Research

Junjie Xu's work spans multimodal generation, affective AI, and language understanding. We aim to turn that research foundation into reusable evidence, methods, and software. The work below is organized around five questions: how to measure, ground, align, adapt, and verify.

Each work retains its publication status and claim boundary. Preprints and manuscripts are not peer-reviewed results; each work page states what is established and what is not claimed.

14 first-author papers among 27 publications, across multimodal generation, affective AI, and language understanding.

The full publication record — Google Scholar profile

Separate demonstrations with different purposes and evidence collections; these are research prototypes.

Project pages

Methods, evidence, and results.

Explore the research in detail: the question, the method, the findings, and the boundaries. Each project page is available in English and Chinese.

Ground / Research project

EviDial

A benchmark that separates agreement with depression symptom labels from the textual evidence available to support an assessment.

BoundaryEvidence support, not diagnostic validity. Candidate references and an exploratory development audit do not establish model rankings.

Submitted · ICLR 2027
Verify / Research project

ANI

Affective Non-Interference studies how input construction, agent behavior, and scoring affect conclusions about emotional expression at authorization boundaries.

BoundaryConditional findings from synthetic tasks and saved records; clean affect-induced expansion of an unchanged authorized action set remains undemonstrated.

Measure / Research project

I'm Fine

A study of how LLM self-reports and a fixed observer differ in score levels, model ordering, and temporal variation across adapted affective instruments.

BoundaryDescriptive results from archived model interactions. No clinical, subjective-experience, or single-mechanism causal conclusions.

Measure

What a signal can actually tell you

  • Chain-of-Affective — Large Language Models have Chain-of-Affective

    Preprint · arXiv 2025

    Across eight LLM families, reported stable, family-specific affective fingerprints under repeated negative exposure — a measurable affective target in model behavior.

    arXiv:2512.12283

  • BDANN — Bidirectional Directed Acyclic Graph Neural Network for Aspect-level Sentiment Classification

    Published · TALLIP 2025

    Sentiment classification that propagates signal along dependency structure, with affective knowledge in the loop — the signal-to-state mapping done in the open.

    DOI 10.1145/3716501

  • AiM — Aesthetic Matters in Music Perception for Image Stylization

    Preprint · arXiv 2025

    Musical emotion treated as a control signal for image stylization, evaluated with image quality, aesthetic judgment, and EEG.

    arXiv:2501.01700

  • I'm Fine

    Submitted · ICLR

    A study of how LLM self-reports and a fixed observer differ in score levels, model ordering, and temporal variation across adapted affective instruments.

    Boundary

    Descriptive results from archived model interactions. No clinical, subjective-experience, or single-mechanism causal conclusions.

    Project details · English

Ground

What a claim is allowed to rest on

  • Evidence Before Severity — Auditing Safety-Evidence Fidelity and Metadata Sensitivity in Depression-Related Dialogues

    Under review · AAAI 2027 — a status, not a result

    A same-text audit — 742 dialogues, 21,628 evaluations, an 869,268-record evidence atlas from 29 sources — of whether LLM safety judgments follow expressed evidence or severity cues.

    Boundary

    Does not claim faithful evidence use, clinician-validated performance, or diagnostic validity.

  • Counterfactual Evidence Fidelity (CEF)

    Exploratory program · ICLR 2027 target

    Defines evidence fidelity as three behavioral tenets, audited across 42,240 exploratory conditions.

  • EviDial

    Submitted · ICLR 2027 — a status, not a result

    A benchmark that separates agreement with depression symptom labels from the textual evidence available to support an assessment.

    Boundary

    Evidence support, not diagnostic validity. Candidate references and an exploratory development audit do not establish model rankings.

    Project details · English

Align

What makes a cross-modal reference valid

  • MARS — Multimodal-Assisted Refined Semantic Alignment

    Published · IPM 2026

    Audio-to-image generation; reported +28.73% semantic relevance and +127.35% aesthetic score over the best AIG baseline, on VGGSound (37,250 records) and VEGAS (9,500 records).

    Boundary

    “Therapeutic or accessibility applications are motivation, not validated outcomes.”

    DOI 10.1016/j.ipm.2025.104292

  • COFFEE — Listen, with Your Eyes: Cross-Modal Fine-Grained Alignment and Fusion Enhanced for Audio-Based Image Generation

    Under review · IEEE TNNLS

    Parallel semantic-alignment and audio-caption detail paths, fused multi-scale for audio-to-image generation.

  • Phoenix — Prioritize the Hierarchical Observation of Auditory Expression for Audio-Based Image Generation

    Under review · IEEE TCSVT

    A two-stage design that keeps layered audio content instead of reducing audio to a single caption sentence.

Adapt

What keeps a response to an inferred state honest

  • Caring Without Feeling — Affective Dynamics as the Control Layer of Human-AI Agent Collaboration

    Preprint · arXiv 2026

    A survey mapping how affective cues alter reliance and oversight in human-AI collaboration — the control-layer framing for systems that adapt to inferred states.

    arXiv:2606.18259

  • HealBench — A Fine-Grained Benchmark for Evaluating Psychological Comfort in Text-to-Image Generation

    Research manuscript — not a reviewed result

    A benchmark scoring text-to-image models on psychological comfort: 19 categories × 92 subcategories across six healing functions, two-stage HealScore, 14 models evaluated.

    Boundary

    “Comfort is measured as reported response, not clinical benefit.” No claim that high HealScore implies therapeutic efficacy.

  • ElysianMV — Bridging Therapeutic Emotions and Aesthetic Imagery via Music-Visual Cross-Modal Synthesis

    Preprint · submitted to PR

    The ElysianEchoes dataset (three-dimensional emotional-semantic alignment), an emotion-driven synthesis method, and a Neuro-Aesthetic evaluation.

Verify

What would change the conclusion

  • ANI — Affective Non-Interference

    Submitted · ICLR

    Affective Non-Interference studies how input construction, agent behavior, and scoring affect conclusions about emotional expression at authorization boundaries.

    Boundary

    Conditional findings from synthetic tasks and saved records; clean affect-induced expansion of an unchanged authorized action set remains undemonstrated.

    Project details · English

  • Before the Score

    Under review · ICLR 2027

    A provenance-based audit across 129 news-interaction runs — 313 temporal contrasts, 79 of 85 runs repeating a rise-then-fall score pattern, with zero new model calls.

Note Works without a public identifier are listed without links; identifiers ship when the venue does.

The record behind this page is one person's: Junjie Xu, first author on 14 of 27 publications.

Google Scholar ORCID

DepressionKG research workbench

Find material in four openly licensed text sources, inspect one historical candidate answer, and record observations. No new model answers. Questions and feedback stay in this browser and can be exported.

research_candidate · not_evaluated

Open the workbench →

DepressionKG consultation research

Record dialogue, freeze the visible information, and export anonymous research notes. Model assessment and automatic follow-up questions are not enabled.

Open consultation workspace →

research_candidate · not_evaluated