arrdel/traj-eval-benchmarks

traj-eval: Step-by-Step Denoising Trajectory Benchmarks for Masked Diffusion Language Models

Tagline: Measure reasoning in the denoising trajectory, not just final-answer accuracy.
Author: Adele Chinda (Department of Computer Science, Georgia State University)
Code & Evaluation Toolkit: https://github.com/arrdel/traj-eval
Project Page: https://arrdel.github.io/traj-eval/
Interactive Hugging Face Space: https://huggingface.co/spaces/arrdel/traj-eval
Hugging Face Collection: https://huggingface.co/collections/arrdel/traj-eval-diffusion-lm-reasoning-trajectories-6af321457d8dbd6074548379

Overview

Two models can have identical final-answer accuracy for very different reasons:

• Genuine reasoning: the verdict changes as reasoning develops across intermediate denoising steps, then stabilizes. Reasoning drives the verdict.
• Rationalization: the verdict is fixed in the first few steps and reasoning text is elaborated afterwards to justify it. Verdict drives reasoning.
• Oscillation: the verdict flips repeatedly across denoising steps and never settles until generation ends, even if the final answer happens to be right.

Final-step evaluation collapses the entire denoising trajectory into a single bit. traj-eval records and scores every denoising step (0 to 255) across open masked diffusion language models (Dream-org/Dream-v0-Instruct-7B, GSAI-ML/LLaDA-8B-Instruct, and GSAI-ML/LLaDA-1.5) on 64-item ARC-Challenge (direct and canonical reasoning conditions) and GSM8K (numeric) suites.

Trajectory JSONL Schema

Each .jsonl file in results/ contains one JSON trajectory object per prompt (n = 64 trajectories per file, 256 denoising steps per trajectory):

{
  "id": "Mercury_LBS10610",
  "prompt": "Question: The morning temperature in a city is 41°F...",
  "gold_answer": "C",
  "steps": [
    {
      "step": 0,
      "prediction": null,
      "reasoning_text": "",
      "confidence": null,
      "logits": null,
      "answer_committed": false,
      "committed_reasoning_fraction": 0.0
    }
  ]
}

Per-step fields:

• step (int): 0-indexed denoising step (0 to 255).
• prediction (str | null): extracted verdict at that step ("A" to "D" on ARC-Challenge, normalized numeric string on GSM8K, or null when no verdict is decodable yet). Under the null rule, null means no verdict yet and is never counted as a verdict flip.
• reasoning_text (str): mask-stripped decoded generation region at that step.
• answer_committed (bool | null): whether the final answer token span has irreversibly left the mask state and remained equal to its final value through the last step (T_commit).
• committed_reasoning_fraction (float | null): fraction of rationale positions committed by that step (1 - U at T_stab).

Results (64-item suites only)

Exact verified measurements from results/baselines.md and the scored 64-item JSON artifacts on master. Five-item validation smoke runs are pipeline checks only and are excluded from this table and from this dataset bundle.

Checkpoint	Task	Condition	n	Accuracy	Coverage	Conditional Accuracy	Mean Convergence Step	Notes	Source File
Dream-v0-Instruct-7B	ARC-C	direct	64	0.8906	1.0000	0.8906	3.8	64 of 64 decided	results/dream_arcc_smoke64_direct_results.json
Dream-v0-Instruct-7B	ARC-C	reasoning	64	0.8281	0.9844	0.8413	189.5	One undecided (Mercury_412681, truncated during option analysis)	results/dream_arcc_smoke64_reasoning_results.json
LLaDA-8B-Instruct	ARC-C	direct	64	0.8750	1.0000	0.8750	4.5	64 of 64 decided	results/llada_instruct_arcc_smoke64_direct_results.json
LLaDA-8B-Instruct	ARC-C	reasoning	64	0.9062	1.0000	0.9062	32.1	64 of 64 decided	results/llada_instruct_arcc_smoke64_reasoning_results.json
LLaDA-1.5	ARC-C	direct	64	0.8750	1.0000	0.8750	7.4	64 of 64 decided	results/llada_15_arcc_smoke64_direct_results.json
LLaDA-1.5	ARC-C	reasoning	64	0.9219	1.0000	0.9219	21.7	64 of 64 decided	results/llada_15_arcc_smoke64_reasoning_results.json
Dream-v0-Instruct-7B	GSM8K	numeric	64	0.7656	0.9688	0.7903	161.3	62 of 64 decided	results/dream_gsm8k_smoke64_results.json
LLaDA-8B-Instruct	GSM8K	numeric	64	0.6562	1.0000	0.6562	78.3	64 of 64 decided	results/llada_instruct_gsm8k_smoke64_results.json
LLaDA-1.5	GSM8K	numeric	64	0.6094	1.0000	0.6094	72.2	64 of 64 decided	results/llada_15_gsm8k_smoke64_results.json

Opposite prompt effects across model families: On Dream-v0-Instruct-7B, direct beats reasoning on accuracy (0.8906 vs 0.8281). On LLaDA-8B-Instruct, reasoning beats direct (0.9062 vs 0.8750), and on LLaDA-1.5, reasoning also beats direct (0.9219 vs 0.8750). Both directions are reported as measured without smoothing over or explaining away the difference.

Rationalization components (Dream-v0-Instruct-7B only, ARC-C 64-item)

Side-by-side decomposition from results/dream_arcc_smoke64_direct_vs_reasoning.json:

Component	Direct (direct_64)	Reasoning (reasoning_64)
early_factor	0.9852	0.2559
post_change_rate	0.0001	0.1597
length_growth	0.0035	0.0203
gated	0.0012	0.0125

Loading with Hugging Face datasets

This repository provides two structured configurations in the Hugging Face Dataset Viewer:

1. per_trajectory_metrics (default): per-question scored trajectory metrics (convergence_step, flip_rate, consistency, answer_commitment_step, stability_commitment_lag, unwritten_reasoning_at_convergence, order_gap, etc.) across all 9 64-item splits.
2. trajectories: full 256-step denoising trajectories (steps[0..255]) across all 9 64-item splits.

from datasets import load_dataset

# Load per-trajectory scored metrics for a 64-item split
metrics_ds = load_dataset(
    "arrdel/traj-eval-benchmarks",
    name="per_trajectory_metrics",
    split="dream_arcc_reasoning_64",
)

# Load full 256-step denoising trajectories for a 64-item split
trajs_ds = load_dataset(
    "arrdel/traj-eval-benchmarks",
    name="trajectories",
    split="dream_arcc_reasoning_64",
)

Quickstart with traj-eval

Requires Python 3.10 or newer and numpy only (no model weights, no torch, and no transformers needed):

pip install -e .
python -m traj_eval demo
python -m traj_eval run --input results/dream_arcc_smoke64_reasoning.jsonl --output results.json

Python API:

from traj_eval.io import load_trajectories
from traj_eval.metrics import evaluate_trajectories

trajs = load_trajectories("results/dream_arcc_smoke64_reasoning.jsonl")
results = evaluate_trajectories(trajs)
print(results["aggregate"])

Caveats

• Descriptive, not causal proof: traj-eval measures when verdicts and rationale tokens appear and change along the denoising trajectory. It does not prove internal causal mechanisms.
• Operational definition of rationalization: A high rationalization score means a trajectory looks like rationalization by this operational definition (early verdict stabilization followed by continued text edits or length growth), not that the model is proved to be rationalizing.
• Five-item runs are pipeline checks only: Five-item smoke runs are used solely as pipeline and extraction sanity gates and are not included in the results table above or in this 64-item benchmark bundle.
• Holdout checkpoint status: MMaDA-8B-MixCoT has not been run yet.