Knowledge improvement in large language models is shifting from predefined knowledge incorporation pipelines toward evaluation-guided updating. Existing methods use evaluation to localize model-specific knowledge weaknesses and guide targeted post-training. However, identifying what to improve does not necessarily provide sufficient information for deciding how to improve it. The same localized weakness can be compatible with distinct operational knowledge states: the relevant knowledge may be unavailable, available but difficult to access reliably, or accessible but ineffective for downstream use. We call this ambiguity knowledge-state aliasing. When compatible states favor different interventions, greater diagnostic resolution alone is insufficient for intervention selection. Moreover, model updates can change the underlying knowledge state before the observed weakness disappears, such that an intervention appropriate in one round may no longer address the bottleneck in the next. To address these challenges, we propose Recursive Knowledge Improvement (RKI), which formulates knowledge improvement as a sequential, state-aware intervention process. RKI organizes knowledge into graph-structured units that provide a shared basis for diagnostic evaluation and targeted intervention across rounds. At each round, RKI localizes knowledge weaknesses, estimates the model's operational knowledge state from diagnostic evidence and intervention history, selects a corresponding intervention, and re-evaluates the updated model before determining the next intervention. On HotpotEval and SafeInternalEval, RKI outperforms existing knowledge incorporation methods by 18.41 points on average. Further analyses show that knowledge states provide additional guidance for intervention selection beyond weakness localization and can change across model updates even when the observed weakness persists.
Figures & tables
Figure 1: From weakness localization to state-aware knowledge improvement. RKI estimates operational knowledge states to guide interventions and re-estimates them after each model update.
Figure 2: Overview of RKI. RKI aligns diagnosis and intervention over graph-structured knowledge units and recursively adapts interventions to the model’s evolving operational knowledge state.
Dataset
Domain
Language
Samples
Avg. Tokens
Max. Tokens
Corpus
Test
Corpus
Test
Corpus
Test
HotpotEval
General
English
9,990
1,000
137.5
21.3
1,984
74
SafeInternalEval
Security & Compliance
Chinese
1,000
600
352.4
41.2
701
139
Table 1: Statistics of the datasets used in our experiments. The corpus serves as the knowledge source for graph construction, diagnostic evaluation, and knowledge incorporation, while the test set is used for closed-book evaluation.
Unit Type
Units
Proportion
Relation
14,439
44.45%
Ego
1,340
4.12%
Contrast
1,066
3.28%
Cycle
975
3.00%
Chain
13,494
41.54%
Tree
1,171
3.60%
Table 2: Distribution of graph-structured knowledge units constructed from SafeInternalEval .
Figure 3: Knowledge-state aliasing under the same localized failure signal. Left: recognition outcomes for recall-failure instances. Right: recognition and recall patterns for understanding-, reasoning-, and application-failure units.
Method
HotpotEval
SafeInternalEval
Avg.
GENIE
23.00
9.75
16.38
LongForm
24.40
11.17
17.79
WRAP
22.90
9.75
16.33
Self-QA
26.80
28.00
27.40
EntiGraph
26.50
21.50
24.00
GraphGen
23.70
22.33
23.02
Table 3: Performance of different knowledge incorporation methods. Bold marks the best result.
Figure 7
Method
Average Score
Random Intervention
29.33
Fixed Access
23.17
Fixed Utilization
48.00
Full-Data Training
50.50
Weakness Only
53.75
Frozen State
39.00
Table 4: Ablation study on intervention selection on SafeInternalEval . Bold marks the best result.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Unit
Role
Relation
Represents a direct relational fact between two entities as the basic unit of relational knowledge.
Contrast
Organizes related facts or local structures along a shared comparison dimension.
Ego
Represents the local relational neighborhood centered on a focal entity.
Cycle
Represents a closed relational path formed by multiple entities and their relations.
Chain
Represents a multi-hop relational path connecting a sequence of entities.
Tree
Represents a branching relational structure extending from a shared local context.
Appendix
Table 5: Graph-structured knowledge units in RKI and their functional roles.
Dimension
Aspect
Interpretation
Acquisition
Availability
Whether the relevant knowledge shows evidence of having been incorporated into the model.
Recognition
Accessibility
Whether the model can identify the relevant knowledge when candidate information or relevant cues are provided.
Recall
Accessibility
Whether the model can reliably retrieve the relevant knowledge without explicit candidate cues.
Understanding
Usability
Whether the model can correctly interpret the relations and meaning expressed by the relevant knowledge.
Reasoning
Usability
Whether the model can combine and derive information from the relevant knowledge through reasoning.
Application
Usability
Whether the model can effectively use the relevant knowledge in downstream tasks or problem settings.
Appendix
Table 6: Diagnostic dimensions, operational aspects, and interpretations for knowledge-state estimation.
Figure 8: Local structure of the knowledge graph constructed from the Hotpot corpus. (a) The complete two-hop neighborhood of Star Wars , with 130 entities and 162 edges. Outlined nodes are expanded in (b). (b) The induced subgraph of 15 selected entities and their 25 edges. Colors and markers denote entity types. All edges are undirected.
Parameter
CPT
SFT
Context Length
2,048
2,048
Global Batch Size
16
64
Epochs
1
1
Peak Learning Rate
5×10−6
2×10−5
LR Scheduler
Cosine
Linear
Warmup Ratio
0.05
0.03
Appendix
Table 7: Training configurations for CPT and SFT.
Dataset
Corpus
Test
HotpotEval
Bolax is a genus of flowering plants in the Apiaceae, with four or five species, and is endemic to temperate South America. Jovibarba heuffelii, commonly known as Hen-and-chickens, is a plant species native to the Balkans and the Carpathians in Europe. …
Question: Which plant has more species, Bolax or Jovibarba? Answer: Bolax.
Appendix
Table 8: Representative source-corpus and evaluation examples from the public benchmarks.
Benchmark
Dimension
Evaluation Criterion
HotpotEval
final_fact
Whether the final answer states the same fact as the reference answer.
completeness
Whether the answer covers the required constituents and qualifiers of the reference answer.
SafeInternalEval Understanding
factual_accuracy
Whether every fact stated in the answer is correct; hallucinated facts lose points.
association_coverage
How many of the required answer points the answer covers. Paraphrase and superordinate summaries count as coverage; a point mentioned with a wrong association does not.
conclusion_correctness
Whether the core judgment (overall positioning, the difference identified, or the misconception clarified) matches the reference, with the correct characterisation basis.
SafeInternalEval Reasoning and Application
factual_accuracy
Whether every fact stated in the answer is correct; hallucinated facts lose points.
Appendix
Table 9: Judging rubrics across benchmarks. Each dimension is scored from 0 to 10, with scores of at least 6 considered passing. For the internal benchmark, understanding evaluates association coverage, while reasoning and application evaluate reasoning coverage.
Benchmark
Pass Agreement
Cohen’s κ
HotpotEval
98%
0.95
SafeInternalEval
96%
0.89
Appendix
Table 10: Agreement between LLM judgments and human annotations on 100 stratified instances per benchmark.
LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.
Boyu Qiao, Sean Guo, Xian Yang +4
Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences · Hong Kong University of Science and Technology +1
Constantly evolving real-world knowledge necessitates models to be updated continuously. Especially in medicine, as clinical evidence changes over time, outdated knowledge can pose safety risks. Existing evaluations of knowledge integration focus on factual recall, offering limited insight into whether newly integrated knowledge is actually usable. Our benchmark MedKIT (Medical Knowledge Integration and Transfer) provides a granular evaluation of how models integrate and apply knowledge under realistic sequences of clinical updates. Each instance corresponds to a factual update derived from clinical evidence, paired with targeted probes that assess transfer across lexical variation, relational transformations, compositional reasoning, and open-ended operationalization, as well as locality tests for knowledge preservation. Using MedKIT, we conduct a large-scale empirical study of 12 knowledge integration strategies across 5 diverse models, including both general-purpose and medical LLMs. Our results reveal a consistent gap between recall and usable knowledge: while most methods achieve strong gains on the original update task and under lexical variation, relational generalization is limited, and no method yields meaningful improvements on compositional or operational tasks. These findings highlight a fundamental challenge in knowledge integration and position MedKIT as a testbed for developing methods that make newly integrated knowledge more consistently usable across tasks and contexts.
Lukas Thede, Yash Kumar Atri, David Chen +4
University of Tübingen, Tübingen AI Center · Helmholtz Munich · Munich Center for Machine Learning (MCML) +4
Large language models (LLMs) often produce incorrect or outdated content after being employed. Efficient and accurate knowledge updates without costly retraining are a major challenge. This problem is particularly challenging in lifelong settings, where complex, unstructured knowledge must coexist without interference. We introduce RILKE (Representation Intervention for Lifelong KnowledgE Control), a robust and scalable method that treats knowledge control as interventions within the model's representation space. Leveraging representation-space expressiveness, we identify two key properties enabling RILKE to achieve fine-grained control over complex, unstructured knowledge while maintaining general utility with frozen base weights. During training, RILKE learns paraphrase-robust and edit-localized modules that limit each update to a low-dimensional subspace to minimize cross-edit interference. At inference, a query-adaptive router selects the appropriate module to guide the model's generation. Across LLaMA and Qwen models, RILKE scales effectively to large-scale benchmarks, demonstrating high edit success and strong paraphrase generalization while preserving general utility with modest memory overhead. These results show RILKE is an effective and scalable solution for lifelong knowledge control in LLMs.
Xuyuan Liu, Shengyu Chen, Xinshuai Dong +6
Dartmouth College · NEC Laboratories America · Carnegie Mellon University