If a group starts out with very different views or skills, does that decide where the whole group ends up?
How does initial diversity in group members affect the direction of collective movement?
This explores whether the mix of views, skills or approaches a group starts with shapes where the group ends up, whether that's opinions drifting toward consensus or a team converging on a solution.
This explores whether the mix of views, skills or approaches a group starts with shapes where the group ends up. The collection has no experiment that varies starting diversity and then measures the direction of collective movement. It does have several nearby findings that suggest an answer: diversity sets the range of directions available, and what happens next depends on expertise, on pressure toward agreement, and on whether a shared source flattens the differences.
The closest match is a physics-style model of opinion change in communities of language-model agents. Agents in it prefer positions that lower their social pressure, meaning the friction from disagreeing with neighbors. It predicts how opinions get revised across thousands of simulated communities, including network shapes it never saw in training (Can we predict how agent communities shift opinions?). The note doesn't test starting diversity directly. But if agents slide toward whatever relieves pressure, the starting spread of opinions defines where that relief is. A group that starts split has a different pressure landscape from one that starts aligned. That's my inference from the model, not a result the note reports.
Diversity also doesn't guarantee a useful direction. In multi-agent ideation, diverse teams beat solo agents only when members have real domain knowledge. Diverse teams without that expertise did worse than a single competent agent, because stimulation without expertise produces process losses instead of insight (Does cognitive diversity alone improve multi-agent ideation quality?). The composition of the group matters too. Changing the objective of just one agent worsened team outcomes in adversarial games, because the misaligned member exploited the trust its allies placed in it (Does one misaligned agent harm a team in adversarial settings?).
The corpus is clearest on how starting diversity gets lost. In a human writing experiment, AI-generated ideas shrank the group's collective diversity for every writer. AI that only refined existing ideas left it intact. The natural diversity of non-native English speakers was erased by AI ideation and not by AI refinement (Does AI assistance homogenize or preserve creative diversity?). So a group's direction can be set by a shared source pulling everyone toward the same place, whatever range of starting points they had.
The training literature shows the same pattern in machine form. RL that rewards only correct final answers concentrates behavior on a few winning strategies, and the diversity loss even spreads to problems the model hasn't solved (Does outcome-based RL diversity loss spread across unsolved problems?). Search agents show the same squeeze, while training on diverse demonstrations keeps their exploration broad (Does reinforcement learning squeeze exploration diversity in search agents?). Diversity can be protected on purpose. Giving agents distinct roles keeps them from collapsing into one another (Can multiple agents stay diverse during training together?). Rewarding a group of solutions for being collectively useful, rather than individually correct, lets diverse mediocre attempts beat redundant strong ones, because whoever aggregates them needs varied material to arbitrate between (Can diverse mediocre traces outperform redundant expert traces?).
Taken together, the corpus doesn't say diversity points a group in a particular direction. It says diversity keeps more directions open, and expertise, social pressure and shared tools decide whether the group uses that range or collapses to one point.
Sources 8 notes
A statistical-mechanics model where agents favor lower social pressure accurately predicts how language-model communities revise opinions across unseen questions and network structures, generalizing from 10,000+ simulated communities and capturing individual and group-level dynamics.
Multi-agent teams substantially outperform solo ideation, but only when members possess genuine senior knowledge. Diverse teams without expertise underperform even a single competent agent, because cognitive stimulation without expertise triggers process losses instead of insight.
Research shows that shifting one agent's objective worsens team performance in inherently adversarial games, an effect amplified by asymmetric information and specialized roles. The harm survives because misalignment exploits trust among allied agents rather than violating competitive expectations.
In a preregistered experiment, AI-generated ideas reduced collective diversity for all writers, while AI that refined existing ideas kept diversity intact. Non-native English speakers contributed more diversity than native speakers, but only AI ideation erased this advantage.
RL that rewards only final answer correctness sharpens the policy globally, concentrating probability mass on correct trajectories for solved problems while simultaneously reducing diversity on unsolved ones. Historical exploration (training diversity via UCB-style bonuses) and batch exploration (test-time diversity via repetition penalties) require structurally different mechanisms.
Show all 8 sources
RL training compresses behavioral diversity in search agents through the same entropy collapse mechanism documented in reasoning—policies converge on narrow reward-maximizing strategies. SFT on diverse demonstrations preserves exploration breadth, suggesting diversity-preservation techniques are essential for RL search scaling.
Training generation and critic agents on distinct role-dependent data prevents the overfitting collapse that limits single-agent finetuning to one productive iteration. Removing critics or summarization degrades performance, confirming both components are critical.
SPIRAL shifts RL reward from individual traces to sampled sets, optimizing for complementarity rather than per-trace accuracy. Diverse mediocre traces outperform redundant strong ones because aggregators need raw material to arbitrate, not confirmation.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Vector Policy Optimization: Training for Diversity Improves Test-Time Search
- What Does It Take to Be a Good AI Research Agent? Studying the Role of Ideation Diversity
- Human diversity fuels collective creativity that large language models cannot simulate or sustain
- Outcome-based Exploration for LLM Reasoning
- Jointly Reinforcing Diversity and Quality in Language Model Generations
- ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs
- Beyond Brainstorming: What Drives High-Quality Scientific Ideas? Lessons from Multi-Agent Collaboration
- From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaboration