TraceBench

This dataset contains the public benchmark data, synchronized aggregate results, trajectories, website data, and accepted submission artifacts for TraceBench. The matching evaluation and reproduction code is available at TommasoBendinelli/TraceBench.

Layout

• questions/BallDrop/, questions/BounceBall/, and questions/MassSlide/ contain the three canonical benchmark environments.
• results.csv and results.parquet contain the same 655 aggregate result rows and 47 columns.
• submissions/ contains six submission bundles, including the 120-run GPT-6 Luna paper-grid submission.
• website/ contains the static website data.

GPT-6 Luna scoring note

GPT-6 Luna adds 120 paper-grid runs across eight conditions, with five seeds for each of three simulators per condition. One BounceBall output used a filename that did not match the expected test sample. The nine exactly matched predictions were scored normally, and the missing sample was scored incorrect. The original output, derived scores, and scoring adjustment are preserved in the submission archive. Two trajectories preserve message and tool steps but omit inconsistent token usage counters. The separate 15-run calibration condition is not included.

DeepSeek V4.1 Flash interim results

The OpenRouter DeepSeek V4.1 Flash submission includes 115 scored paper-grid runs. Six conditions have all five seeds for each of BallDrop, BounceBall, and MassSlide and appear on the leaderboard. Two high-noise, no-context conditions have partial coverage and are not ranked. Their successful per-seed runs remain available in the results table and submission archive.