Social media translation requires preserving meaning while capturing the cultural nuances of user-generated content (UGC). However, gaps in the cultural knowledge of LLM-based evaluators can lead to inaccurate assessments and unreliable training rewards. To improve these signals, we use retrieval-enhanced cultural evaluation to train CUE-MT, a family of social media translation models. A calibrated, domain-specific Judger guides supervised data refinement and is augmented by an Agent for hard-example mining and reinforcement learning. The Agent retrieves relevant external evidence as needed and distills it into concise hints that the Judger uses to score translations and assign rewards. Across 9B, 27B, and 35B-A3B backbones, CUE-MT outperforms the evaluated open-source baselines of comparable size on CUE-MT-Bench and three public benchmarks. Human evaluation on CUE-MT-Bench shows cultural effectiveness gains of 7.44-15.12 percentage points over the strongest external baseline in each size group. Retrieval improves the Judger's four-way classification accuracy by 8.07 percentage points over a matched no-retrieval baseline that uses LLM-generated hints alone. Training ablations further support the combination of hard-example mining and Agent-enhanced rewards.
Figures & tables
Figure 1: An example of a cultural knowledge gap in translation evaluation. “硬控” refers to immobilizing an opponent in gaming, but it can also mean being captivated by content on social media. Applying the gaming meaning may cause a static Judger to overrate “forcibly controlled”. Retrieved evidence helps the Agent-enhanced Judger interpret the expression in context and assign a more appropriate score.
Figure 2: Overview of CUE-MT training. The static Judger guides supervised data refinement (Refine). The Agent-enhanced Judger uses external cultural evidence to score translations for hard-example mining (Mine) and provide cultural-effectiveness rewards during GRPO (Optimize).
Table 1: Evaluation of cultural-effectiveness Judgers on our human-annotated benchmark.
CUE-MT-Bench
CULTURE-MT
RedTrans
Seed-X-Challenge
Model
Agent
Gemini
Human
Agent
Gemini
Agent
Gemini
Agent
Gemini
Model scale: ≤ 9B
Seed-X-7B-SFT
67.96
64.75
77.73
72.55
52.76
78.31
76.24
82.23
76.53
Seed-X-7B-PPO
72.01
71.46
79.78
80.24
69.17
80.26
80.82
84.77
78.68
HY-MT2-7B
68.13
67.14
77.02
79.14
74.45
79.18
80.10
87.82
86.29
Qwen3.5-9B
67.28
56.73
68.74
81.14
75.15
80.65
84.13
89.34
85.28
Table 2: Comparison of CUE-MT with representative translation models on cultural-effectiveness benchmarks. Agent, Gemini, and Human denote evaluations by the GRPO-backed Agent-enhanced Judger, Gemini API Judger, and human annotators, respectively. Best results among comparable-size models are shown in bold.
Figure 3: Human-rated cultural effectiveness on CUE-MT-Bench (percentage scored 2 or 3). CUE-MT combines compact model size with scores approaching those of larger systems.
Training Strategy
Agent
Gemini
Human
Qwen3.5-27B
79.88
61.80
77.53
CUE-MT w/o GRPO
84.08
93.88
90.31
CUE-MT-27B
Hard Mining + Agent Reward
88.27
95.61
92.65
Hard Mining + Static Reward
84.08
94.59
91.32
Random Data + Agent Reward
85.70
95.51
92.04
Table 3: Ablation study on CUE-MT-27B. All GRPO variants start from the same CUE-MT w/o GRPO checkpoint. Hard-example mining and reward evaluator effects are evaluated under four combinations.
Figure 4: Changes in FLORES-zh2en scores relative to the corresponding Qwen3.5 base models.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Qwen3.5-9B
Qwen3.5-27B
Qwen3.5-35B-A3B
Tensor Parallel (TP)
4
4
2
Pipeline Parallel (PP)
1
1
1
Expert Parallel (EP)
1
1
8
Learning rate
1×10−6
1×10−6
5×10−7
Rollout batch size
256
256
128
Global batch size
256
256
128
Appendix
Table 4: Model-specific GRPO training configurations.
Example
Human
Static Judger
Agent-enhanced Judger
843: Contextual slang
3 / 1.0
2 / 1.0
3 / 1.0
938: Named entity error
2 / 1.0
3 / 1.0
1 / 0.1
226: Abbreviation and intent error
1 / 0.1
1 / 0.1
2 / 1.0
Appendix
Table 5: Selected human-annotated cases. Each cell reports score / implied rculture . Rewards are computed according to the translation-training mapping described in Section 3.2 . No claim is made that these examples were used for training.
Model
Agent
Gemini
BLEU
chrF
COMET
Model scale: 9B
Qwen3.5-9B
94.96
95.95
26.83
57.54
87.32
CUE-MT w/o GRPO
93.67
93.37
22.96
54.11
85.82
CUE-MT
94.66
93.48
21.09
52.55
85.22
Model scale: 27B
Qwen3.5-27B
97.04
98.52
28.72
59.06
87.75
Appendix
Table 6: General-domain translation results on FLORES-zh2en. We report Agent and Gemini judgments together with reference-based BLEU, chrF, and COMET scores.
Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.