MARS Reasoning Trace Pools

Pre-generated reasoning-trace pools for MARS - Margin-Adversarial Risk-controlled Stopping, an early-stopping method for parallel LLM reasoning.

Paper: MARS: Margin-Adversarial Risk-controlled Stopping for Parallel LLM Test-time Scaling (arXiv:2606.12935).

Download these traces to exactly reproduce the paper's benchmark numbers using the open-source MARS simulator.

What is stored

Each trace record holds:

• Per-token confidence trajectory (confs: float16 array, ~2k–8k values)
• Intermediate answers probed every 2,048 tokens (probes: answer + is_correct + confidence + raw boxed text at each checkpoint)
• Final extracted answer and ground truth

These are confidence trajectories and intermediate-answer records, not full reasoning text. The raw chain-of-thought is not stored. The text fields hold answer snippets only: raw_text is the boxed answer plus up to ~300 characters that follow it. For a small number of traces (under 100 per file) where final-answer extraction failed, extracted_answer holds a longer span of the trace text.

Contents

3 reasoning models × 3 competition-math benchmarks (30 problems each):

File	Model	Benchmark	Traces
DeepSeek-8B/aime25_deepseek.parquet	DeepSeek-R1-0528-Qwen3-8B	AIME 2025	4,096/problem
DeepSeek-8B/brumo25_deepseek.parquet	DeepSeek-R1-0528-Qwen3-8B	BRUMO 2025	4,096/problem
DeepSeek-8B/hmmt25_deepseek.parquet	DeepSeek-R1-0528-Qwen3-8B	HMMT	4,096/problem
Qwen3-32B/aime25.parquet	Qwen3-32B	AIME 2025	4,096/problem
Qwen3-32B/brumo25.parquet	Qwen3-32B	BRUMO 2025	4,096/problem
Qwen3-32B/hmmt.parquet	Qwen3-32B	HMMT	4,096/problem
Qwen3-next/aime25_thinking.parquet	Qwen3-Next-80B-A3B-Thinking	AIME 2025	4,096/problem
Qwen3-next/brumo25_thinking.parquet	Qwen3-Next-80B-A3B-Thinking	BRUMO 2025	4,096/problem
Qwen3-next/hmmt_thinking.parquet	Qwen3-Next-80B-A3B-Thinking	HMMT	4,096/problem

~35 GB total (zstd-compressed parquet). ~1.1M traces across all files.

A few traces are missing: Qwen3-32B AIME 2025 has 122,878 rows, Qwen3-32B BRUMO 2025 has 122,876, and Qwen3-Next BRUMO 2025 and HMMT have 122,879 each (all other files have 30 x 4,096 = 122,880).

Schema

One row per trace:

Column	Type	Description
qid	int64	Question ID (0-indexed within the benchmark)
trace_idx	int64	Trace index within the question (0 to 4095)
ground_truth	string	Correct answer for the question
extracted_answer	string (nullable)	Final answer parsed from the full trace
confs	list<float16>	Per-token confidence score, full trace length
probes	list<struct>	Intermediate answers at checkpoints (see below)

Each probes element: {position: int64, answer: string, is_correct: bool, raw_text: string, avg_conf: float64}. Positions are checkpoint token indices (multiples of 2048, plus the per-trace termination position).

Ground truth is embedded in every row, so the files are self-contained - no separate answer key is needed.

Reproduce the paper numbers

# 1. Get the simulator
git clone https://github.com/Wenbo11/MARS.git
cd MARS
uv sync

# 2. Download traces into ./data/  (auto-detected by the runner)
hf download wenbochen111/MARS --repo-type=dataset --local-dir ./data

# 3. Run calibrated MARS on any setting (same command for pkl or parquet)
uv run python examples/run_experiment.py \
    --model deepseek-8b --dataset aime-2025 \
    --method sc-mars-cal \
    --warmup-gamma --ucb-z 1.0 --gamma-min 0.5

# All reported runs use:
# --budget 512 --iterations 64 --warmup 16 --window 2048 --seed 42

The runner auto-detects the parquet files - no flag changes needed vs. the pkl format. Results can be diffed against the reference CSVs in repro_reference/.

Reference results

repro_reference/{Model}/{dataset}/{method}/summary_overall.csv and summary_per_question.csv hold the exact per-question accuracy and savings numbers from the paper's benchmark runs. Use these to verify your reproduction matches to the decimal.

Reproducibility notes

Exact reproduction requires:

• The Dynasor version pinned in the code repo's uv.lock (for answer equivalence scoring - dynasor.core.evaluator.math_equal). The answer grouping is frozen in the simulator cache, so downstream numbers are stable once the cache is built.
• NumPy ≥1.24 (for the bootstrap RNG, which is stable across patch versions).

Provenance

Traces were generated using SGLang with the following models:

• DeepSeek-R1-0528-Qwen3-8B - DeepSeek-AI; MIT license
• Qwen3-32B - Qwen; Apache 2.0 license
• Qwen3-Next-80B-A3B-Thinking - Qwen; Apache 2.0 license

Problem sets: AIME 2025, BRUMO 2025, HMMT (competition math - problems are public; no problem text is stored in this dataset).

This dataset is released under CC-BY-4.0.

Citation

If you use this dataset, please cite the MARS paper:

@misc{chen2026mars,
  title         = {{MARS}: Margin-Adversarial Risk-controlled Stopping for Parallel {LLM} Test-time Scaling},
  author        = {Chen, Wenbo and Li, Puheng and Liu, Mengyang and Su, Weijie and Xie, Tianpei},
  year          = {2026},
  eprint        = {2606.12935},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2606.12935}
}