Pricing Shared Mistakes into LLM Decisions

When a language model is used to make a decision under uncertainty, the most reliable recipe so far is to split the job in two. The model first assigns probabilities to the ways the world might turn out, then scores how much each outcome is worth under each action, and finally a calculation picks the action with the highest probability-weighted value. This is expected utility maximization. The obvious upgrade is to ask several models from different labs for the probabilities and average them, on the theory that their mistakes will cancel. They mostly do not. Frontier models trained on the same web text make the same mistakes. Averaging them yields a forecast that looks confident but is wrong in the same direction every model was wrong. The decision built on that forecast inherits the error.

This note proposes treating that shared error as something to be measured and paid for, rather than averaged away or "separated out." We estimate how correlated the models' forecast errors are, using past decisions whose outcomes are already known. That number tells us how many genuinely independent opinions the panel really contains. We then pick the action that does best under every probability distribution the panel cannot rule out at that effective panel size. The test is concrete. On decision benchmarks with known outcomes and lopsided payoffs, does setting the safety margin from measured correlation beat both plain averaging and a safety margin of the same form tuned blindly?

Core Claim

The harm correlated model failure does to an expected-utility decision depends on the effective number of independent forecasters, not the nominal number. That quantity can be estimated from resolved history and used directly to size a robust decision rule.

Consider $M$ models whose forecast errors on a state's probability each have variance $\sigma^2$ and pairwise correlation $\rho$. The averaged forecast then has error variance

$$\sigma^2_{\text{pool}} = \sigma^2\left(\rho + \frac{1-\rho}{M}\right),$$

which follows from summing the variance and covariance terms of an equal-weight average. Two consequences matter. First, as $M$ grows the error does not vanish but floors at $\sigma^2\rho$. This is the same mechanism behind the inverse-U voting curves reported by Are More LLM Calls All You Need?, which studies how majority voting accuracy changes with the number of calls. Second, $\rho$ and $\sigma$ are both estimable from instances whose outcomes have already resolved, so the floor is measurable rather than hypothetical.

We use this to define the decision rule

$$a^* = \arg\max_{a} \min_{p \in B(\bar p,\, r)} \sum_s p(s)\, U(s,a), \qquad r = \kappa\,\sigma\sqrt{\rho + (1-\rho)/M}.$$

Here $\bar p$ is the pooled forecast over states $s$, $U(s,a)$ is the utility of action $a$ in state $s$, $B(\bar p, r)$ is the set of distributions within distance $r$ of $\bar p$, and $\kappa$ is a single confidence multiplier. The rule only changes the answer when two actions are close in expected value under $\bar p$ but differ sharply in some state the panel may be jointly under-weighting. Those are exactly the cases where shared blind spots are expensive. Examples include draws in football, which frontier models under-predicted badly in WorldCup Arena (8 to 14 predicted against 27 that occurred), or rare crop-market shocks.

Relation to Prior Work

The decision pipeline comes from DeLLMa. DeLLMa makes one model list the relevant unknown factors, forecast them, rank outcomes pairwise to fit utilities with a Bradley-Terry model, and compute the expected-utility maximizer. We reuse DeLLMa's pipeline as our method. It works with a single forecaster and says nothing about multiple correlated ones, and that missing piece is what we supply.

Evidence that the missing piece matters comes from four papers.

  • Correlated Errors in Large Language Models audits hundreds of models. It finds that pairs of wrong models pick the same wrong option about 60% of the time on HELM, against 33% by chance, and that more accurate models err more alike.
  • WorldCup Arena shows six frontier models from different labs scoring in lockstep on live forecasts, with 28% of all picks collapsing onto a 2–1 scoreline.
  • Hidden Clones finds the same family-level error clustering across 17 vision-language models.
  • When Individually Calibrated Models Become Collectively Miscalibrated proves that well-calibrated individual classifiers can aggregate into a biased forecast, and repairs this with a payment mechanism that rewards each model's marginal contribution. Its setting is binary prediction with strategic reporters. Ours has non-strategic models, multi-state forecasts and an explicit action space with utilities. We are not designing incentives; we are choosing actions.

Two lines of work constrain the design.

  • CARE removes shared judge biases by assuming that true quality is the dominant common factor across evaluators. Its authors note this fails when the shared bias is stronger than the shared signal. That is plausibly the regime of herded forecasts, where WorldCup models agree with each other more than with reality. We therefore deliberately do not try to identify and subtract the shared error. We only bound its size, which needs resolved outcomes but no dominance assumption. - Complementing Self-Consistency with Cross-Model Disagreement shows that disagreement across model families detects confident hallucinations better than resampling one model. It also notes that cross-model disagreement itself vanishes when families share blind spots. Our correlation floor is the part of the error that disagreement cannot see. Measuring it against outcomes rather than against each other is the core difference.

Game-Theoretic Multi-LLM Routing down-weights redundant models using leverage scores of a disagreement matrix. It optimizes classification accuracy and latency, not decision value.

Research Program

The work has three parts.

Pipeline. We first build a multi-family version of DeLLMa. Five model families each produce state forecasts for every decision instance, while utilities are elicited once with a fixed elicitor so that forecast correlation is isolated. Utility correlation is studied separately later.

Correlation estimation. On resolved instances, for each state factor, we compute each model's error as forecast probability minus the realized indicator. We then estimate $\sigma$ and $\rho$ with cross-fitting so that no instance's own outcome informs its decision. Because 120 instances per benchmark is small, we generate additional resolved instances from the same constructions, for example more stock-month windows from the Yahoo Finance history DeLLMa draws on. We report the variance of the $\rho$ estimates honestly. We also test whether $\rho$ varies enough by factor type to justify estimating it per factor rather than once globally.

Analysis. We measure the realized-utility cost of correlated error as a function of $M$, check whether it flattens at the floor the variance formula predicts, and characterize which instances change action under the robust rule. This last step also addresses an apparent conflict. DeLLMa reports that decisions keep improving with more test-time sampling, while Chen et al. find voting degrades on hard queries. Our formula predicts both: improvement while the $(1-\rho)/M$ term dominates, a plateau once $\rho$ dominates. We will fit this curve rather than assert it.

Utility side. Finally, we repeat the correlation measurement for utilities elicited by multiple families. This tests whether preference elicitation shows the same monoculture as factual prediction, a question no paper in our file has examined.

Baselines

Single-model DeLLMa. The do-nothing reference is single-model DeLLMa using the strongest available family. On Agriculture, DeLLMa reports 73.3% decision accuracy against 33.3% for o1-preview (DeLLMa). For Stocks there is no reported per-benchmark number on Lacuna, only the aggregate "up to 40%" gain over prompting baselines, so we rerun it ourselves.

Pooled DeLLMa. The naive multi-model baseline averages the five families' forecasts and maximizes expected utility under the average, with no correction.

Pooled and recalibrated DeLLMa. The strongest competitor fits a recalibration of the pooled forecast (per-state temperature and bias) on the same resolved data our method uses for estimating $\rho$. This baseline gets identical information and can fix systematic herding such as draw under-prediction. A second strong competitor weights families by inverse total uncertainty following Complementing Self-Consistency with Cross-Model Disagreement, using the same samples per family. For WorldCup we also report the bookmaker-favorite reference, which WorldCup Arena puts at 64.4% accuracy against 63.9% mean model accuracy.

Ablations. Two ablations isolate the key ingredient. The first is the same robust rule with radius set as if errors were independent ($\rho=0$). The second is the same rule with a single radius grid-searched on validation, with no correlation estimate. All methods share models, prompts, sample counts, resolved data and an equal validation tuning budget.

What counts as success. Beating the baselines means lower regret against the oracle action and lower tail loss, measured by the average of the worst 10% of realized utilities. The improvement must hold over pooled-and-recalibrated and over the tuned fixed-radius ablation, with paired bootstrap intervals excluding zero.

Why not just…?

Why not just average forecasts from diverse providers? Cross-provider diversity is largely illusory. Wrong models agree on the same wrong option about 60% of the time (Correlated Errors in Large Language Models), and the WorldCup panel herded jointly. Averaging shrinks only the $(1-\rho)/M$ term. Pooled DeLLMa is in the experiments to show the plateau directly.

Why not just recalibrate the pooled forecast on past outcomes? This is the real threat. If shared error is mostly a consistent bias, such as "always under-predict draws," recalibration removes it and our method adds nothing. Our claim is narrower: recalibration fixes average bias, but the instance-to-instance shared error remains and still flips close decisions. The experiment settles it. We will also run the combination of recalibration followed by the robust rule. If recalibration alone matches the combination, we will report that correlation sizing is unnecessary once outcomes are available.

Why not just use cross-model disagreement as the uncertainty signal? Disagreement measures how much models differ from each other. Shared blind spots are by definition where they do not differ, and Hamidieh et al. acknowledge that disagreement can vanish there. The inverse-uncertainty-weighted baseline tests whether disagreement already captures what $\rho$ measures.

Why not just use a robust decision rule with a tuned margin? If a blindly tuned radius works as well, then the contribution is "be robust," not "measure correlation." The fixed-radius ablation exists to detect exactly this. The distinguishing prediction is transfer. The correlation-derived radius should carry over across panel sizes $M$ and across benchmarks without retuning, while a tuned constant should not.

Proposed Experiments and Decisive Checks

Benchmarks. We use DeLLMa's Agriculture and Stocks benchmarks (120 instances each, outcomes known). Each is run under its original accuracy objective and under two asymmetric variants: loss aversion on revenue, and a downside penalty on returns. Robustness should matter only in the asymmetric variants, and that contrast is itself a check.

From the WorldCup Arena benchmark (104 resolved matches, seven markets), we build a stake-allocation decision. Its utility is a payout table that we fix in advance, including draw outcomes, since we cannot assume the release includes odds.

On the RESUMES data, we use the 450 human-labeled pairs from Correlated Errors in Large Language Models to set up an interview-or-reject decision. Missing a qualified candidate costs more than interviewing a weak one. We measure whether the robust rule reduces the systemic exclusion that paper reports, about 20% of applicants rejected by all 20 simulated firms.

First decisive check. Vary $M$ from 1 to 5 families plus repeated samples. Test whether pooled regret flattens at the level predicted by the estimated $\rho$, as the variance formula requires.

Second decisive check. Compare the full method to the tuned-radius ablation under a train-on-one-benchmark, test-on-another protocol.

Third decisive check. Correlate per-instance action flips with the realized states. The method earns its keep only if the flips cluster on outcomes the panel jointly under-weighted, such as draws.

Failure condition. If, on the asymmetric variants, recalibrated pooling or the tuned radius matches our method within confidence intervals, the central claim is refuted and we will say so.