Skip to content

TRACE definition differs between paper and code #4

Description

@chenxn2020

The TRACE definition in the paper and released evaluator appears inconsistent.

The paper defines citation fidelity using only:

  • Consistency (Con.)
  • Coverage (Cov.)
  • Textual Fidelity (Fid.)

with VEF as a fixed 40% component of TRACE:

$$\mathrm{TRACE} = 100 \left[ 0.4,\mathrm{VEF} + 0.6 \sum_{k \in {\mathrm{Con},\mathrm{Cov},\mathrm{Fid}}} W_k E_k\right]$$

However, scoring_evidence.py includes an additional Diversity dimension:

EVI = 100 * (
    w_con * E_con + w_cov * E_cov
    + w_fid * E_fid + w_div * E_div
)

This changes TRACE/EVI and the final MMDR score.

Could you clarify which definition was used for the paper results: the three-dimensional paper formula or the four-dimensional public implementation?

Paper: https://arxiv.org/abs/2601.12346v1
Code: https://github.com/AIoT-MLSys-Lab/MMDeepResearch-Bench/blob/main/detail/scoring_evidence.py

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions