hmdmahdavi/splice-eval-outputs

SPLICE evaluation outputs

Per-attempt grades behind every number in Splicing Hints into Reasoning Traces: Learning from Demonstrator Failures (arXiv:2610.13996), for the five students: Untutored, Untutored-Correct, SPLICE-NF, SPLICE, SPLICE-HC. Problems, solution clusters and rubrics are in hmdmahdavi/splice-eval; training data in hmdmahdavi/splice-traces.

proof-grades (7,199 rows)

One row per attempt: 45 problems x 32 samples x 5 students (one Untutored and one Untutored-Correct attempt could not be graded). Each attempt was graded 0-7 on its written solution by three judges against the rubric of its matched solution cluster.

Column	Description
arm	student
exam, problem_id, sample_idx	join keys to splice-eval
matched_cluster	solution cluster the attempt was matched to
nemotron_mean	mean of five Nemotron grading draws
gemini25	Gemini 2.5 Pro grade
gemini31	Gemini 3.1 Pro grade
trio_mean	equal-weighted mean of the three
rule_applied	true when Nemotron was the highest judge and the judges disagreed by 4 or more; the paper then uses gemini31 instead of the mean (53 attempts)
score	the score used in the paper: gemini31 if rule_applied, else trio_mean

Table 1 of the paper is score averaged per student (mean), the share of problems with any attempt at score >= T (existence), and the count of attempts at score >= T.

aime (3,600 rows)

One row per attempt: 90 problems x 8 samples x 5 students, scored by the boxed answer. correct, extracted_answer.

Citation

@article{mahdavi2026splice,
  title   = {Splicing Hints into Reasoning Traces: Learning from Demonstrator Failures},
  author  = {Mahdavi, Hamed and Mahdavinia, Pouria and Farhadi, Majid and Honavar, Vasant},
  journal = {arXiv preprint arXiv:2610.13996},
  year    = {2026}
}