A 2024 Hallucination Bound Now Anchors Three 2026 Papers on LLM Training Data, One Extending It to Scientific Discovery
A 2024 theorem tying LLM hallucination to once-seen training facts has drawn three 2026 papers, one linking it to scientific discovery and another challenging standard deduplication practice.
Background
In 2023, Adam Tauman Kalai and Santosh S. Vempala posted a theoretical paper, later published at the 56th Annual ACM Symposium on Theory of Computing in 2024, arguing that a language model’s tendency to “hallucinate” — to generate false but plausible text — is not purely an engineering flaw. For what they call “arbitrary” facts, whose truth cannot be inferred from the training data itself, they showed “there is an inherent statistical lower-bound on the rate that pretrained language models hallucinate certain types of facts, having nothing to do with the transformer LM architecture or data quality,” according to the paper’s abstract. Specifically, they found that a well-calibrated model’s hallucination rate tracks “the fraction of facts that occur exactly once in the training data (a ‘Good-Turing’ estimate), even assuming ideal training data without errors.”
That once-seen-fact statistic has since picked up a shorthand name, “monofact rate,” and become a reference point for a small cluster of 2026 papers that test, extend or apply it to different problems. Two of those papers reach a similar underlying conclusion — that the same statistic governing hallucination also governs something researchers want more of — but push it in different directions: one toward genuine scientific discovery, the other toward a direct challenge to a common data-processing practice.
Current situation
The most direct theoretical follow-up came from Nishant P. Das and Piyush Srivastava, who posted “Innovation: An Almost Characterization of Hallucination” in May 2026. Their abstract frames the open questions left by Kalai and Vempala’s result: “(i) what property of a calibrated LLM makes hallucinations unavoidable? and (ii) can hallucinations be avoided by giving up calibration?” Das and Srivastava answer by defining a property they call “innovation” — “the tendency of a model to produce outputs outside the training data” — and show it is “an almost characterization of hallucination: hallucination implies innovation, and conversely, innovation implies hallucination with high probability.” Their paper extends Kalai and Vempala’s missing-mass bound rather than testing it empirically.
An empirical test came from Miranda Muqing Miao and Michael Kearns, whose paper “Hallucination, Monofacts, and Miscalibration: An Empirical Investigation” was revised in March 2026. Working with classical n-gram models and fine-tuned encoder-decoder Transformers, they write that they “systematically control the monofact rates and establish its positive relationship with hallucination,” confirming the Kalai-Vempala bound holds in practice. But their paper’s central contribution runs the other direction: a technique they call “selective upweighting,” which “strategically repeats as little as 5% of training examples” to deliberately reduce calibration. According to the abstract, “this intervention reduces hallucination by up to 40%, challenging universal deduplication policies.” Miao and Kearns describe a trade-off rather than a free win: “selective upweighting maintains pre-injection levels of accuracy while substantially reducing hallucination, whereas standard training gradually improves accuracy but fails to address persistently high hallucination, indicating an inherent tension in optimization objectives.”
The newest entry, posted November 2, 2026 by Arda Uzunoglu, Mahler Revsine, Michael Schatz and Benjamin Van Durme of Johns Hopkins University (Van Durme is also affiliated with Microsoft), Mehrdad Farajtabar of Apple, and Daniel Khashabi and Eric Nalisnick of Johns Hopkins, takes the same statistic in a different direction still. “Tracing the Scientific Discovery Capacity of LLMs to Their Training Data” opens by summarizing the Kalai-Vempala framework and citing both Miao and Kearns’s and Das and Srivastava’s work, then states its aim: to “extend it from characterizing unseen falsehoods (hallucinations) to characterizing unseen truths (discoveries).” Where the earlier papers treat a high monofact rate as a problem to manage, this paper asks whether the same once-seen-fact statistic that predicts hallucination might also predict a model’s capacity to generate genuinely new, correct knowledge — a scientific discovery rather than a mistake.
To test this, the authors trained two open-weight models, OpenAI’s gpt-oss-20b and Google’s gemma-3-12b-it, on datasets with deliberately controlled monofact rates, then asked them to perform two discovery-style tasks: proposing chemical formulas with specified material properties, using Materials Project data, and proposing DNA sequence edits that achieve a target regulatory effect across three human cell lines. To rule out the possibility that the models were simply reciting memorized answers, the researchers evaluated them on formulas that first appeared in a September 2025 Materials Project snapshot — after the November 2024 snapshot used for training. The result, stated in the paper’s abstract, is that “more monofacts consistently yield more discoveries across two open-weight model families, sampling strategies, and reasoning budgets,” a relationship the paper reports held “across all settings” in both the materials and genomics tasks. The paper also reports that this could not be explained away by one obvious alternative: expected calibration error “remains relatively stable across monofact rates for both model families,” so, in the authors’ words, “the increasing discovery trend cannot be attributed primarily to improved formula-level calibration.”
The paper does not describe monofact rate as a cost-free lever, however. It reports that “higher monofact rate increases both discovery and hallucination in-distribution,” since, as the authors put it, “discovery and hallucination both draw from the unobserved mass” — the same statistical territory that produces a correct new finding can also produce a confident fabrication. The paper formalizes this as a “discovery–hallucination exchange rate,” noting that under one simplifying assumption on frontier posteriors (set out in the paper as α=β=1/2), a model “can expect one discovery per hallucination.”
The paper’s most concrete finding for anyone building training pipelines concerns repetition. The researchers ran the materials-discovery experiment again, training gpt-oss-20b for three and five epochs instead of one, while holding the underlying training sets unchanged. Because repeating a dataset means a fact seen once is now seen multiple times, this intervention — in the paper’s words — “reduc[es] the effective monofact rate to zero across all training sets,” regardless of how rare the facts were in the original, unrepeated data. The reported result: “under one-epoch training, discovery increases with the monofact rate, whereas at three and five epochs, this monotonic relationship disappears,” and at high monofact rates, one-epoch training actually outperformed the multi-epoch runs.
Implications
The points below describe what the cited papers state; they are not the Herald’s predictions.
Taken together, the four papers describe a single statistic — the share of training facts seen exactly once — that multiple independent research groups, working on different problems, now treat as consequential. Kalai and Vempala’s original 2024 result ties it to an unavoidable hallucination rate. Das and Srivastava’s 2026 paper formalizes a near-equivalence between hallucinating and producing anything the model wasn’t trained on. Miao and Kearns’s 2026 paper uses it to argue, on their own stated evidence, against treating deduplication as a universal default. The Johns Hopkins, Microsoft and Apple paper uses the same underlying quantity to argue that it also predicts something researchers want — genuine scientific discovery in materials science and regulatory genomics — not only something they want to avoid.
These findings point toward a tension rather than a single policy prescription. Miao and Kearns report that deliberately repeating a small share of training examples reduces hallucination. The newer paper reports that repeating examples — which happens automatically whenever a dataset is trained for more than one epoch — erases the discovery-boosting effect of monofacts entirely, pushing the effective monofact rate to zero regardless of how the original, unrepeated dataset was composed. Neither paper’s finding contradicts the other on its own terms: both describe repetition as suppressing the statistical signal the monofact framework measures, but one treats that suppression as desirable (less hallucination) and the other as a cost (less discovery). The newer paper’s own exchange-rate framing — that a model which discovers more will also, on the cited theoretical account, hallucinate more — suggests the two outcomes may not be separable by adjusting monofact rate alone, at least within the materials-science and regulatory-genomics tasks the paper tested.
For data curation choices made well before any of these specific findings, the practical stakes described across the four papers concern two decisions already made in most LLM training pipelines: how aggressively a corpus is deduplicated, and how many epochs it is trained for. Both papers that test these levers find they interact with the training data’s monofact rate, and therefore, on these papers’ own accounts, with both how often a model will state something false and how often it will state something newly true.