Chaojun Xiao

Tsinghua University · China

Unclaimed · AI-structured profileORCID OpenAlex

Tsinghua researcher on efficient LLM reasoning, on-policy distillation, and self-evolution evaluation

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

Output metrics for this axis are not yet computed: the current snapshot lacks a verified full publication list and venue breakdown, so no score can be given. This dimension would measure total works, venue tiers, and consistency of production across the whole career.

Impact—

Impact data is currently insufficient to score this axis: verified citation aggregates, citing works, and venue-level influence are not yet available in the snapshot. This dimension would measure citation-based influence, including how widely the distillation and latent-reasoning methods are adopted by others.

Focus0/100

The focus axis scores 0, which reads as breadth rather than concentration: the field distribution splits evenly at 0.5 between efficient reasoning for LLMs and LLM post-training, and the five largest OpenAlex topic clusters — topic modeling (71), NLP techniques (43), advanced neural network applications (23), multimodal machine learning (21), and domain adaptation/few-shot learning (18) — span several subareas. The research program is broad and fast-moving rather than narrowly specialized.

Collaboration71/100

The collaboration axis scores 71. The record shows 55 works with 35 distinct coauthors over the last three years, and twelve recurring collaborators — including Yuxin Zuo, Xiaoang Xu, and Siyuan Liu — point to ongoing team-based work. Co-authorship is substantial but anchored in a moderately sized core group rather than a very large distributed network.

Activity100/100

The activity axis scores the maximum 100. Yearly works in the snapshot rise from 3 in 2018 and 1 in 2019 to 13 in 2023, 18 in 2024, and 28 in 2026, with 55 works in the last three years alone. Three arXiv preprints from 2026 sit in the Evolve catalog, confirming a steep, ongoing ramp-up in research production.

Reproducibility78/100

The reproducibility axis averages 78 across the two catalog papers with reproducibility assessments (reproMean 78, reproPapers 2). Readiness is rated high for the on-policy distillation study and medium for A*-Thought-V2 and StudyBench; Evolve's verdicts flag single runs, missing trained-latent baselines, and small test sets as the main gaps rather than unavailable artifacts.

Research summary

Chaojun Xiao is a researcher at Tsinghua University whose recent work concentrates on making large language models reason and train more efficiently. An OpenAlex-resolved bibliographic profile (ORCID 0000-0001-6039-0942) records an h-index of 15, an i10-index of 21, and 55 works over the last three years with 35 distinct coauthors. Yearly output in the snapshot climbs from 3 works in 2018 and 1 in 2019 to 13 in 2023, 18 in 2024, and 28 in 2026, which the activity axis scores at the maximum 100. The field distribution splits evenly at 0.5/0.5 between natural language processing (efficient reasoning for LLMs) and machine learning (LLM post-training), with the largest topic clusters in topic modeling (71), NLP techniques (43), advanced neural network applications (23), multimodal machine learning (21), and domain adaptation/few-shot learning (18). Three 2026 arXiv preprints anchor the Evolve catalog: A*-Thought-V2 on geometric latent reasoning compression (middle author), Rethinking On-Policy Distillation II showing that one training example recovers most of the distillation gain (last author, high reproducibility readiness), and StudyBench, a physics-olympiad benchmark separating absorption from transfer in self-evolution (last author). Reproducibility averages 78 across the two assessed papers.

Collaboration pattern

The record lists 35 distinct coauthors across 55 works in the last three years, with twelve frequent collaborators, including Yuxin Zuo, Xiaoang Xu, Siyuan Liu, Shuo Wang, and Junlan Feng. On the three catalogued 2026 preprints Chaojun Xiao appears as last author twice and middle author once, consistent with team-based LLM post-training projects rather than solo work.

Research keywords

efficient LLM reasoninglatent reasoningchain-of-thought compressionon-policy distillationknowledge distillationLLM post-trainingdata-efficient trainingone-shot RLVRstate coverageself-evolution of LLMsLLM benchmarkingtopic modeling

Highlights

One training example recovers most of on-policy distillation's gains
Rethinking On-Policy Distillation II reports that a single training query recovers most of the benefit of on-policy distillation, attributed to state coverage and slow absorption. Results are described as solid at 1.5B–7B scale, though from single runs only, and Evolve rates the paper's reproducibility readiness high.
StudyBench separates absorption from transfer in self-evolution
StudyBench is a controlled physics-olympiad benchmark asking whether self-evolution can squeeze textbooks for olympiad capability, and it separates knowledge absorption from genuine transfer using GRPO / label-free RL and in-context guidance. Evolve notes that small test sets and Qwen-anchored filtering limit generality.
A*-Thought-V2 compresses latent reasoning via hidden-state geometry
A*-Thought-V2 uses the geometric dynamics of LLM hidden-state trajectories (PCA geometry) to compress latent chain-of-thought, claiming up to 2.6 points of accuracy at roughly half the length. Evolve finds the direction promising but notes the absence of trained-latent baselines and significance tests.

Timeline

2026StudyBench preprint released: a controlled physics-olympiad benchmark testing whether self-evolution can squeeze textbooks for olympiad capability (last-author position).
2026Rethinking On-Policy Distillation II preprint released, showing one training example recovers most of on-policy distillation's gain (last-author position).
2026A*-Thought-V2 preprint released on efficient latent reasoning via the geometric dynamics of LLMs (middle-author position).

Papers in the Evolve catalog

Frequent co-authors

Yuxin Zuo×2 Xiaoang Xu×1 Siyuan Liu×1 Shuo Wang×1 Junlan Feng×1 Fanyu Meng×1 Zhu Zhang×1 Jixun Wang×1 Xiaorong Wang×1 Zihan Zhou×1 Xin Li×1 Yiming Zhang×1

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 (h-index, i10-index, yearly counts, topics, coauthors) come from an OpenAlex identity snapshot; the profile is name-resolved from public bibliographic data and name ambiguity is possible.
Output and impact radar axes are unscored because the snapshot does not yet provide verified citation and productivity aggregates for those dimensions.
Only the three 2026 arXiv preprints in the Evolve catalog have been individually reviewed; all other works are reflected solely through aggregate bibliometrics.
The reproducibility score rests on 2 assessed papers (mean 78), and all catalog entries are arXiv preprints that may not yet be peer-reviewed.