Research map / Large language models / Language, safety and interpretability
Attribution-Guided Decoding
Piotr Komorowski, Elena Golimblevskaia, Reduan Achtibat, Thomas Wiegand, Sebastian Lapuschkin, Wojciech Samek
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 |
|---|---|---|---|
| Use LRP attribution instead of Input x Gradient in AGD (ablation) | AGD · IHEval | Prompt-level accuracy: 70.8 -> 79.1 (Llama 3.1 8B) | Improved |
| Add entropy-gating to AGD LRP (ablation) | AGD LRP · IHEval | Quality score: 73.2 -> 76.4 (Llama 3.1 8B) | Improved |
| Add entropy-gating to AGD LRP (ablation) | AGD LRP · IHEval | Prompt-level accuracy: 79.1 -> 74.5 (Llama 3.1 8B) | Hurt |
| Add entropy-gating to AGD LRP (ablation) | AGD LRP · Open-book TriviaQA efficiency | Cost relative to greedy: 2.20x -> 1.60x (Llama 3.1 8B) | Improved |
| Apply AGD LRP guidance at every step | Llama 3.1 (8B) · IHEval | Quality score: 81.3 -> 73.2 | Hurt |
| Use AGD with LRP on system prompt embeddings vs greedy decoding | Llama 3.1 (8B) · IHEval | Prompt-level accuracy: 66.0 -> 79.1 | Improved |
| Lower the minimum probability threshold to extremely low values (ablation) | AGD · IHEval | Adherence and quality | Hurt |
| Raise the minimum probability threshold (ablation) | AGD · IHEval | Prompt-level accuracy | Hurt |
| Use only 10 heads instead of 100 (ablation) | AGD LRP heads · TriviaQA closed-book | Recall: 82.4 -> 81.8 | Hurt |
| Use more heads (200 instead of 100) (ablation) | AGD LRP heads · TriviaQA closed-book | Recall: 82.4 -> 82.4 | No clear effect |
| Use AGD toward context embeddings vs greedy decoding in noisy retrieval | Llama 3.1 (8B) · HotPotQA distractor | Recall: 81.3 -> 87.9 | Improved |
| Guide towards parametric knowledge heads in closed-book QA | Gemma 3 (4B) · TriviaQA closed-book | Recall: 61.9 -> 61.5 | No clear effect |
Closest papers by meaning
- Reference Trustable Decoding: A Training-Free Augmentation Paradigm for Large Language Models (NeurIPS 2024 · similarity 0.88)
- Improving Open-Ended Text Generation via Adaptive Decoding (ICML 2024 · similarity 0.87)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.87)
- Reward-Guided Speculative Decoding for Efficient LLM Reasoning (ICML 2025 · similarity 0.87)
- Instructive Decoding: Instruction-Tuned Large Language Models are Self-Refiner from Noisy Instructions (ICLR 2024 · similarity 0.87)
- MBR and QE Finetuning: Training-time Distillation of the Best and Most Expensive Decoding Methods (ICLR 2024 · similarity 0.86)
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.
