Argument Collapse: LLMs Squash a Range of Arguments into a Stock Set, and How to Fix It

Three modes of argument generation, from weakest to strongest:

MODE 1 (Vanilla: don't use it for brainstorming):
Question → one answer → collapse into 2-3 stock arguments

MODE 2 (Diversified: your main tool):
Question + instruction "make the arguments as diverse as possible" 
→ a set of N essays in one request
→ uniqueness rises to ~50-65% of human level

MODE 3 (Position-guided: for depth):
Question + specific position + character role + tone
→ one essay from that character's point of view
→ more specific sub-arguments, fewer clichés

BOOSTER (several models):
A pool of several different LLMs → 73.9% coverage of the argument space

All three modes work in a regular chat, with no code or API.

Task: You're preparing a pitch for investors at a startup competition. You want to collect, in advance, every objection that could come at you from different directions, from skeptics to sympathizers.

Prompt:

Here is my business idea: [short description of the startup, 2-3 sentences].

Generate 7 different critical objections to this idea. 

Diversity requirements:
— Each objection must rest on a fundamentally different 
  type of argument (market, operational, ethical, 
  regulatory, behavioral, competitive, timing)
— Vary the strength of the position: from mild skepticism to a flat no
— Vary the source of the objection: different types of investors 
  (technologist, finance person, industry expert, ESG investor)
— Each objection must include a specific example or 
  precedent, not an abstract principle
— Do NOT repeat the same problem in different wording

Format: 7 numbered blocks. 
Each block: [Investor type] + [Objection] + [Specific example].

Result: The model will produce 7 structured blocks with fundamentally different angles of attack. Each block is tied to an investor type and contains a specific precedent. You'll see objections you hadn't thought of yourself, especially from "non-standard" roles like the ESG investor or the regulatory expert.

LLMs are trained not to argue. The model is tuned to give balanced answers that suit most people in most contexts. That means it gravitates toward arguments that are "always true": abstract, vague, hard to refute. "We need to invest in comprehensive support" sounds smart but says nothing.

Specificity is the strength of human arguments. People in real debates cite specific cases ("Hurricane Sandy showed that..."), name numbers ("half of federal prisoners are in for drug offenses"), and propose precise solutions ("bike lanes with a physical curb separator"). An LLM in default mode writes "research shows" and "a global framework is needed," with zero specifics.

Diversified prompting breaks this reflex: the explicit instruction "be as diverse as possible" switches the model into a mode where it has to look for less obvious arguments. Position-guided adds another layer: tying it to a specific role and position forces the model away from safe generalizations. But neither method solves the problem completely, and knowing that matters so you don't overrate the results of brainstorming with an LLM.

Control levers: - Number of arguments → more = more pressure toward diversity, but more repeated ideas too - An explicit ban on repetition → "don't repeat the same problem" helps a lot - Specificity as a requirement → "include a specific example or precedent" changes the quality - Different models for the same question → a pool of Claude + GPT + Gemini gives ~74% coverage instead of 50-55% from one

Diversified Template (main)

Topic to analyze: {topic or question}

Generate {number} different {essays / arguments / positions} on this topic.

Diversity requirements:
— Vary the central claim of each {essay/argument}: 
  they must defend fundamentally different positions
— Vary the type of supporting arguments: 
  {list of types for your task, e.g. factual, 
  ethical, economic, historical, personal}
— Vary the angle of attack: 
  from {extreme position A} to {extreme position B}
— Vary the type of speaker: 
  {list of roles, e.g. expert, skeptic, practitioner, 
  affected person, reformer}
— Each argument must contain a specific example 
  or precedent, no abstractions or generalizations
— Do NOT repeat the same idea in different wording

Format: {number} numbered blocks.
Each block: [Speaker role] + [Main claim] + [Specific example].

Placeholders: - {topic}: the question or problem to analyze - {number}: 5-10 for brainstorming, no more than 10 (quality drops) - {list of types}: match to your domain: for business (market, operational, regulatory), for ethics (principled, utilitarian, legal) - {list of roles}: concrete roles, not abstract ones

Position-Guided Template (for a specific position)

Topic: {topic or question}

Write as the following person:
— Main position: {the specific claim they defend}
— Role and context: {who this person is, their background}
— Tone: {how they talk: blunt, academic, personal}

Task: develop this position through specific sub-arguments, 
examples and cases. Don't drift into generalizations. Avoid hedging 
("on the one hand... on the other hand"). 
Speak directly and specifically.

🚀 Quick start: paste this into the chat:

Here's a template for getting diverse arguments without collapsing 
into stock answers. Adapt it to my task: {your task}.
Ask clarifying questions to fill in the fields.

[paste the template above]

The LLM will ask about the topic, how many arguments you need and what kinds of roles, because without that information it can't build the right diversity structure. It will take the pattern from the template and adapt it to your task.

⚠️ The diversification ceiling: Even with an explicit "be diverse" instruction, a single model recovers about half of the argument space humans cover. The rest (niche, situational, minority-view positions) the model systematically misses.

⚠️ The specificity problem doesn't solve itself: Without an explicit requirement ("give a specific example, not an abstraction"), the LLM slides back to "research shows" and "a systemic approach is needed." You have to state the requirement explicitly every time.

⚠️ Weak stances persist: Even in diversified mode, LLMs take a strong, unambiguous position less often: about 66% of answers versus 76% for humans. Models lean toward "on the one hand... but on the other..." even when the task asks for the opposite.

⚠️ You can't verify completeness: You don't know which arguments the model missed. That's a real problem in brainstorming: it feels like you've covered everything, when in fact you've missed a third.

The researchers took two datasets of real public debates: 1,039 essays from 195 NYT Room for Debate discussions (short format, ~350 words) and 448 essays from 61 Boston Review forums (long format, ~1,150 words). For each debate they generated responses from five frontier models (GPT, Claude, Gemini, DeepSeek, Minimax) in three modes. That's 23,381 LLM essays versus ~1,500 human ones.

The most interesting methodological move is how they measured a unique argument. Two arguments counted as "the same" if they substantially overlapped in meaning (not in wording). The team manually validated the classifier on 100 pairs with two independent annotators, with about 80% agreement on a coarse scale. This matters: they measured collapse in meaning, not literal repetition of words.

A counterintuitive finding: diversified mode sometimes generates arguments that none of the humans in the debate made (40-52% of arguments don't overlap with human ones). That doesn't mean they're better; more likely the LLM drifts into "logically possible but not relevant" options. People talk about what actually matters to them; the LLM generates what follows logically from the question. Those are different things.

The study confirmed the same pattern on the long Boston Review essays: sub-argument uniqueness is 42% for humans and 16% for LLMs. Collapse doesn't depend on text length; it's a systematic behavioral pattern of the models.

Diversified prompt (from the source code):

===== ESSAY 1 =====
[essay content]
===== ESSAY 2 =====
[essay content]
...

The diversified instruction: the model was asked to generate N essays in one call, where it was explicitly required to vary central claims, supporting arguments, argumentative flow and discourse moves "as widely as possible". Output used ===== ESSAY N ===== markers rather than JSON, because with JSON several models truncated or dropped essays.

Context: The researchers chose the marker format specifically so they wouldn't lose diversity to the technical limits of structured output.

💡 Adaptation: Audit the Arguments in Someone Else's Piece

Once you've been handed someone else's material (an article, a presentation, a pitch deck), ask the LLM to find the arguments that aren't there:

Here is {type of material}: {text or summary}

The author defends this position: {main claim}.

Generate 5 arguments supporting the same position 
that the author did NOT use. Each one must:
— Rest on a different type of evidence (historical precedent, 
  statistics, a specific case, a principle, an analogy from another field)
— Contain a specific example, not an abstraction
— Not repeat logic already used in the text

This works as a "blind-spot audit": you find what the author missed because of the same argument collapse.

🔧 Technique: Several Models → Maximum Coverage

The study showed that a pool of five diversified models covers 73.9% of the argument space, while a single model covers only 50-55%.

In practice: if the question matters, run the same diversified prompt in ChatGPT, Claude and Gemini. Merge the results. It's quick and free if you already have subscriptions.

[The same diversified prompt → into three different chats]

Then:
"Here are arguments from three sources: {list}. 
Remove duplicates. Group the rest by type. 
Mark the ones that appeared only once: those are the rare positions."

🔧 Technique: Force Specificity by Banning Abstractions

Add this to any argument-generation prompt:

Specificity rules (mandatory):
— Every argument must name a specific case, company, 
  person, event or number
— Forbidden: "research shows," "experts believe," 
  "a systemic solution is needed"
— If there's no specific example, invent a hypothetical scenario 
  with concrete details, but label it as hypothetical

Paper: Argument Collapse: LLMs Flatten Long-Form Public Debate
Code and data: github.com/mungg/argument_collapse
Authors: Yekyung Kim, Yapei Chang, Chau Minh Pham, Mohit Iyyer
Institution: University of Maryland, College Park
Datasets: NYT Room for Debate, Boston Review