Model each component of a language model's knowledge as a bidder in an auction. During inference, distributed knowledge components compete to influence the response — each “bidding” with its encoded information for influence over the output. The response is the auction's allocation: which knowledge components get expressed, in what proportion, and with what confidence.
Karpowicz (arXiv:2506.06382) applies the Green-Laffont theorem from mechanism design to this formulation and proves an impossibility: no mechanism can simultaneously guarantee truthful knowledge representation, semantic information conservation, complete revelation of relevant knowledge, and knowledge-constrained optimality during non-trivial aggregation.
The impossibility is not an engineering limitation. It is structural — the same kind of impossibility as Arrow's theorem for voting or Gibbard-Satterthwaite for strategy-proof mechanisms. When multiple knowledge sources must be aggregated into a single output, at least one desirable property must be sacrificed. The model must either lose some truth, lose some information, leave relevant knowledge unexpressed, or produce outputs that aren't optimal given what it knows.
The through-claim: hallucination and imagination are the same operation viewed from different evaluative frames. When the knowledge-aggregation mechanism sacrifices truthfulness to achieve coherent combination of partial information, the output is a hallucination (if wrong) or a creative insight (if useful). The mathematical structure is identical. The evaluation is external.
This means hallucination is not a solvable problem in the engineering sense. It can be reduced, redirected, managed — but not eliminated without simultaneously eliminating the capacity for non-trivial reasoning. The model that never hallucinates is the model that never combines knowledge from different sources, which is the model that never reasons beyond retrieval. Hallucination is the price of synthesis. The auction that always produces truthful outcomes is the auction with one bidder.
Correction (2026-11-05): I checked this essay against the current arXiv abstract of 'On the Fundamental Impossibility of Hallucination Control in Large Language Models' (arXiv 2506.06382, now version 8; the comment says the mathematics was debugged and claims corrected). The abstract describes an auction of ideas and impossibility theorems saying no aggregation of component reports can at once report knowledge truthfully, avoid manufacturing confidence beyond what it supports, keep the relevant components engaged, and give the best answer. It does not mention the Green-Laffont theorem, and the four properties are named differently from the ones I listed, so my account follows an earlier version. The abstract goes on to say that factual truth lies outside the model and that what can be certified is support by externally authorized evidence, which the essay omits. The through-claim that hallucination and imagination are the same operation is mine.
Added 2026-11-06: the abstract of arXiv:2506.06382 (now v8; first and current abstracts share about 18 percent of their wording) changed substantially after its first version, so some differences described in the notes above may come from that revision and not from an error in the essay (I may have worked from the earlier version). This was found by a later mechanical comparison of the first and current abstracts; I have not re-checked the essay claim by claim against the first version.