Incentivizing Truthful Language Models via Peer Elicitation Games

We thank the reviewer for the insightful and supportive comments. Below, we do our best to address the reviewer's questions adequately such that we could receive a better score.

W1: Theoretical assumptions on binary labels and conditionally independent signals

Thanks for pointing out the validity of these assumptions in real-world language settings. We offer the following justifications to clarify our framework’s effectiveness:

First, the generator is not restricted to a specific task type. PEG is compatible with open-ended tasks, as long as the generator can provide a set of candidate answers for evaluation. This makes the framework broadly applicable beyond multiple-choice QA.

Second, our theoretical discussion adopts binary feedback for clarity, as the discriminators provide binary judgments in our setting. However, the DMI mechanism is originally designed for multi-choice settings and still guarantees dominant truthfulness under standard conditions. Specifically, the key requirement is that the batch size must be at least $2C$, where $C$ denotes the number of choices per question.

Third, the conditional independence assumption is used in the theoretical analysis to ensure that the mutual information–based reward does not degenerate to $0$. Intuitively, if agents' signals are highly correlated, then observing others provides little to no additional information, and the utility (measured via a form of mutual information) collapses toward zero. In practice, since we use heterogeneous discriminators, their outputs are diverse and rarely reach consensus in the beginning. This natural diversity prevents the reward signal from collapsing and keeps the system informative.

W2: Clarification on task generality beyond multiple-choice QA

Thank you for the helpful suggestion. While our current experiments focus on multiple-choice QA tasks, PEG is not inherently limited to this setting. For any natural language task, we can use top-$k$ or other sampling methods to generate a candidate set, which PEG can then evaluate and refine. In this sense, the multiple-choice format is a practical instantiation of the candidate generation process, not a fundamental restriction of the framework. We have also discussed this in the limitations section and fully agree that evaluating PEG on broader task types is an important next step. We plan to include additional datasets covering factual, open-ended, and safety-critical tasks in future work.

W3 and Q1: Ablations on number and quality of discriminators

Thank you for raising this valuable point regarding the number and quality of discriminators. To investigate these problems, we conduct additional experiments on the ARC challenge set with 5-discriminators (add Gemma-7B [1] and Mistral-7B [2]) and 7-discriminators (further add Ai-Yi-9B [3] and OpenChat-7B [4]). The results highlight the strong impact of PEG on both individual and collective performance. First, as shown in Tables 1 and 2, individual discriminators show consistent improvement after applying PEG. Secondly, in both the 5-discriminator and 7-discriminator settings, our PEG methods consistently outperforms majority vote and ER-D, as shown in Table 3.

However, we observed a slight drop in PEG’s overall performance when moving to more discriminators. We believe this may be due to a mild violation of the Conditional Independence assumption. The two added models (Ai-Yi-9B and OpenChat-7B [3,4]), while strong in general, may not align well with the existing set in terms of signal diversity or informativeness. This observation further reinforces that it’s not just the number of discriminators that matters, but their quality and compatibility with PEG’s theoretical assumptions. The results show if discriminators are chosen in alignment with our underlying assumptions, a larger number of discriminators is not necessarily required to achieve good performance. This suggests that the quality and relevance of the chosen discriminators are more critical than their sheer quantity for the effectiveness of our PEG mechanism.

Table 1: Accuracy (%) of individual discriminators before and after applying PEG in the 5-discriminator setting.

Table 2: Accuracy (%) of individual discriminators before and after applying PEG in the 7-discriminator setting.

Table 3: Overall accuracy (%) comparison of PEG, Majority Vote, and ER-D under 3-, 5-, and 7-discriminator settings.

Q2: Effect of mixed-domain vs. single-domain batches

We thank the reviewer for pointing out the possibility that each batch may contain a mix of ARC-Challenge and MMLU items. We first clarify that our theoretical guarantee of dominant truthfulness relies on the assumption that tasks within each batch are similar. This homogeneity is important because the utility function aggregates feedback across multiple tasks in a batch; if the tasks are too heterogeneous, the computed reward may become ill-posed and fail to provide a meaningful learning signal for effective updates. From a theoretical perspective, this assumption ensures that agents can adopt a consistent reporting strategy across tasks within the batch. In contrast, if tasks vary significantly in nature or domain, agents may not be incentivized to follow a stable or truthful strategy. Empirically, we enforce this condition by constructing each batch using items from the same domain within a dataset.

Once again, we appreciate the reviewer's precious time. We are eager to engage in further discussions to clear out any confusion.

[1] Team, Gemma, et al. "Gemma: Open models based on gemini research and technology."
[2] Albert Q. Jiang, et al. "Mistral 7B."
[3] Young, Alex, et al. "Yi: Open foundation models by 01. ai."
[4] Wang, Guan, et al. "Openchat: Advancing open-source language models with mixed-quality data."

Thank you very much for your thoughtful follow-up and for raising this important point. The Conditional Independence assumption ensures that the mutual information–based reward remains informative and does not degenerate. Intuitively, if discriminators are too similar, their outputs become redundant, and observing one adds little value beyond the others. To better satisfy this condition, we recommend using heterogeneous model, as combining models from different families or developers helps reduce dependency. In our experiments, we find that using three diverse discriminators provides a good trade-off between performance and adherence to the CI assumption. This setup is lightweight, stable, and less likely to introduce strong correlations.

We agree that further ablation studies are important for verifying our assumption. To this end, we conducted two additional experiments: (1) using 5 discriminators (OQwen, DeepSeek-Qwen, DeepSeek-LLaMA2, Ai-Yi 9B, and OpenChat 7B), and (2) using 7 discriminators (OQwen, DeepSeek-Qwen, DeepSeek-LLaMA2, Gemma 7B, Mistral 7B, Zephyr 7B, and Phi-2). Results are summarized in Tables 4-6. In Table 6, the rows labeled “5-discriminators (new)” and “7-discriminators (new)” correspond to the additional experiments conducted for our ablation study.

Interestingly, as shown in Table 6, even when combining Ai-Yi and OpenChat, the 5-discriminator setting still achieves strong performance (86.92%), and PEG consistently improves each discriminator’s accuracy, as shown in Table 4. However, with 7 discriminators, we observe a noticeable performance drop in Table 6. In Table 5, several models (e.g., OQwen) show limited or even negative improvements after PEG. These results support our hypothesis that adding more discriminators does not necessarily improve performance and may even hinder learning. We attribute this to three possible reasons: (1) Adding more models may not provide sufficiently informative or independent signals and can increase the risk of violating the Conditional Independence assumption due to overlapping knowledge or behaviors, making learning less effective; (2) Optimization becomes more difficult when learning from multiple sources with potentially conflicting or low-quality feedback; (3) The inherent capabilities of the discriminator models also play a role: less capable models may contribute limited value or even introduce noise. For example, the newly added Zephyr-7B and Phi-2 are relatively less powerful, which will hinder effective learning.

Based on these observations, we find our method to be robust with different ablations using up to 5 discriminators (as shown in Table 6). Therefore, we propose a practical guideline: start with 3–5 heterogeneous LLMs, which strikes a strong balance between performance, stability, and computational efficiency. Expanding beyond this should be done cautiously and only when additional diversity and independence can be ensured.

Table 4: Accuracy (%) of individual discriminators before and after applying PEG in the 5-discriminator setting.

Table 5: Accuracy (%) of individual discriminators before and after applying PEG in the 7-discriminator setting.

Table 6: Overall accuracy (%) comparison of PEG, Majority Vote, and ER-D under 3-, 5-, and 7-discriminator settings.

We sincerely appreciate your acknowledgment of the quality, significance, and novelty of our work. Regarding your concerns, we would like to offer further clarification.

W1 and Q1: Clarification on the definition of “Truthfulness” in the LLM setting

We thank the reviewer for raising this important and subtle point. As noted in our submission (Lines 93–102), we adopt the classical notion of incentive compatibility (IC) from mechanism design to define truthfulness. Under IC, an agent is said to be truthful if reporting its private information maximizes its expected utility under the mechanism: $u_i(c_i, c_{-i}) \geq u_i(r_i, c_{-i}) \quad \forall r_i \neq c_i,$ where $c_i \in \\{0,1\\}$ denotes the agent’s true private signal, $r_i$ it's the reported label, and $u_i$ it's the utility based on others' reports $c_{-i}$.

When applied to LLMs, however, the concept of a “true belief” becomes less well-defined. LLM outputs are inherently probabilistic and often miscalibrated, meaning they may not faithfully reflect the model’s internal uncertainty or belief distribution. In our setting, we adopt a practical and operational notion of truthfulness: the report is treated as an output answer rather than a full belief distribution, and we assume the LLM reveals its "true belief" through greedy sampling, without additional exploration. For instance, if the model internally assigns 80% probability to a correct answer and 20% to an incorrect one, a truthful response would be the correct label. However, due to hallucinations or the stochastic nature of generation, the actual output may deviate from this internal belief, occasionally resulting in incorrect or noisy responses.

Returning to the reviewer’s example: suppose the LLM outputs “I believe the content is 60% true” with 80% probability, and “I believe the content is 40% true” with 20% probability. In our framework, we do not treat the textual belief statement itself (e.g., "60% true") as the truthful signal. Instead, we focus on the correctness label associated with the most probable output—namely, the one with hither probability—as the truthful signal. That is, the model is considered truthful if it reports the option it most strongly believes to be correct. The goal of PEG is to encourage reporting behavior that aligns with the model’s strongest internal belief, ensuring that the output reflects its highest-confidence judgment of correctness.

More broadly, we agree that LLM truthfulness can be meaningfully extended to its belief distribution. In this view, truthfulness becomes closely related to calibration: a truthful LLM should produce outputs that accurately reflect its internal confidence or belief distribution. For example, if the model internally believes the answer is 60% likely to be correct and 40% incorrect, then this belief distribution can be treated as its true belief. Miscalibration, a systematic mismatch between internal confidence and the reported output—can result in deviations from incentive-compatible behavior. Additionally, some prior works define LLM truthfulness more narrowly, equating it with factual correctness and treating any deviation (e.g., hallucination) as untruthfulness [1]. While this definition captures practical concerns, it does not incorporate ideas from mechanism design or incentive compatibility. The precise definition of LLM truthfulness remains an open question. We will include this discussion in our revised version.

Q2: Clarification on the covergence rate/speed of the PEG mechanism

We thank the reviewer for raising this important theoretical question. PEG is implemented via policy mirror descent. Specifically, we use exponentiated gradient updates (akin to mirror descent in the simplex) to iteratively refine the distribution over candidate answers.Therefore, PEG has convergence guarantees in the form of a no-regret property over iterations. Specifically, this implies that the average reward across $T$ steps approaches that of the best fixed policy in hindsight at a rate of $\mathcal{O}\left(\frac{1}{\sqrt{T}}\right)$. This result is consistent with standard findings in prior literature [2,3].

In terms of last-iterate convergence rate, we can observe that PEG typically converges and stabilizes within only 10 iterations in our experiments. However, providing theoretical guarantees for the last iterate remains challenging. Recent work such as [4] analyzes last-iterate linear convergence rates under magnetic mirror descent, but these results are restricted to two-player zero-sum games. We leave the extension of such convergence analysis to the multi-agent peer elicitation games used in PEG as a potential future direction.

Once again, we appreciate the reviewer's precious time. We are eager to engage in further discussions to clear out any confusion.

[1] Azaria, Amos, and Tom Mitchell. "The internal state of an LLM knows when it's lying." arXiv preprint arXiv:2304.13734 (2023).
[2] Beck, Amir, and Marc Teboulle. "Mirror descent and nonlinear projected subgradient methods for convex optimization." Operations Research Letters 31.3 (2003): 167-175.
[3] Kiwiel, Krzysztof C. "Proximal minimization methods with generalized Bregman functions." SIAM journal on control and optimization 35.4 (1997): 1142-1168.
[4] Sokota, Samuel, et al. "A unified approach to reinforcement learning, quantal response equilibria, and two-player zero-sum games." arXiv preprint arXiv:2206.05825 (2022).

We extend our sincere appreciation for your valuable feedback and suggestions. Regarding your concerns, we would like to offer further clarification.

W2: Clarification on "without ground-truth labels or model fine-tuning" claim

Thank you for pointing this out. When we state that PEG operates "without relying on ground-truth labels or model fine-tuning," we mean the following:

First, PEG does not require any ground-truth labels. Instead, it leverages signals from a group LLM-based discriminators, which serve as referees to provide feedback. The optimization is driven by agreement among these models, rather than comparison to a gold standard.

Second, PEG does not involve any updating on the model parameters. All models—including both the generator and the discriminators—remain completely frozen throughout the process. PEG operates purely by adjusting a distribution over candidate outputs guided by external feedback, making it distinct from supervised fine-tuning or RLHF approaches.

We will revise our claims in the introduction to clarify that PEG relies on peer evaluation rather than ground truth and no model parameter throughout the process.

W3: Overclaiming due to comparison with older, non-instruction-tuned models

We thank the reviewer for this valuable comment. We agree that some models used in our comparison are older and not instruction-tuned, which may partially explain their weaker performance. Our intention was not to overclaim, but to maintain consistency with prior work in consensus game [1], where these models were commonly adopted as standard baselines. We will remove the text to avoid potential overstatement.

Q1: What prevents the validators from reaching consensus on an incorrect answer?

Thank you for the insightful question. PEG relies on consensus among multiple discriminators as a proxy for truthful and high-quality outputs, without assuming that any single discriminator is always correct. From a theoretical perspective, our approach is inspired by the peer prediction framework, where heterogeneous agents are incentivized to provide informative and truthful signals under proper reward structure. From an empirical perspective, we observe that discriminators often select different correct answers in the early iterations. However, through PEG’s iterative aggregation and update process, these models implicitly “learn from each other” by reinforcing candidates that receive consistent support, while suppressing unstable or weakly supported ones. As a result, the output distribution gradually converges to more coherent and broadly agreed-upon answers, rather than amplifying isolated errors.

Q2: How does the effectiveness relate to prompt “hardness”? (initially have high disagreements)

Thank you for raising this insightful question regarding the relationship between prompt "hardness" and the effectiveness of our approach. To address this, we conducted further analysis by categorizing prompts in each dataset into different levels of difficulty based on the initial level of disagreement among the discriminators (validators). This categorization serves as a proxy for prompt difficulty, with higher disagreement indicating greater uncertainty or ambiguity. We report the accuracy for each method across these groups on both ARC-Challenge and MMLU in Table 1.

Our analysis reveals three key findings. First, PEG consistently improves accuracy across all levels of initial disagreement, suggesting that it enables agents to explore diverse judgments and learn from each other to reach better consensus. Second, a counter-intuitive finding is that cases with initial disagreement among discriminators actually result in higher accuracy compared to those with full agreement. We believe this phenomenon occurs because agreement alone does not ensure correctness: all discriminators may still converge on an incorrect answer. In contrast, disagreement introduces diversity, increasing the likelihood that at least one agent is correct, and serves as a useful signal to trigger further exploration. Third, PEG achieves the highest accuracy across all levels compared to other baselines methods. We will include this discussion in our revised version and leave the extension to settings with more than three discriminators for future work.

Table 1: Accuracy under different initial agreement settings for ARC-Challenge and MMLU datasets.

Q3: How many validators are needed, and is there a tradeoff between iterations to convergence and the cost of each iteration?

Thank you for raising this important point regarding the number of validators and the associated cost-performance tradeoff. Theoretically, PEG only requires 2 discriminators(validators) to achieve dominant truthfulness. Empirically, We conduct additional experiments on the ARC challenge set with 5-discriminators (add Gemma-7B [2] and Mistral-7B [3]) and 7-discriminators (further add Ai-Yi-9B [4] and OpenChat-7B [5]). The results highlight the strong impact of PEG on both individual and collective performance. First, as shown in Tables 2 and 3, individual discriminators show consistent improvement after applying PEG. Secondly, in both the 5-discriminator and 7-discriminator settings, our PEG methods consistently outperforms majority vote and ER-D, as shown in Table 4.

However, as shown in Table 4, we observe a slight drop in performance when increasing the number of discriminators, even though majority vote and consensus game methods show improvement. We attribute this to a potential violation of the PEG mechanism’s conditional independence assumption, which is critical for its theoretical guarantees and empirical effectiveness. Specifically, the two added models (Ai-Yi-9B and OpenChat-7B [4,5]), while strong in general, may not align well with the existing set in terms of signal diversity or informativeness. This observation suggests that the quality and compatibility of discriminators are more important than their sheer quantity. Additionally, introducing additional discriminators does not increase the number of iterations required for convergence, nor does it add overhead to each iteration. However, as performance becomes slightly more sensitive to the composition of the ensemble, a smaller, carefully curated set of discriminators that adheres to the conditional independence assumption can lead to superior and more stable performance.

Table 2: Accuracy (%) of individual discriminators before and after applying PEG in the 5-discriminator setting.

Table 3: Accuracy (%) of individual discriminators before and after applying PEG in the 7-discriminator setting.

Table 4: Overall accuracy (%) comparison of PEG, Majority Vote, and ER-D under 3-, 5-, and 7-discriminator settings.

Q4: Clarification on learning rates in the experiments

Thank you for pointing out the implementation details regarding the learning rate in our experiments. We clarify that the learning rate used in our theoretical analysis is adaptive and depends on a KL divergence term between the initial policy and the optimal policy. However, this term is difficult to estimate in practice. In our empirical studies, we find that the performance of PEG is robust to a range of learning rates. Specifically, we report accuracy results across four datasets using a fixed number of 10 iterations in Table 5. Intuitively, smaller learning rates (e.g., $\leq 0.1$) yield stable performance without compromising convergence speed. In contrast, larger learning rates degrade performance, likely due to the fact that the updates are applied directly to the output distribution, which lies within a bounded space $[0, 1]$. As a result, overly aggressive updates may lead to oscillation or failure to converge. These findings support the practical stability of our approach under reasonable learning rate choices. We will include this experimental results and analysis in our revised version.

Table 5: Accuracy (%) of PEG with different learning rates across four benchmark datasets.

[1] Jacob, Athul Paul, et al. "The consensus game: Language model generation via equilibrium search." ICLR
[2] Team, Gemma, et al. "Gemma: Open models based on gemini research and technology."
[3] Albert Q. Jiang, et al. "Mistral 7B."
[4] Young, Alex, et al. "Yi: Open foundation models by 01. ai."
[5] Wang, Guan, et al. "Openchat: Advancing open-source language models with mixed-quality data."

Thank you for your insightful feedback on our manuscript. We appreciate your comments and have taken them into consideration to improve our paper. Below, we do our best to address all concerns adequately so that we could receive a better score.

W1: Clarification on PEG workflow and ''Training-Free" claim

We appreciate the reviewer’s concern and agree that a clearer explanation of the workflow is necessary. As illustrated in Figure 2 in our submission, the core procedure of PEG consists of four main steps: (1) Given an input question, the generator produces a candidate answer; (2) multiple discriminators independently evaluate this answer, each providing a binary judgment; (3) these judgments are aggregated into a scalar reward signal; (4) a parameterized distribution (policy) over candidate answers is updated using these rewards.

Importantly, the parameters of the generator and discriminators remain entirely frozen throughout the process. All updates occur externally at the level of the output distribution. Thus, we describe PEG as “training-free” in the sense that it requires no gradient-based updates, no backpropagation through model parameters, and no fine-tuning or RLHF. Intuitively, the update rule in Equation 4 in our submission increases the probability of outputs that receive higher rewards without requiring access to model internals. Additionally, this approach is consistent with prior training-free paradigms (e.g., [1,2]) that optimize over output distributions rather than over model parameters. We will add these clarifications to Section 2.2 in the revised version to improve clarity.

W2, W3 & Q2: Clarification on computational cost and infrastructural resources, resource analysis compared to baselines

We thank the reviewer for raising this important point. We respectfully clarify that our method does not require maintaining or re-evaluating multiple discriminators across iterations. In each batch, candidate answers are evaluated once by a fixed set of frozen discriminators. The policy update then occurs entirely outside the models at the level of the output distribution. As a result, during the iterative process, we simply adjust the probability assigned to candidate outputs: there are no additional generations from the generator, no repeated calls to the discriminators, and no model training (e.g., no gradient updates or parameter tuning). This design ensures computational efficiency and maintains a training-free paradigm throughout the PEG procedure.

To be more specific, as detailed in Appendix C.3, the main computational cost arises during the initial policy extraction, i.e., generating candidate answers from the generator and evaluations from the discriminators. This cost scales linearly with the number of questions. In the subsequent PEG iterations, only lightweight computations (such as reward aggregation and softmax updates) are involved, and no additional infrastructure is required to coordinate multiple discriminators. The runtime per iteration of the PEG mechanism for each discriminator is approximately 200 microseconds on average. By comparison, the Consensus Game baseline [1] executes around 40 microseconds per iteration. Notably, PEG processes batches of 8 questions per iteration, enabling efficient parallel updates across multiple inputs.

Importantly, unlike traditional fine-tuning approaches which often require hours to days of compute time, our PEG mechanism is entirely training-free. The full fine-tuning methods generally require thousands of GPU hours and can cost anywhere from $10,000$ to over $35,000$ per run [3]. Even parameter-efficient fine-tuning (PEFT) techniques such as FinLoRA [4] require several hours to days to update model weights. In contrast, PEG operates without modifying model parameters or performing gradient updates, thereby completely eliminating the need for a training phase.

W4: How the number of PEG iterations or number of discriminators affects final performance

We thank the reviewer for this valuable question. We conducted additional experiments to study how performance varies with the number of PEG iterations and the number of discriminators.

First, we evaluated the accuracy of PEG with {10, 20, 30, 40, 50} iterations across all datasets, using a fixed learning rate of 0.1 and a batch size of 8. From Table 1, PEG converges within the first 10 iterations—subsequent iterations result in only minor adjustments on a few questions, which do not noticeably affect the overall accuracy. We attribute this result to the fact that our updates operate directly in the output policy space rather than modifying model parameters, allowing for faster convergence within a few iterations.

Table 1: Accuracy (%) of PEG with different numbers of iterations across four benchmark datasets.

We also conduct additional experiments on the ARC challenge set with 5-discriminators (add Gemma-7B [5] and Mistral-7B [6]) and 7-discriminators (further add Ai-Yi-9B [7] and OpenChat-7B [8]). The results highlight the strong impact of PEG on both individual and collective performance. First, as shown in Tables 2 and 3, individual discriminators show consistent improvement after applying PEG. Secondly, in both the 5-discriminator and 7-discriminator settings, our PEG methods consistently outperforms majority vote and ER-D, as shown in Table 4. However, we observed a slight drop in PEG’s overall performance when adding more discriminators. We believe this may be due to a mild violation of the Conditional Independence assumption. The two added models (Ai-Yi-9B and OpenChat-7B [7,8]), while strong in general, may not align well with the existing set in terms of signal diversity or informativeness. This observation further reinforces that it’s not just the number of discriminators that matters, but their quality and compatibility with PEG’s theoretical assumptions. The results show if discriminators are chosen in alignment with our underlying assumptions, a larger number of discriminators is not necessarily required to achieve good performance. This suggests that the quality and relevance of the chosen discriminators are more critical than their sheer quantity for the effectiveness of our PEG mechanism.

Table 2: Accuracy (%) of individual discriminators before and after applying PEG in the 5-discriminator setting.

Table 3: Accuracy (%) of individual discriminators before and after applying PEG in the 7-discriminator setting.

Table 4: Overall accuracy (%) comparison of PEG, Majority Vote, and ER-D under 3-, 5-, and 7-discriminator settings.

Q1: Clarification on disadvantages of ground truth labels

Thank you for the thoughtful question. Our motivation for avoiding ground-truth labels stems from these limitations:

First, in many specialized domains such as biomedical, legal, or scientific reasoning, high-quality expert annotations are difficult or infeasible to obtain [9]. This makes ground-truth supervision impractical for real-world deployment and model adaptation in new tasks.

Second, even when annotations are available, they are often unreliable. For open-ended prompts, multiple plausible answers may exist, but ground-truth labels typically assume a single correct one, thus penalizing valid alternatives. In subjective tasks like summarization or ranking, labels are inherently ambiguous, and annotators frequently disagree due to personal preferences, leading to noisy supervision [10].

To address these challenges, PEG replaces reliance on a fixed ground-truth label with feedback aggregated from a panel of random referees. This enables learning signals from each other rather than a single “golden” label, making the framework more scalable and robust.

[1] Jacob, Athul Paul, et al. "The consensus game: Language model generation via equilibrium search." ICLR
[2] Jacob, Athul Paul, et al. "Modeling strong and human-like gameplay with KL-regularized search." ICML
[3] Liu, Aixin, et al. "Deepseek-v3 technical report."
[4] Wang, Dannong, et al. "FinLoRA: Benchmarking LoRA Methods for Fine-Tuning LLMs on Financial Datasets."
[5] Team, Gemma, et al. "Gemma: Open models based on gemini research and technology."
[6] Albert Q. Jiang, et al. "Mistral 7B."
[7] Young, Alex, et al. "Yi: Open foundation models by 01. ai."
[8] Wang, Guan, et al. "Openchat: Advancing open-source language models with mixed-quality data."
[9] Zhang, Le, et al. "Learning from multiple annotators for medical image segmentation." Pattern Recognition 138 (2023): 109400.
[10] Davani, Aida Mostafazadeh, Mark Díaz, and Vinodkumar Prabhakaran. "Dealing with disagreements: Looking beyond the majority vote in subjective annotations." Transactions of the Association for Computational Linguistics 10 (2022): 92-110.