Bo Zhang

Chinese Academy of Sciences

Unclaimed · AI-structured profileOpenAlex

LLM efficiency researcher across speculative decoding, continual fine-tuning, and post-training distillation

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

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—

Not yet scored: the underlying productivity data (total publication count, venue record, and yearly output volume) is not available in this bibliometric snapshot, so publication volume and consistency over time cannot be assessed.

Impact—

Not yet scored: no citation counts, h-index, or referencing-record data exist in the input, so this axis — which would measure citation influence and uptake of the three reviewed papers — cannot be estimated.

Focus0/100

Scores 0 on focus because the reviewed output splits perfectly evenly across three subfields — efficient LLM inference, LLM fine-tuning and continual learning, and LLM post-training — each accounting for one third of the cataloged papers. The even three-way split signals breadth rather than a single deep specialty, although all three strands sit within the broader LLM training-and-inference efficiency space.

Collaboration61/100

Scores 60 on collaboration, computed from a co-author network of 19 distinct coauthors, 12 of whom recur as frequent collaborators — including Letian Tao, Jingliang Duan, Wei Xiong, and Shengbo Eben Li. On all three reviewed 2026 papers Bo Zhang appears as a middle-position, non-corresponding author, indicating contribution within larger teams rather than leading authorship. The network size is moderate, and sustained repeated partnerships lift the score.

Activity—

Not yet scored: the snapshot lacks publication-date coverage beyond the three 2026 arXiv entries, so recent publishing cadence and research tempo over time cannot be measured yet.

Reproducibility55/100

Scores 55, the mean repro-readiness rating across three assessed papers, each individually rated medium. Evolve's verdicts flag concrete caveats: thin baselines in the speculative-decoding work, a replay set overlapping the evaluation set in the JANUS study, and single-run results without seeds or confidence intervals for γOPD — fixable gaps that currently temper confidence in independent replication.

Research summary

Bo Zhang is an early-stage machine-learning researcher whose reviewed work centers on making large language models cheaper to run and more stable to train. The Evolve catalog holds three 2026 arXiv papers, one in each of three LLM-oriented subfields. RheoSampling adds a proxy-probability sampled token to EAGLE-3 dynamic draft trees for speculative decoding, with a losslessness proof. A post-hoc JANUS rectification applies Jacobian null-space projection to counter catastrophic forgetting in LLM fine-tuning. And γOPD introduces discounted temporal credit assignment with bounded reward mixing for on-policy distillation. Evolve's verdicts describe the theory across all three as sound but modest: gains are consistent yet incremental, baselines stay thin, test coverage is narrow, results are single-run without seeds or confidence intervals, and the JANUS replay set overlaps its evaluation set. Bibliometrically, output splits evenly — one third each across efficient inference, fine-tuning and continual learning, and post-training — which yields a focus score of 0 and marks breadth over depth. A collaboration score of 60 reflects 19 distinct coauthors, 12 of them recurring. All three catalog entries list Bo Zhang as a middle-position, non-corresponding author, consistent with a contributing role in larger teams. Output, impact, and activity scores are unavailable from current data.

Collaboration pattern

The co-author network spans 19 distinct collaborators, with 12 recurring frequently — including Letian Tao, Jingliang Duan, Wei Xiong, Kehua Sheng, Yang Guan, and Shengbo Eben Li. On all three reviewed 2026 papers Bo Zhang is a middle-position, non-corresponding author, suggesting a steady contributor role inside larger teams across inference, fine-tuning, and post-training projects. The collaboration score is 60.

Research keywords

speculative decodingLLM inference accelerationdynamic draft treesstochastic samplinglossless generationcatastrophic forgettingparameter-efficient fine-tuningJacobian null spaceon-policy distillationtemporal credit assignmentreverse KL divergenceLLM post-training

Highlights

Lossless stochastic sampling for EAGLE-3 speculative decoding
RheoSampling adds a proxy-probability sampled token to EAGLE-3's dynamic draft trees and pairs it with a losslessness proof, resolving the one-hot dilemma in stochastic dynamic-tree speculative decoding; Evolve finds the gains consistent but modest, with baselines staying thin.
Post-hoc Jacobian null-space projection against catastrophic forgetting
The JANUS rectification method claims to eliminate forgetting in LLM fine-tuning via post-hoc Jacobian null-space projection, targeting the stability-plasticity dilemma; Evolve judges the first-order theory reasonable but notes narrow test coverage and an overlap between the replay and evaluation sets.
Reward-compatible temporal credit assignment for on-policy distillation
γOPD extends on-policy distillation with discounted temporal credit assignment and bounded reward mixing, supported by a variance bound; Evolve calls the theory sound but elementary, observing consistent single-run gains without seeds or confidence intervals.

Timeline

2026Published RheoSampling on stochastic dynamic-tree speculative decoding (arXiv).
2026Published the JANUS post-hoc rectification work on the stability-plasticity dilemma in LLM fine-tuning (arXiv).
2026Published γOPD on reward-compatible temporal credit assignment for on-policy distillation (arXiv).

Papers in the Evolve catalog

Frequent co-authors

Letian Tao×2 Jingliang Duan×2 Wei Xiong×2 Kehua Sheng×2 Yang Guan×2 Shengbo Eben Li×2 Shiqi Liu×1 Zeyu He×1 Guojian Zhan×1 Jiaxin Gao×1 Feihong Zhang×1 Bo 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.

Bibliometric scores are computed from a public bibliometric snapshot; 'Bo Zhang' is a common name, so this profile is name-resolved from public bibliographic data and identity should be verified before use.
No institution, country, or ORCID is attached, so affiliation-level context is absent.
Output, impact, and activity dimensions lack underlying data (no citation counts or multi-year publication record) and are left unscored.
All three reviewed papers are 2026 arXiv preprints with middle-author positions; peer-review status and final publication venues are unknown.