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

AdaMCoT: Rethinking Cross-Lingual Factual Reasoning Through Adaptive Multilingual Chain-of-Thought

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 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 AdaMCoT, which fine-tunes LLaMA3.1-8B and Qwen2.5-7B to route factual reasoning through auxiliary "thinking languages" (English, Chinese, Indonesian) or direct generation, selecting training pathways via GPT-4o reward scoring.

Problem

Multilingual LLMs show large performance gaps across languages on factual and logical reasoning, especially for low-resource languages, and existing translation-based or prompt-only cross-lingual methods are costly, noisy, or fail to improve reasoning.

Approach

Instead of translating inputs, the model reasons in an intermediate high-resource or culturally relevant language and then answers in the query language; a GPT-4o reward model scores candidate pathways (cross-lingual CoT vs. direct generation), and only high-scoring (>=9) self-generated pathways are used for LoRA fine-tuning to teach adaptive language routing.

Key results

• LLaMA3.1-8B-AdaMCoT improves TruthfulQA accuracy in 31 of 32 languages, with relative gains of 2.1% in English, 9.0% in Chinese and 12.7% in Indonesian.
• On CrossLogiQA, Qwen2.5-7B-AdaMCoT raises cross-lingual consistency from 59.1 to 85.8 and LLaMA3.1-8B-AdaMCoT raises CrossMMLU consistency from 66.7 to 82.7.
• On CrossAlpaca-Eval 2.0, LLaMA3.1-8B-AdaMCoT improves Chinese from 7.53 to 8.13 and Indonesian from 7.51 to 8.13 with English unchanged at 8.35.

Surprising findings

• Qwen2.5-7B-AdaMCoT predominantly chooses direct answering rather than intermediate reasoning, suggesting implicit cross-lingual reasoning in its shared latent space.
• Qwen2.5-7B-AdaMCoT shows a decline in Chinese on TruthfulQA despite gains in most other languages.

Limitations

• Relies on GPT-4o as reward model and integration model, with about 86% agreement with human preferences on a 50-sample check.
• Evaluation on Cross-MMLU and Cross-LogiQA uses small sets (150 and 176 questions).
• Some languages show declines, e.g., Armenian for LLaMA and Chinese for Qwen on TruthfulQA.

Experimental setup

Datasets: OpenHermes 2.5 Firefly Alpaca GPT-4 Instruction GSM8KInstruct_Parallel Flores-200

Benchmarks: Multilingual TruthfulQA CrossAlpaca-Eval 2.0 Cross-MMLU Cross-LogiQA

Metrics: accuracy GPT-4o judge score (0-10) cross-lingual consistency

Baselines: AutoCAP QAlign LLaMA3.1-8B-Instruct Qwen2.5-7B-Instruct

Builds on: AutoCAP QAlign xCoT Logit Lens

Closest papers by meaning

• When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners (NeurIPS 2025 · similarity 0.92)
• XCOT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning (AAAI 2025 · similarity 0.92)
• Reasoning Transfer for an Extremely Low-Resource and Endangered Language: Bridging Languages Through Sample-Efficient Language Understanding (AAAI 2026 · similarity 0.92)
• Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.91)
• Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought (ICLR 2026 · similarity 0.91)
• Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in LLMs (ICLR 2026 · similarity 0.90)