cross-lingual-reasoning-and-alignment

name

cross-lingual-reasoning-and-alignment

description

Use this skill when the user asks a question in one language (e.g., German, Spanish, Chinese, Arabic) but the information needed is stored in documents of a different language (e.g., English), or split across documents in multiple languages. Trigger it for requests like 'find the answer from these English reports and tell me in German', 'compare what the local news and the international news say', or 'check if the model flips to English when I ask a question in Spanish about an English text'. This skill targets the unique challenge of performing complex reasoning (comparison, aggregation) across language boundaries while maintaining the correct response language.

Skill: cross-lingual-reasoning-and-alignment

1. Capability Definition & Real Case

• Professional Definition: The ability to integrate evidence from documents in multiple languages to satisfy a query in a target language, ensuring both Response Language Correctness (RLC) and accurate information integration. This capability requires the agent to overcome the linguistic bias of retrieved documents (avoiding language-switching) and to perform semantic reasoning—such as comparison, aggregation, or multi-hop chaining—on evidence units originally encoded in different linguistic systems.
• Dimension Hierarchy: Multi-Source Evidence Composition->Cross-Document Synthesis->cross-lingual-reasoning-and-alignment

Real Case

[Case 1]

• Initial Environment: A multilingual RAG corpus containing news articles from 2024. One article is in German discussing German swimmer Lukas Märtens winning the 400m freestyle at 3:41,78 early in the day. A second article is in English discussing the U.S. 4x100m freestyle relay team winning their first gold in the final race of the same day.
• Real Question: Welches Land gewann seine erste Goldmedaille bei den Olympischen Spielen 2024 früher, die Vereinigten Staaten oder Deutschland?
• Real Trajectory: 1. Analyze the German query to identify the intent: a temporal comparison of 'first gold medal' between two countries. 2. Retrieve a German document to find the timing of Germany's first gold (Lukas Märtens on 27 July). 3. Retrieve an English document to find the timing of the U.S. first gold (winning the final swimming race on 27 July). 4. Perform cross-lingual temporal reasoning: recognize that '3:41' or 'early' in the German text corresponds to a timestamp earlier than 'the last swimming race' mentioned in the English text. 5. Formulate the response in German, ensuring the semantic comparison is preserved despite the language mismatch of the sources.
• Real Answer: Deutschland. Beide Länder gewannen am 27. Juli bei den Schwimmwettbewerben ihre ersten Goldmedaillen, aber die USA gewannen ihre Medaille erst im letzten Wettkampf des Tages – später als Deutschland.
• Why this demonstrates the capability: This case requires the agent to reason over retrieved information across languages. The agent must successfully compare a German-language event timestamp with an English-language event timestamp and conclude the correct winner, demonstrating that it can process facts from disparate linguistic sources to reach a single grounded conclusion.

---

[Case 2]

• Initial Environment: A cross-lingual RAG setting where the question is in German, but the available knowledge base is entirely in English (monolingual retrieval). The articles describe Hurricane Helene relief efforts, citing specific pledge amounts from Walmart and Aldi.
• Real Question: Wie viel haben Walmart und ALDI zusammen für die Opfer des Hurrikans Helene 2024 gespendet?
• Real Trajectory: 1. Identify the input query as German and the target documents as English news snippets. 2. Extract the donation figure for Walmart from the English source ($10 million) and the donation figure for Aldi from the English source ($1 million). 3. Perform a mathematical aggregation ($10M + $1M = $11M) as requested by the query. 4. Critically resist the temptation to respond in English (the document's language) and instead translate the findings and the reasoning into German to satisfy the query's language constraint.
• Real Answer: Die gesamten Spenden überstiegen 11 Millionen Dollar.
• Why this demonstrates the capability: This demonstrates 'Response Language Correctness' (RLC). A weak model would often switch to English because both supporting documents are in English. The agent must maintain the target language (German) while successfully performing grounding on mismatched English-only evidence.

---

[Case 3]

• Initial Environment: A multilingual environment where a query is provided in Chinese regarding multiple specific events. Supporting Article 1 is in English covering Event X, and Supporting Article 2 is in Chinese covering Event Y. The question asks for the combined set of specific outcomes.
• Real Question: 2024年6月的这两次国际会议分别达成了哪些主要共识？ (What main consensus was reached in these two international conferences in June 2024 respectively?)
• Real Trajectory: 1. Decompose the Chinese query into two sub-intents: Consensus for Conference 1 and Consensus for Conference 2. 2. Execute a search in English to extract the consensus results for Conference 1 (e.g., climate tech investment). 3. Execute a search in Chinese to extract the results for Conference 2 (e.g., maritime trade rules). 4. Create a unified set of findings by translating the English results into Chinese and combining them with the Chinese results. 5. Verify that all entities and technical outcomes are correctly aligned across the language transition.
• Real Answer: 根据报道，第一次会议（在伦敦举行）就气候技术投资达成了共识，而第二次会议则主要确定了新的海上贸易规则。
• Why this demonstrates the capability: This case tests the agent’s ability to assemble a 'Set' response across a multilingual split-evidence environment. It requires mapping concepts between English and Chinese to ensure that the combined response is coherent and linguistically consistent with the initial user prompt.

Pipeline Execution Instructions

To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:

1. Phase 1: Environment Exploration Read the exploration guidelines to discover raw knowledge seeds: references/EXPLORATION.md

2. Phase 2: Trajectory Selection Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: references/SELECTION.md

3. Phase 3: Data Synthesis Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: references/SYNTHESIS.md