Voting or Consensus? Decision-Making in Multi-Agent Debate

May combine two claims.

Prior work focuses on a single class of decision protocols without systematic comparison, partly because many parameters change between experiments.

“In summary, prior work focuses on a single class of decision protocols without systematic comparison, partly because many parameters change between experiments (Yin et al., 2023;Yang et al., 2024b;Chen et al., 2023).”

• cites Yin et al., 2023 sourceDisputed

  The paper makes the contrary point about its own work: the Conclusion says it develops four communication paradigms, and the Experimental Setups compare their performance across tasks. Figure 6 also compares two termination rules. Our copy does not attribute a lack of comparison to changing parameters between experiments. (LLM, conf=0.83)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured
• cites Yang et al., 2024bNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Chen et al., 2023 sourceDisputed

  The paper does not describe a lack of systematic comparison as a reason for its approach. Instead, its Experimental Setup describes matching the average number of LLM calls for a fair comparison, and Table 12 compares different voting strategies. The cited paper therefore provides a contrary example to the claim about decision protocols; it does not attribute the alleged lack of comparison to changing parameters between experiments. (LLM, conf=0.83)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

May combine two claims.

Exchange-of-Thought (Yin et al., 2023) employs a consensus-based approach but does not compare consensus to other decision protocols.

“Exchangeof-Thought (Yin et al., 2023) employs a consensusbased approach, where agents iteratively refine answers through discussion in reasoning tasks, but they do not compare consensus to other decision protocols.”

• cites Yin et al., 2023 sourceDisputed

  The paper does use majority-consensus termination, but it explicitly compares that approach with consistent-output termination. In the “Termination Condition” results (page unavailable), it reports higher performance for majority consensus across all four communication paradigms, contrary to the claim that it does not compare decision protocols. (LLM, conf=0.93)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

May combine two claims.

The idea of independent answer generation has been previously explored in different multi-agent settings, but these studies did not systematically formalize and evaluate it separately from other multi-agent discussion parameters.

“While the idea of independent answer generation has been previously explored in different multi-agent settings (Du et al., 2023b;Liang et al., 2024), these studies did not systematically formalize and evaluate it separately from other multi-agent discussion parameters.”

• cites Du et al., 2023bNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Liang et al., 2024 sourceDisputed

  The Multi-Agent Debate Framework describes debaters speaking in a fixed order, with each response based on the previous debate history; the negative debater is instructed to disagree with the affirmative debater’s points. That is contrary to citing this paper as an example of independent answer generation. The paper analyzes debate features such as disagreement and adaptive stopping, but does not establish the claimed independent-generation step. (LLM, conf=0.83)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Becker (2024) suggests that extended discussions may lead to decreased accuracy because of agents drifting away from the original task.

“Becker (2024) suggests that extended discussions may lead to decreased accuracy because of agents drifting away from the original task.”

• cites Becker, 2024 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

ReConcile (Chen et al., 2023) integrates a hybrid voting and consensus approach by iteratively refining answers through confidence-weighting until consensus is reached.

“ReConcile (Chen et al., 2023) integrates a hybrid voting and consensus approach by iteratively refining answers through confidenceweighting until consensus is reached.”

• cites Chen et al., 2023 sourcePartly supported

  Section 4 describes multi-round discussion in which agents revise their answers and a confidence-weighted vote produces the team answer after each round. It says discussion stops when consensus is reached or after a maximum number of rounds, so consensus is not guaranteed; answer refinement also occurs through discussion, not confidence-weighting alone. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Research by Yin et al. (2023) suggests that increasing the number of rounds and agents participating in a multi-agent debate is beneficial because of test-time compute scaling.

“Research by Yin et al. (2023) suggests that increasing the number of rounds and agents participating in a multi-agent debate is beneficial because of testtime compute scaling.”

• cites Yin et al., 2023 sourcePartly supported

  The Round Analysis says EoT permits additional exchanges for difficult questions, with some SVAMP cases requiring more than five rounds. But it does not show that increasing rounds generally improves performance. The Limitations section discusses the possibility of involving more models, not a demonstrated benefit from adding agents; the paper also does not attribute such a benefit to test-time compute scaling. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Personas can unlock more knowledge for the LLM on a specific topic.

“This can unlock more knowledge for the LLM on a specific topic (Kim et al., 2024).”

• cites Kim et al., 2024 sourcePartly supported

  Section 2.1 says a domain-specific persona related to a question can improve an LLM’s performance. The paper does not establish that personas unlock more topic-specific knowledge; it measures reasoning performance, and Section 4.2 reports that personas sometimes worsen it. (LLM, conf=0.80)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Yin et al. (2023) found the number of personas (three) to be the most efficient.

“While the number of personas can vary, Yin et al. (2023) found this number to be the most efficient.”

• cites Yin et al., 2023 sourcePartly supported

  The D EoT Prompts section describes three personas—Kitty, Ben, and Peter—and §5.1 says the experiments used three models by default. The Cost Analysis compares EoT with other methods, but does not establish that three personas are the most efficient number. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

This principle is also key to self-consistency approaches and explains why multiple-choice tests are often easier than open-ended ones.

“This principle is also key to self-consistency approaches (Wang et al., 2023) and explains why multiple-choice tests are often easier than openended ones.”

• cites Wang et al., 2023 sourceCould not verify

  The cited paper’s Experiments and Analysis sections discuss Solo Performance Prompting, but the passages found do not address self-consistency or compare multiple-choice with open-ended tests. Searches for both points found no supporting passage. The citing sentence also names Wang et al. (2023), whereas the supplied source is a 2024 paper. (LLM, conf=0.74)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Presenting the task each round mitigates potential drift from the topic of the discussion.

“The prompt language is neutral, and the task is presented each round to mitigate the potential drift from the topic of the discussion (Becker, 2024).”

• cites Becker, 2024 sourcePartly supported

  The discussion prompt in Appendix G.2 includes the task instruction for each agent contribution, supporting the claim that the task is presented each round. Section 5.1.1 describes problem drift during discussions, but the paper does not say that repeating the task mitigates it or that this is why the prompt repeats it. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Each dataset was tested three times to obtain a standard deviation, ensuring reliable and robust performance estimates.

“Each dataset was tested three times to obtain a standard deviation of the results (Reimers and Gurevych, 2017;Chen et al., 2023;Becker, 2024), ensuring reliable and robust performance estimates across multiple evaluations.”

• cites Reimers and Gurevych, 2017 sourcePartly supported

  Section 5, “Robust Model Evaluation,” says that rerunning nondeterministic approaches and comparing score distributions is essential for sound conclusions; the conclusion also discusses studying the standard deviation of scores. The paper does not recommend or establish three tests per dataset as sufficient for reliable estimates. Its reported rerun counts include 86 and 41 for two systems (Section 3), so the specific three-test procedure is not supported by this citation. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites Chen et al., 2023 sourcePartly supported

  The Implementation Details section says the authors reported accuracy and standard deviation and conducted at least three runs for each experiment on the same test samples and prompts. It does not say every dataset was tested exactly three times or establish that these runs ensured reliable and robust estimates. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites Becker, 2024 sourcePartly supported

  The experimental-sampling passage (section “373(×5)”) says the author reported standard deviations across repeated dataset experiments, supporting that part of the claim. It specifies five runs, not three; the Limitations section (6.3) likewise describes five runs on randomized subsets. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

This resembles the effect seen in self-consistency, where increasing the number of sampled answers improves task performance.

“This resembles the effect seen in self-consistency (Wang et al., 2023), where increasing the number of sampled answers improves task performance.”

• cites Wang et al., 2023 sourceCould not verify

  The Experiments section and Table 5 discuss Solo Performance Prompting and compare task scores, but do not address self-consistency or whether sampling more answers improves performance. Searches for self-consistency and multiple sampled answers found no passage in our copy that speaks to the claimed effect; the table’s N values refer to numbers of trivia questions, not sampled answers. Page numbers were not available in the indexed passages. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Using personas is important as it generates agents with expertise in different domains, incorporating diverse viewpoints for the final decision.

“Using personas is important as it generates agents with expertise in different domains, incorporating diverse viewpoints for the final decision.”

• cites Kim et al., 2024 sourcePartly supported

  Sections 3.1 and 4.8 describe generating a question-appropriate persona, with different personas appearing for different questions. Sections 3.2 and 3.3 describe a final decision informed by two perspectives, but these are one persona-based solver and one neutral solver—not multiple agents with expertise in different domains contributing to the same decision. No page numbers are available in our copy. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Huang et al. (2023) questions the findings of Madaan et al. (2023), suggesting that the claimed improvements may not be robust.

“This aligns with recent concerns about the reliability of self-refinement, as Huang et al. (2023) questions the findings of Madaan et al. (2023), suggesting that the claimed improvements may not be robust.”

• cites Huang et al., 2023 sourceSupported
• cites Madaan et al., 2023 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Current research shows that discussion protocols have little impact on downstream task performance.

“Current research by Yin et al. (2023) and Becker (2024) shows that discussion protocols have little impact on downstream task performance.”

• cites Yin et al. (2023) sourcePartly supported

  Table 1 (section labelled COMMUNICATION PROMPTS; page unavailable) shows similar average scores for the four communication paradigms, from 87.20 to 87.50. But the paper also says different paradigms perform best on different tasks, so the broad claim that discussion protocols have little impact on downstream performance is not fully supported. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites Becker (2024) sourcePartly supported

  The Results (§5.1.1; page unavailable) show that some discussion paradigms perform similarly, but also say the report paradigm significantly improves performance on ethical questions. The claim that protocols have little impact on downstream performance is therefore too broad: Becker finds that their effect depends on the task. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

This drift has also been observed by Becker (2024).

“This has also been observed by Becker (2024) and an example can be seen in Section G.3.”

• cites Becker (2024) sourceSupported

May combine two claims.

Consensus protocols perform better on knowledge-based tasks with average improvements of 1.3% on GPQA over the voting average.

“Notably, consensus protocols perform better on knowledge-based tasks with average improvements of 2.3% on MMLU, 4.9% on MMLU-Pro, and 1.3% on GPQA over the voting average.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

An additional experiment comparing solution counting to prompted decision protocols showed similar trends and had little impact on main findings.

“An additional experiment comparing solution counting (selecting the answer that is given the most) to our prompted decision protocols showed similar trends and had little impact on our main findings (see Section D).”

May combine two claims.

Cumulative voting, supermajority, and unanimity consensus perform a bit worse with AAD.

“Cumulative voting, supermajority, and unanimity consensus perform a bit worse.”

May combine two claims.

Consensus protocols often reached a decision faster than voting protocols, particularly on knowledge tasks that also achieved superior performance.

“Notably, consensus protocols often reached a decision faster than voting protocols, particularly on knowledge tasks that also achieved superior performance (Figure 2).”

May combine two claims.

Challenging answers with a changed final score has no positive effect, especially on reasoning tasks, where it leads to worse solutions in 3% of cases.

“It can be seen that this has no positive effect, especially on the reasoning tasks (StrategyQA and SQuAD 2.0), where it leads to worse solutions in 3% of cases.”

May combine two claims.

Agents often simply vote for all answers as they do not have to decide, leading to many ties.

“The agents often simply vote for all answers as they do not have to decide, leading to many ties (see Section A.2 for details).”

May combine two claims.

Agents are agreeable and often only improve the answer from the first agent without proposing an idea based on their own expertise.

“Throughout our experiments, we observe that agents are agreeable and often only improve the answer from the first agent without proposing an idea based on their own expertise.”

May combine two claims.

Approval voting and majority have no real improvement with AAD but are within the standard deviation of the baseline.

“Approval voting and majority have no real improvement but are within the standard deviation of the baseline.”

May combine two claims.

Voting excelled on reasoning tasks despite potentially needing more rounds and compute.

“Conversely, voting excelled on reasoning tasks despite potentially needing more rounds and compute.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Llama 3 70B does not fail to follow the prompt for the MuSR baseline and consensus-based decision protocols.

“This model does not fail to follow the prompt for the MuSR baseline and consensus-based decision protocols.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Decision protocols tend to behave similarly in terms of rounds needed to create a final answer, independent of task and model.

“The decision protocols tend to behave similarly in terms of rounds needed to create a final answer, independent of task and model.”

May combine two claims.

Consensus decision protocols require only one to two rounds to reach consensus and still achieve higher task performance on the MMLU dataset.

“The consensus decision protocols require only one to two rounds to reach consensus and still achieve a higher task performance because these results are based on the MMLU dataset.”

May combine two claims.

Simple voting has the highest agreement rate, but cumulative and ranked voting only need another round in a few cases.

“Simple voting has the highest agreement rate, but also cumulative and ranked voting only need in a few cases another round.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Initial explorations with more diverse prompt structures did not yield consistent improvements over the simpler, reproducible design presented.

“Initial explorations with more diverse prompt structures did not yield consistent improvements over the simpler, reproducible design presented.”

May combine two claims.

Most results are a bit better than the baseline, but multi-agent discussions only outperform the CoT baseline in a few cases.

“Most of the results are a bit better than the baseline, but the multi-agent discussions are only in a few cases able to outperform the CoT baseline.”

May combine two claims.

Llama 3 70B has a much higher baseline for task performance, making it more difficult to improve baseline results.

“In general, Llama 3 70B has a much higher baseline for task performance, making it more difficult to improve baseline results.”

May combine two claims.

SQuAD 2.0 and StrategyQA had the largest performance gains, even outperforming the CoT baseline, similar to the smaller model.

“SQuAD 2.0 and StrategyQA had the largest performance gains, even outperforming the CoT baseline, similar to the results from the smaller model.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Overall, the results were markedly higher than what could be explained by the standard deviation.

“Overall, the results were markedly higher than what could be explained by the standard deviation.”

May combine two claims.

About 14% need another round and the rest is canceled after the fifth round.

“About 14% need another round and the rest is canceled after the fifth round.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the G.3 Bad Voting Discussion example, the agents were tricked by one agent who proposed a solution not provided in the context.

“In this discussion, the agents were tricked by one agent who came up with a solution not provided in the context.”

May combine two claims.

Consensus methods outperform solution counting on knowledge-based tasks (MMLU-Pro, GPQA), and voting notably improves performance on the reasoning task (StrategyQA).

“In particular, consensus methods outperform solution counting on knowledge-based tasks (MMLU-Pro, GPQA), and voting notably improves performance on the reasoning task (Strat-egyQA).”

Probably not a substantial claim: may not pass S2 (checkable).

Decision-making is central to multi-agent processes.

“Thus, we argue that decisionmaking is central to multi-agent processes.”

• Also stated in: § Abstract · boosted

  “This work demonstrates the importance of decisionmaking in multi-agent debates beyond scaling.”

May combine two claims.

With the correct choice, multi-agent debates can be made more error-resistant by using consensus decision protocols and more explorative by using voting decision protocols.

“This demonstrates that with the correct choice, multi-agent debates can be made more error-resistant by using consensus decision protocols and more explorative by using voting decision protocols.”

May combine two claims.

AAD and CI achieve higher answer diversity and therefore higher task performance.

“AAD and CI achieve higher answer diversity and, therefore, higher task performance, but the two response generators struggle to provide reliable results and may even reduce task performance.”

Classified as probably not a substantial claim.

The decision protocol is the only variable changed in the systematic comparison.

“We change only one variable at a time -the decision protocol -to analyze how different methods affect the collaboration between agents and measure differences in knowledge and reasoning tasks.”

• Also stated in: § 4.1 Performance of Decision Protocols · plain

  “This experiment systematically compares a set of seven decision protocols, changing only the decision protocol used and nothing else.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Finding collective solutions markedly impacts human decision-making.

“Finding collective solutions markedly impacts human decision-making (Jones, 1994).”

• cites Jones, 1994 sourceNot checked — no full text of the source

May combine two claims.

More debate rounds harm task performance, suggesting the expected benefits of self-refinement may be unreliable and differ from task to task.

“These findings further support that more debate rounds harm task performance, suggesting the expected benefits of self-refinement may be unreliable and also differ from task to task.”

May combine two claims.

The reasoning response generator decreases answer diversity, which results in a drop in task performance.

“The reasoning response generator decreases answer diversity, which results in a drop in task performance.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Task performance is calculated on a subset of samples with three runs and standard deviation is reported.

“Because of computational constraints, we calculate the task performance on a subset of samples with three runs and report their standard deviation.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Participants can vote for a less preferred but more viable alternative to block an undesired outcome, leading to outcomes that do not reflect the collective will of the group.

“For example, participants can vote for a less preferred but more viable alternative to block an undesired outcome, leading to outcomes that do not reflect the collective will of the group (List, 2022).”

• cites List, 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Yang et al. (2024b,a) introduce multiple voting protocols but do not compare the performance of these voting decision protocols across different tasks.

“Yang et al. (2024b,a) introduce multiple voting protocols (e.g., approval voting) but do not compare the performance of these voting decision protocols across different tasks, focussing more on a comparison of how humans vote compared to LLMs.”

• cites Yang et al., 2024b,aNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Each condition in the scaling experiment is evaluated across three independent runs.

“Both experiments use the simple voting decision protocol, and each condition is evaluated across three independent runs.”

May combine two claims.

Our methods enhance answer diversity and reveal a connection between answer diversity and task performance.

“Our methods enhance answer diversity and reveal a connection between answer diversity and task performance.”

May combine two claims.

Consensus excels in knowledge tasks, and voting works better for reasoning tasks.

“Our results suggest that consensus excels in knowledge tasks, and voting works better for reasoning tasks.”

Classified as probably not a substantial claim.

Three rounds of discussion allow for reasonably strong improvements considering computing resources.

“In the default setting, they have to discuss for three rounds, as current research such as Du et al. (2023b) shows that this allows for reasonable strong improvements considering computing resources.”

• cites Du et al. (2023b)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Larger models tend to hallucinate less, reducing the hallucination effect for Llama 3 70B.

“Larger models tend to hallucinate less, reducing this effect for the Llama 3 70B model (Li et al., 2024).”

• cites Li et al. (2024)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

By changing the way an agent is prompted, a lot of performance can be gained or lost.

“By changing the way an agent is prompted, a lot of performance can be gained or lost.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Reasoning-based datasets (StrategyQA, MuSR, SQuAD 2.0) emphasize multistep reasoning and textual comprehension.

“StrategyQA, MuSR, and SQuAD 2.0.These emphasize multistep reasoning and textual comprehension.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Prompt design and persona selection are relevant parameters for multi-agent debate.

“While this study focuses on decision protocols, we acknowledge that prompt design and persona selection are also relevant parameters.”

May combine two claims.

Smaller models are more likely to hallucinate, which reduces task performance.

“As Li et al. (2024) showed, smaller models are more likely to hallucinate, which reduces task performance.”

• cites Li et al. (2024)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Multi-agent discussions require specific problem structures, or else agents just talk about the same results for multiple rounds and agree with each other.

“This can be taken as evidence that these multi-agent discussions require specific problem structures, or else the agents are just talking about the same results for multiple rounds and agreeing with each other.”

Classified as probably not a substantial claim.

Encouraging agents to be more critical may reduce sycophancy.

“Encouraging them to be more critical may reduce this.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Knowledge-based datasets (MMLU, MMLU-Pro, GPQA) still require some reasoning and domain knowledge.

“MMLU, MMLU-Pro, and GPQA.These still require some reasoning and domain knowledge.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Using sampled subsets of the datasets can lead to some variance.

“Because of the high computational cost and the range of decision protocols and tasks in our work, we used sampled subsets of the datasets, which can lead to some variance.”

Classified as probably not a substantial claim. May combine two claims.

Models like to agree with each other, so they tend to vote for many answers, often leading to a tie.

“This happens because these models like to agree with each other, and therefore they tend to vote for many of the answers, which often leads to a tie.”

Classified as probably not a substantial claim. May combine two claims.

LLMs can show some form of sycophancy, which is not helpful for a constructive discussion.

“Some studies have shown that LLMs can show some form of sycophancy, which is not helpful for a constructive discussion (Sharma et al., 2023).”

• cites Sharma et al., 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Wang et al. (2024a) shows increasing the number of agents may have positive effects.

“Wang et al. (2024a) shows increasing the number of agents may have positive effects.”

• cites Wang et al., 2024aNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The Critical Response Generator encourages agents to critically assess prior answers and propose new solutions, countering sycophantic tendencies.

“The Critical Response Generator encourages agents to critically assess prior answers and propose new solutions, countering sycophantic tendencies (Sharma et al., 2023).”

• cites Sharma et al., 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim. May combine two claims.

The study controlled for variance by sampling with a 95% confidence level and calculating the standard deviation of three independent runs.

“To control for that variance, we sampled with a 95% confidence level and calculated the standard deviation of three independent runs.”

Classified as probably not a substantial claim.

Multi-agent discussions produce multiple results for the same problem because each agent has its own reasoning and ideas.

“Multiagent discussions produce multiple results for the same problem because each agent has its own reasoning and ideas on how to solve the problem.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The solution counting ablation reinforces the main findings and illustrates the added value of decision protocols.

“Thus, this ablation reinforces our main findings and illustrates the added value of decision protocols.”