Source-linked AI summary
Collective Bias Mitigation via Model Routing and Collaboration
Mingzhe Du, Luu Anh Tuan, Xiaobao Wu, Yichong Huang, Yue Liu, Dong Huang, Huijun Liu, Bin Ji, Jie M. Zhang, See-Kiong Ng
TL;DR
LLMs can perpetuate training-data biases, while relying on a single model's intrinsic knowledge may be insufficient to address them. CBM learns fine-grained model behavior to route queries and coordinate diverse LLMs; Debating achieves low bias, while Committee reduces age bias at lower inference cost than Debating.
Problem
LLMs can perpetuate or amplify biases in their training data, and self-debiasing may be insufficient when models lack robust external supervision to recognize deeply rooted bias.
Source: §1 · p. 1Method
CBM collects fine-grained model responses, trains a router to select models for each query, and organizes selected LLMs into collaborative topologies.
Source: §1 · pp. 1–2, §3.3 · p. 3Results
Takeaways & Limitations
Committee offers a balance between bias reduction and computational cost, while Debating achieves the lowest bias in the reported comparisons.
Source: §6 · p. 6, §5 · p. 6Evidence is limited to ambiguous English-language BBQ questions and does not establish general trustworthiness or transfer of bias reductions across languages, cultures, or applications.
Source: F. Scope and Limitations · p. 15
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly deployed in public health, finance, and governance, requiring both accuracy and societal value alignment. Despite recent advances, LLMs often perpetuate or amplify bias embedded in their training data, posing challenges to fairness. While self-debiasing encourages an LLM to identify and correct its own biases, relying on a single model's intrinsic knowledge may be insufficient to address deeply ingrained stereotypes. To address this limitation, we introduce Collective Bias Mitigation (CBM), a framework that alleviates bias by learning fine-grained model behavior and fostering knowledge sharing among diverse LLMs. This work is the first to systematically explore the effective selection and organization of distinct LLMs to cultivate fairer LLM responses. Experiments show CBM substantially outperforms standalone baselines (e.g., in the top-7 setting, Committee lowers the age bias score from 0.25 to 0.10). Our Debating and Committee topologies achieve substantial bias reduction, with the latter balancing mitigation effectiveness and inference cost, highlighting the potential of CBM for fairer LLMs.
1. Introduction
LLMs can reproduce training-data biases, while self-debiasing may be insufficient when models rely on their own intrinsic knowledge. CBM routes queries to selected models and organizes them to collaborate, with experiments reporting reduced bias and a cost-effective Committee topology.
1. Introduction
- LLMs deployed in public health, finance, and governance can perpetuate or amplify training-data biases, raising fairness concerns alongside demands for accuracy and societal values.Self-debiasing may fail to recognize deeply rooted biases without robust external supervision.Source: §1 · p. 1
- CBM collects bias-eliciting responses into CrowdEval, trains a router to detect nuanced biases and select models, then organizes them to exchange knowledge and mitigate bias.The study describes this as the first systematic exploration of selecting and organizing distinct LLMs for more equitable responses.Source: §1 · pp. 1–2
- The framework trains its router on CrowdEval and uses it at inference to detect bias types, select models, and integrate their responses.Figure 2 distinguishes training steps with dashed blue lines from inference steps with solid green lines.Source: Figure 2 · p. 2
- Debating often achieves the lowest bias; Committee reduces age bias from 0.25 for the Single baseline to 0.10 in the top-7 setting and costs less inference than Debating.The experiments cover more than 50 leading LLMs and various social dimensions.Source: §1 · p. 2
2. Related Work
Prior work evaluates bias with benchmarks that often provide holistic model scores, and mitigation approaches commonly focus on individual LLMs. CBM instead combines fine-grained model behavior with multi-model selection and collaboration.
2. Related Work
- CrowS-Pairs and StereoSet compare response pseudo-likelihoods, while BBQ and BiasLens use structured question-answering tasks to probe model bias.These benchmarks often report only holistic scores, obscuring fine-grained model behavior; CrowdEval is introduced to capture per-query behavior.Source: §2 · p. 2
- Prior mitigation spans training-data changes, weight updates, prompting, and output filtering, but primarily targets bias within individual LLMs.CBM is presented as a multi-model collaborative scheme intended to mitigate bias beyond individual-model approaches.Source: §2 · p. 2
- LLM ensembles include pre-inference selection, in-inference token-level fusion, and post-inference integration of decisions made by individual models.CBM selects models for each query and synergizes their decisions in specific topologies.Source: §2 · p. 2
3. Collective Bias Mitigation.
CBM routes each query to selected LLMs, then organizes them into topologies that coordinate responses toward a final answer. Its design combines fine-grained model behavior data, query-specific routing, and different forms of model collaboration.
3.1. CrowdEval Dataset Construction
- CrowdEval captures fine-grained model behavior by collecting LLM responses to bias-eliciting questions from the ambiguous subset of BBQ (Parrish et al., 2021).For most social dimensions, the dataset samples 1,024 questions; dimensions with fewer available instances include all of them.Source: §3.1 · p. 3
3.2. Model Routing
- The router recommends models judged most neutral for a query, using learned distinctions in their behavior and bias patterns.The routing approach fine-tunes a pretrained LLM and ranks candidate models by predicted token probabilities, replacing model names with unique identifiers during training.Source: §3.2 · p. 3
3.3. Collective Bias Mitigation Topologies
- The Single topology is the baseline: the router selects its highest-ranked model, which answers in one turn.The model pool contains over 50 open-source LLMs spanning different sizes, architectures, and training corpora.Source: §3.3 · p. 3, §3.1 · p. 3
- Sequential passes each model’s response to the next model, which can revise the prompt using earlier responses; the last model supplies the final answer.Self-debiasing is a special case that uses the same model throughout.Source: §3.3 · p. 3
- Voting has selected models answer independently and determines the final response by majority vote.Each model generates its own response in parallel.Source: §3.3 · p. 3
- Debating starts with independent responses, then iteratively shares them among models until consensus is reached.The updated prompt incorporates all models’ responses.Source: §3.3 · pp. 3–4
- Committee assigns a coordinator to query the other models, draft a consolidated motion, and seek their approval.The process iterates to consensus with a 50% threshold, and the first model is always the coordinator.Source: §3.3 · p. 4
4. Experiments
The experiments use ambiguous BBQ questions to assess implicit bias and evaluate routing with bias-detection accuracy and model-selection precision. They test a pool of open-source models and specify how candidates are assigned to each topology.
4.1. Bias Benchmark and Metrics
- BBQ provides ambiguous scenarios across nine social dimensions; the experiments exclude disambiguated instances to focus on inherent bias.The Bias Score combines the frequency of non-neutral responses with their tendency toward the bias target; positive and negative scores indicate stereotypical and anti-stereotypical polarity, respectively.Source: §4.1 · pp. 4–5
4.2. Model Routing Metrics
- Router evaluation measures bias detection with accuracy and model selection with the proportion of recommended models that respond neutrally.The selection metric is precision: neutral recommended models divided by all proposed models.Source: §4.2 · p. 5
4.3. Experiment Settings
- Experiments draw from a pool of over 50 models and fine-tune Qwen2.5-32B as the router on one CrowdEval training epoch.Single uses the highest-ranked candidate, Sequential follows router order, and Voting, Debating, and Committee assign models randomly across slots.Source: §4.3 · p. 5
5. Discussion and Key Takeaways
The router can classify bias-related queries and recommend neutral models, while CBM topologies differ in bias reduction, consistency, and inference cost. Debating often achieves the lowest bias scores, whereas Committee offers a more consistent, lower-cost balance.
5. Discussion and Key Takeaways
- The router is fine-tuned to classify the prompt’s social dimension as an auxiliary task for recognizing potential bias.Training uses prompt–social-dimension pairs.Source: §5 · p. 5
- Qwen-2.5-32B achieves routing accuracy of 0.851, with performance stabilizing once router parameters exceed 9B.The accuracy distribution is estimated with 512 bootstrap iterations on the CrowdEval evaluation set; lower variance indicates more consistent routing.Source: §5 · p. 5, Figure 4 · p. 6
- The router achieves higher and more consistent neutral-response precision than random selection, with improvements diminishing once model size reaches 9B.Precision is measured as the proportion of recommended models’ CrowdEval responses classified as neutral.Source: Figure 5 · p. 6, §5 · p. 6
- Sequential chains can increase bias as chain length grows, risking the compounding of earlier models’ biases.Each model response feeds directly into the next in the chain-like topology.Source: §5 · p. 6
- Debating often achieves the lowest bias scores, while Committee delivers more consistent results at lower inference cost than Debating.Debating requires approximately 27 times the computational resources of the Single baseline.Source: §5 · p. 6
- Voting consistently outperforms the Single baseline across eight social dimensions and can perform better under model routing.The topology averages multiple model responses.Source: §5 · p. 6
6. Conclusion
CBM routes queries to selected LLMs and coordinates them in distinct topologies. Debating achieves the lowest bias, while Committee balances mitigation and inference cost.
6. Conclusion
- CBM coordinates distinct LLMs by routing queries to selected models operating in different topologies.Source: §6 · p. 6
- Debating achieves the lowest bias, while Committee balances bias reduction with computational cost.Source: §6 · p. 6
- The framework’s conclusion is that coordinating distinct LLMs can alleviate bias in LLMs.Source: §6 · p. 6
A. Details of the CrowdEval Dataset
CrowdEval collects LLM responses to ambiguous BBQ questions across eight social dimensions, using a pool of diverse open-source models. The dataset’s English, US-norm framing limits the universal applicability of its bias framing.
A. Details of the CrowdEval Dataset
- CrowdEval aggregates model responses to ambiguous-subset BBQ instances covering eight social dimensions.The dimensions listed are age, gender, disability, nationality, race, religion, socioeconomic status, and sexual orientation.Source: A. Details of the CrowdEval Dataset · p. 10
- The candidate pool contains trending, diverse open-source text-generation LLMs ranging from 1 billion to 56 billion parameters.The framework is designed to allow additional models to be integrated into the pool.Source: A. Details of the CrowdEval Dataset · p. 10
- CrowdEval’s English-language, US-norm framing may not apply universally across cultural contexts.Source: A. Details of the CrowdEval Dataset · p. 10
- The prompt template asks a model to choose the most suitable answer given the context, question, and answer choices.Source: A. Details of the CrowdEval Dataset · p. 10
- The dataset distribution table marks social dimensions with fewer BBQ instances, for which all available questions are included.Source: Table 2 · p. 10
A.1. CrowdEval Dataset Guideline
CrowdEval provides fine-grained, per-query records of model responses and bias labels for curated bias-eliciting questions. Its entries support analysis of individual model behavior, and the dataset and code are publicly released.
A.1. CrowdEval Dataset Guideline
- CrowdEval contains responses from over 50 open-source LLMs to curated bias-eliciting questions derived from the ambiguous BBQ subset.Source: §A.1 · p. 10
- Each entry records the context, question, answer choices, a specific model’s response, and a bias label categorized as bias-target, non-target, or neutral.Source: §A.1 · p. 10
- The per-query structure facilitates detailed examination of individual model behavior, and the authors publicly release the code and dataset.Source: §A.1 · p. 10
B. Details of Model Routing
The router detects potential bias in each query and selects models based on their learned behaviors. It also generalizes to bias dimensions excluded from training.
B. Details of Model Routing
- The routing process has two tasks: detecting potential bias in an input and selecting models to reduce the CBM system’s holistic bias.For selection, the router recommends models it has learned to be more neutral for the given query.Source: B. Details of Model Routing · p. 10
- The bias-detection task identifies potential bias attributes including age, disability, gender, nationality, race, religion, sexual orientation, and socioeconomic status.Source: B. Details of Model Routing · p. 10
- The router selects the top-k models for a query by ranking candidate models using their normalized token probabilities.Each model receives a unique identifier; the router ranks models by selection likelihood derived from negative log-likelihood loss.Source: B. Details of Model Routing · p. 10
- With SES and SO excluded from training, the 32B router reaches classification accuracies of 0.883 for SES and 0.809 for SO, above the 0.125 random baseline.The passage also reports that classification accuracy for both dimensions increases with model size.Source: B. Details of Model Routing · p. 11, Table 4 · p. 11
C. Details of CBM Topologies
CBM routes queries to selected models and coordinates their responses through several topologies. The section reports bias, cost, and inference-time comparisons, as well as results for larger-model self-debiasing.
C. Details of CBM Topologies
- In Single, one selected model answers directly; in Sequential, each model can use earlier responses and the last model produces the final answer.Voting models respond independently, with the final response selected by majority vote.Source: C. Details of CBM Topologies · pp. 11–12
- Model ordering affects Sequential performance: placing less biased models later appears to improve resilience to earlier, potentially more biased decisions and yields more neutral outputs.The comparison reverses the usual router-recommended order from less biased to more biased.Source: C. Details of CBM Topologies · p. 12, Table 8 · p. 12, Table 9 · p. 12
- Debating models iteratively refine responses using others’ answers until consensus, while Committee uses a coordinator to consolidate responses and seek approval.Committee consensus uses a 50% threshold, allows models to prefer, reject, or abstain, and defaults to the majority option if consensus is not reached within five iterations.Source: C. Details of CBM Topologies · pp. 12–13
- Debating and Committee costs grow exponentially, with Debating scaling more sharply; Committee requires fewer costs for comparable bias mitigation, while at top-7 the gap narrows because many debates reach maximum consensus iterations.Source: C. Details of CBM Topologies · p. 13
- Model Routing produces lower bias scores than Best Selection in the top-5 setting, while the two strategies are comparable at top-3.Random Selection has limited effect; the passage concludes that selecting diverse, well-matched models fosters more effective bias mitigation.Source: C. Details of CBM Topologies · p. 13
D. CBM Inference Acceleration
- 3.12s is the Single baseline; under optimized inference setups, CBM topology response times closely approach it despite iterative-processing overhead.The paper notes that iterative processing can increase computational overhead and latency.Source: D. CBM Inference Acceleration · pp. 14–15
E. Self-Debiasing with Larger Models
- CBM achieves the lowest average bias score among the compared systems, although individual larger models score lower on some dimensions.Bias scores do not decrease consistently with model size; the passage describes a self-debiasing example in which a biased answer remained unchanged.Source: E. Self-Debiasing with Larger Models · p. 15
F. Scope and Limitations
The paper limits its fairness evidence to ambiguous English-language BBQ questions and cautions against broader trustworthiness claims. It also notes that the datasets reflect stereotypes and require careful use.
F. Scope and Limitations
- The results concern bias mitigation on ambiguous English-language BBQ questions and do not establish general LLM trustworthiness.The bias score does not measure factual accuracy, robustness to distribution shifts or adversarial inputs, privacy, or confidence calibration.Source: F. Scope and Limitations · p. 15
- Because BBQ reflects a particular linguistic and cultural context, observed bias reductions may not transfer to other languages, cultures, or applications.The paper calls for joint evaluation of these dimensions and whether routing and collaboration preserve task accuracy before broader trustworthiness claims.Source: F. Scope and Limitations · p. 15
G. Ethical Considerations
- BBQ and CrowdEval derive from real-world scenarios and reflect existing stereotypes and biases, so their use requires care to avoid reinforcing them.Source: G. Ethical Considerations · p. 15