A reinforcement-learning method for changing facts in a language model reported a stronger balance between applying a requested edit and retaining other knowledge than DR-IKE in benchmark tests. On Llama-3.2-3B-Instruct, MO-IKE’s overall score was 73.5, against 54.9 for DR-IKE. MO-IKE also reported 92% edit success versus 85%, and a retention rate of 57.7% versus 34.7%.
Those measures capture a central tension in in-context knowledge editing. The system must respond to an edited fact, remain consistent when the request is paraphrased, and retain other knowledge. The study labels these targets edit success, paraphrase consistency and retention rate, then combines them in an overall score calculated as their harmonic mean, a type of average that gives less credit when one component lags.
A prompt, not a parameter change
The setup keeps the language model frozen. MO-IKE seeks the desired update through prompt context, without access to the model’s internals, so the experiment tests prompt construction rather than a change to the underlying model. Its reinforcement-learning policy treats prompt construction as a constrained sequence of decisions and trains across COPY, UPDATE and RETAIN demonstration types, allowing the examples to be ordered globally.
Core experiments trained on randomly selected examples and evaluated held-out benchmark subsets: 300 CounterFact examples for training, 300 held-out CounterFact examples, and 100 held-out examples each from ZsRE and WikiCounterfact. The comparisons included DR-IKE and reward ablations in which the constraint design was varied.
Gains varied across tests
The same comparison pattern was reported on Mistral-7B-Instruct-v0.3. MO-IKE’s overall score was 77.4, compared with 62 for DR-IKE. Edit success was 80.1% versus 74.7%, paraphrase consistency was 90.7% versus 88.3%, and retention was 65% versus 42.3%. The listed retention-magnitude values were 32.1 for MO-IKE and -22.5 for DR-IKE.
Across the other core datasets, the picture was more mixed. On ZsRE, MO-IKE recorded an overall score of 42, with edit success at 75%, paraphrase consistency at 79% and retention at 22%. On WikiCounterfact, the figures were 59, 80%, 59% and 47%, respectively. Across both datasets, MO-IKE had the highest listed overall score and retention rate, but not the highest edit success.
A retriever trained on Llama-3.2 and evaluated zero-shot on Mistral-7B produced an overall score of 77.3, compared with 77.1 for the native version. Its edit success was 81% versus 80.1%, and retention was 65.7% versus 65%; paraphrase consistency went the other way, at 89.3% versus 90.7%.
An ablation of the reward design pointed in the same direction. The all-constraints version had the highest overall score, 73.5, while the retention-only version scored 70.1. Both reported retention at 57.7%, but the all-constraints condition had higher edit success, 92% versus 86%, and paraphrase consistency, 79% versus 71%.
What the numbers do not settle
On UniEdit, the reported metrics were an overall score of 24.5, edit success of 51.3%, paraphrase consistency of 28.3% and retention of 14.3%. Those figures were lower than the scores listed in the main Llama and Mistral comparisons, showing that reported performance varied across evaluations.
The authors acknowledge that the dataset scale is relatively small and that using only the first 2,000 CounterFact records may be insufficient to assess the method’s overall capability. The reported comparisons also do not include confidence intervals or statistical tests.
The work is presented as a prompt-construction method for balancing edit reliability, paraphrase consistency and knowledge retention. The document is an arXiv preprint, version 3 dated 2 October 2026. It states that code is publicly available at a GitHub repository and acknowledges support from the National Science Foundation, the Commonwealth Cyber Initiative 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