Research map / Large language models / Retrieval and knowledge
KScope: A Framework for Characterizing the Knowledge Status of Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a five-status taxonomy of LLM knowledge (consistency × correctness) and KScope, a hierarchical statistical testing framework that characterizes it, then uses context feature analysis to guide context summarization that improves knowledge updates.
Problem
Prior work characterizes LLM knowledge mainly through knowledge conflicts and single sampled responses or entropy, which miss the mode structure of the answer distribution and do not fully describe how well a model knows an answer.
Approach
Define five knowledge statuses from the consistency (size of the knowledge mode set) and correctness (whether the ground truth is in the modes) of an empirical response distribution, and assign each question a status via a hierarchical sequence of likelihood-ratio, multinomial, and binomial tests; then use SHAP-based context feature importance to design constrained summarization and credibility au
Key results
- Supporting gold context increases the proportion of consistent correct knowledge across all datasets and models.
- Combined constrained summarization with credibility improves knowledge update success rate by 4.3% on average across statuses and LLMs.
- Naive summarization always hurts update success, while noisy retrieval and open-ended settings substantially reduce update effectiveness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Provide gold supporting context | LLMs evaluated with KScope · Four datasets, multi-choice | proportion of consistent correct knowledge | Helped |
| Use noisy retrieved context instead of gold context | LLMs evaluated with KScope · HotpotQA, multi-choice | successful update to consistent correct knowledge | Hurt |
| Provide gold context | LLMs evaluated with KScope · HotpotQA, open-ended | proportion of consistent correct knowledge | Helped |
| Use constrained summarization guided by context-feature analysis | LLMs evaluated for knowledge updates · Four datasets, multi-choice with gold context; except consistent wrong status | successful knowledge update rate | Helped |
| Use naive summarization | LLMs evaluated for knowledge updates · Four datasets, multi-choice with gold context | successful knowledge update rate | Hurt |
| Add credibility metadata to context | LLMs evaluated for knowledge updates · Consistent wrong knowledge status | successful knowledge update rate | Helped |
| Combine constrained summarization with credibility metadata | LLMs evaluated for knowledge updates · Four datasets, multi-choice with gold context; all statuses | successful knowledge update rate: +4.3% | Helped |
| Increase model size within a model family | Gemma-2, Llama-3, and Qwen-2.5 · Four datasets, multi-choice with and without gold context | proportion of consistent correct knowledge | Helped |
| Use Llama rather than Qwen or Gemma | Evaluated model families · Four datasets, multi-choice without context | proportion of consistent correct parametric knowledge | Helped |
| Provide context in the healthcare datasets rather than the general-domain datasets | Evaluated LLMs · Hemonc and PubMedQA versus NQ and HotpotQA, multi-choice | proportion of consistent correct knowledge | Hurt |
| Use constrained summarization instead of naive summarization | LLMs evaluated for knowledge updates · Context feature analysis on Hemonc | ROUGE-2 recall | Helped |
Surprising findings
- Naive summarization consistently reduces the success rate of knowledge updates.
- Consistent wrong knowledge shows low feature-importance correlation with other statuses and does not respond to constrained summarization, requiring credibility cues instead.
Limitations
- Feature analysis covers only eleven context features; more stylistic features are unexplored.
- Experiments restricted to three-option classification due to computational constraints.
- Noisy retrieval may contain conflicting information not fully addressed.
Experimental setup
Datasets: Hemonc PubMedQA NQ HotpotQA
Benchmarks: Hemonc PubMedQA NQ HotpotQA
Metrics: proportion of knowledge status success rate of knowledge updates Macro-F1 Spearman rank correlation
Baselines: entropy-based uncertainty single-sample most-likely response naive summarization
Builds on: retrieval-augmented generation SHAP Bonferroni correction
Closest papers by meaning
- KoLA: Carefully Benchmarking World Knowledge of Large Language Models (ICLR 2024 · similarity 0.90)
- Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 2024 · similarity 0.89)
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction (ICML 2024 · similarity 0.89)
- PertEval: Unveiling Real Knowledge Capacity of LLMs with Knowledge-Invariant Perturbations (NeurIPS 2024 · similarity 0.89)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.89)
- Knowledge Localization in Mixture-of-Experts LLMs Using Cross-Lingual Inconsistency (NeurIPS 2026 · 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.