Jev for Computational Social Science

Jev makes text annotation cheap, but its accuracy and confidence depend on the task. Results from a computational social science benchmark.
Machine Learning
Author

Paul Simmering

Published

2026-09-27

Decision models like Jev are exciting for computational social science (CSS), as they promise zero-shot classification at lower cost and latency than LLMs. But how accurate is it?

Computational social science classification task benchmark

Researchers from NYU Abu Dhabi put Jev through a CSS text classification benchmark, evaluating it against LLMs and fine-tuned RoBERTa-large (Ibrahim and Zaki 2026). It’s the same CSS tasks compiled by Ziems et al. (2024) that I referenced in our GOR 2024 talk, Where do LLMs fit in NLP Pipelines?. These include detecting humor, emotion, misinformation, toxicity and ideology.

The study covers 18 English-language tasks and 7,977 items. Three tasks were used as a pilot, leaving 15 for the headline comparisons. Jev returns a label and a confidence score directly, without generating a free-text answer that needs parsing.

A useful new point on the frontier

Jev reached a median macro-F1 of 58.1 on 0-100 scale across the 15 evaluation tasks, compared with 66.9 for Gemini 3.8 Flash, the best-performing LLM by this measure. Macro-F1 gives each class equal weight when averaging F1 scores, so frequent classes do not dominate the result. Jev achieved the highest F1 on the dialect task. On the other 14 evaluation tasks, it scored below the per-task best LLM, with a median deficit of 11.6 points. I’m surprised that the relatively cheap 3.8 Flash took the top LLM spot, instead of Fable 5.1. GPT-6 models weren’t available yet for the benchmark.

Horizontal grouped bar chart comparing Jev, fine-tuned RoBERTa-large and Gemini 3.8 Flash across 18 classification tasks. Jev leads on dialect; RoBERTa has a large advantage on empathy.
Figure 1: Macro-F1 by task for Jev, fine-tuned RoBERTa-large and Gemini 3.8 Flash. Recreated from Table 1 of Ibrahim and Zaki (2026). All 18 tasks are shown; implicit hate, stance and discourse are the pilot tasks excluded from the headline results. RoBERTa scores are the historical 2023 results reported by Ziems et al.; the other models were evaluated zero-shot in 2026. RoBERTa’s tropes result was not reported.

Fine-tuning isn’t dead! The fine-tuned RoBERTa-large baseline still beats all tested zero-shot models on half the suite. Labeled training data remains valuable even as zero-shot models improve.

Jev labeled all 7,977 items for just $0.21, cheaper than any tested hosted LLM, with no parsing failures. The paper reports measured inference charges, not the full cost of building and validating an annotation pipeline. Latency was not measured.

Scatter plot of median macro-F1 against measured cost per 1,000 items. Jev has the lowest measured cost, while Gemini 3.8 Flash has the highest median macro-F1.
Figure 2: Median macro-F1 across the 15 evaluation tasks against measured inference cost per 1,000 items. Reproduced from Figure 3A of Ibrahim and Zaki (2026). The cost axis is logarithmic.

Jev scored well on the calibration of its confidence scores, beating 16 of 19 tested LLMs. However, it also showed that Jev was catastrophically overconfident on the empathy classification task. It’s still important to do your own evals on the specific construct.

Overall, Jev is a neat tool that comes into play on easy, high-volume tasks, or in combination with a stronger model that handles cases it’s unsure about. There are also about a dozen open source variants. I’d be excited to see a decision model specifically calibrated for social science tasks and validated in different languages.

References

Ibrahim, Hazem, and Yasir Zaki. 2026. “Evaluating Decision Models for Text Annotation in Computational Social Science.” https://doi.org/10.48550/arXiv.2609.24574.
Ziems, Caleb, William Held, Omar Shaikh, Jiaao Chen, Zhehao Zhang, and Diyi Yang. 2024. “Can Large Language Models Transform Computational Social Science?” Computational Linguistics 50 (1): 237–91. https://aclanthology.org/2024.cl-1.8/.