Research map / Large language models / Language, safety and interpretability

Multilingual LLMs Inherently Reward In-Language Time-Sensitive Semantic Alignment for Low-Resource Languages

AAAI 2025 · Large language models › Language, safety and interpretability · approach: Multilingual LLM data and culture · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces mTEMPREASON (machine-translated multilingual temporal reasoning benchmark for Romanian, German, French) and CLiTSSA, a fine-tuned multilingual Sentence-BERT retriever that selects cross-lingual in-context examples using in-language temporal similarity scores, improving X-ICL temporal QA.

Problem

Cross-lingual in-context learning for time-sensitive questions in low-resource languages relies on multilingual encoders whose embeddings align in-language better than cross-lingually, so retrieved cross-lingual examples are less semantically useful for temporal reasoning.

Approach

Fine-tune a multilingual Sentence-BERT retriever with CoSENT loss on parallel query pairs labeled with similarity scores computed in the low-resource monolingual embedding space, transferring in-language temporal semantic structure into the cross-lingual space for example retrieval in X-ICL.

Key results

• On LLaMA3-8B, CLiTSSA improves mean F1 over X-InSTA by 5.32, 1.62 and 1.43 points for French, German and Romanian (p=0.05).
• Across LLMs on French L1, CLiTSSA raises F1 from 46.62 to 57.15 for LLaMA3-8B and yields a mean L1 F1 gain of 9.68 points across four LLMs.
• In-language semantic similarity retrieval outperforms cross-lingual similarity retrieval for L1 across all three languages (e.g., French F1 60.50 vs 46.62).

Measured effects reported in the paper

Surprising findings

• Retrieving semantically similar examples in-language and then substituting English examples outperforms direct cross-lingual similarity retrieval, suggesting multilingual encoders reward in-language alignment.
• The L1-tuned retriever improves L2 and L3 performance, but the reverse transfer does not hold.

Limitations

• LLM factual knowledge errors in the English monolingual setting propagate to cross-lingual predictions.
• Implicit temporal queries (L3) remain challenging.
• Gains on L3 and some tasks are small or inconsistent; integrated retriever shows no clear advantage.

Experimental setup

Datasets: TEMPREASON mTEMPREASON

Benchmarks: mTEMPREASON

Metrics: F1 exact match Mann-Whitney U test

Baselines: X-ICL X-InSTA Zero-Shot

Builds on: TEMPREASON X-InSTA Sentence-BERT CoSENT loss

Closest papers by meaning

• Soft Token Alignment for Cross-Lingual Reasoning (NeurIPS 2026 · similarity 0.90)
• XCOT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning (AAAI 2025 · similarity 0.88)
• AdaMCoT: Rethinking Cross-Lingual Factual Reasoning Through Adaptive Multilingual Chain-of-Thought (AAAI 2026 · similarity 0.88)
• TimE: A Multi-level Benchmark for Temporal Reasoning of LLMs in Real-World Scenarios (NeurIPS 2025 · similarity 0.88)
• INCLUDE: Evaluating Multilingual Language Understanding with Regional Knowledge (ICLR 2025 · similarity 0.87)
• Context-Alignment: Activating and Enhancing LLMs Capabilities in Time Series (ICLR 2025 · similarity 0.87)