andrw3000/bigr-traces

BIGR: raw particle traces

Tiny recursive models, think BIGR.

Per-step particle traces behind the stored-trace analyses (selection rules, guide corruption, calibration and anytime halting; Tables 1, 8, 10 and 14) of the NeurIPS 2026 paper Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models. Code: https://github.com/digiLab-ai/bigr. Paper: https://openreview.net/forum?id=nZD0FHMUmR.

The repository mirrors assets/runs/analysis/ of the code repository, so python scripts/download_checkpoints.py --traces drops the arrays next to the committed summaries that were derived from them:

ablations/calibration_corruption/raw/{sudoku_test,sudoku_df,maze_test,maze_df}/trace/fold_0*/
    unguided (beta = 0) clouds of S = 16 particles over N = 48 outer steps on the full splits
    (Sudoku-Extreme Test 10,020 and DF 1,427; Maze-Hard Test 1,000 and DF 134), five folds of one seed
ablations/ppbench/{val,golden}/pf_ppbench_{,unguided_}seed3[1-5]/
    guided and unguided clouds on PPBench validation (253 puzzles) and golden (42), five seeds

Each trace directory holds, in original particle order, particle_qlogits.npy (N, 48, 16), particle_weights.npy, particle_correct.npy, particle_ancestors.npy, particle_preds.npy (N, 48, 16, L), map_preds.npy, inputs.npy (the puzzle tokens), trm_preds.npy (the deterministic baseline's answers) and a particle_trace_manifest.json describing the ancestor semantics. They were written by scripts/exps/run_pf.py --save-particle-trace and are analysed by scripts/analysis/run_trace_analyses.py. Total size about 3.7 GB.

Released under CC BY 4.0: credit digiLab Solutions Ltd and cite the paper when you use them.