Early-career NUS researcher spanning agentic speech/audio language models and LLM reliability evaluation
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Publication-volume data for this axis is not yet available in the current snapshot, so no output score can be reported. It would measure cumulative works, yearly production, and venue coverage.
Impact—
Impact cannot be scored: citation data is insufficient for this early-stage profile. Available signals are limited to an h-index of 1 and an i10-index of 0; the axis would measure citations and downstream influence.
Focus0/100
Focus scores 0/100, reflecting a perfectly even split in the field distribution: 50% speech processing and conversational AI (audio language models) and 50% natural language processing (LLM reliability & evaluation). Topic counts reinforce this breadth — speech recognition and synthesis, speech and audio processing, and speech dialogue systems each at 2, alongside ethics and social impacts of AI and adversarial robustness in ML at 1 each.
Collaboration62/100
Collaboration scores 62/100, drawn from 22 distinct coauthors in the bibliometric record, twelve of whom appear as frequent collaborators, including Qian Chen, Chong Deng, Xu Li, and Xiangang Li. Both cataloged papers place this researcher in a middle author position without a corresponding-author role, indicating large-team projects rather than solo-led studies.
Activity40/100
Activity scores 40/100, consistent with an early ramp-up: the record shows 3 works over the recent three-year window, with yearly counts of 1 in 2025 and 2 in 2026. Both cataloged papers appeared on arXiv in 2026, indicating current momentum while overall volume remains small.
Reproducibility60/100
Reproducibility scores 60/100, averaged over the 2 papers Evolve has assessed. The picture is polarized: the type-safety study is rated high readiness for its well-controlled design, while Qwen-Audio-3.1-Realtime is rated low due to missing ablations, confidence intervals, and independent verification of its reported agentic voice gains.
Research summary
Haoxiang Xu is an early-stage researcher at the National University of Singapore whose recorded work to date splits evenly between two fields: speech processing and conversational AI (audio language models) and natural language processing (LLM reliability & evaluation), each at 50% of the field distribution. The bibliometric snapshot, identity-resolved from OpenAlex, shows 3 works within the recent three-year window — 1 in 2025 and 2 in 2026 — with an h-index of 1 and an i10-index of 0, consistent with a career at its outset. Two 2026 arXiv papers are in the Evolve catalog, both with middle authorship positions and no corresponding role. The first, 'Type-Safe Is Not Error-Free', examines whether constrained decision heads in typed models follow option names or their bound rubrics, reporting a Laya AUC drop from 93.8% to 23.2% on yes/no swaps; Evolve judged the study well-controlled and rated its reproducibility readiness high. The second, 'Qwen-Audio-3.1-Realtime', reports broad agentic voice gains for full-duplex interaction via on-policy distillation and GRPO; Evolve notes extensive results but no ablations, confidence intervals, or independent verification, rating reproducibility low. Across 22 distinct coauthors — 12 of them frequent collaborators — the pattern suggests team-embedded research, and Evolve's average reproducibility score across the 2 reviewed papers is 60/100.
Collaboration pattern
The record lists 22 distinct coauthors, 12 of them frequent collaborators, including Qian Chen, Chong Deng, Xu Li, and Xiangang Li. On both cataloged papers Haoxiang Xu holds a middle author position without corresponding-author role, consistent with contribution inside larger engineering-research teams rather than solo-led projects. The collaboration dimension scores 62/100.
Research keywords
Highlights
Constrained decision heads follow option names, not rubrics
Agentic voice gains via on-policy distillation and GRPO
Timeline
- 2025Earliest publication year in the bibliometric snapshot, with 1 recorded work.
- 2026Middle-position authorship on 'Qwen-Audio-3.1-Realtime', an arXiv paper on full-duplex agentic voice interaction.
- 2026Middle-position authorship on 'Type-Safe Is Not Error-Free', an arXiv paper on LLM reliability and rubric following.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Identity is resolved from OpenAlex only; no ORCID is linked, so common-name disambiguation risk applies.
- Bibliometrics rest on a thin early-career record: h-index 1, i10-index 0, and the output and impact dimensions lack sufficient data to score.
- Qualitative assessments rest on 2 Evolve-reviewed 2026 arXiv preprints, not yet peer-reviewed venue publications.
- The 1 work dated 2025 in the yearly counts is not detailed in the catalog, so its content is unknown.