5-agent-llm-debate-where-agents-iteratively-revise-numeric-preference CLM-0531In a 5-agent LLM debate where agents iteratively revise a numeric preference vector, switching from all-to-all to ring communication once aggregate disagreement falls below a threshold reaches consensus using fewer tokens than all-to-all throughout and in fewer rounds than ring-only, but it does not dominate for every backbone: with Gemini 2.5-Flash all-to-all throughout converged in fewer rounds than the adaptive policy.
Capability: Getting more from several models than from one
Observed on
Five role-conditioned agents on one software-architecture planning prompt, 6-dimensional preference vectors, three backbones (GPT-5.4-Mini, Gemini 2.5-Flash, Claude Opus-4.8), one .
Sources
- Validated by: Rounds-to-consensus and total input tokens for adaptive vs complete-only vs ring-only, on one planning task with three LLM backbones; adaptive used 41,727 tokens / 9 rounds for GPT vs 67,593 / 12 complete-only and 60,858 / 17 ring-only; Claude under complete-only exhausted the 100,000-token budget. No repeated trials or statistical test.. Measured, but on a single prompt with one run per cell and no variance reported; token counts were rescaled by hand-chosen constants per model to fit a common budget, so cross-model cost comparison is not meaningful. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When RL-training a central-agent multi-agent harness with 1.5B-3B Qwen2.5 backbones on multi-hop QA, optimizing only the coordinating central policy and leaving sub-agents frozen scores higher than jointly training central and sub-agent policies (average F1 47.79 vs 45.36, accuracy 42.88 vs 40.49 at 3B).Getting more from several models than from one · reviewed by AI
- For LLM agents acting through stateful tools over many turns, unsafe policy-violating completions accumulate with the interaction budget and are roughly doubled when the adversarial user conditions each next request on the agent's actual prior response, so single-turn or short-prefix safety evaluation understates the failure rate.Using the tools it is given · reviewed by AI
- For LLM agents that must compose and then execute domain tool workflows, training separate adapters for the two roles splits the gains cleanly: supervised fine-tuning of the planner on reference tool sequences mainly improves workflow-order fidelity, while node-level RL on argument-match rewards mainly improves tool-argument grounding and final-answer quality, and combining both beats either on a 9B backbone.Planning a sequence of actions before acting · reviewed by AI
- In LLM multi-agent systems that reflect after failure, restricting reflection to the single agent responsible for the first incorrect step — located by a trained step-level correctness scorer and paired with a counterfactually generated corrected step — raises task success more than having every agent reflect on the trajectory, because reflections written by agents that behaved correctly enter their memory as misleading guidance for later tasks.Getting more from several models than from one · reviewed by AI
- For tool-using LLM trading agents with a frozen backbone, periodically rewriting the system prompt from the agent's own decision traces and realized portfolio returns changes tool-use behavior reliably — code-interpreter calls rise from about 1 to 2.6-4.5 per day and regime-relevant metrics (VaR, EMA, RSI, signal normalization) that the fixed-prompt agent never computes start appearing — but the resulting risk-adjusted performance beats the fixed-policy agent in only about half the evaluated model-by-regime settings (best Sharpe and cumulative return among LLM methods in 3 of 6).Improving its own scaffolding from failures · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: Running the same three strategies over many tasks and seeds and finding adaptive switching does not use fewer tokens than all-to-all or fewer rounds than ring-only, or that it is dominated on both axes for most backbones.