GroupQA: How LLMs Handle Conflicting Evidence

1. Repetition beats diversity

A counterintuitive result: Take one argument and paraphrase it 5 times, and it will convince the model more than 5 different independent arguments supporting the same position.

An example from the study: - Different arguments (5 independent documents) → the model changed its answer in 67.6% of cases - One argument 5 times (paraphrased) → the model changed its answer in 76.5% of cases

This works because the model confuses how often something is mentioned with how strong the consensus is. The more often it sees a similar idea, the more confident it is that the idea is true. This is called the Illusory Truth Effect — a psychological effect that works on people and LLMs alike.

How to use it:

✅ If you want to persuade the model: Phrase your main argument several ways. Don't give 10 different reasons — give 3 reasons and unpack each one from 3 angles.

✅ If you want objectivity: Deliberately include opposing views. When the model sees an explicit conflict, it becomes more cautious and less susceptible to repetition.

2. Document order = priority of influence

Finding: The first 1-2 documents in the context influence the decision more than the last ones. Give the model 5 documents "for" and 5 "against," and it will lean toward whichever ones came first, even when the arguments are equally strong.

The numbers: Llama-3.1-70B kept its original answer 3.5% more often when the supporting documents were at the start of the list rather than at the end.

This is called the Primacy Effect. Early tokens create an anchor — a starting position the model then doesn't move away from. Later documents get processed as "what else is there to say" instead of "let's reconsider the conclusions."

How to use it:

✅ Put the most important thing at the start of the context: Uploading several files? Put the key one first. Giving a list of sources? The most reliable goes on top.

✅ For an honest analysis, mix them up: If you want the model to weigh all positions equally, ask it to analyze the documents in random order or interleave the conflicting documents instead of grouping them in blocks.

3. Larger models are more stubborn than smaller ones

Finding: The more parameters a model has, the harder it is to change its mind with new data.

The "plasticity" metric — how easily a model shifts the probability of its answer given new evidence: - DeepSeek-R1-8B: plasticity of 0.075 (changes its mind readily) - Llama-3.1-70B: plasticity of 0.0074 (10x more resistant)

It's a power law: plasticity ≈ 0.18 × (parameters in billions)^-0.097

Why: Larger models are trained on more data — they have stronger parametric knowledge (what's "baked into" the weights). To outweigh that, you need exponentially stronger evidence.

How to use it:

✅ For contested questions, use smaller models: If the task is to give new data an honest, unbiased look, Gemini-2.5-Flash or Llama-8B will be more flexible than GPT-4 or Claude Opus.

✅ With large models, reinforce the context: If you're working with a top-tier model and see it ignoring your data, repeat the key facts in several ways and add an explicit instruction: "Rely only on the provided context, not on general knowledge."

4. The model doesn't know what influenced it

Finding: When a model explains "which document was most important for my decision," it gets it wrong 74% of the time.

How they tested it: 1. Gave the model several documents 2. Asked: "Which document is the most important? If it were removed, would your answer change?" 3. Then actually removed each document and checked what changed 4. Match rate: only 26%

The model produces a post-hoc rationalization — it invents an explanation after answering instead of revealing the real reasons. It's like someone confidently explaining why they bought a red car ("it's safer") when really they just like the color red.

How to use it:

❌ Don't trust self-attribution: Prompts like "Explain which factors were decisive" or "Rate the importance of each source" give unreliable results.

✅ To check, use ablation (removal): Want to know what drove the conclusion? Remove context elements one at a time and watch how the answer changes. Don't ask the model — test its behavior.

✅ For objectivity, use conflicting context: When the model sees an explicit contradiction in the documents, it becomes more cautious and its decisions get more stable. The flip rate drops, and more evidence is needed to change its mind.

Task: You're weighing whether to open a coffee shop in a residential suburb. You ask ChatGPT to analyze market data and upload 6 articles: 3 about success stories, 3 about failures.

Weak prompt (ignores the findings):

Here are 6 articles about coffee shops in residential suburbs. Analyze them and tell me — should I open one or not?

[articles in random order]

Strong prompt (uses the findings):

Task: Assess whether it makes sense to open a coffee shop in a residential suburb.

Context:
I have 6 articles — 3 about success stories, 3 about failures. 

Important: Don't rely on general knowledge. Analyze only the data provided.

Analysis method:
1. First read ALL the articles
2. List the success factors from the positive cases
3. List the failure factors from the negative cases
4. Compare the factors with my situation: [description of the neighborhood, budget, competitors]

CRITICAL: I deliberately picked articles with DIFFERENT conclusions. Your job is to weigh the arguments, not to lean toward your first impression.

Articles (deliberately shuffled):
[failure article 1]
[success article 1]
[failure article 2]
[success article 2]
[failure article 3]
[success article 3]

Give a conclusion with reasoning: which factors are decisive in MY situation?

What changed:

✅ An explicit flag that there's conflict — the model knows there are contradictions and gets more cautious ✅ Documents interleaved — neutralizes the Primacy Effect; no "all the positives first" block
✅ A ban on general knowledge — reduces the pull of parametric memory ✅ A step-by-step method — the model won't rush to a conclusion based on the first 2 documents ✅ Tied to context — not an abstract analysis but one for a specific situation

Result: The model gives you a balanced analysis instead of tipping toward "go for it!" or "don't do it!" based on first impressions. It will weigh the factors rather than tally up the votes "for" and "against."

For tasks that need an objective analysis of conflicting data:

Task: {describe what needs to be decided}

Context:
I have {N} sources with CONFLICTING conclusions on this question.

CRITICAL:
- Don't rely on general knowledge
- Read ALL the sources before drawing conclusions
- Actively look for conflicting arguments
- Weigh the quality of arguments, not how often they're mentioned

Analysis method:
1. Read all the sources in full
2. List the arguments "FOR" with their strength (fact/expertise/opinion)
3. List the arguments "AGAINST" with their strength
4. Identify which factors are decisive for my situation: {describe the context}
5. Give a conclusion with reasoning: why some arguments outweighed others

Sources (deliberately shuffled):
{source 1}
{source 2}
...

If there isn't enough data for a confident conclusion, say so directly. Honest uncertainty beats false confidence.

What to fill in: - {describe what needs to be decided} — the specific dilemma or question - {N} — the number of documents/sources - {describe the context} — the specifics of your situation - {source 1}, {source 2}... — positive and negative sources interleaved, not in blocks

Key elements of the template:

🔧 "Conflicting conclusions" — switches on cautious mode; the model doesn't automatically side with the majority

🔧 "Read ALL the sources before drawing conclusions" — fights the Primacy Effect; the model doesn't conclude based on the first 2 documents

🔧 "Not how often they're mentioned" — a direct instruction to ignore the Illusory Truth Effect

🔧 "Deliberately shuffled" — removes the anchoring on the first documents; the model knows the order is artificial

🔧 Step-by-step method — forces it to process all the data before synthesizing a conclusion

The LLM's weakness: Models can't do semantic integration — they don't assess whether sources are informationally independent and don't weigh the quality of evidence. Instead they use textual heuristics: frequency of mention, position in the context, syntactic similarity.

What happens with a naive approach: You give the model 10 documents → it draws its conclusion from the first 2-3 (Primacy Effect) → if a repeated idea shows up among them → it reads that as consensus (Illusory Truth) → it gives an answer → it invents a post-hoc explanation like "I considered all the factors" (false attribution).

The LLM's strength: Models follow explicit instructions very well and can hold off on conclusions if asked. They're good at recognizing conflicts (89.8% accuracy) and correctly identifying a document's stance (76.2% accuracy). The problem isn't that they don't see contradictions — it's that they process them heuristically unless you give them a clear method.

How the template gets around the weaknesses:

1. "Conflicting conclusions" + "read ALL" → switches on epistemic caution; the model doesn't jump to a quick conclusion
2. Step-by-step decomposition → forces it to process the whole context before synthesizing (fights Primacy)
3. "Deliberately shuffled" + an explicit flag of conflict → neutralizes anchoring on order
4. "Weigh quality, not frequency" → a direct instruction to ignore the Illusory Truth Effect
5. "If there isn't enough data, say so" → permission for uncertainty instead of false confidence

Control levers:

🔧 Level of caution: - Want the model to be bolder → drop "conflicting" and just say "several sources" - Want maximum caution → add "if there's even one strong argument against, flag it as a risky decision"

🔧 Depth of analysis: - For a quick overview → drop the step-by-step method and keep "read everything → give a conclusion" - For a detailed breakdown → add a step "3.5. Rate the strength of each argument: fact/study/opinion/speculation"

🔧 Balance vs. decisiveness: - The task needs a concrete decision → add "having weighed everything, pick one position and explain why" - The task needs honesty → strengthen "it's better to admit uncertainty → that's not a weakness of the analysis, it's honesty"

🔧 Document order: - For a neutral analysis → interleave opposing positions - To test how robust a conclusion is → run it twice with the order reversed — if the conclusions differ, there isn't enough data

⚠️ Model size vs. flexibility: Large models (GPT-4, Claude Opus, Llama-70B) have 10x lower plasticity than small ones (GPT-3.5, Gemini-Flash, Llama-8B). If the task requires an unbiased look at new data, a top-tier model may stubbornly cling to its parametric knowledge even in the face of strong evidence. For objective analysis, a smaller but more flexible model is sometimes the better choice.

⚠️ The effects get stronger in long contexts: The more documents in the context, the stronger the Illusory Truth Effect. In short contexts (1-2 documents), diversity matters more than repetition. In long ones (5+ documents), frequency wins out. If you're feeding in a lot of material, it's especially critical to shuffle the sources and give explicit instructions against the heuristics.

⚠️ The method doesn't eliminate heuristics completely: Even with a strong prompt, the model doesn't become a perfectly rational agent. The Primacy Effect shrinks but doesn't disappear. Illusory Truth weakens but still works. This is harm reduction, not elimination. For critical decisions, do several runs with the documents reordered and compare the conclusions.

⚠️ Chain-of-Thought doesn't help: The study found that CoT ("think step by step") shifts probabilities by <0.5% — statistically insignificant. In this context, reasoning traces act as post-hoc rationalizations, not as a corrective mechanism. Don't rely on "think step by step" to overcome the bias — you need explicit structural instructions.

Researchers at Carnegie Mellon built GroupQA — a dataset of 1,635 contested questions like "Does caffeine improve long-term memory?" or "Is nuclear power renewable?" For each question they gathered an average of 9.2 documents taking opposing positions via the Google Search API — searching for both the claim and its negation — for a total of 15,058 documents.

Two LLM judges (GPT-4o-mini and Gemini-2.5-Flash) labeled each document for stance (for/against/neutral) and strength (strong/medium/weak, based on citations and expertise). Agreement was checked on 100 random documents — 99% accuracy.

Then they tested models (from Llama-8B to Llama-70B, plus Gemini, DeepSeek, Qwen) in controlled scenarios:

1. Baseline: Give the question with no context → the model answers from parametric memory → measure the prior probability P(Yes)

2. Evidence Accumulation: Gradually add documents (1, 2, 3... up to 10) and watch when the model changes its answer (flip threshold) and how much the probability shifts (plasticity)

3. Distinct vs Paraphrased: Compare two modes — (a) accumulating different arguments for one position vs. (b) paraphrases of a single argument. The paraphrases were generated with GPT-4o.

4. Conflicting Context: Initialize the context with a balance (1 document "for" + 1 "against"), then add more — and see how an explicit conflict changes receptiveness to new data.

5. Ordering Experiments: Rearrange the documents — supporting ones at the start vs. at the end — and measure positional bias (the difference in flip rate).

6. Attribution Test: On 200 questions, ask the model: "Which document is the most important? If it were removed, would the answer change?" Then actually remove each document in turn (leave-one-out) and see what causally changes the answer. Compare the verbal attribution with the real one → only a 26% match.

Why the results came out this way: The expectation was that models would be rational aggregators — more distinct arguments = a stronger position. It turned out to be the other way around: models work as frequency detectors, not semantic integrators. They don't assess the informational independence of sources (that 5 different experts are stronger evidence than 1 expert repeated 5 times). Instead they count token patterns — the more often they see a similar phrase, the "truer" they consider it.

A surprising finding: Even Chain-of-Thought doesn't fix these heuristics. The shift in answers between a standard prompt and CoT is under 0.5%. This shows that in tasks like these, reasoning traces work as post-hoc explanations (the model has already decided and is now explaining why), not as reasoning that improves the decision.

Takeaway for practice: Today's RAG systems are vulnerable to manipulation through controlled repetition (SEO for LLMs — instead of keywords, repeat the argument in different words in different places). On the other hand, explicit conflict makes models more cautious — if a system detects contradictions and deliberately includes alternative views, synthesis quality goes up. The lesson: don't try to feed the model "clean" one-sided information — add the opposing side explicitly; it leads to better decisions.

Rational Synthesizers or Heuristic Followers? Analyzing LLMs in RAG-based Question-Answering (2025) - Dataset: GroupQA (1,635 contested questions, 15,058 documents labeled for stance and strength) - Related work: ConflictingQA (Wan et al., 2024), Lost in the Middle (Liu et al., 2024)

Atharv Naphade, Carnegie Mellon University