When a fact changes, say a country gets a new head of government, we want a language model to learn the new answer and still remember the old one. We also want both to hold no matter which language or writing system the user asks in. Two findings suggest this double requirement fails in a structured and predictable way. In English, the standard tools that surgically rewrite a fact inside a model's weights almost completely erase the fact's history. In the multilingual setting, an edit made in English barely reaches Chinese, and it spreads unevenly into other non-Latin scripts. Nobody has put the two together. So nobody knows whether a Chinese or Hindi user of an "updated" model gets the new fact, the old fact wrongly presented as current, or no history at all.
This note proposes to answer that question and to explain the answer with a known internal mechanism. Models process many non-Latin alphabetic scripts, such as Hindi, Russian and Arabic, by passing them through a romanized, English-like intermediate representation before writing the native script. Chinese skips that step. Our hypothesis is that this "latent romanization" bridge decides how a temporal edit behaves in each script. Where the bridge exists, the edit travels across scripts, and so does the destruction of history. Where it is absent, as in logographic Chinese and Japanese, the edit stays put. The result is a split state: the model is updated in English but still treats the obsolete fact as current in Chinese. If the hypothesis holds, the bridge also points to a fix. Computing the edit at the layer where scripts converge should repair the bridged scripts. The logographic scripts would still need their own edits.
Core Claim
The outcome of a temporal edit in a given query language falls into one of four cells: the current fact is updated or not, and the historical fact is kept or erased. We claim that which cell a language lands in can be predicted from how strongly that language's subject representations pass through the shared romanized intermediate space. Script family alone does not predict it. Two things would be new. First, we give the first cross-script measurement of historical forgetting. Second, we give a mechanistic predictor for it that could be falsified. The accompanying editing method is a test of that mechanism, not a standalone contribution. If per-language editing matches it cheaply, the diagnosis still stands.
What Prior Work Establishes and Leaves Open
The temporal side comes from History Matters: Temporal Knowledge Editing in Large Language Model. That paper builds AToKe from time-ordered YAGO fact chains. It shows that ROME, MEMIT, MEND and continual fine-tuning drive historical-knowledge scores (HES, HRS) to near zero. It then proposes METO, which adds an explicit time objective so that past and present facts are optimized together. All of this is in English only. We reuse METO's objective as a component and ask what it does across scripts. ScEdit adds that even in English, temporally edited facts are applied about 27% less reliably when the model generates multi-step plans. This constrains our evaluation: we cannot stop at single-fact lookups.
The cross-script side comes from four papers that together motivate the study. Cross-lingual Editing in Multilingual Language Models finds that script family matters more than language family. Its cross-lingual generability is 90.04% within Latin script and 78.38% when crossing into Indic script. That work used only fact verification with mBERT and BLOOM, and no logographic scripts. MLaKE shows that StableKE reaches 88.4% English editing accuracy but only 5.5% when the edit transfers from English to Chinese. It blames two causes: Chinese subjects are split into many more tokens (11.77 versus 4.04 in Vicuna), and the neurons modified for the same fact in different languages overlap less than 20%. Cross-Lingual Knowledge Editing in Large Language Models documents "locality leaks," where an English edit corrupts unrelated Chinese facts. Language Anisotropic Cross-Lingual Model Editing learns language-specific parameter masks from parallel data. None of these four measures whether history survives the edit.
The mechanism we test comes from RomanLens. It uses the logit lens, which reads a hidden state through the output layer to see which token it "means," and activation patching, which swaps hidden states between runs. With these tools it shows that Hindi, Gujarati, Greek, Ukrainian, Arabic and others produce romanized forms in intermediate layers, specifically layers 20–29 in Llama-2 7B. These romanized forms appear one to two layers before the native script, and patched runs differ from Latin runs by a KL divergence below 0.001. Chinese bypasses this bridge. No editing paper has used the bridge.
Two papers supply methods we reuse. LangEdit projects each new edit onto the null space of earlier edit keys, meaning directions that previous edits did not use, so sequential multilingual edits do not destroy each other. Retrieval-Augmented Multilingual Knowledge Editing (ReMaKE) and Cross-Lingual Multi-Hop Knowledge Editing (CLEVER-CKE) avoid weight edits altogether by retrieving edited facts into the prompt. CLEVER-CKE's CROLIN-MQUAKE-T split already contains real temporal updates across eight languages. However, it scores whether the current answer is correct, not whether the historical answer is retained.
Research Program
The first component is a benchmark, which we call multilingual AToKe. We take AToKe's YAGO chains and render each fact in English, German, Russian, Hindi, Chinese and Japanese. These cover Latin, Cyrillic, Indic and logographic scripts with two languages on the Latin and logographic sides. Entity names come from Wikidata native-script labels rather than machine translation, and relation templates are written once per language and checked by native speakers. Each item carries AToKe's four scores in every language: current edit success, current retaining success, historical edit success and historical retaining success. We add relative-time probes ("the previous office holder"), because English editors already find these harder than explicit years. For generation, we build a small procedural subset in the style of ScEdit. In that subset, a policy fact edited in English must be applied inside a multi-step plan written in Hindi or Chinese.
The second component is the diagnostic. For each model and language, we compute a bridge score $b_\ell$: the fraction of subject-token hidden states in which, at the RomanLens layer range, the logit lens assigns the highest probability to a romanized or English form of the subject rather than to a native-script token. We then edit in English with ROME, MEMIT, MEND, fine-tuning and METO. For every query language $\ell$, we record the transfer of the current fact and the survival of the historical fact. The central test is whether $b_\ell$ predicts both, better than a simple script-family label or the token-count ratio that MLaKE blamed. For Chinese and Japanese, the prediction is a split state: high current-fact failure together with high historical retention, because the edit never reached their parameters.
The third component is a bridge-keyed temporal editor. ROME-style editors change a feed-forward layer by adding a targeted update keyed on the subject's hidden representation, which we call the key vector $k$. Normally $k$ is taken at the last subject token of the English prompt. We instead take it at the layer $L^*$ where cross-script subject representations are most similar, and we average across the languages that share the bridge:
$$k_{\text{bridge}} = \frac{1}{|S|}\sum_{\ell \in S} h^{(L^*)}_{\ell}(\text{subject}),$$
where $S$ is the set of languages with bridge score above a threshold, and $h^{(L^*)}_{\ell}$ is the subject's hidden state at layer $L^*$ in language $\ell$. The update is then fit with METO's time objective, so that "in year $t_0$" still maps to the old object. It is also projected onto LangEdit's null space, so that historical keys are protected. Logographic languages outside $S$ receive a separate native-script key. The mechanism makes an asymmetric prediction. Bridge keys should close most of the gap for Hindi and Russian but not for Chinese. If they help Chinese just as much, the bridge is not what is doing the work.
Baselines
The do-nothing reference is the unedited model, which gives the floor for current-fact accuracy and the ceiling for historical retention in each language. The naive editing reference is ROME and MEMIT applied to English prompts. MLaKE already reports the collapse of transfer, with StableKE falling from 88.4% in English to 5.5% on English-to-Chinese transfer, but it has no historical score. The strongest existing temporal method is METO applied in English. Its reported gain on AToKe-ME is HES* rising from below 0.3 to over 15 (History Matters). We found no reported result on Lacuna for METO or any temporal editor in a non-English query language, so our own cross-script METO runs set that number.
The strongest practical competitor is per-language METO plus LangEdit. This means editing the same fact separately in every language, with null-space projection preventing the edits from interfering with each other. It is the most direct answer to parameter segregation, and our method must match it while using fewer edits, or beat it on relative-time and procedural probes. On the retrieval side, the strongest baseline is ReMaKE-few-bi, which reports 69.26% cross-language reliability on MzsRE against 38.07% for in-context editing. We also run CLEVER-CKE, which reports up to 30% gains over earlier retrieval methods on CROLIN-MQUAKE. For retrieval baselines, the memory stores both the old fact and the new fact with timestamps, so that history retention is a fair test.
All methods run on Llama-2 7B, where the RomanLens layer range is known, as well as Llama3-8B and Qwen2.5-7B. Every method gets the same edit set, the same hyperparameter search budget of equal numbers of configurations tuned on a held-out English split, and the same retriever encoder for the retrieval systems. The decisive ablation keeps everything in our editor fixed but uses native last-token keys instead of bridge keys. A second ablation drops the time objective. Beating a baseline means a gain on the average of current-fact and historical scores in non-Latin languages, with locality no worse than 2 points below the baseline's.
Why not just…?
Why not just edit in every language separately? This is the real threat. MLaKE's finding that neurons overlap less than 20% suggests that separate edits are exactly what the model's storage calls for. We include it as the strongest baseline. If it matches bridge-keyed editing on bridged scripts, we will report that the method adds nothing there, and the contribution reduces to the diagnosis plus cost: $k$ languages mean $k$ edits and $k$ times more pressure on the null space. The diagnostic question of which scripts erase history remains open under either result.
Why not just use multilingual retrieval? ReMaKE's roughly 98% locality makes retrieval attractive. However, CLEVER-CKE finds that retrieval itself is the bottleneck for Chinese, Hindi and Russian, and ScEdit shows that facts supplied in context are still overridden by stale parametric knowledge during long generation. Our procedural subset tests exactly that override. If timestamped retrieval dominates on every score, including procedural consistency, we will say parametric temporal editing across scripts is not worth pursuing.
Why not just translate queries into English? A translate-edit-translate pipeline makes history retention depend on English editing alone. We run it with a fixed machine translation system. It cannot repair the parametric split state, because a user who prompts in Chinese without the pipeline still gets the old fact. It also inherits translation errors on native-script entity names, which our Wikidata-label construction lets us count directly.
Why not just fine-tune with LoRA on multilingual parallel updates? History Matters shows that continual fine-tuning drives historical scores to near zero, and LangEdit shows that sequential multilingual updates degrade general ability. We include LoRA fine-tuning on parallel temporal sentences, which mix old and new states, as a baseline, so that its forgetting is measured rather than assumed.
Decisive Checks
Three results would settle the proposal. The first is whether the bridge score $b_\ell$ explains more variance in cross-script historical retention than script-family labels or token-count ratios, across three models and six languages. If it does not, the mechanistic claim fails. The second is whether Chinese and Japanese actually show the predicted split state after an English edit, meaning a stale current answer with intact history, while Hindi and Russian show transfer together with erased history. The third is whether bridge keys improve bridged scripts significantly more than logographic ones relative to native keys. A uniform improvement, or none, would falsify the idea that the romanized intermediate carries the edit. Qwen2.5-7B, trained heavily on Chinese, is a useful internal control: if its Chinese representations turn out to route differently, the predictor should track that shift.
References
- History Matters: Temporal Knowledge Editing in Large Language Model
- ScEdit: Script-based Assessment of Knowledge Editing
- Cross-lingual Editing in Multilingual Language Models
- MLaKE: Multilingual Knowledge Editing Benchmark for Large Language Models
- Cross-Lingual Knowledge Editing in Large Language Models
- Language Anisotropic Cross-Lingual Model Editing
- RomanLens: The Role Of Latent Romanization In Multilinguality In LLMs
- Mitigating Negative Interference in Multilingual Knowledge Editing through Null-Space Constraints
- Retrieval-Augmented Multilingual Knowledge Editing
- Cross-Lingual Multi-Hop Knowledge Editing