💻 computer science

Factual and Edit-Sensitive Graph-to-Sequence Generation via Graph-Aware Adaptive Noising

The paper introduces DLM4G, a graph-aware adaptive noising diffusion framework that significantly outperforms existing autoregressive and diffusion baselines in factual grounding and edit sensitivity for graph-to-sequence generation by employing an adaptive denoising strategy tailored to graph structures.

Original authors: Aditya Hemant Shahane, Anuj Kumar Sirohi, Tanmoy Chakraborty, Prathosh A P, Sandeep Kumar

Published 2026-04-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aditya Hemant Shahane, Anuj Kumar Sirohi, Tanmoy Chakraborty, Prathosh A P, Sandeep Kumar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Left-to-Right" Traffic Jam

Imagine you are trying to describe a complex map (a graph) to a friend using only words. You have to start at the beginning and speak one word at a time, never looking back to change what you said earlier. This is how most current AI models work (called "autoregressive" models).

The problem? If you make a small mistake early on—like saying "The capital of France is London" instead of "Paris"—you are stuck. Because you can't go back and fix that first word, the rest of your sentence might try to justify the error, or you might forget to mention other important details on the map entirely. This leads to two main issues:

Factual Errors: The AI forgets parts of the map or invents things that aren't there (hallucinations).
Edit Insensitivity: If you change one small thing on the map (e.g., swap "London" for "Paris"), the AI often doesn't change its story correctly. It might keep the old story or rewrite the whole thing chaotically.

The Solution: The "Sculptor" Approach (DLM4G)

The authors propose a new method called DLM4G (Diffusion Language Model for Graphs). Instead of writing a story word-by-word from left to right, imagine a sculptor working with a block of clay.

Start with Noise: The AI starts with a completely scrambled, noisy mess of words (like a block of clay with random shapes).
Iterative Refinement: The AI doesn't just add words; it cleans up the mess. It looks at the whole picture at once and gradually removes the noise, refining the sentence step-by-step until it becomes clear.
The "Graph" Guide: The AI is looking at the original map (the graph) the whole time, using it as a blueprint to ensure the final sculpture matches the plan.

The Secret Sauce: "Adaptive Noising"

This is the paper's most creative innovation. In standard "sculpting" (diffusion), the AI treats every part of the sentence the same way. It adds noise to the word "the" just as much as it adds noise to the name "Albert Einstein."

The authors realized this is unfair. If you scramble the name "Albert Einstein," it's very hard to guess it back correctly. If you scramble the word "the," it's easy.

The Analogy: Imagine you are trying to remember a list of items.

Standard AI: You try to remember the whole list at once, but you get distracted equally by the easy items (like "milk") and the hard items (like "saffron"). You might forget the saffron.
DLM4G (Adaptive Noising): The AI acts like a smart teacher. It knows that "saffron" (the important facts from the graph) is hard to remember. So, it protects the "saffron" tokens by adding less noise to them. It lets the "milk" (grammar words) get a bit more scrambled because they are easier to fix later.

By protecting the important facts (entities and relations) from too much noise, the AI is much better at keeping the facts straight and fixing them if the original map changes.

How They Tested It

The researchers tested this on three different "map-to-story" tasks:

WikiOFGraph: Turning Wikipedia data into sentences.
GenWiki: Turning database graphs into text.
TekGEN: Turning knowledge triples into sentences.

They also tested it on Molecules, turning chemical structures (graphs of atoms) into descriptions.

The Results: Small but Mighty

The paper claims that DLM4G is a "small" model (about 50–63 million parameters), yet it beats much larger models:

It outperforms fine-tuned models that are 12 times larger.
It competes with massive "Zero-Shot" models (models that haven't been trained on this specific task) that are 127 times larger.

Key Wins:

Fact-Checking: It makes fewer mistakes about the facts (Factual Grounding).
Editing: If you change the input graph, the output text changes exactly where it should, without breaking the rest of the sentence (Edit Sensitivity).
Science: It works surprisingly well on describing molecules, proving the method isn't just for text but for any structured data.

The Trade-off

The paper admits there is a cost. Because the AI has to "sculpt" the text over many steps (iterative refinement) rather than just typing it out instantly, it takes a bit more time to generate the answer than the fastest "left-to-right" models. However, the authors argue the extra time is worth it for the massive gain in accuracy and reliability.

Summary

DLM4G is like a smart editor that doesn't just write a story; it constantly checks the original blueprint (the graph) and protects the most important facts from getting lost in the noise. This allows a relatively small AI to write more accurate, reliable, and editable descriptions of complex data than much larger, traditional AI models.

1. Problem Statement

The paper addresses two critical limitations in current Graph-to-Sequence (G2S) generation tasks, where the goal is to convert structured graph data (e.g., Knowledge Graphs, molecular graphs) into natural language text:

Factual Grounding: Fine-tuned autoregressive models (like T5 or GPT) often suffer from hallucinations (introducing unsupported facts) or factual omissions (missing entities/relations). Standard surface-level metrics (BLEU, chrF++) fail to penalize these errors effectively because they focus on n-gram overlap rather than structural fidelity.
Edit Sensitivity: In dynamic settings where the input graph is edited (e.g., adding a node or changing a relation), autoregressive models often fail to produce localized text changes. Due to their left-to-right decoding commitment, small input edits can cause unpredictable, global shifts in the generated text.

Existing non-autoregressive diffusion models offer better global consistency but typically use data-agnostic, uniform noise schedules. This uniform corruption treats syntactic tokens and critical factual tokens (entities/relations) identically, degrading the model's ability to preserve graph structure during the denoising process.

2. Methodology: DLM4G

The authors propose DLM4G (Diffusion Language Model for Graphs), a non-autoregressive diffusion framework designed to generate text through iterative refinement while preserving graph facts.

Core Components:

Graph-Sequence Alignment:
- The input graph (specifically Knowledge Graphs) is serialized into a token sequence using special markers (e.g., [HEAD], [REL], [TAIL], [SEP]).
- During training, an alignment module identifies which output tokens correspond to specific graph entities and relations, creating an alignment set $A$ .
Graph-Aware Adaptive Noising (The Key Innovation):
- Instead of applying a uniform noise schedule to all tokens, DLM4G employs a token-wise adaptive noise schedule.
- Mechanism: The model estimates the denoising error ( $\ell^i_t$ $ℓ_{t}^{i}$ ) for each aligned token $i$ $i$ at timestep $t$ $t$ .
  - High Difficulty: If a token (typically an entity or relation) is hard to recover (high error), the model assigns it less noise (higher signal-to-noise ratio) in subsequent steps to preserve the factual signal.
  - Low Difficulty: Easier tokens (syntactic/functional words) follow the standard baseline schedule.
- Two-Stage Training Process:
  - Stage 1: Periodically estimate token-wise denoising difficulty profiles.
  - Stage 2: Map these difficulty profiles to new cumulative noise schedules ( $\bar{\alpha}^i_t$ ) using a linear mapping function, ensuring that harder tokens retain more information throughout the diffusion trajectory.
Inference-Time Adaptation:
- Since the ground-truth alignment set $A$ is unavailable at inference, DLM4G uses decoder-to-graph cross-attention as a proxy.
- It calculates a weight $w^t_i$ representing the attention mass from a generated token to the serialized graph tokens.
- The final noise schedule is a linear blend between the baseline schedule and an "anchor" schedule (learned from training aligned tokens), weighted by $w^t_i$ .
Training Objective:
- The model optimizes a Variational Lower Bound (VLB) adapted for G2S, including a consistency term (matching clean latent $z_0$ ) and a rounding term (mapping continuous latents back to discrete vocabulary).

3. Key Contributions

Novel Framework: Introduction of DLM4G, a diffusion-based G2S model that integrates graph structure via a soft inductive bias (adaptive noising) rather than hard architectural constraints.
Graph-Aware Adaptive Noising: A strategy that dynamically modulates noise levels based on token difficulty, specifically protecting factual entities and relations from excessive corruption.
Task-Grounded Metrics: The proposal of two new evaluation metrics to address the gaps in standard benchmarks:
- Factual Grounding Metric (FGT): Measures the precision of entity realization and penalizes hallucinations.
- Edit Sensitivity Rate (ESR): Quantifies how well localized graph edits are reflected as localized text changes.
Generalization: Demonstration of the method's applicability beyond knowledge graphs to scientific domains, specifically molecule captioning.

4. Experimental Results

The model was evaluated on three datasets: WikiOFGraph, GenWiki, and TekGEN, plus a molecule captioning task.

Performance vs. Autoregressive Baselines:
- DLM4G (63M parameters) outperformed fine-tuned T5-Large (770M) and GPT-2 across surface metrics (BLEU, METEOR) and embedding-based metrics (BERTScore, MAUVE).
- It achieved up to 12x parameter efficiency compared to T5-Large while delivering superior results.
Performance vs. Diffusion Baselines:
- DLM4G consistently outperformed other diffusion models (DiffuSeq, FlowSeq) trained on identical splits.
- It showed significant gains in Edit Sensitivity Rate (ESR) (+23.6% over the best diffusion baseline) and Factual Grounding (FGT).
Zero-Shot Comparison:
- DLM4G was competitive with zero-shot LLMs (LLaMa-3, GPT-o4-mini) that are 127x larger, particularly excelling in the TekGEN dataset.
Ablation Studies:
- Removing the graph-aware schedule or applying it to all tokens (not just aligned ones) resulted in performance drops, confirming the necessity of selective, difficulty-based noise modulation.
- The linear mapping function for difficulty-to-noise conversion proved superior to polynomial or cosine alternatives.
Molecule Captioning:
- On the M3-20M dataset, DLM4G-2.o achieved new State-of-the-Art (SOTA) results, outperforming specialized models like MolT5 and GitMol while being significantly smaller (63M vs. 700M+ parameters).

5. Significance

This work represents a significant shift in how structured data is verbalized into text:

Reliability: By explicitly protecting factual tokens during the diffusion process, DLM4G mitigates the hallucination issues common in autoregressive models, making it more suitable for high-stakes applications like scientific reporting or QA.
Editability: The model's inherent ability to perform localized updates makes it ideal for interactive systems where knowledge graphs evolve over time.
Efficiency: It demonstrates that specialized, smaller diffusion models can outperform massive, general-purpose LLMs on structured generation tasks, offering a more computationally efficient path to factual text generation.
Scientific Applicability: The successful extension to molecule captioning suggests the framework is robust enough for complex scientific domains where structural accuracy is paramount.

In summary, DLM4G bridges the gap between the flexibility of diffusion models and the structural rigidity required for graph-to-text tasks, offering a new paradigm for generating faithful, editable, and factually grounded text from structured data.