The easiest thing to say—and hardest to deliver—to a model is: 'That statement of yours is wrong; correct it.' A 2024 survey paper authored by 22 researchers opens by highlighting a contradiction: training incurs prohibitively high computational costs, yet the world constantly changes, requiring outdated information to be corrected on demand—making efficient, lightweight, real-time modification methods an active research direction. Directly rewriting the erroneous entry appears elegant, but the material one writes upon differs fundamentally from paper: paper-based entries are explicit, and the pen lands unambiguously on a specific line; weights contain no such collection of discrete entries. This problem admits a more precise formulation: when new evidence arrives, which old beliefs should be abandoned—and how? In the literature, this is termed belief revision. Over forty years ago, during the era of symbolic systems, two answers were proposed—one engineering-oriented, the other logic-based. Today’s large models implement neither.
## Evidence Impacts the Whole System
Why does 'abandoning one belief' require an entire set of rules? The philosophical stance known as holism provides the foundational rationale (paraphrased; no verbatim source obtained): a single observational test never confronts an isolated hypothesis, but rather an entire bundle comprising the hypothesis together with auxiliary assumptions and instrument theories; when a prediction fails, logic only tells us that something within this bundle is mistaken—not which specific element should be revised. One may revise the core hypothesis, adjust auxiliary assumptions, or even question the instrumentation. Revision always permits multiple options; selecting 'which part to change' requires additional structure to decide. This is also the deep structure underlying 'discussing specific judgments': evidence never shakes just an isolated sentence. Without structures like importance rankings, supporting reasons, or hierarchical organization, a belief system cannot support the notion of 'reasonable revision.'
## Attaching Reasons to Beliefs
In June 1979, a technical report (Memo 521) from MIT’s Artificial Intelligence Laboratory—later published in the journal Artificial Intelligence, Volume 12, pages 231–272—directly embedded revision into a machine. The system was called the Truth Maintenance System (TMS). Its official abstract lists seven components: the system’s representation and structure; mechanisms for revising the current belief set; how dependency-directed backtracking modifies the current set of assumptions; summary techniques for belief explanations; how to organize problem solvers into modules engaged in 'dialectical argument' (quoting the paper itself); how to revise models of others’ belief systems; and how to embed control structures within assumption modes.
Read from the perspective of revision, the abstract clarifies the core idea: the inference procedure first makes assumptions, then revises beliefs upon detecting contradictions with those assumptions; the TMS implements these functions by recording and maintaining the reasons underlying the program’s beliefs. The abstract also explicitly states: 'the problem solver must choose among alternative belief systems.' This admission directly acknowledges the holism dilemma: multiple revision paths exist, so the machine must possess a selection mechanism. The full text remains unavailable through public channels; the above paraphrase relies entirely on the official abstract, and no further mechanistic details beyond the abstract are elaborated here.
## Formalizing 'Correct Revision' as Postulates
In June 1985, a 21-page paper appeared in the Journal of Symbolic Logic (received October 18, 1983; revised May 1, 1984; pages 510–530), co-authored by three researchers and subsequently referred to collectively by the initials of their surnames—the AGM framework. This article uses 'AGM' solely as a framework name. The paper formalizes theory change into three operations: expansion (adding a new proposition directly), contraction (removing an existing proposition), and revision (adding a proposition inconsistent with the original theory while preserving consistency and logical closure of the result). The paper also draws a legal analogy: when a theory is viewed as a codified legal system, contraction corresponds to derogation in legal theory, and revision corresponds to amendment. Legal systems had already confronted this very same problem.
The paper then reviews a set of rationality postulates, explicitly stating they are drawn from earlier literature: six for contraction and six for revision, each bearing a name. As an example from contraction: closure (the result remains a theory), inclusion (only removal, no addition), vacuity (if the proposition is irrelevant to the theory, the theory remains unchanged), success (the proposition to be removed is indeed no longer derivable), preservation (propositions entailing identical logical consequences are treated identically), and recovery (if the removed proposition is re-accepted, the original theory can be recovered). Recovery merits a pause: it ensures deletions are reversible—a logical cousin to today’s engineering term 'rollback.' Revision can be reduced to contraction: first contract the negation of the new proposition, then expand with the new proposition. The paper adopts prior work calling this the Levi identity.
The paper’s own contribution lies in construction. Taking the intersection of *all* maximal subsets that do not entail the proposition to be deleted yields too little: the contracted result contains only what can be derived solely from the negation of the target proposition; applying the Levi identity to convert this into revision yields only the logical consequences of the new proposition itself—both cases judged in the original text as "far too small." Selecting *any single* maximal subset removes too much: the paper terms this selection strategy 'inflationary,' noting that when the new proposition conflicts with the original theory, the revision result becomes a complete theory—assigning truth values to *every* proposition. The paper’s solution is partial meet contraction: using a selection function to pick several 'most important' (quoting the paper) candidate subsets, then taking their intersection; how 'importance' is concretely realized depends on relational structures defined over the candidate subsets—a topic one author had previously explored in depth. A representation theorem proves that such constructions satisfy the aforementioned rationality postulates and are sufficiently general to provide representations for those postulates. Henceforth, 'what counts as a correct revision' acquired a mathematical characterization.
## Versions and Rollback
The same intuition has long been institutionalized in human knowledge organization. Wiki-style community knowledge systems record every edit: editor, timestamp, discussion history; old versions remain preserved and retrievable on demand; revision thus becomes a first-class, traceable event (background knowledge, not dependent on any single source). Juxtaposed with the 1979 reason-recording approach and the 1985 recovery postulate, this idea is strikingly simple: to make revision reasonable, one must first make revision recordable. As for where knowledge originates, that is addressed elsewhere; here we focus solely on *how* knowledge changes.
## Locating and Writing into Weights
Starting in 2022, large models advanced the problem using entirely different materials. A February 2022 paper (comment field notes NeurIPS 2022) first provided evidence that factual knowledge resides in 'specific places' within models: causal interventions identified neurons whose activation decisively influenced factual predictions, and feedforward modules in intermediate layers mediated factual predictions when processing subject tokens. If there is a 'place,' direct editing becomes possible: the method is called ROME (Rank-One Model Editing), updating weights associated with specific factual relations via rank-one updates. Evaluation proceeds at two levels: performance matches existing methods on standard zero-shot relation extraction (zsRE) editing tasks; on a newly constructed counterfactual assertion dataset designed for more sensitive evaluation, ROME simultaneously maintains specificity and generalization—whereas other methods trade off one against the other.
In October 2022, MEMIT scaled writing from single facts to batches: extending to thousands of associations, validated on GPT-J (6B) and GPT-NeoX (20B), exceeding prior work by several orders of magnitude. Its abstract explicitly states the motivation: replacing outdated information or adding domain-specific knowledge.
## Ripple Effects
A July 2023 paper exposed a blind spot in this evaluation paradigm (comment field notes acceptance by TACL in 2024): prior evaluations mainly tested only two things—whether a single fact was successfully written in, and whether predictions for similar subjects remained unchanged. But knowledge is relational. The paper’s example is direct (paraphrased anonymously per its abstract): inserting 'Person A is the son of Celebrity B' into the model should also cause statements like 'Person A is a sibling of another child of Celebrity B' to change accordingly. The paper terms this the ripple effect and constructs a dedicated diagnostic benchmark, RippleEdits, containing 5,000 fact edits covering multiple types. The result: existing editing methods fail to induce consistent changes across the model’s knowledge; the top-performing baseline is surprisingly simple in-context editing—inserting the new fact into the context without modifying weights at all.
A January 2024 survey (22 authors, revised to version 5 by November 2024, comment field notes it remains ongoing) organizes this field into three lines: leveraging external knowledge, integrating knowledge into the model, and editing intrinsic knowledge; it also introduces a new evaluation benchmark, KnowEdit, and knowledge localization analysis. A detail easily overlooked in its version-5 revision notes: the authors fixed computational errors in companion tools related to two editing methods, optimized parts of certain methods, and updated comparative results tables accordingly. A survey *about revision* repeatedly undergoing revision is perhaps the most honest footnote this field could have.
## Three Judgments
First, belief revision was formalized by the mid-1980s: postulates provide criteria for 'correct revision'; constructions supply procedures satisfying those criteria; and the Truth Maintenance System made belief reasons explicit records (synthesizing two threads here). Today’s large model knowledge systems lack counterparts: no proposition set named a 'belief set' exists within weights; editing weights or swapping retrieval sources are *not* belief set revision—they are local patches applied to a system lacking any such set.
Second, ripple errors indicate that 'changing one fact' is harder in continuously superimposed weights than in explicit symbolic structures (this article’s judgment). Symbolic systems pre-encode 'what else must change' and 'whether changes can be undone' into rules: reason records support revision; the recovery postulate guarantees reversibility; weight-based systems can only detect inconsistency *post hoc*, during evaluation. The 5,000-edit diagnostic benchmark and the finding that 'existing methods fail to introduce consistent changes' constitute the tangible cost of this absence.
Third, the crux of the revision problem lies not in write-in technology, but in the organization of evidence and sources—and in the assignment of revision authority (this article’s judgment): *what* may be changed, *why* it may be changed, and *how* to revert if changed incorrectly—all require explicit recording as a prerequisite. Evidence for this judgment remains thin; the strongest empirical fact currently available is likely this: the top performer on the ripple benchmark is precisely the method that leaves weights untouched. This remains a hypothesis awaiting verification.
## What Different Readers Can Take Away
For teams maintaining large-model applications: treat each knowledge edit as a software release—after a point edit, run consistency regression tests on related facts; for lightweight updates, prioritize retrieval-side and context-side approaches; the strong performance of the context-based baseline on the ripple benchmark deserves serious attention.
For evaluation researchers: single-point injection success rates are no longer sufficient; ripple effects and consistency dimensions now have ready-made diagnostic benchmarks; also remember the survey’s lesson—once computational errors in evaluation tools themselves are fixed, reported numbers shift accordingly.
For readers concerned with platform governance: auditability and rollback for knowledge revision both depend on explicit records; how weight-based systems can acquire a 'version history as ordinary as Wikipedia’s' remains both a regulatory gap and a technical gap.
For general readers: the everyday intuition that 'AI hallucinates' reflects an old epistemological problem—evidence does not alter isolated beliefs, but reshapes the entire web of belief; humans manage revision by recording reasons and preserving versions; models today have learned neither.
Have you ever tried correcting an AI’s mistake? After your fix, did it acknowledge the correction—or did it 'change elsewhere' as a side effect? Share your experiences in the comments.
Scope and Boundaries: The 1985 paper was read in full, based on a publicly available scanned image (21 pages total; OCR contains noise; English quotations transcribed directly from the scan; page numbers 510–530 inferred from headers—the second page reads 511, the twentieth reads 529; DOI 10.2307/2274239 verified to resolve to the publisher’s article page). The 1979 report is cited at the abstract level only; repeated attempts to access the full text failed (403 Forbidden and rate-limiting 429 errors); no mechanisms omitted from the abstract are described. Four arXiv papers are cited at the abstract level only; numbers and conclusions follow the abstract wording—MEMIT’s scale is cited as 'thousands of associations' per its abstract, not larger numbers appearing in the main text. The holism section is paraphrased; primary sources were not obtained, and no verbatim quotation is used. The wiki-versioning section reflects background knowledge; how such systems manage versions is covered elsewhere and is not elaborated here. 'AGM' and the 'Levi identity' are conventional names固化 by surnames and personal names, used here solely as framework and term names; authors’ surnames do not appear in the main text, and papers are referenced by year and publication venue. The ripple-effect example is paraphrased anonymously per its abstract. The two editing studies are ordered chronologically by arXiv posting date only, with no attribution to research teams. Papers contain no conflict-of-interest disclosures; no claims about presence or absence are made. ROME, MEMIT, RippleEdits, KnowEdit, GPT-J, GPT-NeoX, zsRE, and TMS are proper nouns retained verbatim, not implying product evaluation. 'Over forty years ago' is a rounded figure. 'Truth Maintenance System', 'dependency-directed backtracking', 'partial meet contraction', and 'ripple effect' are translated with English originals provided at first occurrence. Mechanisms of retrieval-augmentation are discussed elsewhere; this article mentions them only at the division-of-labor level.
Sources: 'On the Logic of Theory Change: Partial Meet Contraction and Revision Functions', The Journal of Symbolic Logic, Vol. 50, No. 2 (June 1985), pp. 510–530, DOI 10.2307/2274239; 'A Truth Maintenance System', Artificial Intelligence, Vol. 12 (1979), pp. 231–272, plus MIT AI Lab Technical Report Memo 521 (June 1979); arXiv:2202.05262 'Locating and Editing Factual Associations in GPT', posted February 10, 2022, comment field notes NeurIPS 2022 (https://arxiv.org/abs/2202.05262); arXiv:2210.07229 'Mass-Editing Memory in a Transformer', posted October 13, 2022; arXiv:2307.12976 'Evaluating the Ripple Effects of Knowledge Editing in Language Models', posted July 24, 2023, comment field notes acceptance by TACL 2024; arXiv:2401.01286 'A Comprehensive Study of Knowledge Editing for Large Language Models', posted January 2, 2024, revised to v5 on November 17, 2024, comment field notes it remains ongoing.