Audio ML researcher at Beijing Institute of Technology: LALM hallucination correction and speech token LM training
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
1
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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—
No output score is computed yet: the OpenAlex record is too sparse to measure publication volume reliably. This axis would capture total works, per-year counts (e.g., the 7 recent-3-year works), and author-role distribution across the record.
Impact—
No impact score is available: citation coverage is too thin for a reliable score. This axis would measure citation-based influence (citations, h-index, i10 index); the only raw signals are an h-index of 1 and an i10 index of 0.
Focus0/100
Focus scores 0 because the record splits evenly across fields: the field distribution assigns 0.5 to audio machine learning (speech and audio processing) and 0.5 to speech processing / machine learning (speech token language models). The two are closely related, so the split may partly reflect labeling rather than genuine dispersion; topic counts also list power-electronics items, consistent with possible identity mixing.
Collaboration51/100
Score 51 reflects a moderately collaborative pattern: 12 distinct coauthors appear in the last three years, and the frequent-coauthor list includes Jiasheng Kuang, Xinyu Yang, Qiuyu Fang and nine others. Both catalog papers are first-author but not corresponding, consistent with contributing the core technical work inside larger speech and audio teams rather than leading sole-author efforts.
Activity61/100
Activity scores 61 on the strength of recent output: OpenAlex records 7 works in the last three years, with per-year counts rising from 2 in 2024 and 1 in 2025 to 4 in 2026, the most productive year on the record. Both reviewed first-author preprints appeared in 2026, indicating a concentrated burst of research activity rather than a long cumulative record.
Reproducibility78/100
Reproducibility scores 78, computed as the mean readiness across 2 reviewed papers. The length-based training study is rated high, a well-controlled comparison of batch composition, loss normalization, and data ordering, though limited to one small dataset. REVE is rated medium: its encoder-state reuse method is concrete, but evaluation rests on small test sets (n=500). The overall picture favors careful, checkable experimental design.
Research summary
Hongjin Song is a researcher at Beijing Institute of Technology whose profile is name-resolved from OpenAlex bibliographic data. The recent record is concentrated in audio machine learning: the field distribution splits evenly, 0.5 to speech-and-audio processing and 0.5 to speech token language models. Two first-author 2026 arXiv preprints are reviewed in the Evolve catalog. REVE reuses large audio-language model encoder states to verify audio event mentions, removing 92.9% of unsupported ones at 0.756 recall, although the evaluation rests on small test sets (n=500). A second study shows that short-to-long training gains in speech token language models come from batch composition and loss normalization rather than data ordering, a well-controlled but single-dataset result. OpenAlex records 7 works in the last three years, 4 of them in 2026, with 12 distinct coauthors; the h-index is 1 and the i10 index 0, so citation impact remains early-stage. Evolve's reproducibility mean across the 2 reviewed papers is 78 (one high, one medium readiness). Topic labels also include power-electronics items, suggesting possible identity mixing; treat this as an early-stage profile.
Collaboration pattern
Hongjin Song works in team settings: 12 distinct coauthors over the last three years, with a frequent-coauthor roster including Jiasheng Kuang, Xinyu Yang, Qiuyu Fang, Xiang Xie and Runwu Shi. Both catalog papers are first-author but not corresponding-author works, consistent with core technical contributions inside larger speech and audio research groups rather than independently led projects.
Research keywords
Highlights
REVE: reusing encoder states to correct audio hallucinations
Reevaluating short-to-long training in speech token LMs
Above-average reproducibility at an early career stage
Evolve's reproducibility mean is 78 across 2 reviewed papers (one high, one medium readiness), alongside an h-index of 1 and 7 recent-3-year works, a small but methodologically careful record.
Timeline
- 2026First-author arXiv preprint REVE proposes reusing large audio-language model encoder states for hallucination correction.
- 2026First-author arXiv study attributes short-to-long training gains in speech token language models to batch composition and loss normalization.
- 2026OpenAlex records 4 works in 2026, the most active year on the profile, within a recent-3-year total of 7 works.
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.
- Bibliometrics come from an OpenAlex snapshot only; no ORCID is linked, so identity resolution for this name carries some uncertainty.
- Output and impact scores are not computed (dimMissing); the only raw signals are h-index 1, i10 index 0, and 7 recent-3-year works.
- OpenAlex topics include power-electronics items (advanced DC-DC converters, multilevel inverters), which may indicate profile conflation with another same-name author.
- The reproducibility mean (78) is based on only 2 reviewed papers, so evaluation coverage is narrow and may not represent the full body of work.