An arXiv preprint reports a higher edit-success score for MO-IKE than for DR-IKE in a Llama-3.2 benchmark comparison. The reported scores were 92.0 for MO-IKE and 85.0 for DR-IKE, in an evaluation where the language-model parameters remained frozen.
On a separate UniEdit evaluation, MO-IKE's reported overall Score was 24.5, with edit success at 51.3, paraphrase consistency at 28.3 and retention rate at 14.3. The paper reported that row as the best overall Score and retention rate among the listed methods in that evaluation.
Choosing which examples to show
The study focuses on how to construct the prompt for in-context knowledge editing. MO-IKE treats retrieval as a constrained decision process that selects and orders demonstrations for each editing request, while the evaluated language-model parameters remain frozen.
Its action space spans three kinds of demonstrations: COPY, UPDATE and RETAIN. The reinforcement-learning setup applies reward penalties for degradation in paraphrase performance and retention, two of the concerns the system is designed to balance.
The retriever was trained on 300 randomly selected examples. Evaluation used held-out sets of 300 CounterFact samples and 100 samples each from ZsRE and WikiData CounterFact.
Results varied across models and benchmarks
For Llama-3.2, reported edit success was 85.0 for DR-IKE and 92.0 for MO-IKE. For Mistral-v0.3, the reported retention-rate margin over DR-IKE was +22.7, while retention magnitude was -22.5 with DR-IKE and +32.1 with MO-IKE.
Across the listed datasets, MO-IKE had the highest reported overall Scores on ZsRE and WikiData CounterFact, at 42.0 and 59.0, respectively, and the strongest reported retention rate.
The native and zero-shot results were similar. In the native setting, the reported overall Score, edit success, paraphrase consistency and retention rate were 77.1, 80.1, 90.7 and 65.0. In the zero-shot setting, they were 77.3, 81.0, 89.3 and 65.7.
Retention remained the weaker measure
An ablation that removed the reward constraints reported a lower overall Score of 65.5 and a retention rate of 50.0. With all constraints included, the reported overall Score was 73.5, with edit success at 92.0, paraphrase consistency at 79.0 and retention at 57.7.
The action-space comparison also favored the broader choice of demonstrations. Allowing the retriever to use the full set of COPY, UPDATE and RETAIN choices was reported to perform better across all three evaluated metrics than optimizing solely for RETAIN demonstrations.
The all-constraints result still put retention at 57.7, below edit success at 92.0 and paraphrase consistency at 79.0. In the reported balance of objectives, retaining unrelated information remained a weak point.
A benchmark-specific result
The highest overall rates in the model-scale comparison were reported for models with parameter scales of at least 7 billion. An additional model evaluation reported MO-IKE scores of 70.2 overall, 98.3 for edit success, 94.3 for paraphrase consistency and 53.0 for retention.
The evidence covers benchmark examples and outputs from frozen models, with the retriever trained and tested on the reported held-out sets. It therefore describes performance in those comparisons, rather than establishing how the method would behave in broader datasets or real-world deployments.
The remaining questions include how to select or adapt the constraint weights, whether the retention bottleneck can be overcome, and whether the approach transfers to other in-context-learning applications and model families. The acknowledgement reports support from the National Science Foundation under IIS-2348405 and the William & Mary Semester Research Grant.
Paper data and sources
Original title: Towards Reliable, Generalizable, and Specific In-Context Knowledge Editing via Multi-Objective Reinforcement Learning
Authors: Xuzhong Wang, Maiqi Jiang, Tejal Nair et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text