mlsanityData to 14 Oct 2026

Semantic Fibers and Cross-Gram Interference: A Calculus of Safety Drift in Overcomplete Representations

arXiv 2609.14861 ↗First posted 14 Sep 2026cs.LG cs.AIcited by 0

On the map: LLM safety and security → Safety alignment, refusal and guardrails

Abstract (arXiv)

A deployed language model may refuse a harmful request in English yet comply with its faithful translation, revealing a cross-lingual safety failure that cannot be characterized reliably by output behavior alone. We formalize this phenomenon through an audited equivalence relation and show that, for a declared quotient, representation, metric, feature dictionary, scoring head, threshold, and contrast model, the resulting safety drift admits an exact linear-algebraic characterization. Specifically, the drift is a cross-Gram functional of the within-fiber contrast; its worst admissible value is a support function, while margin invariance is characterized by an annihilator condition. We introduce an intrinsic calibrated exposure measure, governed by the leverage duality chi2=1/ℓ-1, which separates observed drift into three diagnostically distinct regimes: a reader fault removable by recalibration, an exact correction that is too ill-conditioned to be reliable, and a representation-level collision that no readout-only intervention can remove. Thus, identical observed exposure can lead to fundamentally different remediation verdicts. The framework also extends to cone-valued safety heads. An untied order-swap identity provides a diagnostic for the linear control interface; its calibration-state residual predicts a distinct three-control composition error on unseen states and targets, achieving median Spearman correlation 0.964, compared with 0.269 for a static cross-Gram baseline. etc.....

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security. Builds on 24 of its 62 references that are in the map. First posted on arXiv 14 Sep 2026.

Builds on

24 of 62 references.

Closest ancestors · 3

Same direction, newest first

  1. 2026BabelSteering: Multilingual Safety Alignment via English Steering Vectors17 Aug
  2. 2025The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence24 Feb
  3. 2024Refusal in Language Models Is Mediated by a Single Direction17 Jun

Other references · 21

  1. 2014Explaining and Harnessing Adversarial Examplescited by 7,364
  2. 2013Intriguing properties of neural networkscited by 5,155
  3. 2015Domain-Adversarial Training of Neural Networkscited by 3,044
  4. 2019Invariant Risk Minimizationcited by 1,854
  5. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  6. 2019Universal Adversarial Triggers for Attacking and Analyzing NLPcited by 722
  7. 2023Low-Resource Languages Jailbreak GPT-4cited by 309
  8. 2023LEACE: Perfect linear concept erasure in closed formcited by 267
  9. 2023Multilingual Jailbreak Challenges in Large Language Modelscited by 240
  10. 2023Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretabilitycited by 205
  11. 2026Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnetcited by 191
  12. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 143
9 more
  1. 2022Linear Adversarial Concept Erasurecited by 101
  2. 2022Polysemanticity and Capacity in Neural Networkscited by 77
  3. 2025Understanding (Un)Reliability of Steering Vectors in Language Modelscited by 47
  4. 2025Refusal Direction is Universal Across Safety-Aligned Languagescited by 38
  5. 2025Adversarial Examples Are Not Bugs, They Are Superpositioncited by 13
  6. 2026Steering Vector Fields for Context-Aware Inference-Time Control in Large Language Modelscited by 13
  7. 2026Multilingual Steering by Design: Multilingual Sparse Autoencoders and Principled Layer Selectioncited by 4
  8. 2022Generalized Leverage Scores: Geometric Interpretation and Applicationscited by 3
  9. 2026How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordingscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Hard Negatives Reveal What Easy Negatives Hide: Cross-Lingual Harmfulness Representations Degrade with Resource Tier Under Hard Negatives
  2. 2026MMAligner: Safeguarding Multimodal Large Language Models through Representation Calibration
  3. 2025Steering Multimodal Large Language Models Decoding for Context-Aware Safety
  4. 2026Lost in Translation? A Comparative Study on the Cross-Lingual Transfer of Composite Harms
  5. 2026Transfer Safety Awareness for Cross-Modal Safety Drift in Multimodal Large Language Models
  6. 2025LLM Probing with Contrastive Eigenproblems: Improving Understanding and Applicability of CCS
  7. 2025Read the Scene, Not the Script: Outcome-Aware Safety for LLMs
  8. 2025Understanding and Mitigating Over-refusal for Large Language Models via Representation Intervention