Microsoft researcher in LLM distillation and post-training — agent memory transfer and on-policy distillation analysis.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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—
The output axis measures total publication volume and productivity over time. This profile's OpenAlex snapshot flags it as missing: only a single 2026 count (6 works) is recorded, so no reliable cumulative output score can be computed.
Impact—
The impact axis captures citation-based influence such as h-index and i10-index. The snapshot lists both as 0 with insufficient citation coverage, so this profile reports no impact score rather than estimating one.
Focus0/100
The focus score of 0 reflects maximal dispersion in this snapshot: fieldDist splits exactly 50/50 between 'machine learning — LLM distillation and post-training' and 'machine learning / NLP (LLM agents and memory systems)', and OpenAlex topics scatter across five areas (topic modeling 5, domain adaptation and few-shot learning 4, information retrieval and search behavior 2, NLP techniques 1, RL in robotics 1). Both main fields are distillation-adjacent, but the even split produces the low focus reading.
Collaboration54/100
Collaboration scores 54 on moderate evidence. The snapshot records 14 distinct coauthors across the 6 works in its three-year window, and the frequent-collaborator list — Chetan Bansal, Menglin Xia, Huaxiu Yao, Taylor W. Killian, Weitong Zhang, Pirzada Suhail, Yuxiao Yang, Tianrun Yu, Shangzhe Li, Kaixiang Zhao — suggests a stable Microsoft lab team. Both cataloged papers carry middle-position, non-corresponding authorship, fitting team-based industrial research.
Activity57/100
Activity scores 57: countsByYear records 6 works in 2026, and the same 6 works constitute the entire trailing three-year window, so all captured recent output is concentrated in a single year. Both cataloged papers appeared on arXiv in 2026, consistent with an active, recently intensified industrial research tempo; longer multi-year trend data is absent from the snapshot.
Reproducibility55/100
Reproducibility scores 55, the mean across the 2 Evolve-assessed papers (reproPapers 2, reproMean 55), both rated 'medium' readiness. PSD ships black-box self-distillation for agent memory, but Evolve notes its LoCoMo wins lack significance tests; the EOS-token length-inflation study is judged plausible yet small-scale. arXiv sourcing helps, though evaluation-rigor gaps temper the score.
Research summary
Xuchao Zhang is a researcher at Microsoft (United States) working at the intersection of LLM distillation, post-training, and agent memory systems. Evolve's catalog holds two of his 2026 arXiv papers, both with middle-position authorship: 'PSD: Pseudo Self-Distillation of Memory Representation Capabilities for LLM Agents', which uses privileged-prompt self-distillation to move GPT-4.1-mini memory-construction capability into Qwen3 models, and 'When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation', which traces on-policy distillation length inflation to student–teacher EOS-token mismatch and mitigates it with probability-level alignment across 3 model families. The OpenAlex snapshot (name-resolved, no ORCID) records 6 works in the trailing three-year window, all in 2026, with 14 distinct coauthors; topics span topic modeling (5), domain adaptation and few-shot learning (4), information retrieval and search behavior (2), plus single-count topics in NLP techniques and RL in robotics, while the field distribution splits evenly between LLM distillation/post-training and LLM agents/memory, yielding a focus score of 0. Radar scores stand at collaboration 54, activity 57, and reproducibility 55 — the mean over 2 assessed papers, both rated medium readiness — while output and impact remain unscored, with h-index and i10-index recorded as 0 under insufficient coverage. Evolve's verdicts flag evaluation-rigor gaps — missing significance tests on LoCoMo and small-scale EOS experiments — making this a coherent but early-visibility industrial research profile.
Collaboration pattern
The snapshot lists 14 distinct coauthors across the 6 recorded works. Chetan Bansal appears twice in the frequent-collaborator list, alongside Menglin Xia, Huaxiu Yao, Taylor W. Killian, Weitong Zhang, Pirzada Suhail, Yuxiao Yang, Tianrun Yu, Shangzhe Li, and Kaixiang Zhao — a Microsoft-industry-lab pattern with a few academic collaborators. Both cataloged papers are middle-position, non-corresponding authorships, consistent with team-based industrial research rather than solo leadership.
Research keywords
Highlights
Pseudo self-distillation for LLM agent memory
Diagnosing length inflation in on-policy distillation
A converging distillation research line
Both 2026 papers attack knowledge distillation from complementary angles — transferring capabilities across model boundaries (agent memory) and repairing its failure modes (length inflation) — with both rated medium reproducibility readiness, averaging 55 across Evolve's 2 assessments.
Timeline
- 2026Posted 'PSD: Pseudo Self-Distillation of Memory Representation Capabilities for LLM Agents' to arXiv, applying privileged-prompt self-distillation to LLM agent memory.
- 2026Posted 'When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation' to arXiv, linking length inflation to EOS-token mismatch across 3 model families.
- 2026OpenAlex countsByYear records 6 works in 2026 — the only year present in the snapshot — with 14 distinct coauthors across the three-year window.
- 2026Evolve assessed both papers at medium reproducibility readiness, averaging 55 across the 2 reviews.
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; no ORCID is linked, so this identity is name-resolved from public bibliographic data and may mix homonymous authors named Xuchao Zhang.
- Output and impact axes are unscored: h-index and i10-index read as 0 with insufficient citation coverage in the snapshot.
- countsByYear contains only 2026 (6 works), so multi-year productivity trends cannot be assessed.
- Qualitative findings and verdicts reflect only the 2 Evolve-reviewed 2026 arXiv preprints, both with middle-position, non-corresponding authorship.