When Agents Disagree: The Selection Bottleneck in Multi-Agent LLM Pipelines

Does multi-agent LLM interaction help or hurt? Some work reports gains from debate (Du et al., 2024), critique loops (Chen et al., 2025), and mixture-of-agents synthesis (Wang et al., 2025), while other work finds that interaction adds cost without improving quality under equal budgets (Tran & Kiela, 2026; Xu et al., 2026; Jarrett et al., 2025), or that independent sampling already captures multi-agent gains (Li et al., 2024). We argue this contradiction partly reflects a missing distinction, because not all multi-agent communication is equal. Different model families find structurally different solutions, but when agents read each other's complete outputs, their proposals converge within one round, erasing the diversity that motivates using multiple models. We call this the interaction tax. We test 11 verifier-scored optimization tasks under matched budgets and find that full-solution interaction is a weak default. Independent proposal generation avoids this collapse. Full-solution interaction mainly makes agents stay close to the first solution they see instead of trying different approaches, and critique helps only if the violated rule is easy for the LLM to find and fix. These results suggest that multi-agent performance depends less on the number of agents than on the information they exchange, and interaction helps only when agents share the right information at the right time.

The performance ceiling of an LLM team is constrained not only by individual model capabilities, but also by inter-member error resonance and predictive differences. Although heterogeneous teaming is often observed to be effective in practice, existing approaches lack complementarity metrics that are computable, interpretable, and optimizable, leaving team composition to rely on heuristics. We propose a heterogeneity-driven team selection framework that performs offline profiling to characterize individual capability along with two complementary signals: one captures decorrelation in error patterns to reduce co-failures, while the other measures divergence in predictive behavior to capture strategy diversity. We formulate team selection as a standardized quality--complementarity combinatorial objective and apply an efficient greedy search to select a small team from a candidate pool. Experiments across multiple benchmarks demonstrate that our framework consistently outperforms quality-only baselines under controlled candidate pools and team sizes, establishing reusable selection principles for multi-LLM systems.

Multi-agent debate (MAD) reportedly improves reasoning and factuality over single-model inference, but prior work treats agents as symmetric peers, leaving open what drives the gains. We test the hypothesis that cognitive diversity among agents is the driver, in the setting where the question is still measurable: small open-weight models with benchmark headroom. Across 23 models from eleven vendor families, five tasks, and 5,500+ debate and control runs, we vary diversity along three axes - personas, sampling temperature, and model identity - pairing every debate configuration with a generation-budget-matched majority-vote control. The hypothesis is rejected on every axis. Debate beats single-agent inference (3--7 points where tasks have headroom) but at matched budget conditions it ties or even loses to self-consistency sampling at 1.6$\times$ the wall-clock and 3.4$\times$ the token cost. Persona prompting reduces accuracy and a dose-response experiment over each model's full combinatorial persona space shows the cost is a persona tax, not a diversity tax: redundant personas hurt most, while maximally-diverse teams recover part of the loss. Furthermore, mixed-model teams lose to majority votes over their own rosters, with accuracy tracking member capability rather than heterogeneity, and nearly all of debate's benefit comes from the first exchange of answers. We further identify a pervasive measurement hazard in which debate transcripts silently overflow serving context windows, whose correction alone moves our debate-versus-sampling comparison from $-1.8$ points to parity. Our results recast reported MAD gains as an ensemble-sampling effect and provide the budget-matched, contamination-checked baseline bar that future debate mechanisms should be required to clear.