ForesightBench

A model is given a researcher-voiced description of an experiment from a recent arXiv ML preprint, including the baseline numbers, and must predict the numeric outcome of the technique or ablation. The results are withheld; the answer key is read directly off the paper's own tables.

Every sample is one controlled, within-paper comparison: same codebase, same compute, same authors. One preprint yields several samples, one per ablated component, introduced technique, or experimental setting.

The 71 source papers were submitted to arXiv after the latest published training cutoff (June 2026) of every model evaluated in the paper: 43 between 8 and 24 August 2026 and 28 between 5 and 21 October 2026. This submission window is the contamination control.

This dataset mirrors github.com/ruhzi/benchmark file for file. The construction pipeline lives at github.com/ruhzi/autoresearchrl.

Configs

config	split	rows	contents
eval (default)	test	71	the evaluation set: one sample drawn per paper (split seed 20260815)
corpus	train	457	every sample in the corpus, over 71 papers, 3,343 graded slots

Every eval row is also in the corpus; in_eval_set on each corpus row marks it. The corpus split is named train only because Hugging Face requires a split name. It was not used to train anything.

from datasets import load_dataset

eval_set = load_dataset("ruhzi/foresightbench", "eval", split="test")
corpus   = load_dataset("ruhzi/foresightbench", "corpus", split="train")

Record format

field	meaning
sample_id	<arxiv_id>-s<nnn>
arxiv_id, title	source paper
source_license	the license the source paper carries on arXiv (see License)
prompt	the question given to the model; baseline values are stated, outcome values are withheld
asked	the metrics to predict, in order
target	answer key, positionally aligned with asked. Each entry has benchmark, metric, unit, baseline_value (appears in the prompt) and method_value (the hidden outcome)
meta	base_id, title, area (research-area tag, see below), component ablated, experimental setting, and a flagged review bit
in_eval_set	corpus only: whether the row is in the eval config

The model is asked only for values. Direction (helps / hurts) is derivable from a predicted value and the metric's polarity and is deliberately not a graded target.

Research areas

meta.area is a per-paper tag, assigned by hand from the paper's title and abstract and applied to every sample of that paper. It is the area split used in the paper's per-area results. Six values:

area	meaning	papers	corpus rows	eval rows
vision	Vision & multimodal	16	135	16
safety-interp	Safety, attacks, interpretability & auditing	12	78	12
agents	LLM agents, memory & context systems	15	56	15
applied-sci	Applied & scientific ML (time series, control, neuro, audio & speech, reliability, recsys)	17	108	17
llm-training	LLM pre-/post-training & architecture	6	71	6
llm-eval	LLM evaluation & measurement studies	5	9	5

Scoring

Per predicted slot, capped relative error:

slot_error = min(|pred - true| / max(|true|, 1e-6), 1.0)

true is method_value. A slot the model leaves blank scores 1.0. A sample's score is the mean over its slots, and a model's score is the mean over samples. Lower is better. Provider or harness failures are excluded from the mean rather than penalised.

Reference points under this metric on the eval set: copying the baseline value into every slot scores 0.327.

License

ForesightBench has mixed licensing; LICENSE has the full terms.

• The dataset author's contribution is licensed CC BY 4.0. This covers the sample selection and decomposition, the question framing, the schema, the asked / target / meta annotations and the area tags.
• Material from the source papers stays under the license its authors chose on arXiv. Each prompt adapts passages of one paper, and title is that paper's title. The license is recorded per sample in source_license:

source_license	papers	corpus rows	eval rows
cc-by-4.0	34	227	34
cc-by-sa-4.0	2	7	2
cc-by-nc-sa-4.0	1	2	1
cc-by-nc-nd-4.0	3	7	3
arxiv-nonexclusive-distrib-1.0	31	214	31

If you use a sample, you are responsible for respecting its source license. To keep only material whose source is CC BY 4.0:

cc_by = corpus.filter(lambda r: r["source_license"] == "cc-by-4.0")

Authors of a source paper who want its samples removed can email ruhaanchopra2005@gmail.com with the arXiv id. When you use ForesightBench, please also cite the source papers listed in the paper's Appendix D.

Citation

@misc{foresightbench2026,
  title  = {ForesightBench},
  author = {Chopra, Ruhaan},
  year   = {2026},
  url    = {https://github.com/ruhzi/benchmark}
}