RL for LLMs at NTU Singapore: senior author on test-time RL, tool-use agents, and data-efficient RLVR for code
4
papers here
4
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
No output score: the snapshot offers no reliable career publication record — only recent3yWorks = 11 (countsByYear 2026 = 11) under the OpenAlex author record. This axis would measure cumulative and per-year output, venue mix, and authorship position once disambiguation improves.
Impact—
No impact score: citation counts, h-index and i10-index are absent or zeroed in this OpenAlex snapshot, so influence cannot be assessed. This axis would measure citations per paper, h-index trajectory, and external uptake of the four 2026 preprints.
Focus0/100
Focus scores 0 because work is spread perfectly evenly across four Evolve fields — LLM agents, RLVR post-training systems, LLM reinforcement learning, and LLMs for code — each at 25%. This is breadth rather than drift: all four clusters share one agenda of applying RL to LLM post-training and deployment, though OpenAlex topic labels (robotics-led) do not match and may reflect namesake records.
Collaboration68/100
Collaboration scores 68, from 30 distinct coauthors across 11 recent works. Twelve recur in the Evolve catalog network — including Zhiyi Lyu, Longtao Zheng, Shengtian Yang, Lang Feng, Lei Feng, Qingpeng Cai and Kun Gai — and the author sits in the last position on three of four catalog papers, a senior/supervisory pattern typical of a group lead rather than a lone-author style.
Activity72/100
Activity scores 72: recent3yWorks = 11 and countsByYear records 11 works in 2026, matching the four arXiv preprints in the Evolve catalog, all stamped 2026. Output is concentrated in a single recent burst rather than a long visible tail — consistent with an actively producing group whose earlier record sits outside this snapshot's coverage.
Reproducibility55/100
Reproducibility scores 55, the mean readiness across 4 reviewed papers (reproMean 55, reproPapers 4); all four are rated medium. Evolve verdicts repeatedly flag the same gap — strong ablations and real gains, but no variance reporting or significance tests — and one framework paper leans on self-citations for key methods, limiting independent reimplementation confidence.
Research summary
Bo An is a researcher at Nanyang Technological University, Singapore, whose visible recent work centers on reinforcement learning for large language models. The Evolve catalog holds four 2026 arXiv preprints spanning four evenly weighted subfields — LLM agents (25%), LLM post-training/RLVR systems (25%), LLM reinforcement learning (25%), and LLMs for code (25%) — with last-author position on three of the four, a senior-authorship pattern consistent with group leadership. The papers form one coherent agenda: making RL-based post-training and deployment of code-focused LLMs more reliable and data-efficient. Entropy-Regularized Rank-Masked Policy Optimization tackles test-time RL for code generation with consensus rewards that resist reward hacking; AgentBrew learns offline tool-use agents from raw real-world MCP trajectories; a two-stage RL method generates sound, adversarial tests; and DE-Venus is a verl-based framework for data-efficient RLVR. OpenAlex records 11 works in the recent three-year window, 30 distinct coauthors, a collaboration score of 68 and an activity score of 72, while a focus score of 0 reflects the perfectly even four-way field split. Citation-based output and impact scores are missing: the snapshot's h- and i10-index fields read zero, most plausibly an author-disambiguation artifact rather than a measured result. Evolve rates reproducibility at 55/100, with all four papers marked medium readiness and verdicts consistently noting absent variance and significance reporting.
Collaboration pattern
The record shows 30 distinct coauthors across 11 recent works, with 12 recurring collaborators including Zhiyi Lyu, Yewen Li, Longtao Zheng, Shengtian Yang, Lang Feng, Lei Feng, Peng Jiang, Kun Gai, Qingpeng Cai, Shenzhi Yang, Guangcheng Zhu and Kai Tang. On the four catalog papers the author takes the last position three times and a middle position once, with no corresponding-author flags — a supervisory senior pattern exercised through shared group authorship rather than named correspondence.
Research keywords
Highlights
Entropy-regularized, rank-masked test-time RL for code generation
AgentBrew: offline tool-use agents from real-world trajectories
Two-stage RL for sound, adversarial test generation
DE-Venus: a data-efficient RLVR framework
Timeline
- 2026Last-author preprint introduces entropy-regularized, rank-masked policy optimization for test-time RL in code generation.
- 2026AgentBrew preprint trains offline tool-use agents from raw real-world MCP trajectories.
- 2026Two-stage RL preprint targets sound and adversarial test generation in code LLMs.
- 2026DE-Venus paper contributes a unified, verl-based, data-efficient RLVR framework.
- 2026OpenAlex logs 11 works in the recent three-year window with 30 distinct coauthors, all indexed in 2026.
Papers in the Evolve catalog
Entropy-Regularized Rank-Masked Policy Optimization for Test-Time Reinforcement Learning in Code Generation
AgentBrew: Offline Tool-Use Agent Learning from Raw Real-World Trajectories
Two-Stage Reinforcement Learning for Sound and Adversarial Test Generation in Code LLMs
DE-Venus: A Data-Efficient RLVR Framework for Large Language Models
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.
- Bibliometrics come from an OpenAlex snapshot only (identity: openalex); its h-index and i10-index fields read zero, most plausibly reflecting incomplete author disambiguation rather than measured zero impact.
- 'Bo An' is a common name: this profile is name-resolved from public bibliographic data, and OpenAlex topic labels (e.g., reinforcement learning in robotics) conflict with the reviewed LLM papers, so namesake conflation cannot be excluded.
- All catalog evidence is four 2026 arXiv preprints with no peer-review venue information and no citation counts, which is why the output and impact dimensions are unscored.
- The reproducibility score of 55 is computed from only four reviewed papers, all rated medium readiness, with Evolve verdicts consistently noting missing variance and significance reporting.