Letian Tao

Vision pour la Robotique

Unclaimed · AI-structured profileOpenAlex

LLM post-training researcher: continual-learning rectification and credit assignment for on-policy 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—

Insufficient data: total publication count, venue list, and yearly output volumes are not available in the current snapshot. This axis would measure the scale and consistency of the researcher's publication record across the full career, and is left unscored here.

Impact—

No citation counts, h-index, or downstream adoption signals are available yet. This axis would measure how often the work is cited and taken up by other researchers and its influence on the field, and cannot be scored from the current data.

Focus0/100

The focus axis scores 0/100, reflecting an even field split: the recorded distribution assigns 0.5 to machine learning (LLM post-training) and 0.5 to machine learning (LLM fine-tuning and continual learning). In practice the two cataloged papers are thematically adjacent — both target post-training of large language models — so the low score captures formal breadth across two labeled fields rather than genuinely scattered research interests.

Collaboration58/100

Collaboration scores 58/100, with 17 distinct co-authors recorded in the bibliometric snapshot. The recurring partner list spans Jingliang Duan, Wei Xiong, Kehua Sheng, Bo Zhang, Yang Guan, Shengbo Eben Li, Shiqi Liu, Zeyu He, Guojian Zhan, Jiaxin Gao, and Feihong Zhang, suggesting stable work within a mid-sized network. On both cataloged 2026 papers the author sits in a middle position rather than as leading or corresponding author.

Activity—

Recent publication cadence and year-by-year activity levels are not available in the snapshot. This axis would measure output tempo over time — for example papers per year and gaps in the record — and is therefore left unscored rather than estimated.

Reproducibility55/100

Reproducibility averages 55/100 across the 2 reviewed papers (reproPapers = 2, reproMean = 55), both rated medium readiness. The review of the JANUS rectification paper notes a reasonable first-order theory but narrow test scope and replay/evaluation-set overlap; the γOPD review reports consistent but single-run gains without seeds or confidence intervals — concrete, fixable gaps for future work.

Research summary

Letian Tao, affiliated with Vision pour la Robotique, works on the post-training side of large language models. The recorded field distribution splits evenly between machine learning (LLM post-training) and machine learning (LLM fine-tuning and continual learning), at 0.5 each. Two 2026 arXiv preprints are reviewed in the Evolve catalog, in both cases with the author in a middle position. The first, 'Past, Future, All at Once', proposes post-hoc JANUS rectification, using Jacobian null-space projection to mitigate the stability-plasticity dilemma in LLM fine-tuning; Evolve's review finds the first-order theory reasonable but flags a narrow test scope and overlap between the replay and evaluation sets. The second, 'Beyond Token-Local Imitation', introduces γOPD, which adds discounted temporal credit assignment and bounded reward mixing to on-policy distillation; the review calls the theory sound but elementary and notes consistent single-run gains without seeds or confidence intervals. The collaboration footprint covers 17 distinct co-authors (score 58/100), including Jingliang Duan, Wei Xiong, Kehua Sheng, Bo Zhang, Yang Guan, and Shengbo Eben Li. Reproducibility averages 55/100 over the two reviewed papers, both rated medium readiness. Because output, impact, and activity metrics are unavailable, this remains an early-stage profile grounded in a limited bibliometric snapshot.

Collaboration pattern

Collaboration scores 58/100 with 17 distinct co-authors in the snapshot. Recurring partners include Jingliang Duan, Wei Xiong, Kehua Sheng, Bo Zhang, Yang Guan, Shengbo Eben Li, Shiqi Liu, Zeyu He, Guojian Zhan, Jiaxin Gao, and Feihong Zhang, indicating a stable mid-sized network. Both cataloged 2026 papers place the author in a middle position without corresponding-author status, consistent with contributing within a larger team rather than leading small-group efforts.

Research keywords

LLM post-trainingon-policy distillationtemporal credit assignmentreverse KL divergencereinforcement learning with verifiable rewardscatastrophic forgettingcontinual learningparameter-efficient fine-tuningLoRAJacobian null spacestability-plasticity dilemmapost-hoc weight rectification

Highlights

Post-hoc JANUS rectification targets forgetting in LLM fine-tuning
The 2026 preprint 'Past, Future, All at Once' proposes post-hoc Jacobian null-space projection to mitigate the stability-plasticity dilemma in fine-tuned LLMs. Evolve's review finds the first-order theory reasonable but flags a narrow test scope and overlap between the replay and evaluation sets; reproducibility readiness is medium.
γOPD brings temporal credit assignment to on-policy distillation
The 2026 preprint 'Beyond Token-Local Imitation' adds discounted temporal credit and bounded reward mixing to on-policy distillation for LLM post-training. The review judges the theory sound but elementary and reports consistent single-run gains without seeds or confidence intervals; reproducibility readiness is medium.
A coherent post-training profile built on two medium-readiness reviews
Across both cataloged papers the author works in a middle position on LLM post-training questions, with a reproducibility mean of 55/100 and a 50/50 field split between LLM post-training and LLM fine-tuning with continual learning. Reviewer concerns cluster on evaluation rigor — test scope, replay overlap, and missing seeds — a consistent theme across the two studies.

Timeline

2026Publishes 'Past, Future, All at Once' on arXiv (middle author position), proposing post-hoc JANUS rectification via Jacobian null-space projection for the stability-plasticity dilemma in LLM fine-tuning.
2026Publishes 'Beyond Token-Local Imitation' on arXiv (middle author position), introducing γOPD with reward-compatible temporal credit assignment for on-policy distillation in LLM post-training.
2026Evolve coverage consolidates around the two 2026 preprints: 17 distinct co-authors recorded, reproducibility averaged 55/100, and the field split evenly between LLM post-training and LLM fine-tuning with continual learning.

Papers in the Evolve catalog

Frequent co-authors

Jingliang Duan×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 scores come from a public bibliometric snapshot (OpenAlex-style metadata) only; the output, impact, and activity axes lack sufficient data and are left unscored.
Both cataloged papers are 2026 arXiv preprints; peer-review status and final publication venues are not recorded in this data.
No ORCID or country is provided; the profile is name-resolved from public bibliographic data, and 'Letian Tao' may correspond to more than one researcher.
The reproducibility mean (55) rests on only two reviewed papers, with the author in a middle position on both; the frequent-coauthor list also contains a duplicated entry (Bo Zhang).