When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models

Weekly Paper Notes — one of the top picks from the 2026-06-27 CS paper digest. Area: AI / ML.

Authors: Josef Chen arXiv: 2606.27288 · PDF

TL;DR

The paper formalizes a hard accuracy ceiling for any multi-model LLM system whose final output is one member’s answer — routing, voting, cascades, fusion, mixture-of-agents. That ceiling is 1 − β, where β is the co-failure rate: the fraction of queries on which every model in the pool is wrong simultaneously. The author then shows the usual community diagnostic, average pairwise error correlation ρ, cannot recover β: two ensembles with identical marginals and identical ρ can have wildly different all-wrong rates. A Clopper–Pearson bound on β supplies a finite-sample certificate before anyone trains a router. Empirically, across 67 models from 21 providers, Gaussian-copula calibration underprices the co-failure tail by roughly 2.5×, and on checkable tasks, combining models rarely beats the single best member without a strong query-level routing signal. The headline takeaway: ensembles win when models fail on different questions, not when you add more of them.

What problem is the paper actually attacking?

The “combine more models, win more benchmarks” intuition has been a steady drumbeat for two years. Self-consistency, Mixture-of-Agents (MoA), routing (RouteLLM, FrugalGPT), cascades (Chen et al. 2023), and the recent wave of Self-MoA / heterogeneous-MoA work all rest on the implicit claim that diversity + aggregation monotonically reduce error.

The lineage this paper is reacting to runs through three threads:

1. Pairwise-correlation framing (Dietterich’s classical ensemble theory, ported into the LLM era by Wang et al. and the Self-MoA team) — measure ρ between member errors, treat low ρ as a sufficient diversity signal.
2. Empirical MoA gains that depend on whose paper you read: some report large gains, some report none. Nobody pre-registers a ceiling.
3. Router-training papers that fit query → model classifiers and report Pareto improvements, but rarely report how close they are to the achievable frontier.

The single observation that motivates the new mechanism: ρ is a second-moment quantity, and a second moment cannot pin down a tail event like “all 67 models wrong on the same query.” Two joint error distributions can match on means and pairwise covariances and still disagree on the all-wrong probability by an order of magnitude. So the field has been navigating with the wrong instrument.

The mechanism: the co-failure ceiling

The argument is short and information-theoretic. Let M = {m₁, …, mₖ} be the model pool, and define

β = Pr( m₁ wrong ∧ m₂ wrong ∧ … ∧ mₖ wrong )

For any selection policy π whose output on query q is the answer of some member πq ∈ M, the policy is wrong on q whenever every member is wrong. Hence

Pr( π wrong ) ≥ β ⟹ accuracy(π) ≤ 1 − β.

This holds for routing (pick one), voting (majority of members), cascades (escalate until one passes a check), and MoA-style answer-fusion as long as the fused output collapses to one member’s answer (which, on checkable tasks, it must). Crucially the bound is policy-free: no training, no router, no oracle escapes it.

The second contribution is showing that the Clopper–Pearson interval on β, computed from a small held-out set, is already a certificate on the maximum gain any future router could deliver — so practitioners can stop training routers when the ceiling is too low. The third contribution is the tetrachoric-calibrated single-factor copula stress test: even the canonical Gaussian copula, the textbook tool for joint binary outcomes, underestimates β by ~2.5× on open-ended math. The tail is fatter than the second moment lets on.

Why it works on real model pools

The author runs the analysis across 67 frontier models from 21 providers on open-ended mathematics, execution-graded code, and GPQA-Diamond reformatted as free-response. Numbers worth memorizing:

The MC→free-response flip is the most pointed result: the tail “closes” under multiple choice (members co-fail on the same wrong option because the option set anchors them) and reopens under free-response. Co-failure lives in the answer format, not the subject. That has direct implications for benchmark design — MC benchmarks systematically under-report ensemble headroom because they manufacture artificial co-failure modes.

A panel of five judges (κ = 0.73–0.92) grades the free-response runs, so the result isn’t an artifact of brittle string-match scoring.

Results in one paragraph

At matched quality, low-ρ heterogeneous ensembles beat high-ρ Self-MoA — but the win shrinks fast. On checkable tasks (code execution, math with verifiers) combining models rarely beats the single best member without a strong query-level routing signal. The gain is concentrated in regimes where (a) the best member’s error distribution differs structurally from the second-best, and (b) a cheap classifier can predict which query belongs to which regime. Absent (b), the ceiling is set by the joint tail, not by the marginals — and the joint tail, in practice, is wider than the field has been assuming.

Why this matters

The paper does for ensemble-LLM theory what bias-variance did for classical ML: it gives you a quantity to report. Three downstream consequences:

1. Router papers should publish β before training. A 1 − β ceiling of 0.94 with a single-best baseline of 0.91 leaves 3 points of headroom; a ceiling of 0.92 leaves 1 point and probably isn’t worth the routing latency.
2. Benchmarks built on multiple-choice systematically over-promise ensemble methods. GPQA-Diamond, MMLU, and the rest mask the co-failure tail. Free-response variants are the honest measurement.
3. Diversity ≠ accuracy. Adding the 68th model to your pool does nothing if it co-fails on the same queries as the worst three. The ceiling is a property of the intersection of failure sets, not the union of capabilities.

The next obvious follow-up is to extend the bound to answer-synthesis MoA (where the fused output isn’t a verbatim member answer) and to multi-turn pipelines where members get to see each other’s drafts — both regimes break the “output is one member’s answer” assumption, and may legitimately escape the ceiling.

Read alongside

• Wang et al., Self-Consistency Improves Chain-of-Thought Reasoning (2022) — the canonical voting baseline.
• Chen et al., FrugalGPT (2023) — cascade policy reference.
• Self-MoA / Mixture-of-Agents papers (2024–2025) — the targets of the empirical comparison.
• Dietterich, Ensemble Methods in Machine Learning (2000) — the classical bias–variance / correlation framing the paper supersedes.
• Clopper & Pearson (1934) — the binomial CI used for the certificate.

Links

📄 arXiv abstract · 📄 PDF

Part of the Weekly CS Paper Digest series. Summary written from a close read of the preprint abstract; full-paper figures will be incorporated in a follow-up revision.