Token-level credit assignment and rollout efficiency in RL for LLM post-training of reasoning models
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-08) 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—
Publication-volume data for this axis is not in the current snapshot, which covers only the 2 Evolve-reviewed papers. The axis would measure total output, annual publication counts, and authorship positions across the full bibliographic record; no score is estimated here.
Impact—
Citation counts, h-index, and downstream adoption are absent from the provided data, so impact is unscored. This axis would measure citations to the two reviewed papers and their uptake in reinforcement-learning post-training practice; nothing is inferred or estimated here.
Focus0/100
Evolve's focus score is 0, computed from a field distribution split exactly 50/50 between reinforcement learning for LLM post-training and machine learning systems for LLM training. Against the two reviewed 2026 papers, this even split reads as deliberate dual coverage rather than scattered interests: PACT attacks the algorithmic question of token credit, while the rollout-efficiency taxonomy addresses the systems cost of the same training loop. Both papers target one pipeline stage: post-training reasoning LLMs.
Collaboration58/100
Evolve scores collaboration at 58, derived from 17 distinct coauthors, of whom 12 recur frequently. The roster splits into an RL/post-training cluster including Jiayan Fu, Hang Xu, Zhaokai Luo, Yao Hu, and Dongyan Zhao, and a systems cluster including Marcos Assuncao, Rajkumar Buyya, Niloofar Gholipour, and Julien Gascon-Samson, matching the 50/50 field split. Both reviewed papers place the researcher in a middle position without corresponding-author role, consistent with contributing inside larger teams.
Activity—
Only two 2026 arXiv entries are available, so recent-activity signals such as publication cadence across prior years are insufficient. This axis would measure yearly output trends and sustained research engagement over a multi-year window; it cannot be measured from the current data and remains unscored.
Reproducibility38/100
Reproducibility averages 38 across the 2 papers Evolve has assessed. The rollout-efficiency survey rates low, since a taxonomy inherently offers no code or experiments to rerun, while PACT rates medium: Evolve's verdict calls its reported gains strong but flags single-run evaluations as limiting confidence. The 38 therefore reflects one partially replicable methods paper plus one survey, not failed reproductions; multi-seed reporting on PACT would be the clearest way to lift this axis.
Research summary
Yong Zhang works at the junction of two halves of the same problem: making reinforcement learning for large language model post-training both algorithmically principled and computationally efficient. The Evolve catalog holds two 2026 arXiv papers with the researcher in middle authorship. The first, PACT: From Credit Assignment to Critic Alignment, develops token-level credit assignment within actor-critic reinforcement learning, proving axioms that fix token credit as value martingale differences in the PPO/GAE and GRPO/RLOO tradition. The second is a survey organizing 80 rollout-efficiency methods for reasoning-model RL into a dual taxonomy with an explicit gap analysis and a critique of reporting practice. The two entries split the field distribution evenly at 50/50 between reinforcement learning for LLM post-training and machine learning systems for LLM training; the focus score of 0 reflects that even split rather than scattered interests. Evolve scores collaboration at 58 across 17 distinct coauthors and reproducibility at 38 over the two reviewed papers, with PACT rated medium and the survey low for reproduction readiness. Output, impact, and activity metrics are not yet available in this snapshot.
Collaboration pattern
Seventeen distinct coauthors appear in the record, twelve of them recurring. They divide into an RL/post-training group (Jiayan Fu, Hang Xu, Zhaokai Luo, Yao Hu, Mu Chuan, Dongyan Zhao) and a systems/cloud group (Niloofar Gholipour, Marcos Assuncao, Gursimran Singh, Timothy Yu, Rajkumar Buyya, Julien Gascon-Samson). This two-cluster structure mirrors the 50/50 field split, and both reviewed papers list the researcher as a middle author without corresponding role, indicating a contributor inside larger teams rather than a solo lead.
Research keywords
Highlights
Three axioms uniquely fix token credit as value martingale differences
A dual taxonomy of 80 rollout-efficiency methods
Algorithm and systems coverage of one training loop
The two reviewed papers complement each other: PACT improves what the critic learns at token level, while the taxonomy lowers the cost of generating rollouts. Together they span the full post-training loop, which is why the field distribution splits evenly at 50/50 between the two ML subfields.
Timeline
- 2026PACT: From Credit Assignment to Critic Alignment appears on arXiv, a middle-authorship contribution on token-level credit assignment for LLM post-training.
- 2026Rollout Efficiency in Reinforcement Learning for Reasoning LLMs: A Taxonomy and Future Directions appears on arXiv, surveying 80 methods in the machine learning systems (LLM training) field.
- 2026Both Evolve-cataloged papers fall in the same year, a concentrated recent burst at the algorithm/systems boundary of LLM post-training; the provided data covers no earlier years.
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.
- Bibliometric signals come from Evolve's OpenAlex-based snapshot; no citation counts, h-index, or publication-volume data was provided, so output, impact, and activity are unscored.
- 'Yong Zhang' is a common name; this profile is name-resolved from public bibliographic data, and institution, country, and ORCID are unrecorded.
- Scope covers only the 2 papers reviewed by Evolve; the researcher's full publication record is not represented.
- Both cataloged entries are 2026 arXiv preprints; peer-review status and venue acceptance are not captured.