Research map / Large language models / Reasoning and chain of thought

Reasoning Transfer for an Extremely Low-Resource and Endangered Language: Bridging Languages Through Sample-Efficient Language Understanding

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 1 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.

Proposes English-Pivoted CoT Training, which fine-tunes LLMs to reason in English while receiving input and producing answers in a low-resource language, and releases LC2024, an Irish math benchmark.

Problem

Chain-of-thought reasoning gains largely apply to high-resource languages; extremely low-resource languages such as Irish suffer from misunderstanding and failed CoT in the target language.

Approach

Supervised fine-tuning with the chain-of-thought constrained to English while the input and final answer remain in the target language, separating language understanding from reasoning and exploiting the English-centric latent reasoning space of LLMs.

Key results

• On Irish AIME2024, English-Pivoted CoT Training reaches 35.00% for r1-distill-Llama-8B versus 6.67% zero-shot, an improvement of up to 28.33 points.
• On LC2024 (Irish), r1-distill-Llama-8B improves from 63.64% (zero-shot) to 73.33%.
• On French MGSM, r1-distill-Llama-8B improves from 54.8% to 83.2%, while Chinese gains are marginal (65.2% to 70.0%).

Surprising findings

• On Chinese, English pivoting yields only marginal gains over baseline because the model naturally reasons in Chinese, so forcing English reasoning can cause interference.
• Native CoT training reduces DeepHermes-3-Llama-3-8B accuracy on LC2024 concepts & skills from 45.45 to 13.64.

Limitations

• LC2024 is small (55 samples), and AIME2024 Irish results are based on few problems, making variance high.
• Chinese results show English pivoting can hurt when the model already reasons well in the target language.
• Evaluated only on math tasks and a few languages (Irish, French, Chinese).

Experimental setup

Datasets: Bespoke pensez congliu

Benchmarks: AIME2024 MGSM MATH-hard LC2024

Metrics: accuracy

Baselines: Zero-shot Prompting Few-shot Prompting Native CoT Training TIES-merging Layer Swap

Builds on: chain-of-thought DeepSeek-R1 distillation NLLB TIES-merging layer swap

Closest papers by meaning

• Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.93)
• Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought (ICLR 2026 · similarity 0.92)
• AdaMCoT: Rethinking Cross-Lingual Factual Reasoning Through Adaptive Multilingual Chain-of-Thought (AAAI 2026 · similarity 0.92)
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.92)
• XCOT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning (AAAI 2025 · similarity 0.92)
• Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization (AAAI 2026 · similarity 0.90)