TRACE-DeBERTa-v3-base

A fine-tuned DeBERTa-v3-base model that labels each sentence with Constructive Elements of reasoning, used as a component of the TRACE framework.

TRACE: Toulmin-based Reasoning Assessment through Constructive Elements for LLM CoT Evaluation

This model is a multi-label sentence classifier, not a scoring model. It assigns one or more Constructive Element tags to each input sentence; the downstream TRACE pipeline aggregates these labels into a reasoning quality score.

• Developed by: Korea Institute of Science and Technology Information (KISTI)
• Model type: Multi-label text classification (sentence-level)
• Base model: microsoft/deberta-v3-base
• Language: English

Role in the TRACE Pipeline

TRACE evaluates Chain-of-Thought (CoT) reasoning of LLMs in two stages:

1. Sentence labeling (this model). Reasoning text is split into sentences with spaCy, and each sentence is multi-labeled with Constructive Elements.
2. Score extraction (rule-based). A separate rule-based component computes State Validity and Transition Coherence from the resulting label sequence to produce the final TRACE score.

This model is responsible only for step 1.

Labels

The model outputs 8 independent confidence scores (sigmoid). A label is assigned when its score is ≥ 0.5.

Based on Toulmin's argumentation model:

• Claim — a conclusion, assertion, or answer being argued for
• Data/Evidence — concrete facts, observations, or given information
• Warrant — reasoning that connects evidence to the claim
• Backing — support for the warrant (definitions, theorems, principles)
• Qualifier — expressions of certainty or uncertainty
• Rebuttal — counterarguments, exceptions, or alternative considerations

Extended with Flavell's metacognition theory:

• Monitoring — self-checking, tracking progress, noticing errors
• Evaluation — judging the quality or correctness of reasoning

Usage

Quick start

from transformers import pipeline

clf = pipeline(
    "text-classification",
    model="hyyangkisti/TRACE-DeBERTa-v3-base",
    top_k=None,  # return all label scores
)

clf("Therefore, I conclude that the hypothesis is correct.")
# [[{'label': 'Claim', 'score': 0.95}, {'label': 'Qualifier', 'score': 0.82}, ...]]

Inputs and Outputs

• Input: a single English sentence (max 512 tokens, DeBERTa limit).
• Output: 8-dimensional vector of independent sigmoid probabilities, one per label.

Training Data

The model was fine-tuned on approximately 100K reasoning sentences with multi-label annotations grounded in Toulmin's argumentation model and Flavell's metacognition theory. Sentences were segmented via spaCy.

Acknowledgment

This work has been supported by the Korea Institute of Science and Technology Information (grant K26L2M3C7).

Citation

@misc{kim2026tracetoulminbasedreasoningassessment,
      title={TRACE: Toulmin-based Reasoning Assessment through Constructive Elements for LLM CoT Evaluation}, 
      author={Yundong Kim and Heyoung Yang},
      year={2026},
      eprint={2605.29656},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.29656}, 
}