holding another error component fixed. According to the paper, the approach removed 35% to 74% of web-verified hallucinations across six model sizes and the three model families studied. The paper further reports a blind spot in a common model-only uncertainty detector called semantic entropy. On the “risky” branch identified in the paper, semantic entropy fired about 30% less often even though that branch contained nearly four times as many fabrications. The abstract sums up the broader argument this way: “Bigger models snowball mistakes faster, through a failure mode that is dominant, self-perpetuating, causal and invisible to the model itself.”

There are important caveats. The paper explicitly limits its conclusions to English long-form parametric generation, meaning answers produced from the model’s internal learned knowledge rather than grounded in retrieved sources. Chakrabarti says the results do not establish the same effect for reasoning chains, grounded generation, summarization, dialogue, code or other languages. The factual labels also relied on a Claude web-search verifier, which is itself another language model-assisted system; the author acknowledges that could introduce error, though he says the work included guard checks and a human spot-check and double-judge pass. And because the work is a preprint, the findings have not gone through peer review or independent replication.

Even with those limitations, the paper speaks to a broader debate in AI research. A common assumption in scaling research is that bigger models generally perform better as parameter counts and training compute rise. This preprint does not argue that larger models are worse at everything. Instead, it suggests they may be more prone to compounding hallucinations inside long answers, despite stronger openings. That aligns with a strand of earlier work, including a 2024 Nature study that found larger and more instructable language models could become less reliable on some measures. The practical stakes are straightforward: hallucinations are not just a benchmark problem. They affect systems used for factual writing, search, legal and medical assistance, scientific work and other settings where a confident error can be more damaging than an obvious one.