Tsinghua ML researcher advancing RL-based post-training and multimodal reasoning for LLMs and vision-language models
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
4
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Insufficient data: OpenAlex does not provide venue-resolved publication coverage for this record, so cumulative output cannot be scored. This axis would measure total peer-reviewed volume, venue tiers, and author-position share; per-year work counts (4 in 2023 to 8 in 2026) exist but do not establish venue quality.
Impact—
Insufficient data: citation-based impact cannot be scored yet — the record shows only h-index 4 and i10-index 2. This axis would measure citation totals, the quality of citing works, and downstream adoption of the RL post-training methods introduced in the cataloged papers.
Focus100/100
Scores at the maximum because the Evolve-cataloged record sits in a single OpenAlex field — machine learning (multimodal reasoning, RL for LVLMs). Both 2026 preprints pursue one agenda, anchor-based credit assignment for RL post-training, spanning token-level vision weighting in multimodal RLVR and response-level rubric evolution for LLM alignment; the wider topic list hints at earlier breadth.
Collaboration48/100
A mid-range score reflecting a compact network: 10 distinct coauthors in the record, with Siyuan Li, Xinxin Song, and Jinli Suo recurring most often in the frequent-collaborator list, alongside Chen Ruinian, Jingjing Fan, Yangen Hu, and Ke Zeng. Both cataloged papers are middle-position, non-corresponding works, consistent with team-based rather than lead-author collaboration.
Activity87/100
High and rising: OpenAlex year counts run 4 (2023), 6 (2024), 5 (2025), and 8 (2026), the 2026 peak coinciding with both reviewed preprints, and 19 works fall in the most recent three-year window. The profile depicts an early-career researcher in a sustained, high-tempo publishing phase.
Reproducibility55/100
Grounded in one assessed paper (reproMean 55, reproPapers 1), with both cataloged preprints rated medium repro-readiness. Evolve's verdicts temper confidence: PIVOT's gains over DAPO rest on single-seed evidence, and CARE's benchmark claims come from a single run with LLM-judged evaluation, so multi-seed replication remains unpublished.
Research summary
Tingxiong Xiao is a researcher at Tsinghua University whose Evolve-reviewed work centers on reinforcement-learning-based post-training for large language models (LLMs) and large vision-language models (LVLMs). Two 2026 arXiv preprints define the current agenda. The first, PIVOT ('Anchoring What Matters'), introduces a dual-level framework that adds replay anchors and vision-weighted token advantages to multimodal RLVR; Evolve notes solid gains over DAPO while flagging single-seed evidence. The second, CARE, anchors rubric evolution to frontier-model responses for LLM post-training, claiming state-of-the-art results on three benchmarks and stability across 300 training steps — though from a single run with LLM-judged evaluation. A shared signature across both papers is anchor-based credit assignment: stabilizing RL by pinning learning signals to reference responses, at token level for multimodal reasoning and response level for rubric-based alignment. Bibliometric context from OpenAlex (h-index 4, i10-index 2, 10 distinct coauthors) indicates an early-stage record with accelerating activity — 19 works in the most recent three-year window, rising from 4 works in 2023 to 8 in 2026. Xiao's OpenAlex topic profile (video coding, data compression, time-series analysis, model reduction, Bayesian modeling and causal inference) is adjacent to but not identical with the reviewed LLM work, so the Evolve catalog should be read as the most current view of the research direction.
Collaboration pattern
Collaboration is concentrated in a small, stable team: 10 distinct coauthors appear in the record, and the frequent-collaborator list repeats Siyuan Li, Xinxin Song, and Jinli Suo — the closest working ties — with Chen Ruinian, Jingjing Fan, Yangen Hu, and Ke Zeng also present. Both cataloged 2026 papers place Xiao in a middle, non-corresponding author position, consistent with contributing inside a group rather than leading an independent lab.
Research keywords
Highlights
PIVOT: replay anchors and vision-weighted token advantages for multimodal RLVR
CARE: anchor-based rubric evolution for LLM post-training
One agenda: anchor-based credit assignment across token and response levels
Both 2026 preprints share a unifying idea — pinning RL learning signals to reference anchors, token-level in multimodal reasoning and response-level in rubric-based alignment — to reduce credit-assignment noise and reward overoptimization.
Timeline
- 2023Earliest year in the provided OpenAlex counts: 4 works recorded.
- 2024Output rises to 6 works, the second-highest annual total in the record.
- 20255 works recorded, sustaining the pace within the recent three-year window that totals 19 works.
- 2026Peak year with 8 works; both Evolve-reviewed preprints — PIVOT and CARE — are dated 2026.
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 OpenAlex only, with identity name-resolved from public bibliographic data and linked to ORCID 0000-0002-0943-4097.
- Output and impact axes are unscored because venue-resolved publication and citation coverage is insufficient; only h-index 4 and i10-index 2 are available.
- The OpenAlex topic profile (video coding, data compression, time series, model reduction, Bayesian modeling and causal inference) diverges from the reviewed LLM post-training papers; this may reflect name disambiguation or arXiv-only work not yet consolidated.
- Both reviewed papers are arXiv preprints with single-seed or single-run evidence per Evolve verdicts, and reproducibility is assessed on one paper only.