Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection

Technical Summary: LOONG – A Human-Like Long Document Translation Agent

Problem Statement
Document-level machine translation (DocMT) remains a significant challenge for Large Language Models (LLMs). While LLMs excel at sentence-level translation, they struggle with long documents due to limited context windows, which impede global cohesion, and the presence of redundant contextual information that degrades translation quality. Existing approaches often either fail to maintain strict sentence-level alignment (Doc2Doc methods) or indiscriminately incorporate extraneous context, introducing noise that leads to suboptimal performance. Furthermore, standard chunking methods often fail as document length increases, causing context capacity limits to be exceeded and resulting in translation failure.

Methodology
The authors propose LOONG, a human-like long-document translation agent designed to mimic the cognitive processes of professional human translators. The methodology is built upon three core components:

1. Human-Like Memory Module (3E):
   LOONG employs a multi-granularity memory architecture updated after processing each document segment. This module stores three distinct types of historical information:

   • Essence: Global summaries of preceding segments to capture semantic progression and genre signals.
   • Exemplars: Bilingual sentence pairs (source-target) to maintain stylistic consistency and provide in-context demonstrations.
   • Entities: Structured records of entities (e.g., characters, organizations) with bilingual names and domain-specific attributes to ensure terminological precision.

2. Observe-and-Act Adaptive Context Selection:
   Instead of passively attending to all retrieved history, LOONG utilizes a deep reasoning mechanism to distinguish useful context from redundancy. The agent executes a sequential 3-step reasoning process for each segment:

   • Observe: The agent observes the current segment alongside candidate items retrieved from the Essence, Exemplar, and Entity memories.
   • Act: The agent generates a reasoning thought analyzing the semantic relevance of each candidate and selects the optimal subset for translation.
   • Sequential Decomposition: To manage the combinatorial search space, the agent explores context selection strategies for each memory type (Essence, then Exemplar, then Entity) sequentially rather than jointly.

3. Reinforcement Learning and Alignment:

   • Preference Data Construction: The system samples diverse "observe-and-act" trajectories to construct preference datasets. It identifies preferred and dispreferred context selection actions and translation outputs based on quality metrics (sCOMET).
   • Training: The agent is optimized via a two-stage process: Cold-start Supervised Fine-Tuning (SFT) on preferred instances, followed by Direct Preference Optimization (DPO) using the full preference dataset. Low-Rank Adaptation (LoRA) is used to improve training efficiency.
   • Alignment-Enforced Inference: To ensure strict sentence-level alignment between source and target (crucial for evaluation and memory updates), LOONG employs a recursive algorithm. If an LLM output fails to align with source sentence boundaries, the segment is recursively split until alignment is achieved.

Key Contributions

• LOONG Agent: The development of a DocMT agent that retrieves and adaptively identifies useful contextual information through deep reasoning, rather than relying on static context windows.
• Sequential Exploration Strategy: A novel method for constructing preference data by decomposing the context selection process into sequential steps, enabling fine-grained feedback on the utility of specific context types (Essence, Exemplar, Entity).
• Alignment-Enforced Algorithm: A specialized algorithm for document-level translation that guarantees strict sentence-level correspondence between source and generated target documents, combining the strengths of Doc2Sent and Doc2Doc approaches.

Experimental Results
Empirical evaluations were conducted on English $\Leftrightarrow$ Chinese, German, and French translation tasks across various backbone models (Qwen2.5/3, Llama3.1) and domains (News, Speech, Literary).

• Performance Gains: LOONG achieved substantial improvements over baselines (Sentence, Segment, Doc2Doc, and the memory-based agent DELTA). It demonstrated average gains of up to 13.0 points across sentence-level (sCOMET), document-level (dCOMET), and LLM-as-a-Judge metrics.
• Robustness: The agent exhibited strong generalization across unseen domains (e.g., web novels, TED talks) and languages (including low-resource pairs like Czech, Portuguese, and Russian $\to$ English).
• Noise Resistance: In experiments with injected pseudo-context noise, LOONG maintained stable performance, whereas baseline methods suffered significant degradation.
• Ultra-Long Document Stability: In a case study translating the first 12 chapters of Journey to the West (Chinese $\to$ Portuguese), LOONG successfully completed the task without context overflow, while standard full-history methods failed once the document exceeded context limits.

Significance
The paper claims that LOONG addresses the longstanding bottlenecks of document-level translation, specifically cohesion and coherence in ultra-long texts. By bridging the gap between deep reasoning capabilities and document-level translation, LOONG demonstrates that adaptive, reasoning-driven context selection is superior to indiscriminate context inclusion. The work suggests that such agents can be deployed directly in diverse domains without additional tuning, offering a robust solution for high-quality, long-document translation tasks where maintaining global consistency and terminology precision is critical.