mlsanityData to 14 Oct 2026

Shortcut Before Circuit: Document Statistics Time In-Context Conflict Resolution

arXiv 2608.24460 ↗First posted 25 Aug 2026cs.CLcited by 0

Not on the map. Why?

Abstract (arXiv)

When a context asserts two values for one fact, a model commits to a cue -- recency, repetition, position -- but natural data rarely makes these disagree, so behavior cannot reveal which. We train 26M-parameter transformers on a synthetic language where recency and rarity are exactly coextensive, and separate them with a minimal causal edit that inverts one cue while holding the truth, token count and answer position fixed. All 75 runs reach accuracy >= 0.999, including where the trivial heuristic fails, so no held-in evaluation distinguishes them. Under intervention the per-cell readout does not replicate: 13 of 25 cells differ by more than 0.3 in sign fraction across three seeds, the largest by 0.879 against a standard error of 0.025. The construction predicts this -- coextensive rules leave the objective indifferent between them -- and the variance is ordered by how much of the optimization each comparison releases. What replicates is timing: escape from a positional shortcut with a closed-form ceiling, monotone in redundancy. Probed before that escape, attribution reverses sign in 32 of 75 runs at unchanged accuracy, and gating on circuit formation is necessary but not sufficient. The corpus fixes when a mechanism appears, not which one -- a criterion for when mechanistic attribution to data is available at all, and our construction makes the unavailable case exact.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 35 of its 37 references that are in the map. First posted on arXiv 25 Aug 2026.

Builds on

35 of 37 references.

References · 35

  1. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,455
  2. 2020Shortcut Learning in Deep Neural Networkscited by 1,969
  3. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  4. 2019Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inferencecited by 876
  5. 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
  6. 2022Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetscited by 629
  7. 2023Are Emergent Abilities of Large Language Models a Mirage?cited by 523
  8. 2020Underspecification Presents Challenges for Credibility in Modern Machine Learningcited by 496
  9. 2021Causal Abstractions of Neural Networkscited by 422
  10. 2021Entity-Based Knowledge Conflicts in Question Answeringcited by 356
  11. 2023Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflictscited by 341
  12. 2022Data Distributional Properties Drive Emergent In-Context Learning in Transformerscited by 309
23 more
  1. 2023Physics of Language Models: Part 3.1, Knowledge Storage and Extractioncited by 270
  2. 2023Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representationscited by 209
  3. 2019Predictive Multiplicity in Classificationcited by 166
  4. 2023Pretraining task diversity and the emergence of non-Bayesian in-context learning for regressioncited by 146
  5. 2022Rich Knowledge Sources Bring Complex Knowledge Conflicts: Recalibrating Models to Reflect Conflicting Evidencecited by 144
  6. 2023The mechanistic basis of data dependence and abrupt learning in an in-context classification taskcited by 118
  7. 2024Tug-of-War Between Knowledge: Exploring and Resolving Knowledge Conflicts in Retrieval-Augmented Language Modelscited by 93
  8. 2024Cutting Off the Head Ends the Conflict: A Mechanism for Interpreting and Mitigating Knowledge Conflicts in Language Modelscited by 91
  9. 2024What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formationcited by 89
  10. 2023The Transient Nature of Emergent In-Context Learning in Transformerscited by 76
  11. 2024Competition Dynamics Shape Algorithmic Phases of In-Context Learningcited by 54
  12. 2023Understanding In-Context Learning via Supportive Pretraining Datacited by 41
  13. 2025PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runscited by 35
  14. 2025Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable?cited by 34
  15. 2024Studying Large Language Model Behaviors Under Context-Memory Conflicts With Real Documentscited by 27
  16. 2023To grok or not to grok: Disentangling generalization and memorization on corrupted algorithmic datasetscited by 24
  17. 2025When can in-context learning generalize out of task distribution?cited by 16
  18. 2026Whose Facts Win? LLM Source Preferences under Knowledge Conflictscited by 12
  19. 2026All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMscited by 9
  20. 2025How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Modelscited by 4
  21. 2026Distinct mechanisms underlying in-context learning in transformerscited by 4
  22. 2026Explanation Multiplicity: Circuit-Level Interpretability Evidence Does Not Survive Defensible Analytic Variationcited by 2
  23. 2025Overcoming State Inertia: Minimally Invasive Temporal Alignment for Evolving Contextscited by 1

Built on by

Not cited yet.