acceptodds
Under review as a conference paper at ICLR 2027

Beyond Equal Negotiation: Authority-Aware Multi-Agent Collaboration and Adaptive Distillation for Cultural Alignment

Abstract

Multi-agent negotiation has emerged as a promising approach to cultural alignment. However, existing methods generally rely on equal-weight negotiation and overlook the asymmetric distribution of cultural expertise, which may lead to culturally uninformed majorities overriding informed minority judgments. Moreover, the high inference cost of multi-agent systems limits their practical deployment. To address these challenges, we propose **A**uthority-Aware **M**ulti-**A**gent **C**ollaboration and Adaptive **D**istillation (**AMACD**) for cultural alignment. First, AMACD constructs an authority-aware multi-agent teacher via **H**ome-**F**ield **C**ulture-**A**uthority **C**ollaboration (HFCAC). The teacher dynamically activates a target-culture Guardian agent and resolves disagreements through authority-aware arbitration, with a cultural affinity-based fallback when the Guardian fails. Second, its distillation component, **H**ybrid **A**daptive **D**istillation (HAD), transfers the teacher's heterogeneous cultural expertise into a single model through role-specific SFT and RL branches. The Judge's final answers provide outcome-level supervision, while the Guardian's cultural guidance provides preference-level direction, with both signals jointly optimized using mastery-adaptive weighting. Experiments on three cultural benchmarks demonstrate that AMACD achieves stronger cultural alignment than equal-weight negotiation approaches while substantially reducing inference cost compared with multi-agent systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Cultural Alignment, Multi-Agent Collaboration, Beyond Equal Negotiation, Authority-Aware, Culture-Aware Distillation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3TDUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses two limitations in existing multi-agent frameworks for cross-cultural LLM alignment. The major one is that multiple agents with less relevant cultural knowledge (the majority) may surpass the specialized minority. The minor, common one is the high inference overhead of multi-round deliberation. The authors propose AMACD (a method combined cultural authority and knowledge distillation) to address these two limitations. The method is evaluated on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct across three datasets (NormAd, CulturalBench, and BLEnD), to show the performance and speed advantage.

Critical strengths
  1. The paper identifies a clear and practical problem regarding knowledge asymmetry in cross-cultural multi-agent deliberation. The authors observe that multiple agents with less relevant cultural knowledge (the majority) may surpass the specialized minority, which can lead to accurate, culturally grounded judgments being overridden by a less-informed consensus. Addressing this is relevant to building culturally-aligned LLMs.

  2. By integrating multi-agent deliberation with adaptive distillation, the proposed framework bridges the gap between reasoning performance and real-world deployment constraints. The authors adopt a pragmatic teacher-student paradigm to internalize collaborative capabilities into a single LLM. The results demonstrate the practical value of this design, achieving a reduction in inference cost while preserving the vast majority of the teacher ensemble's performance.

  3. The evaluation is comprehensive, spanning two LLMs (Qwen2.5-7B and Llama-3.1-8B), three cultural reasoning tasks (social norms, general cultural knowledge, and everyday practices), and several baselines ranging from zero-shot prompting to multi-teacher distillation. Furthermore, the authors provide ablation studies that isolate the incremental contributions of both the role-based collaborative mechanisms within HFCAC and the hybrid SFT–RL training paradigms within HAD.

Critical weaknesses
  1. Table 6 exhibits numerical inconsistencies. For instance, in NormAd with Qwen, the table reports a disagreement sample count of 102 and a Guardian win rate of 85.49%, which implies 87.20 correct instances. If win rate is defined as the fraction of disagreement samples where the Guardian is correct, the count of correct instances must be an integer; 87 out of 102 yields 85.29%, while 88 yields 86.27%, making 85.49% mathematically impossible on a single run. Similar discrepancies appear in NormAd with Llama (reported 86.00% over 265 samples implies 227.90 instances), CulturalBench with Qwen (74.61% over 88 samples implies 65.66 instances), and CulturalBench with Llama (72.68% over 66 samples implies 47.97 instances).

  2. Multiple references contain factual errors. For example (incorrect authors): Qiming Tan, Long Jiang, Yuzhen Zeng, and et al. AgentArk: A multi-agent distillation framework for cultural alignment. arXiv preprint arXiv:2601.12092, 2026b ; Qiming Tan, Long Jiang, Yuzhen Zeng, Sheng Ding, Xiang Xu, and et al. Mitigating cultural bias in LLMs via multi-agent cultural debate. 2026a

  3. The Guardian's cultural authority stems primarily from role prompts rather than verified expertise. While the authors conceptualize the Guardian as an agent possessing deeper target-culture familiarity and higher cultural authority, both Guardian and Auditors are instantiated from the same underlying LLM via prompting. The paper does not provide empirical evidence demonstrating that an agent assigned the Guardian persona holds verifiable, independent expertise in the target culture over Auditors.

  4. The HAD ablation does not isolate the independent contributions of Guardian guidance and mastery-adaptive weighting. HAD combines two core components—Guardian guidance and mastery-adaptive weighting—yet the current ablation in Table 5 does not cleanly separate their individual effects.

  5. Guardian guidance functions as answer-level reward shaping. Although the paper characterizes it as preference-level cultural guidance, its operational definition relies exclusively on binary answer matching between student rollouts and the Guardian's output. The Guardian's structured cultural rationales, evidentiary arguments, and debate trajectories do not directly contribute to this guidance signal.

Questions for authors

See weakness

6 comments

Regarding your first question,

(1) Statistical scope of Table 6. A disagreement sample is a sample where the Guardian's answer differs from the Auditor majority vote. Win rate = the number of samples where the Guardian is correct but the majority is wrong ÷ the total number of disagreement samples. Since the test sets are small, and this analysis is purely an evaluation analysis with no training or model selection involved, we compute it on the full datasets.

(2) All experimental results in the paper are the mean of five independent runs with different random seeds. For example, the aggregation rule of Table 6 is: for each run, we first count the number of disagreement samples and the win rate within that run; we then take the arithmetic mean over the five independent runs. Therefore, each Disagree cell is the mean of the counts over the five runs, and each Win rate cell is the mean of the per-run win rates. We will add a detailed explanation in the revised version.

Regarding your second question, thank you for pointing this out. We will correct the author and title information of the affected entries, and we will check all references one by one to ensure such errors do not happen again.

Regarding your third question,

(1) The cultural authority of the Guardian has already been directly tested in the original paper. The results are reported in Table 6. The logic of the test is as follows. The Guardian and the Auditors are instantiated from the same base LLM, and the only difference is the cultural role assignment. Table 6 compares the Guardian with the Auditor majority vote on the samples where their judgments disagree. The win or loss is determined by the gold labels. As a counterfactual baseline, a single agent without informative role assignment should in theory win against the majority at a rate below 50%. In practice, all six settings far exceed this baseline. This shows that the same LLM systematically outperforms its own majority vote after being assigned the target cultural role. Therefore, the Guardian authority is a tested mechanism premise.

(2) To directly show that the agent assigned the Guardian role has verifiable and independent expertise in the target culture over the Auditors, we list the answer accuracy of each cultural role agent on each cultural sphere. We split all questions from the three datasets into six subsets according to their target cultural sphere. We then compute the accuracy of each agent on each sphere. Under the same base model, the role agent that matches the target cultural sphere of a question serves as the Guardian for that question. The results are shown in the tables below. Each row is a question sphere. Each column is a cultural role agent. The bold value marks the Guardian of that sphere, i.e., the diagonal.

Results under the Llama-3.1-8B base model:

Question sphere \ AgentWesternLatin AmericanSub-Saharan AfricanEast-AsianIslamic & Middle-EasternSouth & Southeast Asian
Western and Anglo-Saxon72.0464.9564.0265.7764.9664.25
Latin American60.6066.2462.1360.6462.0462.34
Sub-Saharan African46.1148.6252.9445.8750.4049.77
East-Asian59.4959.1258.2065.4258.5160.32
Islamic and Middle-Eastern56.0656.6959.1656.2161.3059.22
South and Southeast Asian51.4653.5456.5851.1855.0357.92

Results under the Qwen2.5-7B base model:

Question sphere \ AgentWesternLatin AmericanSub-Saharan AfricanEast-AsianIslamic & Middle-EasternSouth & Southeast Asian
Western and Anglo-Saxon69.2466.5162.0463.5661.0662.04
Latin American61.3763.6659.1657.3459.6759.97
Sub-Saharan African56.6959.0063.4856.2761.3260.52
East-Asian62.0561.4060.5167.7260.6863.82
Islamic and Middle-Eastern58.9162.1263.9559.0766.1264.03
South and Southeast Asian59.8963.4260.6063.9262.8967.09

On both base models, each agent achieves the highest accuracy on its own cultural sphere subset, showing a clear diagonal advantage. This indicates that the cultural role prompts inject verifiable expertise into the agents. The agent assigned the Guardian role consistently outperforms the other role agents acting as Auditors on questions of its target culture.

Regarding your fourth question, the ablation of the two components' individual contributions is already provided in Table 5 (Ablation Results section). Specifically, the SFT Joint RL (λg=0) row only disables Guardian guidance, and the SFT Joint RL (w_sft=0.5) row only replaces the mastery-adaptive weighting with a fixed weight of 0.5. The rest of their configurations are identical to the full HAD. Each row disables exactly one component. Therefore, the difference from HAD reflects the individual contribution of the corresponding component. The SFT-only, RL-only, and SFT-then-RL rows serve as basic references without these two components. The results show that the full HAD outperforms both rows on all six settings, so both components make positive individual contributions. Both rows also outperform SFT-only and RL-only. This indicates that the contribution of each component holds independently without the other.

Regarding your fifth question, the structured cultural rationales, evidentiary arguments, and debate trajectories of the Guardian and the multi-agent teacher have already been incorporated into the distillation. The guidance signal is not limited to binary answer matching. Specifically, there are three signal paths:

(1) The Guardian's judgment enters the guidance signal of the guide branch. In the RL stage, the advantage term uses the consistency between a rollout and the Guardian's judgment as the guidance term. The guidance weight also adapts to the student's mastery of each culture. The less the student masters a culture, the stronger the guidance becomes.

(2) The Guardian's evidentiary reasoning enters the process reward. The process reward R_process measures whether the reasoning direction of a rollout is consistent with the reasoning direction of the Guardian. In this way, the evidence usage and the cultural argumentation logic in the Guardian's reasoning are explicitly included in the guidance signal.

(3) The multi-agent debate trajectory enters the anchor branch through the Judge's verdict. The Judge gives the verdict after reading the structured cultural rationales of all six roles and completing the debate. This verdict answer serves as the SFT supervision target of the anchor branch.

The value of debate and evidentiary reasoning is also supported by empirical evidence. Table 6 shows that the Guardian is significantly more accurate on inconsistent samples. In the teacher construction ablation, multi-agent variants such as MAD and MACD are all inferior to HFCAC (Table 1).

Reviewer_KPSJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

AMACD gives a target-culture Guardian priority in a multi-agent discussion, then distills the resulting teacher into a single model. The student combines supervised learning from the Judge's answer with a Guardian-agreement reward, weighted by an estimate of mastery for each culture. Results on NormAd, CulturalBench, and BLEnD with two model families show reported accuracy gains over debate and distillation baselines, while greatly reducing inference cost relative to the teacher.

Critical strengths
  1. The question behind the authority mechanism is sensible: a majority need not be the most reliable source on a culturally specific question. Table 4 attempts to separate the effects of persona, judge, authority, and fallback. Checking Guardian accuracy specifically on disagreements is also a useful diagnostic, provided the counts and aggregation can be clarified.
  2. Distillation addresses the cost of the proposed teacher. The student stays close to teacher accuracy while using roughly 12.6–17.5 times fewer tokens in the reported experiments (Tables 1–2). Testing two backbones and three datasets is a useful starting point for assessing whether this carries across settings.
Critical weaknesses
  1. I cannot reconcile the reported evaluation counts. An 8:1:1 split of 2,633 NormAd examples leaves about 263–264 test examples, but Table 6 reports 265 Llama disagreement examples. The same table gives 85.49% accuracy on 102 disagreements, which cannot be the rounded fraction of an integer number of correct answers out of 102. Averaging across runs might explain this, but that is not stated. Table 5 also says “10% subsample” while reproducing all six full HAD scores from Table 1. There may be straightforward explanations, but exact denominators and aggregation rules are necessary before these results can be assessed.
  2. The mechanism claims need stronger controls and a complete reward specification. The incremental authority gains in Table 4 are small, with no uncertainty estimates. The Guardian's expertise is assigned through a cultural prompt, and it is unclear how the Judge verifies the claimed cultural evidence. HAD's Guardian signal is exact answer agreement, so simpler reward-shaping or weighted-SFT controls would help establish what the role-specific formulation adds. Appendix D.3 also introduces a process reward without fully defining how it is computed. Without these details, it is difficult to tell which part of the design explains the gains.
Questions for authors
  1. Please provide exact train/validation/test counts, split seeds, and evaluation units for every table. How can Table 6 contain 265 NormAd test disagreements, and how is 85.49% computed from 102 samples? Are counts or rates aggregated over multiple runs?
  2. What does “10% subsample” in Table 5 refer to, and why does the HAD row exactly match Table 1? Please report repeated-run uncertainty and paired comparisons on the same held-out examples.
  3. Please fully specify outcome reward, process reward, mastery-hit-rate targets, and the source of their labels. Does HAD retain its advantage over a compute-matched joint SFT/RL baseline with the same reward information and a simpler adaptive weighting rule?
  4. How is Guardian reliability verified, how often is the affinity fallback used, and what are its conditional error rates? Please distinguish target-culture relevance from genuine evidence of expertise.
  5. In the efficiency experiment, what does “context length 64” mean alongside approximately 825–872 student tokens per query (lines 373–377; Tables 2–3)? Please separate prompt and generated tokens and state the concurrency, generation limit, hardware allocation, and timing boundaries.
7 comments

Regarding your first question,

(1) Statistical scope of Table 6. A disagreement sample is a sample where the Guardian's answer differs from the Auditor majority vote. Win rate = the number of samples where the Guardian is correct but the majority is wrong ÷ the total number of disagreement samples. Since the test sets are small, and this analysis is purely an evaluation analysis with no training or model selection involved, we compute it on the full datasets.

(2) All experimental results in the paper are the mean of five independent runs with different random seeds. The aggregation rule of Table 6 is: for each run, we first count the number of disagreement samples and the win rate within that run; we then take the arithmetic mean over the five independent runs. Therefore, each Disagree cell is the mean of the counts over the five runs, and each Win rate cell is the mean of the per-run win rates. We will add a detailed explanation in the revised version.

(3) AMACD and HAD are essentially the same model, named from different perspectives: AMACD refers to the single model finally distilled by our framework, and HAD refers to the distillation method itself. Table 1 names this configuration from the overall performance perspective, and Table 5 names it from the component-ablation perspective. Both are evaluated on the same test set, i.e., the held-out test split obtained by the 8:1:1 split of the full data.

Regarding your second question,

(1) All results reported in the paper are the mean of five independent runs with different random seeds, not the result of a single run; they reflect stable improvements rather than random fluctuations. We will add this statement in the revised version, and report the mean ± standard deviation and the 95% confidence interval.

(2) The relevant prompts are all given in Appendix F. The Judge verification procedure is written in Appendix F.4: it checks whether the evidence exists, whether it points to the target cultural context of the question, and whether the evidence is consistent with the conclusion. The Guardian failure condition (explicitly expressing uncertainty, or a broken output format) is clearly defined; upon failure, the system switches to the affinity fallback instead of the majority vote. We have explained this arbitration mechanism in Section 3.1 of the main paper; see Appendix F for details.

(3) The simpler reward-shaping and weighted-SFT controls are already given in Table 5, and the extra contribution of the role-specific formulation is exactly the margin of HAD over these controls. In our implementation, the Guardian guidance itself is a reward-shaping term added to the GRPO advantage: Ai=(ri−rˉ)+λgαcSguardian,iA_i = (r_i - \bar{r}) + \lambda_g \alpha_c S_{guardian,i}, where SguardianS_{guardian} is the exact match between the rollout and the Guardian's answer. Compared with simple shaping, it differs in only two design choices: αc=1−hitratec\alpha_c = 1 - hitrate^c adapts to the student's mastery instead of being fixed; and the Guardian guidance is jointly optimized with the Judge-supervised SFT instead of in separate stages. The ablation rows of Table 5 correspond to these factors: λg=0\lambda_g = 0 is the simpler reward-shaping control, wsft=0.5w_{sft} = 0.5 is the weighted-SFT control, and SFT-only / RL-only / SFT-then-RL are simpler training paths. HAD outperforms all of these controls, and each control is trained with the same framework, the same number of rollouts (G=8G = 8), and the same number of training steps as HAD (compute-matched).

(4) The appendix defines Rtotal=αRoutcome+(1−α)Mean(Rprocess)R_{total} = \alpha R_{outcome} + (1-\alpha) Mean(R_{process}). RoutcomeR_{outcome} is 1 if the final answer parsed from the rollout matches the answer of the training sample, and 0 if the output cannot be parsed; RprocessR_{process} indicates whether the reasoning direction of the rollout is consistent with the reasoning direction of the Guardian in the multi-agent teacher.

Regarding your third question,

(1) context length 64 refers to the upper limit on the number of tokens of the question text itself for each question in the datasets used in the inference experiment. This number does not include the system prompt, template instructions, or other input content.

(2) Table 2 reports the total number of input and generated tokens in a full call. Note: the teacher-side prompt consists of three parts — the question, the system prompts of the six cultural roles, and the back-filled upstream answers (each Phase 2 Auditor prompt carries the Guardian's full answer, and the Phase 3 Judge prompt carries all six answers).

(3) Generation limit. The generation limit of each agent is set to max_tokens = 1024.

(4) Hardware allocation. We deploy the models with vLLM on 2 NVIDIA V100 GPUs (32GB) with tensor parallelism (tensor parallel size 2). The teacher and the student use the same hardware and inference framework configuration.

(5) Timing boundary. Latency is measured from the moment the prompt is built (the request is sent) to the moment the final answer is parsed and produced.

Reviewer_bu84OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work aim to ask agents systems to solve culture alignment tasks. Current methods focus on simple majority voting where each agent has the same weight of vote. They propose AMACD. It builds a multi-agent teacher which does well in culture alignment, which gives higher weights to those Guardian agents that knows the target culture well. Then they distill the teacher to a single agent student through SFT and RL.The results show that their AMACD shows slightly stronger performance with less inference cost.

Critical strengths
  1. Valid limitation of current methods detected. I think majority voting can indeed lead to worse perfromance if most of the agents are not experts in the domain.
  2. The inference time reduction is amazing. After distilation, the student can achieve the answer with 12-18 times fewer tokens and answer 10 times faster. This is practical since the users are sensitive to the severe latency.
  3. They performed thorough ablation study to test on each component's effectiveness.
Critical weaknesses
  1. The Guardian (i.e., the domain expert) is simply achieved by prompt instruction, asking a ordinary agent to play the role. However, I feel like there's no guarantee that an agent can be more familiar with Korean culture if you instruct them that they are Korean-culture specialist. There might need some experiments to verify that with/wothout instruciton prompt on some specific Korean-culture QAs.
  2. Its representation of culture is too broad. Six cultural groups cannot fully represent differences within countries, generations, communities etc. Maybe more principled ways of culture grouping can be used, e.g., from broadly adopted socialogy literatures.
  3. The performance improvement is too little. For Qwen, it's only from 69.2% (euqal voting) to 70.28%. Withoout error bar or CI, I don't know if we can consider this a valid improvement.
Questions for authors
  1. Can you verify if "instruction prompting" can elicit domain knowledge by extensive quantitative experiments. I think this is a foundation / premise for the work?
  2. Is there any justification on your culture grouping and its granularity? If not, can you try other grouoping and validate performance effectiveness?
  3. Can you add error bar to your results?
3 comments

Regarding your first question, previous work has shown that role assignment can improve a model's understanding of the corresponding culture (Kamruzzaman et al., 2024; Araujo et al., 2025, EMNLP). In the baselines of our paper, Role-play differs from Base by adding a cultural role assignment: the target cultural identity is specified in the system prompt. As shown in Table 1, assigning a cultural role brings a consistent gain on questions that match the target culture.

To further verify this conclusion, we added a new role-instruction controlled experiment. Its purpose is to separate the expert role setting from specific cultural knowledge, and to test whether assigning a specific cultural role makes the model better understand the target culture. All three settings: (1) Base, with no role assignment, which is the baseline in Table 1; (2) General expert, told only that it is a cultural expert, without specifying any particular culture; (3) Role-play, assigned as an expert of the target culture, which is the baseline in Table 1.

Results on the full datasets are as follows:

MethodNormAd (Qwen)CulturalBench (Qwen)BLEnD (Qwen)Avg (Qwen)NormAd (Llama)CulturalBench (Llama)BLEnD (Llama)Avg (Llama)
Base (no role)59.78±0.0672.13±0.1366.09±0.0066.00±0.0658.15±0.0667.21±0.1565.95±0.0063.77±0.07
General expert (told only it is a cultural expert)59.39±0.0971.20±0.2366.27±0.0265.62±0.0858.36±0.0267.33±0.1565.87±0.0263.86±0.05
Role-play (assigned a specific cultural role)59.51±0.0771.31±0.1366.52±0.0065.78±0.0758.64±0.0668.03±0.1565.87±0.0064.18±0.07

The sample sizes of the three Korean subsets are 27 (NormAd), 41 (CulturalBench), and 2,512 (BLEnD). Results on Korean QA are as follows:

MethodNormAd (Korea, Qwen)CulturalBench (South Korea, Qwen)BLEnD (South Korea, Qwen)Avg (Qwen)NormAd (Korea, Llama)CulturalBench (South Korea, Llama)BLEnD (South Korea, Llama)Avg (Llama)
Base (no role)66.67±0.0063.41±1.7265.50±0.0565.19±0.5866.67±2.6256.59±1.0965.90±0.0563.05±0.96
General expert (told only it is a cultural expert)66.67±2.6263.90±1.0966.18±0.0265.58±0.9567.41±1.6657.07±1.3465.84±0.0663.44±0.54
Role-play (assigned a specific cultural role)67.41±3.1065.37±2.0466.44±0.0366.40±1.0568.89±2.0359.02±1.0965.98±0.0464.63±0.89

All results above are the mean ± standard deviation over five independent runs with five different random seeds. The results indicate that the gain mainly comes from the cultural information injected by assigning a specific cultural role, rather than from the expert setting itself.

Regarding your second question:

(1) Section 4 (Agent configuration) of our paper explicitly states that the six cultural spheres are derived from the WVS Inglehart–Welzel cultural map by merging adjacent cultural clusters, citing Inglehart & Welzel (2005). That is, the cultural grouping is based on the Inglehart–Welzel cultural map. Appendix G (Construction Method and Usage of the Cultural Affinity Matrix) further provides the complete construction process.

(2) The sensitivity to the number of groups and to the grouping scheme is supported by experiments. We added two groups of ablation studies in Appendix A (Structural Analysis of HFCAC). The first is an ablation on the number of cultural groups (Appendix A.1): it shows that six groups is the most robust configuration in our design, and that the six spheres of the WVS map already reach the upper limit of cultural coverage. The second is a comparison of grouping schemes (Appendix A.2, HFCAC-6 / Diverse-6 / Random-6 / Same-Sphere-6): the pipeline is kept unchanged, and only the way of assigning cultural backgrounds to the six agents is replaced. The ordering HFCAC-6 > Diverse-6 > Random-6 > Same-Sphere-6 holds consistently on all three datasets and both backbones. This shows that the WVS-based grouping is a better realization of diversity.

The original results of Appendix A.1 are as follows (in table form):

# of groupsNormAd (Qwen)NormAd (Llama)CulturalBench (Qwen)CulturalBench (Llama)BLEnD (Qwen)BLEnD (Llama)
266.8464.4173.9469.4466.6165.13
369.9664.4974.1869.6067.3765.31
466.3564.8774.3569.9365.6265.53
565.5865.5175.4970.3367.5165.22
666.8165.8675.5772.3768.4567.25

The original results of Appendix A.2 are as follows:

SchemeQwen NormAdQwen CulturalBenchQwen BLEnDLlama NormAdLlama CulturalBenchLlama BLEnD
HFCAC-6 (WVS six spheres)66.8175.5768.4565.8672.3767.25
Diverse-666.4675.2268.3165.0272.0567.16
Random-665.3674.4168.1864.6471.3167.08
Same-Sphere-664.9174.1668.1264.2271.2366.97

Regarding your third question, we would like to clarify: all results reported in the paper are averaged over five independent runs with different random seeds, not obtained from a single run. Therefore, this improvement holds after averaging over multiple independent evaluations. It reflects a stable improvement, not random fluctuation. Due to the page limit, the original version does not report confidence intervals. In the revised version, we will add the mean ± standard deviation and the 95% confidence interval, to present the stability of the results more completely.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.