acceptodds
Under review as a conference paper at ICLR 2027

Multi-Expert Collaborative Optimization for Low-Resource Machine Translation through Debate and Ant Colony Search

Abstract

Despite recent advances in neural machine translation, low-resource translation remains constrained by insufficient parallel supervision and model-specific generation biases. Heterogeneous experts can provide complementary translation hypotheses, yet conventional multi-expert methods often rely on direct reranking or fixed fusion, limiting iterative collaboration. We propose a collaborative translation optimization framework that combines hierarchical multi-agent debate with ant colony search. Frozen heterogeneous experts first generate initial translations to construct an input-dependent collaboration graph, over which ant agents explore refinement and fusion paths. Intra-group debate identifies and corrects local translation errors, while inter-group interaction exchanges elite paths and complementary evidence. Debate feedback is further incorporated into transition estimation and pheromone updating, jointly guiding subsequent search without modifying the expert models or training an additional reinforcement-learning policy. Experiments on low-resource Mongolian–Chinese translation achieve 30.8 BLEU, 55.1 chrF++, and 0.463 COMET, improving over the strongest controlled baseline by 0.6 BLEU, 0.7 chrF++, and 0.009 COMET, while requiring fewer model calls and lower inference latency. Ablation and convergence analyses further validate the effectiveness of the proposed collaborative search framework.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Low-resource machine translation, Multi-expert collaboration, Multi-agent debate, Ant colony optimization, Collaborative search

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Ga3ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a multi-expert collaborative optimization framework for low-resource Mongolian–Chinese machine translation, combining hierarchical multi-agent debate with ant colony optimization (ACO). The framework constructs a translation collaboration graph from candidate translations generated by heterogeneous expert models and explores iterative refinement and fusion paths. Intra-group debate identifies local translation errors, while inter-group interaction facilitates the exchange of high-quality paths and complementary information. Debate feedback is incorporated into transition estimation and pheromone updating to guide subsequent search.

Experiments on Mongolian–Chinese translation demonstrate improvements over several baselines, including single-expert translation, quality-estimation reranking, static multi-expert fusion, standard ACO, and flat multi-agent debate. Ablation studies and convergence analyses further examine the contributions of the proposed components.

Critical strengths
  1. An interesting integration of multi-agent debate and collaborative search. The paper combines hierarchical multi-agent debate with ACO, allowing linguistic feedback to influence both transition estimation and pheromone updating. This provides a potentially useful mechanism for iterative translation refinement beyond conventional candidate selection and static fusion.

  2. Consistent improvements over the evaluated baselines. The proposed method achieves the best BLEU, chrF++, and COMET scores among the evaluated approaches. The ablation results suggest that hierarchical debate, debate-aware search, and candidate fusion each contribute to the overall performance. The method also requires fewer model calls and lower inference latency than the flat multi-agent debate baseline.

Critical weaknesses
  1. Limited consideration of Mongolian-specific linguistic characteristics. Although the paper focuses on Mongolian–Chinese machine translation, the proposed framework is largely language-agnostic. Mongolian has rich morphology, agglutinative word formation, and syntactic structures that differ substantially from Chinese. However, these linguistic characteristics are not explicitly incorporated into the debate or search mechanisms. Consequently, the task-specific motivation and novelty of the proposed approach remain insufficiently established.

  2. Insufficient comparison with modern state-of-the-art language models. The expert pool consists of LLaMA-7B, BLOOM-7B1, and Qwen-7B, which are not representative of the capabilities of recent high-performing language models. Although the method outperforms the selected baselines, it remains unclear whether the proposed collaborative framework offers meaningful advantages over directly using a stronger contemporary model. Furthermore, the paper does not investigate whether combining modern efficient language models as heterogeneous experts could achieve a better translation quality–cost trade-off.

Questions for authors

I would appreciate clarification and additional evidence regarding the following two directions. Substantively addressing at least one of them would significantly affect my assessment of the paper.

  1. Mongolian-specific linguistic characteristics: Given the morphological and syntactic differences between Mongolian and Chinese, how does the proposed hierarchical debate mechanism address translation errors related to morphology, grammatical relations, word order, and semantic omissions? Could the authors provide a linguistically grounded error analysis or demonstrate whether explicitly incorporating Mongolian-specific linguistic information into the debate mechanism improves translation quality?

  2. Comparison with modern SOTA models: Why were LLaMA-7B, BLOOM-7B1, and Qwen-7B selected as the expert models instead of more recent and capable alternatives? Could the authors compare their framework against strong contemporary language models on the same Mongolian–Chinese test set? Alternatively, could they construct a heterogeneous expert pool using modern efficient or Flash-style language models and evaluate whether the proposed collaborative optimization framework can achieve competitive or superior translation quality under comparable inference costs?

I do not necessarily expect both directions to be fully explored. However, convincing evidence addressing at least one of them is important for establishing either the linguistic relevance or the practical advantage of the proposed approach.

Reviewer_ekzSOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for Mongolian-to-Chinese translation. It combines multiple translation experts, multi-agent debate, and ant colony optimization. The agents revise and combine different translations through a search process. The method achieves 30.8 BLEU and 0.463 COMET.

Critical strengths
  1. The paper explores an interesting way to combine multi-agent debate with ant colony optimization.
  2. The experimental evaluation includes several relevant baseline categories, component ablations, convergence analysis, and computational cost measurements.
  3. The method improves translation quality without updating the expert models during search.
Critical weaknesses
  1. Important implementation details are missing. The method relies heavily on a reference-free quality estimator. However, the paper does not clearly state which model is used. It also gives limited details about the debate prompts and graph construction. This makes the method difficult to reproduce.
  2. The low-resource setting needs further justification. The experiments use about one million parallel sentence pairs. It would be helpful to clarify how this dataset size relates to the low-resource setting. Additional experiments with smaller training sets could further strengthen the evaluation.
  3. The improvement over the strongest baseline is small. The method improves BLEU by 0.6 and COMET by 0.009 over Flat MAD, but the paper does not report standard deviations or significance tests. It is unclear whether these gains are reliable.
  4. The computational cost could be explained in more detail. The paper reports 9.7 model calls on average, but does not provide a detailed breakdown. It is also unclear whether the cost of quality estimation is included in the reported latency. More details would help assess the efficiency comparison.
  5. The case study could better illustrate the proposed method. Figure 2 compares the final translations, but does not show the debate process or search paths. An example with intermediate steps would help explain the benefit of ant colony search.
Questions for authors
  1. Which model is used for reference-free quality estimation?
  2. How are new translation states generated during graph search? Can the authors provide a detailed example?
  3. Can the authors test the method with smaller training datasets?
  4. Are the improvements over Flat MAD statistically significant?
  5. Can the authors provide a breakdown of model calls and report the cost of quality estimation?
Reviewer_MrYtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a reference-free, training-free inference-time framework for low-resource Mongolian-Chinese translation by combining hierarchical multi-agent debate with ant colony optimization . Without updating model parameters or training an RL policy, the method structures initial translations from K frozen heterogeneous LLMs into a graph, where edges represent refinement or fusion actions validated by a reference-free QE score. Nine logical ant agents traverse this graph, using multi-dimensional intra-group critiques and elite inter-group exchanges to compute a debate-calibrated score that dynamically guides ACO transition heuristics and pheromone updates. Evaluated on ~1M sentence pairs, the framework achieves 30.8 BLEU / 55.1 chrF++ / 0.463 COMET, outperforming Flat MAD by +0.6 BLEU with fewer model calls and lower latency, while demonstrating peak convergence at three iterations.

Critical strengths
  1. A principled coupling of linguistic feedback with population-based search, validated by ablations. Debate-derived signals (quality, agreement, refinement gain) condition both the ACO transition heuristic and pheromone updating, giving ACO the linguistic evaluation it lacks and giving MAD the persistent search memory it lacks. The ablation study supports this synergy rather than merely asserting it: removing debate-calibrated pheromone updating causes the largest degradation, exceeding the effect of removing either debate level. This LLM feedback as heuristic guidance inside a classical metaheuristic may transfer to other structured generation tasks.

  2. A practical, training-free, reference-free, and cost-aware design. No parameter updates, trained rerankers, or reference translations are needed, and the utility function explicitly penalizes model calls. The method is simultaneously more accurate and cheaper than Flat MAD, and the convergence analysis is used constructively to justify the three-iteration budget.

Critical weaknesses

1 Narrow evaluation and weak statistical evidence. Experiments cover a single language pair with ~1M parallel sentences so the paper's central motivation is not directly tested. The headline gains over Flat MADare small, and no significance tests or per-seed standard deviations are reported, leaving the core empirical claim statistically unestablished. There is also no comparison to strong decoding-time baselines such as MBR decoding.

2 Severe underspecification, and the necessity of ACO at this scale is unclear. The refinement/fusion operators, the QE estimator, debate prompts, and all hyperparameters are unspecified, making reproduction impossible. With only 9 ants, path length 3, and 3 iterations, the reachable search space is tiny; without an enumeration or beam-search control over the same graph, one cannot tell whether the ant-colony machinery contributes anything beyond exhaustive search. Moreover, all ants share one frozen instruction model, so debate diversity is questionable

Questions for authors
  1. Have you compared against exhaustive enumeration or beam search over the same collaboration graph (same operators, same QE, same budget)? Given the small search space, does ACO find better paths than enumeration, or is its benefit purely the early-stopping cost saving?

  2. Which QE model is used for Q_QE (and is it the same metric family as the COMET evaluation, which could inflate COMET)? How are A_ref and A_fus implemented, and which model executes them? Can you release the debate prompts and the values of all hyperparameters (alpha, beta, rho, lambda_q/d/c, lambda_1/2/3, lambda_cost, r, epsilon_stop)?

  3. How does the method perform on a genuinely low-resource split (e.g., 50k-100k training pairs for the experts) and on at least one additional language pair? Does the relative ordering against Flat MAD and Standard ACO hold?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.