- The paper presents cross-lingual inference strategies that significantly boost knowledge transfer and factual recall using autonomous language selection.
- It introduces multi-path reasoning with minority-aware aggregation, outperforming fixed English pivots for localized and culture-specific queries.
- The approach achieves over 2x compute efficiency improvements, setting a new benchmark for inference-time parametric knowledge retrieval.
Cross-Lingual Exploration for Parametric Knowledge: Mechanisms and Impact
LLMs exhibit substantive parametric knowledge, but this knowledge is not equally accessible across languages. Standard inference often fails to retrieve localized or culture-specific factual content, impeding cross-lingual knowledge transfer and resulting in inconsistencies. This paper interrogates the latent mechanisms governing parametric knowledge retrieval, proposing explicit cross-lingual exploration strategies at inference-time to address observed retrieval failures.
Dimensions of Cross-Lingual Exploration
The paper identifies four orthogonal dimensions governing multilingual parametric exploration:
- Language Selection: Optimal retrieval frequently requires reasoning in the language most aligned with the fact's source context. Autonomous selection, where the model predicts the most suitable language based on fact origin, outperforms fixed language pivots (notably English). This challenges conventional assumptions regarding universal English pivoting for factual tasks.
- Exploration Routing: Routing strategies include single-path (fixed or dynamically selected language) and multi-path (parallel reasoning in multiple candidate languages). Multi-path aggregation consistently yields superior factual recall and knowledge transfer gains.
- Answer Selection and Aggregation: When multiple paths generate candidate answers, effective aggregation becomes pivotal. Majority-vote is optimal for broad facts, but a minority-aware selection is superior for highly localized facts, aligning with the hypothesis that correct knowledge may be accessible only within a narrow linguistic cohort.
- Inference Budget: The efficiency of cross-lingual exploration is evaluated in terms of compute cost (token usage). Cross-lingual strategies dominate conventional native-language scaling on the compute-accuracy Pareto frontier, demonstrating greater factual recall per token.
Methodological Advances
The authors instantiate a suite of strategies over the design space, applying them to the ECLeKTic and CLIKE benchmarks covering 17 typologically diverse languages. These benchmarks include localized and broad factual queries. Every prompt template is anchored in the query's native language, ensuring isolation of cross-lingual effects independent of prompt-induced code-switching.
Models evaluated include Gemini 2.5, GPT-4o, Qwen 3 235B, and Grok 4.1, covering both proprietary and open LLMs. Evaluation metrics combine mean accuracy, cross-lingual knowledge transfer (targeting facts unknown in the query language), and cross-lingual consistency.
Empirical Findings
Knowledge Transfer and Factual Recall
Cross-lingual exploration confers substantial gains in both knowledge transfer (+21% over baseline) and factual recall (+16% over monolingual reasoning). Notably, gains are achieved without model parameter updates or external sources, strictly via inference-time strategy.
Autonomous origin-aware selection achieves 53.1% accuracy, outperforming fixed English pivot (47%) and the oracle source-language routing (51%). Multi-path strategies provide additional gains (+4.7% factual recall, +6.2% knowledge transfer), demonstrating higher upper bounds for potential knowledge unlocked versus realized via aggregation.
The mechanism is validated by qualitative analysis: fixed English routing hallucinates answers for culture-specific queries, whereas autonomous selection identifies the appropriate language origin, yielding precise factual retrieval.
Separation from Reasoning Scaling
Performance gains persist when controlling for reasoning budget—cross-lingual strategies yield >2x the improvement over monolingual reasoning at the same compute cost, confirming that the language shift, not inference-time scaling, is the dominant retrieval mechanism.
Aggregation Bottleneck
The primary bottleneck in parametric knowledge access is the aggregation step. Minority-aware aggregation is necessary for highly localized facts, whereas majority-vote suffices for broadly represented facts. The difference between the potential upper bound (at least one path contains the answer) and realized accuracy underscores inefficiency in answer selection, marking a critical domain for algorithmic innovation.
Compute Efficiency
Pareto frontier analysis confirms that cross-lingual exploration constitutes a more efficient use of compute resources compared to scaling inference within the native language. Token cost versus accuracy curves consistently favor multi-path cross-lingual trajectories.
Intrinsic Cross-Lingual Consistency
Statistical hypothesis testing (CMH, logistic regression) demonstrates that all cross-lingual strategies yield significant intrinsic consistency improvements, beyond what can be explained by accuracy gains alone.
Practical and Theoretical Implications
Unlocking latent parametric knowledge via cross-lingual exploration enables more robust and consistent multilingual systems. The results indicate that LLMs inherently partition factual content in language-dependent subspaces, with cultural/corpus factors influencing accessibility. Thus, strategic inference-time language selection serves as a powerful tool for mitigating the recall gap, especially in data-scarce or low-resource settings.
The aggregation dimension emerges as a major locus for future research—developing approaches for optimal answer selection across reasoning paths will be essential for maximizing the utility of cross-lingual exploration. Furthermore, findings suggest that training models to explicitly reason about source language origin may enhance generalization and factual capacity.
Future Directions
- Autonomous Language Routing: Algorithmic advances in automatic language origin identification and selection could further enhance factual retrieval, particularly for zero-shot or low-resource entities.
- Aggregation Optimization: Research into probabilistic aggregation, outlier detection, and minority alignment will be vital for maximizing realized accuracy in multi-path settings.
- Training for Cross-Lingual Exploration: Explicitly incorporating cross-lingual exploration as a supervised or reinforcement learning objective may yield models with higher inherent factual accessibility and consistency, enabling scalable multilingual QA systems.
Conclusion
This paper establishes cross-lingual exploration as an efficient, principled approach for surfacing latent parametric knowledge in LLMs. Explicit characterization of methodological dimensions and comprehensive empirical validation demonstrate significant improvements in factual recall, knowledge transfer, and cross-lingual consistency. The results have direct implications for inference-time QA reliability in multilingual contexts and provide a foundational framework for future research into robust and accessible factual systems in AI (2606.24579).
