Dynamic Debate Relationship Regulation (DEAR)

Mutual referencing among debating agents is modelled as an explicit, mutable graph of Debate Relationships. A Selection RL-Agent dynamically rewires which peers each agent attends to each round based on measured group-level consensus/divergence evidence, and a Behavior RL-Agent adapts each agent's generation strategy in response, the two optimised jointly as a sequential multi-agent RL problem.

Properties

family: Deliberation & Debate
maturity: frontier
diagram: debate
whyMultiAgent: Blind conformity is a property of who-references-whom between distinct agents, not any single agent's confidence score; regulating a reference graph is meaningless with one model, since there is no independent peer to select, ignore, or reweight, and the group evidence (consensus vs divergence) the framework perceives only exists across multiple separately generated positions. Distinct from the already-tracked dynamic-topology-routing (DyTopo), which rewires the graph per round for stage-appropriate information flow and efficiency via a semantic need/offer matcher: DEAR rewires against social conformity, deliberately restricting which peers' positions an agent sees to preserve independent reasoning, driven by measured consensus/divergence and optimised end-to-end by reinforcement learning rather than semantic matching. Also distinct from the already-tracked localized-multi-agent-debate, which keeps a fixed set of debaters and narrows debate to the point of conflict inside a shared reasoning trace: DEAR instead changes who debates whom each round, not how much of the trace is re-examined.
mast: [Cat 2]
useCases: [Multi-round LLM debate resistant to blind conformity][Group decision-making tasks prone to correlated or homogenised reasoning][Token-efficient debate at scale (reported to reduce token consumption while improving accuracy)]
source: Dynamic Debate Relationship Regulation (DEAR)

Provenance

Source: https://arxiv.org/abs/2608.03648

Dynamic Debate Relationship Regulation (DEAR)

Properties

family

Deliberation & Debate

maturity

frontier

diagram

debate

whyMultiAgent

Blind conformity is a property of who-references-whom between distinct agents, not any single agent's confidence score; regulating a reference graph is meaningless with one model, since there is no independent peer to select, ignore, or reweight, and the group evidence (consensus vs divergence) the framework perceives only exists across multiple separately generated positions. Distinct from the already-tracked dynamic-topology-routing (DyTopo), which rewires the graph per round for stage-appropriate information flow and efficiency via a semantic need/offer matcher: DEAR rewires against social conformity, deliberately restricting which peers' positions an agent sees to preserve independent reasoning, driven by measured consensus/divergence and optimised end-to-end by reinforcement learning rather than semantic matching. Also distinct from the already-tracked localized-multi-agent-debate, which keeps a fixed set of debaters and narrows debate to the point of conflict inside a shared reasoning trace: DEAR instead changes who debates whom each round, not how much of the trace is re-examined.

mast

[Cat 2]

useCases

[Multi-round LLM debate resistant to blind conformity][Group decision-making tasks prone to correlated or homogenised reasoning][Token-efficient debate at scale (reported to reduce token consumption while improving accuracy)]

source

Dynamic Debate Relationship Regulation (DEAR)