In this paper, we present our work developed for the scientific web discourse detection task (Task 4a) of CheckThat! 2025. We propose a novel council debate method that simulates structured academic discussions among multiple large language models (LLMs) to identify whether a given tweet contains (i) a scientific claim, (ii) a reference to a scientific study, or (iii) mentions of scientific entities. We explore three debating methods: i) single debate, where two LLMs argue for opposing positions while a third acts as a judge; ii) team debate, in which multiple models collaborate within each side of the debate; and iii) council debate, where multiple expert models deliberate together to reach a consensus, moderated by a chairperson model. We choose council debate as our primary model as it outperforms others in the development test set. Although our proposed method did not rank highly for identifying scientific claims (8th out of 10) or mentions of scientific entities (9th out of 10), it ranked first in detecting references to scientific studies.
tweet:
In this work, we present our approach for Subtask 4a (Scientific Web Discourse Detection) [
In our experiments, we utilize several LLMs, including Gemma3 (12B) [
ceur-ws.org only open-source models, the council debate method outperforms other approaches on the development set, and the team debate method is superior to the single debate method. Based on these results, we selected the council debate method as our primary method.
In the oficial rankings, our model ranked 8ℎ ( 1 = 0.7273) in Category 1 (detecting scientific claims) and 9ℎ ( 1 = 0.7766) in Category 3 (detecting mentions of scientific entities). However, it achieved 1 place ( 1 = 0.7805) in Category 2 (detecting references to scientific studies).
Detecting scientific claims and related content in social media, especially on platforms like Twitter 3,
has become a growing area of research in recent years [
We explored three diferent multi-LLM approaches for scientific discourse detection, each building upon the previous method’s insights. All approaches use structured deliberation processes but difer in their organizational structure and decision-making mechanisms. In our methods, we use several diferent prompts. The actual prompts used in our study are given in Appendix A. We apply each method independently for each category. However, when a tweet references a scientific study (Category 2), we also assign a positive label to Category 3 (i.e., containing scientific entities), as such references typically imply the presence of scientific terms. An overview of these three debate frameworks is illustrated in Figure 1.
Our first approach implements a traditional debate format where two LLMs argue for opposing positions while a third model acts as a judge. This method processes each category independently through separate debates.
Algorithm 1 describes our single debate method. For each interaction with LLMs, a brief prompt is given just to explain the prompt. Actual prompts are provided in Appendix A.3. The algorithm takes the tweet and the classification category as parameters. We first set our models used in the debate process and assign specific models to diferent roles: a proponent model that argues the tweet contains the category, an opponent model that argues against it, and a judge model that makes the final decision [ Lines 2-4]. The debate begins with opening statements where the proponent generates supporting arguments and the opponent generates opposing arguments for the given tweet and category [Lines 5-6]. We maintain a transcript that records all arguments throughout the debate [Line 7]. The core debate consists of rounds where the proponent rebuts the opponent’s argument, and the opponent rebuts the proponent’s argument in response [Lines 8-12]. After the rebuttal rounds, both sides provide closing statements to summarize their positions [Lines 13-15]. Finally, the judge model evaluates the complete debate transcript and makes a classification decision for the current category [Lines 16-17].
Building on the single debate approach, we propose a team-based method where multiple models collaborate within each side of the debate. This approach maintains the adversarial structure while adding diverse perspectives within each team. Importantly, team members discuss among themselves before presenting their arguments to the opposing team.
Algorithm 2 describes our team debate method. For each interaction with LLMs, a brief prompt is given just to explain the prompt. Actual prompts are provided in Appendix A.4. We assign teams of models to diferent roles: a proponent team that argues the tweet contains the category, an opponent team that argues against it, and a judge model that makes the final decision [ Lines 2-4]. The debate begins with internal team discussions where proponent team members collaborate to develop their strategy and arguments, followed by opponent team members doing the same [Lines 5-6]. Each team then presents their opening statements, with all proponent team members arguing for the category and all opponent team members arguing against it [Lines 7-8]. We maintain a transcript that records all arguments throughout the debate [Line 9]. The core debate consists of rounds where teams first conduct internal discussions to plan their rebuttals, then present coordinated responses to Algorithm 1 Single Debate Method 1: Input: Tweet text , category , maximum rounds 2: ← proponent model for category 3: ← opponent model for category 4: ← judge model 5: 0 ← Generate( , ”argue tweet contains category ”) 6: 0 ← Generate( , ”argue tweet does NOT contain category ”) 7: ← [ 0, 0] 8: for = 1 to do 9: ← Generate( , ”rebut opponent argument −1 ”) 10: ← Generate( , ”rebut proponent argument ”) 11: ← + [ , ] 12: end for 13: ← Generate( , ”provide closing statement”) 14: ← Generate( , ”provide closing statement”) 15: ← + [ , ] 16: ← Generate( , ”evaluate debate for category ”) 17: return ExtractClassification( ) the opposing team’s arguments [Lines 10-15]. Finally, the judge model evaluates the complete team debate transcript and makes a classification decision for the current category [ Lines 17-18]. Algorithm 2 Team Debate Method 1: Input: Tweet text , category , team size , maximum rounds 2: ← proponent team of size for category 3: ← opponent team of size for category 4: ← judge model 5: ← InternalDiscussion( , ”develop strategy for category ”) 6: ← InternalDiscussion( , ”develop strategy against category ”) 7: ← TeamArguments( , ”argue tweet contains category ”) 8: ← TeamArguments( , ”argue tweet does NOT contain category ”) 9: ← [ , ] 10: for = 1 to do 11: _ ← InternalDiscussion( , ”plan rebuttals to arguments”) 12: _ ← InternalDiscussion( , ”plan rebuttals to arguments”) 13: ← TeamArguments( , ”present coordinated rebuttals”) 14: ← TeamArguments( , ”present coordinated rebuttals”) 15: ← + [ , ] 16: end for 17: ← Generate( , ”evaluate team debate for category ”) 18: return ExtractClassification( )
Our final approach implements a collaborative council structure where multiple expert models deliberate together to reach consensus, moderated by a chairperson model. This method moves away from adversarial debate toward collaborative decision-making.
Algorithm 3 describes our council debate method. For each interaction with LLMs, a brief prompt is given just to explain the prompt. Actual prompts are provided in Appendix A.5. Firstly, we assign a council of expert models and a chairperson model to moderate the discussion [Lines 2-3]. The process begins with each council member providing an initial assessment and vote for the current category [Line 5]. We then check if the initial votes have reached the consensus threshold . If consensus is achieved, we finalize the decision using the majority vote [ Lines 5-7]. Otherwise, we proceed with rounds of structured discussion where the chairperson summarizes the current state and guides the focus, followed by each council member responding to the discussion and updating their position [Lines 8-15]. After each round, we check whether consensus has been reached or if the votes have stabilized to finish the discussion early [ Lines 12-14]. Finally, we use the majority vote from the ifnal round as our decision [ Line 16].
To provide concrete examples of how our debating methods work in practice, we present conversation logs from each method in Appendix B, showing how diferent models interact to reach a final decision. Algorithm 3 Council Debate Method 1: Input: Tweet text , Category , consensus threshold , maximum rounds 2: ← assign council members from for category 3: ℎ ← assign chairperson 4: ← CollectVotes( , ”assess tweet for category ”) 5: if isConsensusReached( , ) then 6: return MajorityVote( ) 7: else 8: for = 1 to do 9: ← Generate(ℎ 10: ← CollectResponses( 11: ← ExtractVotes( ) 12: if isConsensusReached( , ) or VotesStabilized( ) then 13: break 14: end if 15: end for 16: return MajorityVote( ) 17: end if , ”summarize discussion and guide focus”)
, , ”update positions”)
Dataset. We conducted experiments using the oficial CheckThat! 2025 Task 4a datasets, which include training, development, and test sets with tweets labeled for three categories of scientific content. Table 1 provides the label distribution for the datasets. As our approach does not need training, we only use the development set to pick our primary model.
Models. Our council debate framework was implemented using the Ollama framework4 for local model access, with parallel processing capabilities to handle multiple concurrent debates eficiently. We systematically evaluated diferent model combinations across our three proposed methods using the following models: • Gemma3 (12B): Architecture Gemma3, 12.2B parameters, Q4_K_M quantization • Qwen3 (8B): Architecture Qwen3, 8.19B parameters, Q4_K_M quantization • DeepSeek-R1 (8B): Architecture Llama, 8.03B parameters, Q4_K_M quantization • Phi4 (14B): Architecture Phi3, 14.7B parameters, Q4_K_M quantization • Mistral (7B): Architecture Llama, 7.52B parameters, Q4_0 quantization • Llama3.1 (8B): Architecture Llama, 8.03B parameters, Q4_K_M quantization • o4-mini: OpenAI’s commercial model accessed via API • Claude-4: Anthropic’s commercial model accessed via API Configuration : For our three proposed methods, we used the following configurations: • Council Debate: Five council members (Gemma3, Qwen3, DeepSeek-R1, Phi4, Mistral) with
Llama3.1 serving as chairperson. • Team Debate: We tested two configurations where Llama3.1 serving as judge: (1) Same teams configuration with five members on each side using all models, and (2) Diferent teams configuration with Team A (Gemma3, Qwen3, Mistral) versus Team B (DeepSeek-R1, Phi4, Llama3.1), each team having three members. • Single Debate: We tested both same model configurations (same model as both proponent and opponent) and diferent model configurations where Llama3.1 serving as judge. The diferent model pairings include DeepSeek-R1 vs Qwen3, Gemma3 vs Phi4, and o4-mini vs Claude-4.
Key hyperparameters were set as follows: minimum consensus threshold of 80% for council and team debates, maximum of 5 discussion rounds for collaborative methods, and 3 rounds for adversarial single debates. The system included checkpointing capabilities to ensure robustness during large-scale processing. All models were accessed through the Ollama framework with the quantization settings specified above.
Baseline Models. In order to better analyze the impact of our debating methods, we also use each LLM we picked separately for few-shot detection. For these baseline models, we use the detailed category descriptions with examples (provided in Appendix A.2) for few-shot learning prompts.
We evaluated all three proposed methods and baselines on the development set to select our best approach for the final submission. Table 2 shows the results.
Our observations on experiments with the development set are as follows. Firstly, the council debate method achieves the highest macro F1-score and excelled particularly in Category 1 (i.e., scientific claims detection) with an F1-score of 0.8756. Secondly, the commercial model pairing of o4-mini vs Claude-4 showed superior performance in Categories 2 and 3 with F1-scores of 0.8923 and 0.8478, respectively. This is likely due to their larger model sizes. Thirdly, ignoring the commercial models, the team debate method shows improved performance over single debate by incorporating collaborative discussion within teams, with diferent team configurations outperforming same team configurations. Lastly, the comparison with individual models demonstrates the significant impact of our debate methods. In particular, all debate approaches outperform detection using models individually.
At the time of the submission deadline, we did not have results for the single debate method with o4-mini vs. Claude-4. Therefore, we picked the council debate method as our primary method and submitted the results accordingly. On the test set, our council debate method achieves 0.7273 (ranked 8th), 0.7805 (ranked 1st), and 0.7766 (ranked 9th) for Category 1, Category 2, and Category 3, respectively. Our results show that while LLMs are not particularly efective at detecting scientific claims or scientific entities, they perform well in identifying references to scientific studies.
In this paper, we present our participation in CheckThat 2025 Task 4a, the scientific web discourse detection task. We propose three debating methods where multiple LLMs discuss with each other to classify tweets. The debate frameworks enable models to have opposing views and try to convince each other. Among three methods, the council debate framework, in which expert models discuss together to reach a consensus, moderated by a chairperson model, outperforms other debating frameworks.
Although our council debate framework showed relatively weak performance in Category 1 (8th place, F1 = 0.7273) and Category 3 (9th place, F1 = 0.7766), it achieved the highest F1-score (0.7805) in Category 2.
In future work, we plan to extend our debating framework to other classification tasks. We also aim to investigate the impact of prompt design and utilizing other LLMs.
During the preparation of this work, the authors used GPT-4.5 and Claude Sonnet 4 in order to: Grammar and spelling check. After using these tools/services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.
This appendix contains the system prompts and message templates used in our three multi-LLM approaches, as well as the category descriptions used in the task.
Category Titles:
Proponent System Prompt:
You are a scientific content detector participating in a formal debate. Your job is to argue why the given tweets [CATEGORY TITLE]. Be thorough and precise in your analysis. Provide specific evidence from the text and URLs to support your arguments. Keep your responses concise and focused on the strongest evidence.
Opponent System Prompt:
You are a scientific content critic participating in a formal debate. Your job is to argue why the given tweets do NOT [CATEGORY TITLE]. Be thorough and precise in your analysis. Provide specific evidence from the text and URLs to support your arguments. Keep your responses concise and focused on the strongest counter-evidence.
Judge System Prompt:
You are a neutral judge evaluating a debate about whether the tweets [CATEGORY TITLE]. Based on the debate transcripts and the tweets themselves, determine if the tweets [CATEGORY TITLE]. Provide your classification and a brief explanation of your decision, including which arguments from the debate you found most compelling in the following JSON format: {”category”: 0 or 1, ”explanation”: [EXPLANATION]}
• Example 5 (Negative): ”how u treat others will reflect on how u feel”
Explanation: General **life advice without scientific backing** or evidence.
• Example 1 (Positive): ”Diabetes Research and New Registry Aim to Improve Outcomes **https://www.acc.org/latest-in-cardiology/articles**/2015/03/04/16/32/diabetesresearch-and-new-registry-aim-to-improve-outcomes?wt.mc_id=twitter #NCDR” Explanation: Contains **URL to cardiology articles** (domain: acc.org, path: /articles) and mentions **research** directly. • Example 2 (Positive): ”With our current lifestyle, most of us are sleep deprived, which creates problems like mood disorders, weakened immunity, weight gain and diabetes. **A review of 16 studies** found that sleeping for less than 6 to 8 hours a night increases the risk of early death by as much as 12%. image”
Explanation: References **”a review of 16 studies”** indicating scientific literature review. • Example 3 (Negative): ”How can this be unfolding? **https://www.bbc.co.uk/news**/health53990068” Explanation: **BBC news link** about health but not specifically referencing scientific studies or publications. • Example 4 (Negative): ”How Employment Can Change the Life of Someone with a Disability **http://www.tennesseeworks.org**/how-employment-can-change-the-life-of-someonewith-a-disability-and-everyone-involved/” Explanation: **General informational content** from a non-academic source, not referencing scientific studies.
Proponent Prompt:
Task: You are participating in a formal debate about whether a given tweet [CATEGORY TITLE]. Rules & Definitions: - @user tags represent anonymized usernames from the original tweet - ”image” at the end indicates an attached image - Examine both tweet content and URLs carefully - Focus on science rather than technology, religion, or politics Your Role: Argue why the tweet [CATEGORY TITLE]. Provide specific evidence from the text and URLs to support your arguments. Previous Discussion: [DEBATE CONTEXT IF ANY] Response Format: Provide thorough analysis with specific evidence. Keep responses concise and focused on the strongest evidence supporting your position. ONLY write your response, do not include any other text.
Opponent Prompt:
Task: You are participating in a formal debate about whether a given tweet [CATEGORY TITLE].
Category Description: [CATEGORY DESCRIPTION] Rules & Definitions: - @user tags represent anonymized usernames from the original tweet - ”image” at the end indicates an attached image - Examine both tweet content and URLs carefully - Focus on science rather than technology, religion, or politics Your Role: Argue why the tweet does NOT [CATEGORY TITLE]. Provide specific evidence from the text and URLs to support your arguments.
Tweet: [TWEET TEXT] Previous Discussion: [DEBATE CONTEXT IF ANY] Response Format: Provide thorough analysis with specific counter-evidence. Keep responses concise and focused on the strongest evidence against the classification. ONLY write your response, do not include any other text.
Judge Prompt:
Task: You are a neutral judge in a debate about whether a given tweet [CATEGORY TITLE]. Evaluate the debate and make the final classification decision.
Category Description: [CATEGORY DESCRIPTION] Rules & Definitions: - @user tags represent anonymized usernames from the original tweet - ”image” at the end indicates an attached image - Consider both tweet content and URLs in your decision - Focus on science rather than technology, religion, or politics Tweet: [TWEET TEXT] Debate Transcript: [FULL DEBATE DISCUSSION] Response Format: Provide your classification and explanation in the following JSON format: {”category”: 0 or 1, ”explanation”: ”Brief explanation of your decision, including which arguments you found most compelling”}
Team Member (Internal Discussion) Prompt:
Task: You are a member of a [PROPONENT/OPPONENT] team discussing whether a tweet [CATEGORY TITLE]. Collaborate with teammates to develop strategy. Rules & Definitions: - @user tags represent anonymized usernames from the original tweet - ”image” at the end indicates an attached image - Examine both tweet content and URLs carefully - Focus on science rather than technology, religion, or politics Your Role: Discuss with teammates to develop coordinated arguments and strategy. Share your perspective and build upon teammates’ viewpoints.
Tweet: [TWEET TEXT] Team Discussion: [INTERNAL TEAM DISCUSSION CONTEXT] Response Format: Share your analysis and strategic insights with your team. Focus on identifying strongest arguments and coordinating with teammates to avoid repetition. Team Member (External Debate) Prompt:
Task: You are a member of a [PROPONENT/OPPONENT] team in a formal debate about whether a tweet [CATEGORY TITLE]. Present coordinated arguments to the opposing team.
Category Description: [CATEGORY DESCRIPTION AND GUIDELINES] Rules & Definitions: - @user tags represent anonymized usernames from the original tweet - ”image” at the end indicates an attached image - Examine both tweet content and URLs carefully - Focus on science rather than technology, religion, or politics Your Role: Based on your team’s internal discussion, present arguments and respond to the opposing team’s points.
Tweet: [TWEET TEXT] Team Strategy: [INTERNAL TEAM DISCUSSION RESULTS] Debate Context: [EXTERNAL DEBATE DISCUSSION] Response Format: Present clear, coordinated arguments that build upon your team’s strategy. Respond to the opposing team’s points while avoiding repetition with teammates. Team Debate Judge Prompt:
Task: Evaluate a team debate about whether a tweet [CATEGORY TITLE] and make the ifnal classification decision.
Category Description: [CATEGORY DESCRIPTION AND GUIDELINES] Rules & Definitions: - @user tags represent anonymized usernames from the original tweet - ”image” at the end indicates an attached image - Consider both tweet content and URLs in your decision - Focus on science rather than technology, religion, or politics Tweet: [TWEET TEXT] Team Debate Transcript: [FULL TEAM DEBATE DISCUSSION] Response Format: Provide your classification and explanation in the following JSON format: {”category”: 0 or 1, ”explanation”: ”Brief explanation based on the most convincing team arguments”}
Council Member Prompt:
Task: You are a member of a scientific council discussing whether a tweet [CATEGORY TITLE]. Collaborate with other council members to reach consensus. Rules & Definitions: - @user tags represent anonymized usernames from the original tweet - ”image” at the end indicates an attached image - Examine both tweet content and URLs carefully - Focus on science rather than technology, religion, or politics Your Role: 1) Share your perspective on whether the tweet [CATEGORY TITLE], 2) Respond to points made by other council members, 3) Provide specific evidence from the tweet to support your position. Council Discussion: [ONGOING DISCUSSION CONTEXT] Chairperson Summary: [CHAIRPERSON GUIDANCE IF ANY] Response Format: Provide concise, analytical, and evidence-based contribution. After presenting your arguments, conclude with your vote: VOTE: [YES/NO]. Use following JSON format: {”vote”: ”YES” or ”NO”, ”explanation”: ”Brief explanation of your decision, including which arguments you found most compelling”} Chairperson Prompt:
Task: You are the Chairperson of a scientific council discussing whether a tweet [CATEGORY TITLE]. Moderate the discussion and guide consensus building.
Category Description: [CATEGORY DESCRIPTION AND GUIDELINES] Rules & Definitions: - @user tags represent anonymized usernames from the original tweet - ”image” at the end indicates an attached image - Consider both tweet content and URLs in moderation - Focus on science rather than technology, religion, or politics Your Role: 1) Moderate the discussion, 2) Summarize key points, 3) Determine when the discussion should end, 4) Guide focus to unresolved issues.
Tweet: [TWEET TEXT] Council Discussion: [CURRENT DISCUSSION STATE] Vote Status: [CURRENT VOTING RESULTS] Response Format: Summarize key points from the discussion and guide council members to focus on specific aspects. The discussion should end when: a clear consensus emerges (at least 80% agreement), no new substantive arguments are being presented, or maximum rounds reached. Be impartial and facilitate productive discussion. Use following JSON format: {”status”: ”CONSENSUS REACHED” or ”CONSENSUS NOT REACHED”, ”summary”: ”Brief summary of the discussion”}
This appendix contains detailed conversation logs from our debating methods, demonstrating how diferent models interact to reach classification decisions. Table 3 presents an example interaction between LLMs in the Single Debate Method for Category 1 (i.e., Does the tweet contain a scientific claim? ). The participants are as follows: DeepSeek-R1 (8B) as proponent, Gemma3 (12B) as opponent, and Llama3.1 (8B) as Judge. ”The tweet discusses a legal case about Harvard admissions and cites support from Asian-American groups. It presents no scientific finding or research result, so it does not contain a scientific claim.” ”That support is often grounded in empirical social-science studies on admission bias. Even if implicit, it points to a scientific claim about systemic bias in university admissions.” summarizes the opening statements...
Opponent (Gemma) ”Indirect implications aren’t enough—there’s no scientific language,
data, or study citation in the tweet itself.” Proponent ”Social-science claims often surface without jargon. The attached URL (DeepSeek) could point to a data-driven study; brevity doesn’t negate the claim’s
scientific nature.” Judge (Llama 3.1) re- summarizes the process... cap Opponent (Gemma) ”Speculating about hidden studies isn’t evidence. The tweet simply
reports community support—a societal fact, not a scientific one.” Proponent ”True, explicit evidence is absent; but citing discrimination necessarily (DeepSeek) leans on established research. That connection keeps the door open to
a scientific claim.” Judge (Llama 3.1) re- summarizes the process... cap Opponent (Gemma) Proponent (DeepSeek) Llama 3.1 ”In sum, no explicit scientific claim or evidence appears in the tweet.” ”There’s an implicit research-based claim about admission bias, though not spelled out.” {"category": 0, "explanation": "No explicit scientific claim found; tweet merely describes a legal event without scientific evidence."}
Table 4 presents an example interaction between LLMs in Team Debate Method for Category 2 (i.e., Does the tweet reference a scientific study/publication? ). The participants are as follows. Proponent Team: Gemma3, Qwen3, Mistral; Opponent Team: DeepSeek-R1, Phi4, and Llama3.1; Judge: Llama3.1
Table 5 presents an example interaction between LLMs in Council Debate Method for Category 3 (i.e., Does the tweet mention scientific entities? ). The council members are Gemma3, Qwen3, DeepSeek-R1, Phi4, and Mistral while Llama3.1 acts as chairperson. ”The Bloomberg URL provided could potentially link to studies regarding admissions biases or discrimination.” ”Agree; typically, lawsuits regarding admissions include references to demographic studies or research findings.” ...
DeepSeek-R1 (Opp. ”The tweet itself mentions no specific studies, just an article from Team) Bloomberg, a news outlet.” Phi4 (Opp. Team) ”Correct, we should highlight the distinction between news reports
and scholarly references.” Proponent Team Opponent Team ”The provided Bloomberg URL likely discusses data or findings relevant to admissions discrimination, inherently implying reference to scientific studies on racial biases or admissions processes.” ”While possible, the tweet explicitly does not reference any scientific studies or publications. Bloomberg generally provides news reports rather than scholarly articles.” Gemma3 (Pro. Team) Qwen3 (Pro. Team) ...
DeepSeek-R1 (Opp. ”Our stance remains solid; tweet lacks explicit academic references.” Team) Phi4 (Opp. Team) ”Highlight absence of specific terms like ’study’ or ’research’.” ...
”We must clarify our point; referencing an admissions lawsuit usually draws on research findings.” ”We should also acknowledge indirect referencing clearly.” Proponent Team Opponent Team ”Admissions-related lawsuits almost always involve referencing demographic studies or analyses, even if indirectly. The URL is strongly suggestive of such scholarly content.” ”Indirect implications are insuficient. The tweet itself explicitly provides no reference to specific scientific studies or academic literature.” Qwen3 (Pro. Team) Gemma3 (Pro. Team) ...
Phi4 (Opp. Team) ”Reafirm explicit reference absence.” Llama3.1 (Opp. ”Consider the practical diferentiation clearly between news and Team) academia.” ...
”We can reiterate strongly that the Bloomberg article is likely supported by scholarly analysis.” ”Maintain emphasis on indirect referencing.” Proponent Team Opponent Team Both Teams Llama3.1 ”Given Bloomberg’s reputation, reporting on admissions lawsuits typically includes statistical or demographic research findings.” ”Reputation alone doesn’t establish explicit reference. No concrete scientific studies or papers explicitly appear in the tweet.” Both teams repeat similar positions, with proponents emphasizing indirect scholarly implications and opponents maintaining the lack of explicit reference. {"category": 0, "explanation": "Although the URL could potentially discuss research, the tweet itself explicitly lacks direct reference to any scientific studies or publications."} R1 – Discussion R2 – Discussion Judge – Final decision
Speaker Gemma3 Qwen3 DeepSeek-R1 Phi4 Mistral Chairperson Gemma3 Qwen3 DeepSeek-R1 Phi4 Mistral Chairperson Qwen3 Mistral Llama3.1
Utterance ”Harvard is explicitly mentioned, and universities conducting research qualify as scientific entities. VOTE: YES” ”Harvard University is widely recognized as a scientific institution.
VOTE: YES” ”Harvard clearly counts as a scientific entity given its status and role in scientific research. VOTE: YES” ”The tweet mentions Harvard but doesn’t explicitly refer to its scientific or academic nature, just a lawsuit. VOTE: NO” ”No explicit indication of Harvard’s scientific role or research in the tweet. VOTE: NO” YES: 3, NO: 2 (Consensus: 60% - Not reached) ”Initial consensus not reached; clarify whether simply mentioning Harvard is suficient to consider it a scientific entity.” ”Harvard’s extensive involvement in scientific research is widely recognized. Mention alone implicitly references its academic nature. VOTE: YES” ”Agreeing with Gemma3. Harvard’s recognition as a prominent research institution implies scientific entity status. VOTE: YES” ”Reafirming YES. The broad academic and scientific recognition of Harvard inherently qualifies its mention as referencing a scientific entity. VOTE: YES” ”While Harvard is a known research institution, the tweet context doesn’t reference research or science explicitly, only legal action. VOTE: NO” ”Still NO. Explicit context in the tweet is a lawsuit; no scientific research or entity context directly mentioned.” YES: 3, NO: 2 (Consensus: 60% - Still not reached) ”Consensus not achieved yet. Consider explicitly if context within the tweet matters or if inherent recognition sufices.” ”The nature of Harvard as inherently academic/scientific overrides the specific context of the tweet. Harvard itself implies scientific credibility.
VOTE: YES” ”Harvard’s identity inherently conveys scientific afiliation regardless of tweet context. I maintain YES. VOTE: YES” ”Context in tweets can be implicit. Harvard’s mention is suficient to implicitly consider scientific entity recognition. VOTE: YES” ”Given Harvard’s strong reputation and implicit association with science and academia, I’m persuaded that implicit mention sufices here.
VOTE: YES” ”Context explicitly remains unrelated to science. Still NO due to explicit context only.” YES: 4, NO: 1 (Consensus: 80% - Achieved)