Research map / Large language models / Retrieval and knowledge
Leveraging Pretrained Knowledge at Inference Time: LoRA-Gated Contrastive Decoding for Multilingual Factual Language Generation in Adapted LLMs
Gwangseon Jang, Hongseok Choi, Chanuk lim, Kyong-Ha Lee, Mun Yong Yi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LGCD contrastive decoding to the language-adapted model | 12 language-adapted LLMs · Global MMLU 0-shot | accuracy: 0.441 -> 0.477 | Improved |
| add LGCD to the language-adapted model | 12 language-adapted LLMs · Global MMLU 5-shot | accuracy: 0.481 -> 0.498 | Improved |
| add LGCD to the language-adapted model | 6 language-adapted models · Multilingual TruthfulQA 0-shot | accuracy: 0.334 -> 0.376 | Improved |
| add LGCD to the language-adapted model | 6 language-adapted models · Multilingual TruthfulQA 5-shot | accuracy: 0.367 -> 0.435 | Improved |
| add LGCD to the language-adapted model | 9 language-adapted models · Multi-FAct | factuality score: 0.272 -> 0.312 | Mixed |
| use DoLa decoding instead of nucleus sampling | language-adapted models · Multilingual TruthfulQA 0-shot | accuracy: 0.334 -> 0.315 | Hurt |
| approximate PTM with attention Q/V layers instead of FFN layers (ablation) | LGCD · Global MMLU 0-shot | accuracy: 0.465 -> 0.477 | Improved |
| approximate PTM with all layers (FFN + attention) instead of FFN only (ablation) | LGCD · Global MMLU 0-shot | accuracy: 0.472 -> 0.477 | Improved |
| use online adaptive confidence threshold instead of heuristic language-tier threshold (ablation) | LGCD · Global MMLU 0-shot | accuracy: 0.477 -> 0.473 | No clear effect |
| use online adaptive confidence threshold instead of heuristic language-tier threshold (ablation) | LGCD · Global MMLU 5-shot | accuracy: 0.498 -> 0.495 | No clear effect |
| raise confidence threshold from 0.2 to 0.8 (ablation) | LGCD · Decoding throughput | tokens/s: 14.37 -> 10.22 | Hurt |
| apply LGCD to extremely low-resource languages | llama-8b-south-africa · Translated Global MMLU 0-shot | accuracy: 0.452 -> 0.459 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DoLa — DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models (ICLR 2024) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models (ICLR 2024 · similarity 0.90)
- Refine Knowledge of Large Language Models via Adaptive Contrastive Learning (ICLR 2025 · similarity 0.89)
- AdaMCoT: Rethinking Cross-Lingual Factual Reasoning Through Adaptive Multilingual Chain-of-Thought (AAAI 2026 · similarity 0.88)
- SLED: Self Logits Evolution Decoding for Improving Factuality in Large Language Models (NeurIPS 2024 · similarity 0.88)
- Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators (AAAI 2025 · similarity 0.88)
- When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
