AI models agreed with each other more than with reality in a new test

AI models agreed more closely with one another than with laboratory measurements in a new test of AI predictions. Pooling their rankings improved results on average, but several systems could still share the same mistake.

The Science Prediction Universe study, led by UniPat AI and submitted to arXiv on Nov. 11, tested 13 agent configurations spanning 12 models on 100 public tasks and 16 private tasks. Each configuration combined a model, reasoning setting and agent software. The study asked them to rank specified experiments across chemistry, materials science, biology and medicine.

Science Prediction Universe researchers reported that, on the public tasks, average agreement between scored configurations reached a Spearman rank correlation of 0.52, compared with 0.33 against measured outcomes. Agreement between agents exceeded their correlation with measurements on 81 of the 100 tasks. These averages use available scored predictions, with coverage varying by configuration.

Those figures describe how closely orderings match, rather than percentages of answers correct. Agents might rank reaction conditions by yield or materials by conductivity without having to predict numerical values. A perfect ordering without measurement ties scores 1; random rankings average 0.

Consensus improved predictions

The researchers found a benefit to pooling predictions. Averaging each experiment's position across configurations produced a consensus ranking that scored 0.08 points above the average submission, improving results on 73 of the 100 public tasks. The reported 95% confidence interval for the gain was 0.05 to 0.11.

On one oncology task, eleven configurations agreed at 0.9 but averaged just 0.27 against measurements. The benchmark’s best individual system also beat the consensus of its peers on matched tasks.

More broadly, tasks with greater agreement tended to have better predictions. The paper's agreement analysis therefore presents consensus as an informative but fallible signal. The authors explain that averaging can reduce errors that differ between systems while retaining common ones.

Compare agreement with measured outcomes

Mean pairwise agreement was 0.52, while the mean score against measurements was 0.33, using available scored predictions.

Agreement contains useful signal, but shared errors can survive a consensus.

Available scored predictions have unequal configuration coverage. The average covers 100 public tasks; the oncology view is one task with eleven configurations. Correlations are not percentages correct. Neither view measures human scientific ability.

Sources and evidence

Sources

1. Science Prediction Universe, Appendix E.8
2. UniPat AI: results and scored coverage

Supporting facts

• In the study's 100 public tasks, mean pairwise agreement between scored configurations was a Spearman rank correlation of 0.52, versus a mean correlation of 0.33 with measured outcomes; agreement was higher on 81 tasks. [1]
• Averaging item positions into a consensus ranking raised the score by 0.081 over the average submission, with a 95% confidence interval of 0.054 to 0.110, and improved results on 73 of the 100 public tasks. [1]
• Agreement contained useful signal in SPU, but could be jointly wrong: across tasks agreement and accuracy had a Spearman correlation of 0.56, while eleven configurations agreed at 0.90 on one oncology task and averaged 0.27 against measurements. [1]
• The authors interpret shared errors as agents relying on simple assumptions about how outcomes rise or fall; averaging can cancel errors that differ between configurations while retaining their shared errors. [1]
• Scored coverage was unequal: Claude Opus 5.5 and Claude Fable 5.1 were scored on 44 and 30 public tasks respectively, largely because of safety refusals, and only 28 tasks across the full benchmark were scored by all 13 configurations. [1] [2]
• The reported evaluation provides no human prediction baseline, so it does not establish whether these configurations outperform scientists, and its specified ranking tasks do not measure every kind of AI-assisted science. [1]

Method

Reported public-task agreement and measurement correlations share a -1 to 1 axis. Dot x = left + (correlation + 1) / 2 × axis width. Pooling paths are labelled schematic. Consensus gain is 0.081 score points (95% CI 0.054 to 0.110), higher on 73/100 public tasks.

Scope

Available scored predictions have unequal configuration coverage. The average covers 100 public tasks; the oncology view is one task with eleven configurations. Correlations are not percentages correct. Neither view measures human scientific ability.

Infographic by CryptoSlate

The authors connect those shared errors to simple assumptions about how experimental outcomes should rise or fall. If several agents rely on the same mistaken relationship, collecting more of their opinions can preserve that error.

The configurations were scored on different task sets. Claude Opus 5.5 and Claude Fable 5.1 had scores on only 44 and 30 public tasks, respectively, largely because of safety refusals. Across the full benchmark, all 13 configurations shared just 28 scored tasks: 27 public and one private.

The public tasks also looked backward at published measurements and were selected after pilot evaluations. The private set contained only 16 tasks across four unpublished experimental series, limiting how broadly its findings can be applied.

There was useful predictive ability in that private set: twelve of the 13 configurations ranked the unpublished experiments significantly better than chance overall, according to UniPat's Nov. 12 results release. Of 208 official private runs, 159 were scored.

The study used no human prediction baseline. Its findings apply to ranking the experimental outcomes it tested, with lab measurements still needed to check the answers.