Build1 publisher2 min readPublished
Polling-style aggregation often amplified shared misconceptions across five benchmarks
An arXiv preprint ran majority voting, confidence weighting and the Surprisingly Popular algorithm over several open models, and the errors were correlated enough that none of them beat a single sample.
The Engineer · Build desk
What happened
- An arXiv preprint reports that polling-style aggregation over model samples produced no consistent accuracy gain over single-sample baselines on five benchmarks, and often amplified shared misconceptions.
- The strategies tested were majority voting, confidence weighting and the Surprisingly Popular algorithm, run over several open-source models on factual, commonsense, expert-level and forecasting questions.
- Results moved in both directions, with a method that helped on one benchmark often hurting on another, and none of them beating the single-sample baseline consistently.
- On forecasting questions whose outcomes postdate the models' knowledge cutoffs, every aggregation method performed at chance.
- Conditioned on out-of-distribution random strings and asked for pseudo-random outputs, different models still produced correlated outputs.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Best-of-N on unverified factual work multiplies generation spend per answer and returns agreement instead of accuracy, and the bill lands on whoever pays per token.
- constraint Mixing vendors to buy error independence is not the escape hatch it is usually treated as, since the correlation showed up on inputs that contain no fact to share.
- decision The design question before adding samples is whether a reject step exists at all: with a test or a solver, scale sampling; without one, spend the effort somewhere other than more samples.
- exposure Any product surfacing inter-model agreement or a self-reported confidence score as a trust signal is showing users a number the paper ties to social prediction.
When nothing outside the model can reject a candidate, selection has to run on an internal signal: agreement, self-reported confidence, or predicted popularity [21]. Each of the three strategies in the paper leans on one of those, and each rests on an assumption the paper reports fails.
Majority voting needs errors that are at most weakly correlated. Human crowds meet that condition because individuals draw on different experiences and information sources [10]. The paper attributes shared priors and blind spots in language models to overlapping training corpora and similar training objectives [11], and compares the pattern to adversarial transferability, where different models fail in similar ways because they learned similar features [12]. Confidence weighting needs calibration, and the paper reports that self-reported confidence did not reliably distinguish correct answers from incorrect ones [8]. The Surprisingly Popular algorithm needs a model to predict what the crowd will say and then favour the answers that beat that prediction; under uncertainty, the paper found models are better at predicting what other models will say than at identifying what is true [9].
To test whether the correlation comes only from shared facts, the authors conditioned models on randomly generated ASCII strings and forced a multiple-choice answer [16]. There is no fact in that input for a model to be wrong about.
The paper does not claim self-consistency was oversold in the places it was measured. "These results do not contradict prior successes of self-consistency; they identify the regime in which those methods stop working," the authors wrote [14]. Pass@ and related inference-scaling methods improve performance in domains with external verifiers, including mathematics and code, because incorrect candidates can be filtered reliably there [13]. Self-consistency, repeated sampling and search are the methods the paper credits in those domains [17]. The boundary it draws is the presence of a filter: in verified domains more samples give a verifier more candidates to reject, and in unverified domains more samples reinforce shared misconceptions [15].
For that finding to transfer to a production stack, your domain has to lack a cheap reject step and your panel has to resemble the population the paper polled. The excerpt describes several open-source models and benchmark categories covering factual knowledge, commonsense reasoning, expert-level questions and forecasting, and does not name the individual models [20]. If the panel is three checkpoints of one family, the overlap the paper blames for the correlation is larger [11].
The work is a preprint posted on arxiv.org. The failure it reports is specific: language model errors are strongly correlated, so aggregation does not give a robust truth signal [1][3].
What to watch
- Whether the full paper names the models polled, and whether any closed frontier model sat on the panel alongside the open-source ones.
- Whether the no-gain result survives peer review, and whether other groups reproduce it on their own benchmark mix.
- Whether adding an external reject step, such as retrieval with a citation check or a tool call, restores the gains the paper's verified-domain boundary predicts.