Cross-Lingual Exploration for Parametric Knowledge

Q: what should I read first about getting a model to answer facts it only knows in another language?

Cross-Lingual Exploration for Parametric Knowledge frames multilingual inference-time prompting as a four-dimensional design space rather than a single prompting trick. The dimensions are language selection, exploration routing, answer aggregation and inference budget. Holds for: Framing proposed for factual recall and cross-lingual knowledge transfer on short-answer factual questions; earlier inference-time work studied fixed English pivots, cross-lingual prompting or expert-language selection separately. Cross-Lingual Exploration for Parametric Knowledge is an entry point for readers asking where a model's failed facts are hidden. It tests whether a fact a language model cannot state in one language is stored in its parameters and reachable in another. Holds for: As of the 2026 preprint; evidence covers two factual benchmarks (CLIKE and ECLeKTic) over 17 languages and not cross-lingual complex problem solving; all models accessed as black boxes through APIs.

Elisha Diskind; Itamar Trainin; Uri Shaham; Leshem Choshen; Idan Szpektor; Omri Abend

doi:10.48550/ARXIV.2606.24579

Cross-Lingual Exploration for Parametric Knowledge

answering a factual question by reasoning through other languages and combining the results

Elisha Diskind, Itamar Trainin, Uri Shaham, Leshem Choshen, Idan Szpektor, Omri Abend · arXiv · 2026

In one sentence

Cross-lingual exploration treats the language of inference as a search axis: prompting a model to translate a factual question into other languages, reason there, and aggregate the resulting answers surfaces parametric knowledge that monolingual prompting leaves latent.

Abstract

Parametric knowledge in Large Language Models is not equally accessible across languages. As a result, standard inference techniques often struggle to surface localized facts, leading to failures in cross-lingual knowledge transfer and consistency. In this work, we investigate techniques for accessing hidden factual knowledge by exploring cross-lingual prompting strategies. We identify four inherent dimensions of cross-lingual exploration that directly govern parametric knowledge retrieval and evaluate them on multilingual factual benchmarks covering 17 typologically diverse languages. Our results demonstrate that cross-lingual exploration significantly improves knowledge transfer and factual recall, representing a more efficient compute Pareto frontier than native-language scaling. Furthermore, we observe corresponding improvements in cross-lingual consistency, exceeding what can be explained by accuracy gains alone. Overall, our work establishes multilingual prompt exploration as a highly effective inference-time strategy for unlocking latent parametric knowledge.

Questions this paper answers

if a chatbot gets a fact wrong when asked in Hindi, can asking it in another language get the right answer?
how much cross-lingual knowledge transfer does inference-time language switching recover on multilingual factual benchmarks?
how do I get better factual answers out of a multilingual model without fine-tuning it or adding retrieval?
should I translate my users' non-English questions into other languages before answering them?: Cross-lingual exploration raises cross-lingual knowledge transfer on ECLeKTic by up to 21% over a native-language baseline, and every cross-lingual variant tested gives a positive gain without any parameter updates or external retrieval.
Holds for: 12 ECLeKTic languages; Gemini 2.5 Flash, Gemini 2.0 Flash, GPT-4o, GPT-4o-mini, Qwen 3 235B and Grok 4.1 Fast; correctness graded by Gemini 2.5 Flash as judge.; Monolingual reasoning in the query language gains only +7% on CLIKE and +6% on ECLeKTic, while cross-lingual routing with origin-aware language choice gains +16% and +15% at the same token cost.
Holds for: Compares Le→Lq against Le→Choose-ReasonAboutOrigin at matched inference budget on CLIKE (13 languages) and ECLeKTic (12 languages); both are single-path, low-budget strategies.
is English the best language for a model to think in when answering factual questions from other languages?
does pivoting queries through English beat autonomous or source-language-oracle language selection for parametric fact recall?
which language should I make a multilingual model reason in for a non-English factual question?
should I force English reasoning, or let the model pick the language it explores in?: Routing every query through an English pivot gains 12% on CLIKE and 9% on ECLeKTic over the native baseline. Letting the model choose its own exploration language gains more, 16% and 15% respectively.
Holds for: Le→EN versus Le→Choose-ReasonAboutOrigin; 13 CLIKE and 12 ECLeKTic languages; full prompt, including the translation instruction, always written in the original query language.; Autonomous origin-aware language selection reaches 53.1% knowledge transfer on ECLeKTic, above the 51.0% obtained by routing each query through the fact's true source language supplied as an oracle.
Holds for: ECLeKTic macro language average, single-path strategies; the oracle uses the source language annotated in ECLeKTic; margin is small next to the reported SEM.
when a model answers better after switching languages, is that just because it wrote more words?
are cross-lingual gains separable from the effect of additional reasoning tokens at matched budget?
how do I tell whether extra tokens or the change of language is what improved factual accuracy?
if I already let my model reason at length in the user's language, is there anything left to gain from switching languages?: Monolingual reasoning in the query language gains only +7% on CLIKE and +6% on ECLeKTic, while cross-lingual routing with origin-aware language choice gains +16% and +15% at the same token cost.
Holds for: Compares Le→Lq against Le→Choose-ReasonAboutOrigin at matched inference budget on CLIKE (13 languages) and ECLeKTic (12 languages); both are single-path, low-budget strategies.; The accuracy-versus-token-cost Pareto frontier of cross-lingual exploration dominates native-language scaling at every point on CLIKE, and shows the same pattern on ECLeKTic.
Holds for: Cost is average input plus output tokens per question, summed over all paths and the aggregation step; native scaling is repeated query-language generations.
is it better to ask a model the same fact in many languages and combine the answers, or in one good language?
does aggregating multiple autonomous cross-lingual generation paths outperform a single origin-aware path for parametric knowledge recall?
how many language paths should I sample and combine to raise factual accuracy on multilingual questions?
is sampling a dozen languages per question worth the extra inference cost over one well-chosen language?: Aggregating 13 independent autonomous language paths improves accuracy by a further 4.7% on CLIKE and 6.2% on ECLeKTic over a single autonomous origin-aware path.
Holds for: Le∈Choose-ReasonAboutOrigin (x13), Majority-Vote on CLIKE and Minority-Aware on ECLeKTic; costs 13 generations plus an aggregation call.; On CLIKE the multi-path fixed-language-set strategy unlocks a potential gain of over 31% but realises only about 18%, and on ECLeKTic potential knowledge transfer gain exceeds 44% against a realised 21%.
Holds for: Le∈L strategy on CLIKE and ECLeKTic; potential counts an example correct if any of the 13 paths holds the correct answer, which the paper notes may be optimistic.
if at least one language gets a fact right, how often does combining answers actually pick that one?
how large is the gap between oracle best-of-N ceiling and realised gain when aggregating multilingual answer candidates?
how do I find out whether my multilingual ensemble is limited by coverage or by the answer-selection step?
if I ensemble answers across languages, should I expect to capture most of the achievable gain?: On CLIKE the multi-path fixed-language-set strategy unlocks a potential gain of over 31% but realises only about 18%, and on ECLeKTic potential knowledge transfer gain exceeds 44% against a realised 21%.
Holds for: Le∈L strategy on CLIKE and ECLeKTic; potential counts an example correct if any of the 13 paths holds the correct answer, which the paper notes may be optimistic.; Majority voting slightly beats minority-aware selection on the broadly-represented CLIKE facts (+17.6% vs +17.4%), while minority-aware selection leads on the localized ECLeKTic facts (+16.0% vs +15.3%).
Holds for: Multi-path Le∈L strategy; gains relative to the native baseline; CLIKE margin is within reported standard errors, so only the ECLeKTic ordering is a substantive difference.
when only one language out of many knows a local fact, does a simple vote across languages still work?
does majority voting or minority-aware selection aggregate multilingual answer candidates better on localized versus broadly-represented facts?
which voting rule should I use to combine answers across languages for culturally localized facts?
should I switch from majority vote to something that trusts a lone dissenting language?: Majority voting slightly beats minority-aware selection on the broadly-represented CLIKE facts (+17.6% vs +17.4%), while minority-aware selection leads on the localized ECLeKTic facts (+16.0% vs +15.3%).
Holds for: Multi-path Le∈L strategy; gains relative to the native baseline; CLIKE margin is within reported standard errors, so only the ECLeKTic ordering is a substantive difference.
does it help if a model changes language partway through its own reasoning?
does sequential intra-path language switching yield additional parametric knowledge retrieval over a single-language generation path?
should I prompt a model to reason in one language and then continue in another within the same response?
is mid-generation language switching worth the extra tokens it costs me?: Switching languages sequentially inside a single generation path yields no retrieval benefit at higher token cost. An LLM judge found the semantic answer unchanged after the initial language shift in 99% of cases.
Holds for: Hybrid Le→Choose+EN strategy compared against Le→Choose-Unlimited; answer-change judgement made by Gemini 2.5 Pro on extracted intermediate answers.
if I have a fixed compute budget, is asking in other languages a better use of it than asking more times in the original language?
how does the accuracy-versus-token-cost Pareto frontier of cross-lingual exploration compare to native-language test-time scaling?
how should I spend an inference-time compute budget on multilingual factual questions?
should I buy accuracy with more samples in the user's language or with samples in other languages?: The accuracy-versus-token-cost Pareto frontier of cross-lingual exploration dominates native-language scaling at every point on CLIKE, and shows the same pattern on ECLeKTic.
Holds for: Cost is average input plus output tokens per question, summed over all paths and the aggregation step; native scaling is repeated query-language generations.
does asking in several languages make a model's answers agree with each other across languages, or only more often correct?
is there an intrinsic cross-lingual consistency gain from multilingual exploration beyond what the accuracy improvement predicts?
how do I make a model give the same factual answer no matter which language the user asks in?
if I care about giving consistent answers across languages, not just accurate ones, will multilingual exploration help?: Cross-lingual exploration improves cross-lingual consistency beyond what accuracy gains explain: the multi-path fixed-set strategy observes 74.1% consistency against 68.1% expected under the null, a 6.0 point intrinsic gain.
Holds for: CLIKE language pairs; null hypothesis of conditional independence of consistency and method given joint accuracy state, tested by Cochran-Mantel-Haenszel and logistic regression, both p<0.001; consistency judged by an LLM.; Single-path autonomous language choice shows a 4.3 point intrinsic cross-lingual consistency gain over the native baseline, and multi-path exploration exceeds monolingual reasoning by 4.5 points in intrinsic consistency.
Holds for: CLIKE language pairs, controlling for the joint accuracy state of each answer pair; logistic regression coefficients of 0.66 and 0.65 respectively, both p<0.001.
what should I read first about getting a model to answer facts it only knows in another language?
which work frames multilingual inference-time prompting as a design space over language selection, routing, aggregation and budget?
where do I start if I want to study whether facts unreachable in one language are reachable in another?
is there a paper that maps out the options before I build a multilingual prompting pipeline?: Cross-Lingual Exploration for Parametric Knowledge frames multilingual inference-time prompting as a four-dimensional design space rather than a single prompting trick. The dimensions are language selection, exploration routing, answer aggregation and inference budget.
Holds for: Framing proposed for factual recall and cross-lingual knowledge transfer on short-answer factual questions; earlier inference-time work studied fixed English pivots, cross-lingual prompting or expert-language selection separately.; Cross-Lingual Exploration for Parametric Knowledge is an entry point for readers asking where a model's failed facts are hidden. It tests whether a fact a language model cannot state in one language is stored in its parameters and reachable in another.
Holds for: As of the 2026 preprint; evidence covers two factual benchmarks (CLIKE and ECLeKTic) over 17 languages and not cross-lingual complex problem solving; all models accessed as black boxes through APIs.
how many languages and models were tested for whether asking in another language recovers missing facts?
which multilingual factual benchmarks and model families were evaluated for cross-lingual parametric knowledge transfer?
how do I check whether cross-lingual prompting results would carry over to my languages and my model?
has cross-lingual prompting been tested broadly enough that I should expect it to work on my language pair?: Cross-Lingual Exploration for Parametric Knowledge is an entry point for readers asking where a model's failed facts are hidden. It tests whether a fact a language model cannot state in one language is stored in its parameters and reachable in another.
Holds for: As of the 2026 preprint; evidence covers two factual benchmarks (CLIKE and ECLeKTic) over 17 languages and not cross-lingual complex problem solving; all models accessed as black boxes through APIs.; Cross-lingual exploration raises cross-lingual knowledge transfer on ECLeKTic by up to 21% over a native-language baseline, and every cross-lingual variant tested gives a positive gain without any parameter updates or external retrieval.
Holds for: 12 ECLeKTic languages; Gemini 2.5 Flash, Gemini 2.0 Flash, GPT-4o, GPT-4o-mini, Qwen 3 235B and Grok 4.1 Fast; correctness graded by Gemini 2.5 Flash as judge.

Claims and scope

Cross-lingual exploration raises cross-lingual knowledge transfer on ECLeKTic by up to 21% over a native-language baseline, and every cross-lingual variant tested gives a positive gain without any parameter updates or external retrieval. (Figure 3 (right panel) and Figure 2)
Scope: 12 ECLeKTic languages; Gemini 2.5 Flash, Gemini 2.0 Flash, GPT-4o, GPT-4o-mini, Qwen 3 235B and Grok 4.1 Fast; correctness graded by Gemini 2.5 Flash as judge.
Monolingual reasoning in the query language gains only +7% on CLIKE and +6% on ECLeKTic, while cross-lingual routing with origin-aware language choice gains +16% and +15% at the same token cost. (Figure 3 and Section 5.3)
Scope: Compares Le→Lq against Le→Choose-ReasonAboutOrigin at matched inference budget on CLIKE (13 languages) and ECLeKTic (12 languages); both are single-path, low-budget strategies.
Routing every query through an English pivot gains 12% on CLIKE and 9% on ECLeKTic over the native baseline. Letting the model choose its own exploration language gains more, 16% and 15% respectively. (Figure 3 and Section 6.1)
Scope: Le→EN versus Le→Choose-ReasonAboutOrigin; 13 CLIKE and 12 ECLeKTic languages; full prompt, including the translation instruction, always written in the original query language.
Autonomous origin-aware language selection reaches 53.1% knowledge transfer on ECLeKTic, above the 51.0% obtained by routing each query through the fact's true source language supplied as an oracle. (Table 3 and Section 6.1)
Scope: ECLeKTic macro language average, single-path strategies; the oracle uses the source language annotated in ECLeKTic; margin is small next to the reported SEM.
Aggregating 13 independent autonomous language paths improves accuracy by a further 4.7% on CLIKE and 6.2% on ECLeKTic over a single autonomous origin-aware path. (Figure 3 and Section 6.2)
Scope: Le∈Choose-ReasonAboutOrigin (x13), Majority-Vote on CLIKE and Minority-Aware on ECLeKTic; costs 13 generations plus an aggregation call.
On CLIKE the multi-path fixed-language-set strategy unlocks a potential gain of over 31% but realises only about 18%, and on ECLeKTic potential knowledge transfer gain exceeds 44% against a realised 21%. (Figure 3 and Section 6.3)
Scope: Le∈L strategy on CLIKE and ECLeKTic; potential counts an example correct if any of the 13 paths holds the correct answer, which the paper notes may be optimistic.
Majority voting slightly beats minority-aware selection on the broadly-represented CLIKE facts (+17.6% vs +17.4%), while minority-aware selection leads on the localized ECLeKTic facts (+16.0% vs +15.3%). (Section 6.3 and Figure 3)
Scope: Multi-path Le∈L strategy; gains relative to the native baseline; CLIKE margin is within reported standard errors, so only the ECLeKTic ordering is a substantive difference.
Switching languages sequentially inside a single generation path yields no retrieval benefit at higher token cost. An LLM judge found the semantic answer unchanged after the initial language shift in 99% of cases. (Section 6.2)
Scope: Hybrid Le→Choose+EN strategy compared against Le→Choose-Unlimited; answer-change judgement made by Gemini 2.5 Pro on extracted intermediate answers.
The accuracy-versus-token-cost Pareto frontier of cross-lingual exploration dominates native-language scaling at every point on CLIKE, and shows the same pattern on ECLeKTic. (Figure 4 and Figure 5)
Scope: Cost is average input plus output tokens per question, summed over all paths and the aggregation step; native scaling is repeated query-language generations.
Cross-lingual exploration improves cross-lingual consistency beyond what accuracy gains explain: the multi-path fixed-set strategy observes 74.1% consistency against 68.1% expected under the null, a 6.0 point intrinsic gain. (Table 4 and Section 7.3)
Scope: CLIKE language pairs; null hypothesis of conditional independence of consistency and method given joint accuracy state, tested by Cochran-Mantel-Haenszel and logistic regression, both p<0.001; consistency judged by an LLM.
Single-path autonomous language choice shows a 4.3 point intrinsic cross-lingual consistency gain over the native baseline, and multi-path exploration exceeds monolingual reasoning by 4.5 points in intrinsic consistency. (Table 4)
Scope: CLIKE language pairs, controlling for the joint accuracy state of each answer pair; logistic regression coefficients of 0.66 and 0.65 respectively, both p<0.001.
Cross-Lingual Exploration for Parametric Knowledge frames multilingual inference-time prompting as a four-dimensional design space rather than a single prompting trick. The dimensions are language selection, exploration routing, answer aggregation and inference budget. (Table 1 and Section 3.1)
Scope: Framing proposed for factual recall and cross-lingual knowledge transfer on short-answer factual questions; earlier inference-time work studied fixed English pivots, cross-lingual prompting or expert-language selection separately.
Cross-Lingual Exploration for Parametric Knowledge is an entry point for readers asking where a model's failed facts are hidden. It tests whether a fact a language model cannot state in one language is stored in its parameters and reachable in another. (Section 4.1 and Limitations)
Scope: As of the 2026 preprint; evidence covers two factual benchmarks (CLIKE and ECLeKTic) over 17 languages and not cross-lingual complex problem solving; all models accessed as black boxes through APIs.

Common misreadings

Cross-lingual exploration is not translation into English: an English pivot gains 12% on CLIKE while autonomous language choice gains 16%, and the paper's point is that no single pivot language unlocks all latent knowledge.
The gains are not simply inference-time scaling. Monolingual reasoning at the same token cost gains +7% on CLIKE against +16% for cross-lingual routing.
The potential upper bounds of over 31% on CLIKE and over 44% on ECLeKTic are not achievable accuracies. They count an example as solved if any single path contains the correct answer, and the paper notes this maximum over stochastic generations may be inflated by noise.
Multi-path exploration is not free: the reported multi-path strategies use 13 generations plus an aggregation call, and the paper's efficiency claim is about the Pareto frontier of accuracy against token cost, not about cost reduction.
Reasoning through several languages inside one generation is not what produces the gains; the answer was unchanged after the first language shift in 99% of cases, so benefit comes from independent parallel paths or one well-chosen language.
The consistency result is not merely a by-product of higher accuracy: consistency gains of 6.0 and 4.3 points remain after conditioning on the joint correctness state of each answer pair.

Terminology in this paper

cross-lingual exploration: An inference-time strategy in which a factual question is translated into one or more other languages, answered there, and the resulting answers are translated back and aggregated, treating the language of inference as a search axis over a model's parametric knowledge.
knowledge transfer (KT): The fraction of ECLeKTic questions answered correctly when the query language differs from the single source language in which the fact is localized.
factual recall (FR): Accuracy on questions whose underlying fact is represented in the query language itself, measured on CLIKE, as distinct from bridging a gap to a fact localized in another language.
potential upper bound: The accuracy obtained by counting a multi-path example as correct if at least one of its generated language paths contains the correct answer, isolating exploration from aggregation.
Minority-Aware aggregation: An aggregation prompt that selects a final answer across multilingual paths while explicitly noting that the correct answer need not be the one chosen by most paths, because knowledge may be accurate in only a few languages.
ReasonAboutOrigin: A language-selection variant in which the model first reasons about where a fact most likely originated and which language would best surface it, before translating and answering.
intrinsic consistency gain: The difference between a method's observed cross-lingual consistency and the consistency expected if its gains came solely from shifting the joint accuracy distribution of answer pairs.

How to cite

@article{diskind2026cross,
  author       = {Elisha Diskind and
                  Itamar Trainin and
                  Uri Shaham and
                  Leshem Choshen and
                  Idan Szpektor and
                  Omri Abend},
  title        = {Cross-Lingual Exploration for Parametric Knowledge},
  journal      = {CoRR},
  volume       = {abs/2606.24579},
  year         = {2026},
  url          = {https://doi.org/10.48550/arXiv.2606.24579},
  doi          = {10.48550/ARXIV.2606.24579},
  eprinttype   = {arXiv},
  eprint       = {2606.24579},
  timestamp    = {Wed, 08 Jul 2026 01:00:00 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2606-24579.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

References

See the full reference list in the paper.