Jingliang Duan

Vision pour la Robotique

Unclaimed · AI-structured profileORCID OpenAlex

LLM post-training researcher: continual-learning rectification and temporal credit assignment for stable fine-tuning

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—

No publication-volume data is available in this snapshot, so output is left unscored. This axis would measure the count and type of indexed publications — preprints and peer-reviewed papers — across the researcher's career.

Impact—

Citation and influence indicators are absent from the provided data, so impact is unscored. This axis would measure citation counts and aggregate influence, including uptake of the two catalogued 2026 preprints.

Focus0/100

The focus dimension scores 0, with the raw field distribution split exactly 50/50 between two adjacent subfields: LLM post-training and LLM fine-tuning with continual learning. Substantively, both catalogued papers target one shared problem — mitigating training pathologies in LLM fine-tuning — so the zero score reads as the classifier splitting a narrow research band across two labels rather than genuinely scattered interests.

Collaboration58/100

The collaboration dimension scores 58, based on 17 distinct coauthors in the underlying bibliometric record. Named frequent collaborators include Letian Tao, Wei Xiong, Kehua Sheng, Bo Zhang, Yang Guan, and Shengbo Eben Li, among a dozen-plus repeat partners. On both catalogued 2026 papers Duan sits in a middle author position without corresponding-author status, indicating contribution within larger teams rather than independent leadership.

Activity—

Recent publication cadence cannot be computed from a snapshot holding only two 2026 records and no year-by-year counts. This axis would measure publishing rate and continuity over recent years.

Reproducibility55/100

Reproducibility scores 55, a mean computed over exactly two scored papers. Both carry medium reproducibility readiness: the JANUS rectification review flags narrow test scope and replay–evaluation overlap, while the γOPD review notes consistent single-run gains reported without seeds or confidence intervals. Method-level theory is judged sound in both cases, but empirical reporting falls short of replication-grade standards.

Research summary

Jingliang Duan (ORCID 0000-0002-3697-1576) is affiliated in this snapshot with Vision pour la Robotique and works at the intersection of LLM post-training and LLM fine-tuning with continual learning. The Evolve catalog holds two 2026 arXiv papers on which Duan is a middle author and not the corresponding author. The first, 'Past, Future, All at Once,' proposes post-hoc JANUS rectification — a Jacobian null-space projection intended to suppress catastrophic forgetting during fine-tuning. Evolve's review judges the first-order theory reasonable but flags narrow test coverage and overlap between the replay and evaluation sets. The second, 'Beyond Token-Local Imitation,' introduces γOPD, adding discounted temporal credit assignment and bounded reward mixing to on-policy distillation; the analysis is sound but elementary, and consistent single-run gains are reported without seeds or confidence intervals. Both papers carry medium reproducibility readiness, and the reproducibility mean across two scored papers is 55. The field distribution splits evenly, 50/50, between the two subfields, and 17 distinct coauthors appear in the collaboration graph (dimension score 58). Output, impact, and activity metrics are not yet available, so this should be read as an early-stage, method-focused profile.

Collaboration pattern

Duan's collaboration footprint spans 17 distinct coauthors (dimension score 58), with a stable repeat core: Letian Tao, Wei Xiong, Kehua Sheng, Bo Zhang, Yang Guan, Shengbo Eben Li, Shiqi Liu, Zeyu He, Guojian Zhan, Jiaxin Gao, and Feihong Zhang. On both catalogued 2026 papers Duan contributes from a middle author position without corresponding-author status, consistent with specialized method work inside larger LLM-training teams.

Research keywords

catastrophic forgettingparameter-efficient fine-tuningLoRAJacobian null spacestability-plasticity dilemmapost-hoc weight rectificationon-policy distillationtemporal credit assignmentreverse KL divergencereinforcement learning with verifiable rewardsLLM post-trainingvariance bound

Highlights

JANUS: post-hoc Jacobian null-space rectification against catastrophic forgetting
A 2026 arXiv paper proposes rectifying fine-tuned weights after training via Jacobian null-space projection to mitigate the stability–plasticity dilemma in LLM fine-tuning. Evolve's verdict: the first-order theory is reasonable, but the test scope is narrow and the replay set overlaps the evaluation set. Reproducibility readiness: medium.
γOPD: temporal credit assignment for reward-compatible on-policy distillation
The second 2026 arXiv paper extends on-policy distillation with discounted temporal credit and bounded reward mixing, analyzed through reverse KL divergence and a variance bound in the RL-with-verifiable-rewards setting. Evolve calls the theory sound but elementary, with consistent single-run gains and no seeds or confidence intervals. Reproducibility readiness: medium.
A shared agenda: suppressing training pathologies in LLM fine-tuning
Across both reviewed papers the recurring move is post-processing or reweighting an existing training signal — null-space rectification of updated weights, discounted credit in distillation — rather than new architectures. This gives the profile a coherent methodological signature, split 50/50 between LLM post-training and LLM fine-tuning with continual learning.

Timeline

2026Posts 'Past, Future, All at Once: Mitigating Stability-Plasticity Dilemma via Post-hoc JANUS Rectification' to arXiv (LLM fine-tuning and continual learning), as a middle author.
2026Posts 'Beyond Token-Local Imitation: Reward-Compatible Temporal Credit Assignment for On-Policy Distillation' to arXiv (LLM post-training), also as a middle author.
2026Both 2026 papers enter the Evolve catalog with medium reproducibility readiness, yielding a reproducibility mean of 55 over two scored papers.

Papers in the Evolve catalog

Frequent co-authors

Letian Tao×2 Wei Xiong×2 Kehua Sheng×2 Bo Zhang×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 dimension scores are computed from public bibliographic data only; citation counts, h-index, and publication-volume metrics are absent from this snapshot.
The Evolve catalog holds only two reviewed papers (both 2026 arXiv preprints, middle-author, non-corresponding); the researcher's full publication record is not covered here.
Identity is name-resolved from public bibliographic records tied to ORCID 0000-0002-3697-1576; no biographical, award, or career facts are asserted.
Affiliation metadata is limited to the listed affiliation 'Vision pour la Robotique'; country and career history are missing from the input data.