---
title: "Competence Collapse in Code-Mixed Generation: Spectral"
description: "To measure it, the paper introduces Instructional Density, a score that counts imperative sentences… 27 references, cited by 2 papers here."
authors: ["Tanushree Ravindra Pratap Yadav"]
year: 2026
venue: "Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)"
source: "https://aclanthology.org/2026.loreslm-1.3/"
doi: "10.18653/v1/2026.loreslm-1.3"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/yadav-2026-competence-collapse-code-mixed-generation/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/yadav-2026-competence-collapse-code-mixed-generation/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/7/) / Competence Collapse in Code-Mixed Generation: Spectral Evidence and Mechanistic Recovery via Cross-Lingual Activation Steering

# Competence Collapse in Code-Mixed Generation: Spectral Evidence and Mechanistic Recovery via Cross-Lingual Activation Steering

**Authors:** Tanushree Ravindra Pratap Yadav · **Year:** 2026 · **Venue:** [Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)](https://md.research.elamur.ai/free/en/venues/loreslm/) ([Workshop on Language Models for Low-Resource Languages 2026](https://md.research.elamur.ai/free/en/venues/loreslm/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.loreslm-1.3/) · [PDF](https://aclanthology.org/2026.loreslm-1.3.pdf) · [DOI 10.18653/v1/2026.loreslm-1.3](https://doi.org/10.18653/v1/2026.loreslm-1.3) · [OpenAlex](https://openalex.org/W7140123408)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · [2](https://md.research.elamur.ai/free/en/papers/yadav-2026-competence-collapse-code-mixed-generation/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
This paper asks why LLMs that reason well in English give much weaker answers to the same request written in Hinglish (Hindi-English code-mixing). Instead of expert guidance, they often return vague tips or defensive disclaimers. The author calls this Competence Collapse. To measure it, the paper introduces Instructional Density, a score that counts imperative sentences and structural blocks. Native bilingual annotators validated it on a 600-prompt benchmark covering 10 high-stakes domains. An audit of 9 models shows a consistent quality drop, termed the Service Gap. Layer-wise spectral analysis suggests the cause: Hinglish and English representations drift apart along a low-rank direction in the middle layers.

The proposed fix is Cross-Lingual Activation Steering (CLAS). It averages English-minus-Hinglish final-token activation differences over 50 paired prompts. It then adds this vector, norm-scaled, to the residual stream at prompt tokens only. Across 6 open-weight models, CLAS improves utility with no weight updates, keeps outputs code-mixed and reportedly strengthens safety. This offers a cheap, training-free route to fairer service for code-mixed language users.

## Key points
- **Service Gap:** Instructional Density falls by about 11.3% (p \< 0.001) on Hinglish prompts across 9 models, including Gemini 1.5 Flash, Llama 3.1 8B, Qwen 2.5 7B and Phi-3.5 Mini. The benchmark has 600 prompts (100 seeds × 6 persona variants), each rated by 3 annotators. Agreement between the Llama3 judge and humans is Spearman ρ = 0.87.
- **Steerability Window:** The "Steerability Window" spans layers 11–15. PC1 explains about 26% of the variance at layer 13. Steering is applied at layer L/2 with α = 1.0.
- **CLAS results:** The global gain is about +2.1 to +2.2 D (the abstract reports d = 0.60; Table 1 reports d = 0.57). Phi-3.5 Mini gains the most (+5.61, d = 2.89). Qwen 2.5 7B shows no gain (−0.71, described as an "alignment ceiling"). The Code-Mixing Index (CMI) stays at about 0.4, and jailbreaks drop by 52.5%.
- **Baselines on Phi-3.5 Mini (ΔD):**
  - CLAS: +5.78
  - Global CAA: +5.60
  - LoRA SFT (N = 50): +4.93, but with a spike in assertiveness, which the paper flags as a risk of safety unlearning
  - Translate-first: +0.12
  - Random vector: −2.45
- **Trade-offs and limits:**
  - Higher α makes the tone more clinical and less empathetic. The paper suggests α of 0.3–0.5 for counseling and 0.8–1.0 for medical triage.
  - At α > 2.0, outputs drift into plain English.
  - The method was validated mainly on Hinglish.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 2,869 works

🧭 Topics (2): [🌍 Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) · 828 works · [🔬 Interpretability](https://md.research.elamur.ai/free/en/topics/interpretability/) · 130 works

## Cite

Cite as: Tanushree Ravindra Pratap Yadav. 2026. Competence Collapse in Code-Mixed Generation: Spectral Evidence and Mechanistic Recovery via Cross-Lingual Activation Steering. Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026).

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{yadav2026competence,
  title = {{Competence Collapse in Code-Mixed Generation: Spectral Evidence and Mechanistic Recovery via Cross-Lingual Activation Steering}},
  author = {Tanushree Ravindra Pratap Yadav},
  booktitle = {Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.loreslm-1.3},
  url = {https://doi.org/10.18653/v1/2026.loreslm-1.3}
}
```

- [← previous Comparing Human and Large Language Model Interpretation of Implicit Information](https://md.research.elamur.ai/free/en/papers/santis-2026-comparing-human-large-language-model/)
- [next → Confidence over Time: Confidence Calibration with Temporal Logic for Large Language Model Reasoning](https://md.research.elamur.ai/free/en/papers/mao-2026-confidence-over-time-confidence-calibration/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/7/)

## References

27 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Gustavo Aguilar, Thamar Solorio, Mona Diab, and Julia Hirschberg. 2020. [LinCE: A centralized benchmark for linguistic code-switching evaluation](https://md.research.elamur.ai/free/en/papers/aguilar-2020-lince-centralized-benchmark-linguistic-code/). In Proceedings of the 12th Language Resources and Evaluation Conference (LREC), pages 1803–1813.
2. Kalai Mani, Pierre Godey, and Maarten Sap. 2025. Bridging the script gap: Transliteration-augmented decoding for low-resource language generation. arXiv preprint arXiv:2512.10780. — [arXiv:2512.10780](https://arxiv.org/abs/2512.10780) [find](https://scholar.google.com/scholar?q=Kalai+Mani%2C+Pierre+Godey%2C+and+Maarten+Sap%2E+2025%2E+Bridging+the+script+gap%3A+Transliteration%2Daugmented+decoding+for+low%2Dresource+language+generation%2E+arXiv+preprint+arXiv%3A2512%2E10780%2E+%E2%80%94+%5BarXiv%3A2512%2E10780%5D)
3. Roee Aharoni, Melvin Johnson, and Orhan Firat. 2019. [Massively multilingual neural machine translation in the wild: Findings and challenges](https://md.research.elamur.ai/free/en/papers/arivazhagan-2019-massively-multilingual-neural-machine-translation/). In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3874–3888.
4. Tyler A Chang, David Deutsch, and Graham Neubig. 2023. The curse of multilinguality in language models. arXiv preprint arXiv:2305.12038. — [arXiv:2305.12038](https://arxiv.org/abs/2305.12038) [find](https://scholar.google.com/scholar?q=Tyler+A+Chang%2C+David+Deutsch%2C+and+Graham+Neubig%2E+2023%2E+The+curse+of+multilinguality+in+language+models%2E+arXiv+preprint+arXiv%3A2305%2E12038%2E+%E2%80%94+%5BarXiv%3A2305%2E12038%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2305%2E12038%29)
5. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Prafulla Mishkin, Chak M Zhang, and Dario Amodei. 2022. [Training language models to follow instructions with human feedback](https://md.research.elamur.ai/free/en/papers/ouyang-2022-training-language-models-follow-instructions/). arXiv preprint arXiv:2203.02155.
6. Seonglae Cho, Tae Ham, Bosung Min, and Jaesik Park. 2025. Corrsteer: Generation-time llm steering via correlation estimation. arXiv preprint arXiv:2508.12535. — [arXiv:2508.12535](https://arxiv.org/abs/2508.12535) [find](https://scholar.google.com/scholar?q=Seonglae+Cho%2C+Tae+Ham%2C+Bosung+Min%2C+and+Jaesik+Park%2E+2025%2E+Corrsteer%3A+Generation%2Dtime+llm+steering+via+correlation+estimation%2E+arXiv+preprint+arXiv%3A2508%2E12535%2E+%E2%80%94+%5BarXiv%3A2508%2E12535%5D%28https%3A%2F%2Farxiv%2Eorg%2Fab)
7. Seungwoo Park, Elman Mansimov, Rajesh Joshi, and Minh Le-Khac. 2025. Cross-lingual alignment via contrastive direct preference optimization. arXiv preprint arXiv:2505.12584. [find](https://scholar.google.com/scholar?q=Seungwoo+Park%2C+Elman+Mansimov%2C+Rajesh+Joshi%2C+and+Minh+Le%2DKhac%2E+2025%2E+Cross%2Dlingual+alignment+via+contrastive+direct+preference+optimization%2E+arXiv+preprint+arXiv%3A2505%2E12584%2E)
8. Sasha Cui, Cengiz Demircigil, Cem Anil, Samuel Hedetniemi, and Lin Tong. 2025. Painless activation steering. arXiv preprint arXiv:2509.22739. — [arXiv:2509.22739](https://arxiv.org/abs/2509.22739) [find](https://scholar.google.com/scholar?q=Sasha+Cui%2C+Cengiz+Demircigil%2C+Cem+Anil%2C+Samuel+Hedetniemi%2C+and+Lin+Tong%2E+2025%2E+Painless+activation+steering%2E+arXiv+preprint+arXiv%3A2509%2E22739%2E+%E2%80%94+%5BarXiv%3A2509%2E22739%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2509%2E22739%29)
9. Nina Rimsky, Alex Turner, Tao Conerly, Nathan Tsoi, Calum Rager, Sumit Arora, and Jacob Steinhardt. 2024. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), pages 11000–11020. [find](https://scholar.google.com/scholar?q=Nina+Rimsky%2C+Alex+Turner%2C+Tao+Conerly%2C+Nathan+Tsoi%2C+Calum+Rager%2C+Sumit+Arora%2C+and+Jacob+Steinhardt%2E+2024%2E+Steering+llama+2+via+contrastive+activation+addition%2E+In+Proceedings+of+the+62nd+Annual+Meetin)
10. Yue Deng, Wenxuan Li, Paul Alphonse, and Yuanshun Zhang. 2023. [Multilingual jailbreak challenges in large language models](https://md.research.elamur.ai/free/en/papers/deng-2023-multilingual-jailbreak-challenges-large-language/). arXiv preprint arXiv:2310.06474.
11. Joshua Robinson, Rowan Zellers, and Yonatan Bisk. 2024. Safety risks in code-mixed language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics: Findings (ACL Findings), pages 12456–12472. [find](https://scholar.google.com/scholar?q=Joshua+Robinson%2C+Rowan+Zellers%2C+and+Yonatan+Bisk%2E+2024%2E+Safety+risks+in+code%2Dmixed+language+models%2E+In+Proceedings+of+the+62nd+Annual+Meeting+of+the+Association+for+Computational+Linguistics%3A+Findings)
12. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. [BERT: Pre-training of deep bidirectional transformers for language understanding](https://md.research.elamur.ai/free/en/papers/devlin-2019-bert-pre-training-deep-bidirectional/). In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT), pages 4171–4186.
13. Rajvee Sheth, Aman Ganu, Rajesh Sinha, and Priya Kumari. 2025. COMI-LINGUA: Expert annotated large-scale dataset for multitask NLP in Hindi-English code-mixing. In Findings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7973–7992. [find](https://scholar.google.com/scholar?q=Rajvee+Sheth%2C+Aman+Ganu%2C+Rajesh+Sinha%2C+and+Priya+Kumari%2E+2025%2E+COMI%2DLINGUA%3A+Expert+annotated+large%2Dscale+dataset+for+multitask+NLP+in+Hindi%2DEnglish+code%2Dmixing%2E+In+Findings+of+the+2025+Conference+on+E)
14. Deep Ganguli, Danny Hernandez, Liane Lovitt, Michael Sellitto, Danielle Rai, Tao Conerly, Amanda Askell, Danny Drain, Gal Farquhar, Ethan Fort, and 1 others. 2022. [Red teaming language models with language models](https://md.research.elamur.ai/free/en/papers/perez-2022-red-teaming-language-models-language/). arXiv preprint arXiv:2202.03286.
15. Thamar Solorio, Elizabeth Blair, Suraj Maharjan, Steven Bethard, Mona Diab, Debbie Ghosh, Abdelati Hawwari, Julia Ho, Chanda Kambhampati, Sepideh Mansouri, and 1 others. 2014. [Overview for the first shared task on language identification in code-switched data](https://md.research.elamur.ai/free/en/papers/solorio-2014-overview-first-shared-task-language/). pages 62–72.
16. Amir Karimi, Gilles Barthe, Borja Balle, and Isabel Casamayor. 2021. Certifying some distributional robustness with principled adversarial training. In Proceedings of the International Conference on Learning Representations (ICLR), pages 5652– 5665. [find](https://scholar.google.com/scholar?q=Amir+Karimi%2C+Gilles+Barthe%2C+Borja+Balle%2C+and+Isabel+Casamayor%2E+2021%2E+Certifying+some+distributional+robustness+with+principled+adversarial+training%2E+In+Proceedings+of+the+International+Conference+on+L)
17. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Arindam Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. [Llama 2: Open foundation and fine-tuned chat models](https://md.research.elamur.ai/free/en/papers/touvron-2023-llama-2-open-foundation-fine/). arXiv preprint arXiv:2307.09288.
18. Simran Khanuja, Sandipan Dandapat, Gholamreza Sinha, and Sunayana Sitaram. 2020. [GLUECoS: An evaluation benchmark for code-switched NLP](https://md.research.elamur.ai/free/en/papers/khanuja-2020-gluecos-evaluation-benchmark-code-switched/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5123–5135.
19. Alex Turner, Nathan Tsoi, Calum Rager, Tao Conerly, and David Atkinson. 2023. [Activation addition: Steering language models without optimization](https://md.research.elamur.ai/free/en/papers/turner-2023-steering-language-models-activation-engineering/). arXiv preprint arXiv:2308.10248.
20. Manurag Khullar, Utkarsh Desai, Poorva Malviya, Aman Dalmia, and Zheyuan Ryan Shi. 2025. Script gap: Evaluating llm triage on indian languages in native vs roman scripts in a real world setting. arXiv preprint arXiv:2512.10780. — [arXiv:2512.10780](https://arxiv.org/abs/2512.10780) [find](https://scholar.google.com/scholar?q=Manurag+Khullar%2C+Utkarsh+Desai%2C+Poorva+Malviya%2C+Aman+Dalmia%2C+and+Zheyuan+Ryan+Shi%2E+2025%2E+Script+gap%3A+Evaluating+llm+triage+on+indian+languages+in+native+vs+roman+scripts+in+a+real+world+setting%2E+arXiv)
21. Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Ekin D Cubuk, Quoc V Chi, and Denny Zhou. 2022. [Emergent abilities of large language models](https://md.research.elamur.ai/free/en/papers/wei-2022-emergent-abilities-large-language-models/). arXiv preprint arXiv:2206.07682.
22. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Valeriu Ves, and Luke Schwab. 2019. [BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension](https://md.research.elamur.ai/free/en/papers/lewis-2020-bart-denoising-sequence-sequence-pre/). arXiv preprint arXiv:1910.13461.
23. Shuyuan Xie, Suraj Maharjan, and Thamar Solorio. 2023. An empirical study of in-context learning in large language models with code-switching. arXiv preprint arXiv:2305.09066. — [arXiv:2305.09066](https://arxiv.org/abs/2305.09066) [find](https://scholar.google.com/scholar?q=Shuyuan+Xie%2C+Suraj+Maharjan%2C+and+Thamar+Solorio%2E+2023%2E+An+empirical+study+of+in%2Dcontext+learning+in+large+language+models+with+code%2Dswitching%2E+arXiv+preprint+arXiv%3A2305%2E09066%2E+%E2%80%94+%5BarXiv%3A2305%2E09066%5D%28htt)
24. Result: In a sweep with a conservative steering coefficient (α = 0.05), Phi-3.5 showed clear utility gains (∆D = +0.40, ∆Assertiveness = +0.13). Zephyr 7B exhibited milder improvements (∆D = +0.20), confirming cross-lingual transferability. [find](https://scholar.google.com/scholar?q=Result%3A+In+a+sweep+with+a+conservative+steering+coefficient+%28%CE%B1+%3D+0%2E05%29%2C+Phi%2D3%2E5+showed+clear+utility+gains+%28%E2%88%86D+%3D+%2B0%2E40%2C+%E2%88%86Assertiveness+%3D+%2B0%2E13%29%2E+Zephyr+7B+exhibited+milder+improvements+%28%E2%88%86D+%3D+%2B0%2E20%29%2C+c)
25. Yilun Yang and Yekun Chai. 2025. Codemixbench: Evaluating code-mixing capabilities of llms across 18 languages. arXiv preprint arXiv:2507.18791. [find](https://scholar.google.com/scholar?q=Yilun+Yang+and+Yekun+Chai%2E+2025%2E+Codemixbench%3A+Evaluating+code%2Dmixing+capabilities+of+llms+across+18+languages%2E+arXiv+preprint+arXiv%3A2507%2E18791%2E)
26. Andy Zou, Long Xiao, Ying Mu, Wei Liu, Liane Sandoval, Abhay Ramasamy, Alex Tamkin, Amanda Askell, Gabriel Fournier, Jeff Wu, and 1 others. 2023. [Representation engineering: A top-down approach to ai transparency](https://md.research.elamur.ai/free/en/papers/zou-2023-representation-engineering-top-down-approach/). arXiv preprint arXiv:2310.01405.
27. A • Comparison to Hinglish: The magnitude of recovery is lower than in Hinglish (∆D ≈ +2.22). We hypothesize this is because Spanish and English share higher lexical overlap and a more convergent latent topology than Hindi-English, offering less geometric leverage for steering to act upon at low α v… [find](https://scholar.google.com/scholar?q=A+%E2%80%A2+Comparison+to+Hinglish%3A+The+magnitude+of+recovery+is+lower+than+in+Hinglish+%28%E2%88%86D+%E2%89%88+%2B2%2E22%29%2E+We+hypothesize+this+is+because+Spanish+and+English+share+higher+lexical+overlap+and+a+more+convergent+late)

*Updated 31 October 2026*