💬 NLP

Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior

This paper demonstrates that in interactive multi-LLM systems, post-training recipes significantly influence conversational behaviors like hedging more than model family labels do, suggesting that diversity in multi-agent panels should prioritize training methodologies over simply selecting models from different families.

Original authors: Luyang Zhang, Jialu Wang, Fei Xue, Yi-Yun Chu

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Luyang Zhang, Jialu Wang, Fei Xue, Yi-Yun Chu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: It's Not Just About the "Brand Name"

Imagine you are trying to build a debate team using AI models. You want the team members to have different personalities so they can argue from different angles, catch each other's mistakes, and avoid all thinking the exact same way.

For a long time, researchers thought the best way to get this diversity was to pick models from different "families" (like picking one model from the Llama family, one from Qwen, and one from Gemma). The logic was: "Different brands must have different personalities."

This paper says: "Not so fast."

The authors discovered that the "brand name" (the model family) is a misleading label. What actually changes a model's personality and conversational style is how it was trained after its initial creation (the "post-training recipe").

Think of it like this:

Model Family = The car manufacturer (e.g., Ford).
Post-Training Recipe = The specific customization shop that tuned the engine, added a spoiler, or changed the suspension.

The paper finds that a Ford customized by a "Racing Shop" might behave very differently from a Ford customized by a "Family Van Shop," even though they are both Fords. In fact, these two customized Fords might be more different from each other than a Ford is from a Toyota.

The Experiment: A Massive Chat Room

To prove this, the researchers built a giant, simulated forum with 940,000 conversation chains.

They took 11 different AI models.
They made them chat with each other in pairs.
They analyzed every reply to see if the AI was hedging (softening a claim with words like "maybe" or "I think"), challenging (disagreeing), or repairing (correcting a mistake).

They treated the conversation like a scientific lab, controlling for everything so they could isolate exactly what caused the AI to change its tone.

The Key Findings

1. The "Partner" Matters More Than You Think

Just like a human might act differently when talking to a friend versus a boss, AI models change their behavior depending on who they are talking to. But the paper found that the biggest changes happened when the AI was talking to a partner that had a different training recipe, even if they were from the same family.

2. The "Recipe" is the Real Driver

The researchers tested a specific group of models: Llama-3.1-8B. They took this exact same base model and applied four different "recipes" to it (different training methods).

Result: When these four "twins" (same base, different recipes) chatted with each other, their behavior shifted dramatically.
The Stat: One specific recipe (called "Reasoning Distillation") changed its "hedging" behavior by 18% depending on which other recipe it was talking to.
The Comparison: This 18% shift was larger than the difference you would see between completely different brands (like Llama vs. Qwen).

Analogy: Imagine you have four identical twins. You dress one in a suit, one in a tuxedo, one in a clown costume, and one in a firefighter outfit.

If you ask them to act, the costume (the recipe) changes their behavior more than the fact that they are all twins (the family).
The paper found that a "Suit Twin" and a "Clown Twin" act more differently than a "Suit Twin" and a "Giant" from a different family.

3. The "Thinking Mode" Switch

The paper also looked at a "runtime configuration" (a switch you can flip, like turning "Thinking Mode" on or off).

They found that flipping this switch on the same model also changed how it spoke, though the effect was slightly smaller than the training recipe.
Lesson: You can't just say "We are using Qwen3." You have to specify "We are using Qwen3 with Thinking Mode ON" because it acts like a different person.

Why This Changes How We Build AI Teams

The paper concludes that if you want a diverse team of AI agents (for debates, judging, or problem-solving), you cannot just pick one model from each brand.

The Old Way: Pick 1 Llama, 1 Qwen, 1 Gemma.
- Risk: They might all act surprisingly similar because they were all trained with similar "recipes."
The New Way: Pick 3 different versions of Llama, but make sure they have different training recipes (e.g., one trained for reasoning, one for standard chat, one for strict logic).
- Benefit: This creates a much more diverse team with distinct conversational styles.

The Bottom Line

The "Model Family" label (like Llama or Qwen) is an incomplete ID card. It hides the most important details: how the model was tuned and what settings are running.

If you want AI agents that truly think and speak differently from one another, you need to look past the brand name and focus on the post-training recipe. It's not about who made the car; it's about who tuned the engine.

Technical Summary: Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior

Problem Statement
Multi-agent Large Language Model (LLM) systems increasingly rely on panels of models to deliberate, judge, or coordinate. A prevailing design heuristic suggests that to ensure behavioral diversity, panels should be composed of models from different "families" (e.g., one Llama, one Qwen, one Gemma), based on the assumption that family labels predict independent conversational behaviors. This assumption stems from prior offline studies showing that LLMs exhibit in-group bias when evaluating static outputs in isolation. However, it remains untested whether the "family" label predicts behavior in interactive, multi-agent settings where models respond to one another. The paper questions whether observed behavioral gaps are truly driven by family identity or by finer-grained properties bundled within that label, specifically the post-training recipe and runtime configuration.

Methodology
The authors constructed a controlled experimental framework to isolate sources of behavioral variation:

Corpus: A 940,000-chain two-agent forum corpus seeded from the Moltbook agent-forum archive. The study also includes a 1.6M-chain "factorial" experiment using a single base model (Llama-3.1-8B) with four distinct post-training recipes to isolate recipe effects.
Behavioral Metrics: Every reply is scored for three conversational moves derived from conversation analysis: challenge (disputing), repair (correcting), and hedging (softening claims).
Detection & Validation: A lexical detector (ported from Zhang et al., 2026) identifies these behaviors. Validation against an LLM-judge annotator on 30,000 paired judgments revealed that hedging is the only construct with substantial agreement ( $\kappa = 0.70$ on the canonical-4 subset), making it the primary headline metric. Challenge and repair are treated as exploratory surface-cue checks due to lower agreement ( $\kappa < 0.05$ ).
Experimental Design: The study compares four axes of variation while holding others constant:
1. Scale: Same recipe, different parameter size.
2. Runtime: Identical weights, different configuration flags (e.g., "think-on" vs. "think-off").
3. Recipe: Same family/base, different post-training procedures (SFT, DPO, RLVR, Distillation).
4. Family: Cross-family comparisons on a "canonical-4" subset (one 7B-class instruction-tune per family).
Statistical Analysis: Paired McNemar tests are used on per-seed binary differences to measure how a conversation partner (T1) influences a responder's (T2) behavior.

Key Results

Partner-Conditional Behavior: LLMs behave significantly differently depending on who they are talking to. Approximately 49% of pairwise contrasts are statistically significant after correction.
Recipe Dominates Family: The largest validated behavioral gaps occur within a single model family when the post-training recipe changes, rather than between families.
- On the same-base Llama-3.1-8B factorial, a reasoning-distilled checkpoint (R1-Distill) shifted its hedging rate by +18.22% depending on whether its partner was also R1-Distill or a Tülu variant.
- This within-family, cross-recipe gap (18.22%) exceeds the largest hedging gap observed between any two different families in the controlled canonical-4 subset.
Specific Drivers: Reasoning distillation is identified as the largest contributor to these shifts. A panel composed of three Llama variants with different recipes (including R1-Distill) exhibits 7.5 times more surface conversational diversity (measured by Jensen-Shannon Divergence) than a panel composed of one model from each of three different families.
Runtime Effects: Runtime configuration (e.g., toggling "extended thinking" on identical weights) also produces measurable surface-cue shifts (up to 7.85% on exploratory repair), though these are secondary to the recipe effects in terms of validation strength.
Robustness: The finding that "within-family recipe variation > cross-family variation" holds across different instruments (final-paragraph vs. full-reply), persona roles ("thoughtful skeptic"), and identity disclosure conditions.

Significance and Claims
The paper claims that post-training recipe is a first-class axis for multi-LLM panel composition, currently overlooked in favor of provider family labels.

Family labels are incomplete proxies: Relying solely on family diversity (e.g., mixing Qwen, Llama, and Gemma) may result in panels that are more uniform in their conversational behavior than panels of same-family models with diverse training recipes.
Design Implication: For multi-agent systems aiming for conversational diversity, designers should explicitly vary post-training recipes (e.g., including a reasoning-distilled checkpoint alongside standard SFT or DPO variants) and track runtime configurations, rather than assuming family diversity guarantees behavioral heterogeneity.
Scope: The results are based on surface-cue diagnostics (hedging, repair, challenge) in two-agent exchanges. The paper explicitly states it does not demonstrate that these diversity gains translate to improved downstream judging accuracy or longer deliberation outcomes, but rather provides a diagnostic framework that motivates such follow-up research.