Competence Collapse in Code-Mixed Generation: Spectral Evidence and Mechanistic Recovery via Cross-Lingual Activation Steering
Year: 2026
Type: conference-paper
Abstract: As Large Language Models (LLMs) approach human-level reasoning in English, their performance in low-resource, code-mixed languages remains surprisingly brittle.We identify Competence Collapse, a distinct pathology where models capable of complex reasoning in English exhibit severe utility degradation when prompted in Hinglish (Hindi-English).We quantify this as a Service Gap, observing a statistically significant decline in instructional quality (∆D ≈ -11.3%, p < 0.001) across 9 diverse architectures.Spectral analysis suggests that this stems from a representational divergence between the model's High-Utility Direction and its Generation Subspace.To bridge this gap, we propose Cross-Lingual Activation Steering (CLAS), an inferencetime intervention that injects a "Competence Gap Vector" into the residual stream.Evaluated across 6 open-weight models (using a lightweight calibration set, N = 50), CLAS recovered utility by ∆D = +2.22(d = 0.60) while preserving code-mixed fidelity (CMI ≈ 0.4) and reinforcing safety protocols.
Author: Tanushree Ravindra Pratap Yadav
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7140123408