Life of Research · open-access repository · LOR-00237
Two Answers Are Not Two Memories: What the Evidence for Intra-Memory Knowledge Conflict Certifies, Why Head Interventions That Settle Trained-In Conflicts Lose Their Grip on Real Entities, and the Controls That Would Separate Competing Stores From Unstable Retrieval
- SONYTECH
Preprint
DOI10.5281/zenodo.24326760 LOR-00237
Permanent link: https://lifeofresearch.com/preprints/237
Abstract
(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0.
Life of Research A language model that gives different answers to the same factual question is commonly said to hold an intra-memory knowledge conflict: two incompatible facts about one subject, both written into its weights by contradictory training data. In this analysis the author reviews and synthesises published results to show that the evidence offered for this construct does not, in most cases, distinguish it from four other states that produce the same spread of answers: a single stored fact retrieved unreliably, no adequate stored fact at all, a fact that genuinely has several correct values, and interference between true facts that share a cue. Where conflict is deliberately trained in, it behaves as the construct says: a 2026 mechanistic study finds it resolved in the last layers, and its most robust head-level intervention steers it at success rates of 0.84 to 0.99. On real entities the same intervention reaches 0.48 to 0.78, heads that back only one answer become scarcer, and filtering down to those heads lowers zero-ablation success in all four models, where on synthetic conflicts it raised it, and costs clean-model patching 0.08 to 0.20. The author shows that this synthetic-versus-real contrast bundles at least four differences, including the similarity of the competing answers, which alone predicts the loss of single-answer heads. On natural data, semantic entropy, the main behavioural instrument, does not rise with how often a fact has changed, in the original study or its reproduction, and, with the reproduction's corrected scoring, the models tested answer only about one question in five to one in eight correctly without context. The paper consolidates 25 published values in a table and a figure, states what each reading predicts, gives a procedure for classifying an observed inconsistency, and specifies the factorial injection design and the corpus, ranking and paraphrase tests that would settle which reading a given conflict belongs to.
Keywords
Files (1)
Cite this record
Guides How to cite a preprint Preprints vs journal articles What CC BY 4.0 allows