Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DiffER: Diffusion Entity-Relation Modeling for Reversal Curse in Diffusion Large Language Models

Shaokai He, Kaiwen Wei, Xinyi Zeng, Xiang Chen, Xue Yang, Zhenyang Li, Jiang Zhong, Yu Tian

TL;DR

This work addresses the reversal curse in diffusion LLMs, showing that bidirectional architectures still exhibit unidirectional biases in relational reasoning. It introduces DiffER, consisting of Whole-Entity Masking to preserve entity integrity, Symmetric Alignment to balance directional data, and Inverse Relation Modeling to learn inverse predicates. Across parent–child and company–ceo benchmarks, DiffER improves both forward recall and backward inference, with gains extending to the Dream backbone, and hinges on optimizing $P(B|A)$ and $P(A|B)$ under symmetric data. The results suggest that mitigating fragmentation, asymmetry, and sparsity through coordinated training strategies enhances bidirectional relational reasoning in DLLMs.

Abstract

The "reversal curse" refers to the phenomenon where large language models (LLMs) exhibit predominantly unidirectional behavior when processing logically bidirectional relationships. Prior work attributed this to autoregressive training -- predicting the next token inherently favors left-to-right information flow over genuine bidirectional knowledge associations. However, we observe that Diffusion LLMs (DLLMs), despite being trained bidirectionally, also suffer from the reversal curse. To investigate the root causes, we conduct systematic experiments on DLLMs and identify three key reasons: 1) entity fragmentation during training, 2) data asymmetry, and 3) missing entity relations. Motivated by the analysis of these reasons, we propose Diffusion Entity-Relation Modeling (DiffER), which addresses the reversal curse through entity-aware training and balanced data construction. Specifically, DiffER introduces whole-entity masking, which mitigates entity fragmentation by predicting complete entities in a single step. DiffER further employs distribution-symmetric and relation-enhanced data construction strategies to alleviate data asymmetry and missing relations. Extensive experiments demonstrate that DiffER effectively alleviates the reversal curse in Diffusion LLMs, offering new perspectives for future research.

DiffER: Diffusion Entity-Relation Modeling for Reversal Curse in Diffusion Large Language Models

TL;DR

This work addresses the reversal curse in diffusion LLMs, showing that bidirectional architectures still exhibit unidirectional biases in relational reasoning. It introduces DiffER, consisting of Whole-Entity Masking to preserve entity integrity, Symmetric Alignment to balance directional data, and Inverse Relation Modeling to learn inverse predicates. Across parent–child and company–ceo benchmarks, DiffER improves both forward recall and backward inference, with gains extending to the Dream backbone, and hinges on optimizing and under symmetric data. The results suggest that mitigating fragmentation, asymmetry, and sparsity through coordinated training strategies enhances bidirectional relational reasoning in DLLMs.

Abstract

The "reversal curse" refers to the phenomenon where large language models (LLMs) exhibit predominantly unidirectional behavior when processing logically bidirectional relationships. Prior work attributed this to autoregressive training -- predicting the next token inherently favors left-to-right information flow over genuine bidirectional knowledge associations. However, we observe that Diffusion LLMs (DLLMs), despite being trained bidirectionally, also suffer from the reversal curse. To investigate the root causes, we conduct systematic experiments on DLLMs and identify three key reasons: 1) entity fragmentation during training, 2) data asymmetry, and 3) missing entity relations. Motivated by the analysis of these reasons, we propose Diffusion Entity-Relation Modeling (DiffER), which addresses the reversal curse through entity-aware training and balanced data construction. Specifically, DiffER introduces whole-entity masking, which mitigates entity fragmentation by predicting complete entities in a single step. DiffER further employs distribution-symmetric and relation-enhanced data construction strategies to alleviate data asymmetry and missing relations. Extensive experiments demonstrate that DiffER effectively alleviates the reversal curse in Diffusion LLMs, offering new perspectives for future research.
Paper Structure (27 sections, 4 equations, 4 figures, 5 tables)

This paper contains 27 sections, 4 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Reversal Curse
  4. Diffusion LLMs
  5. Pilot Experiments
  6. Experimental Setup
  7. Stage 1: Knowledge Injection via Continued Pre-training.
  8. Stage 2: Instruction Alignment via Prompt-Conditioned SFT.
  9. Verification and Root Cause Analysis
  10. Evaluation on Qualitative Errors.
  11. Evaluation on Forward Consistency ($A \to B$).
  12. Evaluation on Backward Generalization ($B \to A$).
  13. Motivation.
  14. Methodology
  15. Whole-Entity Masking
  16. ...and 12 more sections

Figures (4)

  • Figure 1: The identified 3 factors of DLLM reversal curse: 1) entity fragmentation (split representation), 2) data asymmetry (directional bias), and 3) relationship sparsity (missing logic).
  • Figure 2: Overview of DiffER. (1) WEM: entity-level denoising for structural integrity; (2) Symmetric Alignment: bidirectional balancing via distribution-symmetric $D_{\text{sym}}$ augmentation; (3) Inverse Relation Modeling: logical commutativity via relation prediction on relational dataset $D_{\text{rel}}$.
  • Figure 3: Error-type breakdown of inference failures under a unified evaluation setting.
  • Figure 4: Typical predictions from LLaDA and DiffER.