Research map / Large language models / Retrieval and knowledge
Bridging the Knowledge-Prediction Gap in LLMs on Multiple-Choice Questions
Yoonah Park, Haesung Pyun, Yohan Jo
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 |
|---|---|---|---|
| apply KAPPA alignment of prediction and knowledge subspaces at 6 layers | Llama 3.1 8B Instruct · TruthfulQA (MCQ) | accuracy: 56.7 -> 73.5 | Improved |
| apply KAPPA at 6 layers | Mistral v0.3 7B · TruthfulQA (MCQ) | accuracy: 40.7 -> 58.3 | Improved |
| apply KAPPA at 6 layers | Qwen2.5 7B Instruct · GSM8k (MCQ) | accuracy: 47.8 -> 49.2 | Improved |
| use CAA activation steering instead of KAPPA | Llama 3.1 8B Instruct · TruthfulQA (MCQ) | accuracy: 56.7 -> 62.3 (CAA) vs 73.5 (KAPPA) | Improved |
| use DoLA decoding as baseline intervention | Llama 3.1 8B Instruct · TruthfulQA (MCQ) | accuracy: 56.7 -> 55.6 | No clear effect |
| increase number of intervened layers from 1 to 6 (ablation) | KAPPA · TruthfulQA (MCQ) | accuracy: 67.8 -> 73.5 | Improved |
| apply KAPPA on knowledge-intensive benchmarks with negligible gap | KAPPA · knowledge MCQ benchmarks (e.g., MMLU, ARC) | accuracy | No clear effect |
| increase alpha or beta in extended alignment (ablation) | KAPPA · TruthfulQA | agreement rate | Improved |
| reduce probe training data to 10% (ablation) | KAPPA · probe data size sweep | accuracy vs base | No clear effect |
| transfer probes across datasets (ablation) | KAPPA · cross-dataset agreement (Llama 3.1 8B) | agreement improvement: +4.81 points BBH-NLP to GSM8k; +5.72 points TruthfulQA to BBQ-Age | Mixed |
| apply KAPPA to free-form generation using MCQ-derived subspaces | Qwen2.5 7B Instruct · TruthfulQA free-form | accuracy: 41.7 -> 44.2 | Improved |
| apply KAPPA to free-form generation | Qwen2.5 7B Instruct · GSM8k free-form | accuracy: 91.6 -> 90.7 | Hurt |
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 2 of 2 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
- Potemkin Understanding in Large Language Models (ICML 2025 · similarity 0.90)
- Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models (NeurIPS 2026 · similarity 0.90)
- Geometry of Decision Making in Language Models (NeurIPS 2025 · similarity 0.90)
- Evaluating language models as risk scores (NeurIPS 2024 · similarity 0.90)
- Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions (ICLR 2025 · similarity 0.89)
- Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · similarity 0.89)
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.
