Jiexiong Liu

—

Unclaimed · AI-structured profileORCID OpenAlex

Researcher across LLM post-training (RL) and autonomous agent orchestration, with two 2026 arXiv preprints.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) 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 volume cannot be scored yet: the catalog holds only two 2026 arXiv preprints, and no publication counts, venue history, or yearly output totals are available. This axis would measure total papers and sustained publication rate.

Impact—

Impact data is insufficient to score: no citation counts, h-index, or downstream adoption metrics are provided for either preprint. This axis would measure citations, influence on follow-up work, and recognition within the field.

Focus0/100

Focus scores 0 because the profile is split exactly evenly: the field distribution assigns 50% to machine learning — LLM post-training and reinforcement learning, and 50% to AI systems engineering (autonomous agent orchestration). With only two reviewed papers, one in each area, no concentrated core topic has emerged yet, though both strands connect through coding-agent training and agent pipeline execution.

Collaboration32/100

Collaboration scores 32 on a 0-100 scale, based on a small pool of 4 distinct coauthors: Cheng Li, Yixuan Chen, Chi Hong, and Ziheng Huang. Both reviewed papers list Jiexiong Liu in a middle author position without corresponding-author status, indicating contribution within a compact, stable team rather than an independent or large-network collaboration pattern.

Activity—

Activity cannot be scored yet: only two preprints, both dated 2026, are in scope, with no year-by-year publication record to show cadence or continuity. This axis would measure publication frequency, recency, and sustained research engagement over time.

Reproducibility20/100

Reproducibility scores 20, the mean readiness rating across 2 reviewed papers, both tagged low. The incremental-execution paper reports n=1 case studies with no code release and misaligned citations, while the C-DPPO post-training paper's headline +3.0 gain rests on single runs below the noise floor. Neither preprint currently provides the artifacts a third party would need to replicate results.

Research summary

Jiexiong Liu is an early-stage researcher whose public footprint centers on two 2026 arXiv preprints split evenly across two fields: machine learning (LLM post-training and reinforcement learning) and AI systems engineering for autonomous agent orchestration. The first, 'Train What You Deploy,' couples post-training to the exact deployed production coding agent at the token level and layers C-DPPO, a trust-region certification theory built on total variation divergence, on top; Evolve's review judges the theory sound while cautioning that the headline +3.0 gain rests on single runs below the noise floor. The second, 'Incremental Consistency Execution for Autonomous Intelligent Systems,' targets agent pipelines, using perturbation-built invariant domains, task fact contracts, and field-level dependency masks to confine re-execution to what actually changed; reviewers call the design promising but note n=1 case studies, no code release, and misaligned citations. The two strands share a common instinct — eliminating redundant computation while keeping deployed behavior faithful — though neither preprint yet provides replication-ready artifacts, driving a reproducibility score of 20 across both reviewed papers. Collaboration is compact: 4 distinct coauthors (Cheng Li, Yixuan Chen, Chi Hong, Ziheng Huang), with Liu in middle author positions on both papers and no corresponding-author role. With no institution, citation, or publication-count data available, output, impact, and activity cannot be scored; this remains a name-resolved, preprint-stage profile.

Collaboration pattern

Jiexiong Liu works within a small, stable collaboration pool of 4 distinct coauthors — Cheng Li, Yixuan Chen, Chi Hong, and Ziheng Huang — yielding a collaboration score of 32. On both reviewed 2026 preprints Liu appears in a middle author position and is not the corresponding author, consistent with contributing specialist work inside a compact team rather than leading standalone efforts or maintaining a broad external network.

Research keywords

incremental executioninvariant domaintask fact contractfield-level dependency maskLLM agent orchestrationchange propagationreinforcement learningLLM post-trainingcoding agentstoken fidelitytrust regiontotal variation divergence

Highlights

Invariant-domain incremental re-execution for LLM agent pipelines
The 2026 preprint 'Incremental Consistency Execution for Autonomous Intelligent Systems' builds perturbation-derived invariant domains, task fact contracts, and field-level dependency masks so agent pipelines re-execute only what changed. Evolve's review calls the design promising but notes n=1 case studies, no code release, and misaligned citations.
C-DPPO certification for token-faithful post-training
'Train What You Deploy' (2026) couples agent training with the deployed production coding model at token level and adds C-DPPO certification theory built on trust-region control of total variation divergence. Evolve finds the theory sound but flags that the headline +3.0 gain comes from single runs below the noise floor.
A shared avoid-redundant-work theme across two strands
Both 2026 preprints attack wasted computation from different angles: one confines agent pipeline re-execution to changed segments, the other aligns post-training tokens with what is actually deployed. Both score low on reproducibility readiness, making artifact release and multi-run evaluation the clearest next steps.

Timeline

2026Publishes 'Train What You Deploy', introducing token-faithful post-training of a production coding agent with C-DPPO certification theory.
2026Publishes 'Incremental Consistency Execution for Autonomous Intelligent Systems', proposing perturbation-built invariant domains for agent pipeline re-execution.
2026Both preprints enter the Evolve catalog, establishing an early footprint split evenly between LLM post-training and autonomous agent systems engineering.

Papers in the Evolve catalog

Frequent co-authors

Cheng Li×2 Yixuan Chen×2 Chi Hong×1 Ziheng Huang×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.

Radar scores and raw counts come from a limited bibliometric snapshot; no citation counts, h-index, or publication-volume data were available.
Both papers in scope are 2026 arXiv preprints; no peer-reviewed venue publications are included in this profile.
Institution and country are unavailable; this profile is name-resolved from public bibliographic data via ORCID 0009-0009-7431-505X.
The reproducibility score averages only 2 reviewed papers; Evolve verdicts reflect review-time assessments and may change as code or multi-run results appear.