In banking, healthcare, law, insurance, and government, stale answers carry real consequences. Consider a regional bank. A lending regulation changes overnight. By morning, the bank’s internal AI assistant is still citing the old rule to loan officers, confidently and in full paragraphs.

Retraining the model would take weeks of computing time the team does not have.

A new paper from researchers from our Data Science & System Security Department, written with collaborators at Dartmouth College and Carnegie Mellon University, offers a way to correct that kind of error in place. “Representation Interventions Enable Lifelong Knowledge Memory Control in LLMs” appears in the Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026).

The authors are Xuyuan Liu, graduate researcher at Dartmouth College, who completed the work during an internship at NEC Laboratories America; Shengyu Chen, Researcher, NEC Laboratories America; Xinshuai Dong, Carnegie Mellon University; Yanchi Liu, Senior Researcher, NEC Laboratories America; Xujiang Zhao, Researcher, NEC Laboratories America; Haoyu Wang, Researcher, NEC Laboratories America; Yujun Yan, Assistant Professor of Computer Science, Dartmouth College; Haifeng Chen, Department Head, Data Science & System Security, NEC Laboratories America; and Zhengzhang Chen, Senior Researcher, NEC Laboratories America.

Why Model Edits Collapse

Large language models (LLMs), the AI systems behind chatbots and enterprise assistants, learn facts during training and then stop learning. Full retraining is expensive and often causes catastrophic forgetting, where a model loses older knowledge as it takes in new material. Retrieval-augmented generation (RAG), which feeds relevant documents to the model at question time, helps. Retrieved text can still conflict with what the model already believes, and answer quality depends on the search step. A third option, knowledge editing, rewrites specific facts inside the model.

Most editing tools handle a handful of short facts well. Real organizations need thousands of edits over months, often in long, free-form passages. The researchers call this lifelong knowledge control. Under those conditions, existing methods break down. Methods that change the model’s internal weights suffer edit collapse, a steady decline in output quality as edits accumulate. In the team’s tests, most competing methods began to degrade after roughly 10 sequential edits. After 1,000 edits, two recent methods produced strings of repeated letters and stray symbols in place of sentences.

Steering Thoughts Instead of Weights

The team’s method, RILKE (Representation Intervention for Lifelong KnowledgE Control), leaves the original model untouched. It works on the model’s hidden representations, the internal numerical states an LLM produces as it processes a question, much like the notes a person jots down while reasoning. For each new piece of knowledge, RILKE trains a small intervention module that nudges those internal states toward the correct answer. Each module operates within a low-dimensional subspace, a narrow slice of the model’s internal space, which keeps one edit from spilling into unrelated knowledge.

The approach rests on two properties the team identified and validated. The first, semantic locality, means questions with the same meaning land close together inside the model even when the wording differs. The second shows that related facts tend to share the same internal subspace, making them easier to group.

Robust training handles rephrased questions. During training, RILKE adds small random disturbances to each question’s internal state and penalizes any resulting change in the model’s answer. The result is an edit that holds across paraphrases, not only the exact wording used in training.

A query-adaptive router decides when to apply an edit. It compares each incoming question with a stored index of edited topics and activates the matching module only when the similarity passes a set threshold. Unrelated questions go straight to the unmodified model.

Shared-subspace clustering keeps storage manageable. Instead of one module per fact, RILKE groups similar facts with hierarchical agglomerative clustering, repeatedly merging the closest items, then trains one shared module per group.

Stable After 1,000 Edits

The team tested RILKE on two open models, Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct, against six established editing methods. The main benchmark, UnKE, contains unstructured, long-form knowledge. Scores use BERTScore, a 0-to-1 measure of how closely generated text matches the target answer’s meaning.

After 1,000 sequential edits on Llama, RILKE scored 1.000 on the original questions and 0.963 on paraphrased versions. The strongest competitor, WISE, scored 0.681 and 0.673. UnKE fell to 0.013, and MEMIT fell to 0.033. General reasoning held steady. On MMLU (Massive Multitask Language Understanding), a broad test of knowledge and reasoning, the edited model scored 0.622, compared with 0.633 for the unedited base model. WISE dropped to 0.584, and single-layer fine-tuning dropped to 0.226. On Qwen, RILKE maintained an MMLU score of 0.712, slightly below the base of 0.713.

On ZsRE, a structured question-answering benchmark, RILKE averaged 0.88 across 3,000 edits compared with 0.73 for WISE. The router sent more than 93% of paraphrased questions to the correct module and filtered out more than 98% of unrelated questions. RILKE needed 96.1 MiB of storage, compared with 224.0 MiB for WISE. The clustered version cut that figure to 29.4 MiB. Peak training memory reached 19 GB compared with 59 GB for UnKE. Average response time rose from 5.64 to 5.68 seconds, a less-than-1 % increase.

Real-World Applications

Banks, insurers, and other financial services firms operate under rules that change often. A compliance assistant built on RILKE could absorb each regulatory update as a small, separate module, leaving the rest of its knowledge intact. Hospital systems and clinical software vendors face a similar need when drug guidance, dosing recommendations, and formulary lists change. Keeping each edit in its own module lets teams trace exactly which updates a model carries.

Enterprise IT departments and telecommunications operators run assistants that answer questions about internal systems, configurations, and procedures, all of which change frequently. RILKE’s low memory overhead and minimal added response time suit on-premises deployments, where hardware budgets are fixed. Government agencies and legal teams that need precise control over what an AI says about policies and people could also benefit from targeted, contained edits.

“Most editing methods try to rewrite what a model knows by changing its weights. That works for a few facts, then the model starts to fall apart. We found that the model’s own internal representations already organize knowledge by meaning, and RILKE uses that structure to keep each update precise, contained, and scalable to thousands of edits.” Xuyuan Liu, Dartmouth College and NEC Laboratories America Intern.

Looking Ahead

The authors note that a systematic risk analysis, including protection against malicious edits and unintended bias, remains future work. As organizations rely on AI systems longer, methods that keep model knowledge current without retraining will play a growing role in keeping those systems accurate and accountable.