Bin Chong

Peking University · China

Unclaimed · AI-structured profileOpenAlex

Peking University researcher in RL for LLM agents and AI safety — credit assignment and reward-hacking detection

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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 scoring needs full per-year publication counts, which this snapshot lacks beyond a single 2026 tally (17 works). The axis would measure cumulative productivity and its growth; no score is assigned.

Impact—

Impact scoring requires citation-based metrics (h-index, i10-index, citations), which are not yet populated for this OpenAlex identity (both indexes show 0). This axis would measure citation influence; no score is estimated.

Focus0/100

The focus score is 0, reflecting a wide rather than narrow footprint: the OpenAlex field split is an even 50/50 between reinforcement learning for LLM agents and machine learning / AI safety, and the five weighted topics — topic modeling (6), adversarial robustness (5), NLP techniques (4), security and verification (4) and AI planning (3) — spread across several sub-areas instead of one lane. This reads as breadth-oriented early work.

Collaboration69/100

The collaboration score is 69, supported by 31 distinct co-authors in the recent three-year window — a wide network for an early-stage record. Recurring names include Jianshen Zhang, Shang Luo, Yingguang Yang, Yulin Huang, Congjing Ran, Haoxuan Jia, Yongzhi Qi, Taoran Liang, Yang Liu, Rongrong Zhang and Yingzong Min. In the Evolve catalog he appears as a middle author on one paper and last author on two, none as corresponding author.

Activity84/100

The activity score is 84, the strongest axis in this profile: OpenAlex records 17 works in the recent three-year window, with the counts-by-year data placing all 17 in 2026. Three of those outputs are arXiv preprints already in the Evolve catalog, spanning LLM-agent RL, AI safety and agent memory — a dense, current publishing cadence.

Reproducibility78/100

The reproducibility score is 78, averaged over the 2 papers Evolve has assessed for reproducibility readiness. The GACA credit-assignment paper is rated high, though its review notes headline baselines were quoted rather than re-run; the reward-hacking detection paper is rated medium, with validation on a single prompt-level host and a reported 0.434 false positive rate as flagged limits.

Research summary

Bin Chong is a researcher at Peking University whose recorded work splits evenly — 50/50 in the OpenAlex field distribution — between reinforcement learning for LLM agents and machine learning / AI safety. Name-resolved from OpenAlex bibliographic data, the profile reflects an active early stage: 17 works in the recent three-year window, all tallied in 2026, alongside 31 distinct co-authors. Weighted topics point to topic modeling (6), adversarial robustness in machine learning (5), NLP techniques (4), security and verification in computing (4), and AI-based problem solving and planning (3). Three 2026 arXiv preprints appear in the Evolve catalog. Granularity-Adaptive Credit Assignment (middle author) mixes step- and episode-level advantages per step via policy NLL in the GRPO/GiGPO line, reporting clear gains on ALFWorld and WebShop, though headline baselines are quoted rather than re-run; Evolve rates its reproducibility high. Harness-agnostic detection and immunization of reward hacking (last author) proposes black-box monitoring of self-evolving language models using a held-out probe bank and change-point detection, validated so far on one prompt-level host with a 0.434 false positive rate. Hindsight Memory-PRM (last author) supervises agent memory management with auditable hindsight credit and awaits review. Citation metrics are not yet populated, so output and impact scores are pending.

Collaboration pattern

Bin Chong works in team-based settings: 31 distinct co-authors appear in the recent three-year OpenAlex window, and the catalog papers list him as a middle author once and last author twice, never as corresponding author. Recurring collaborators include Jianshen Zhang, Shang Luo, Yingguang Yang, Yulin Huang, Congjing Ran, Haoxuan Jia, Yongzhi Qi, Taoran Liang, Yang Liu, Rongrong Zhang and Yingzong Min. The collaboration score is 69, indicating a broad but still-forming network.

Research keywords

credit assignmentLLM agentsreinforcement learningGRPOadvantage estimationreward hackingGoodhart's lawself-evolving language modelsblack-box monitoringchange-point detectionmemory managementprocess reward models

Highlights

Granularity-adaptive credit assignment for long-horizon LLM agents
2026 arXiv preprint with Bin Chong as a middle author. GACA mixes step- and episode-level advantages per step via policy NLL, in the GRPO/GiGPO line, reporting clear gains on ALFWorld and WebShop. Evolve notes headline baselines are quoted, not re-run, yet rates reproducibility high.
Black-box detection and immunization of reward hacking
Last-authored 2026 preprint proposing harness-agnostic detection and immunization of reward hacking in self-evolving language models, built on a held-out probe bank and change-point detection. Evolve flags validation on a single prompt-level host with a 0.434 false positive rate; reproducibility is rated medium.
Hindsight Memory-PRM for agent memory management
Last-authored 2026 arXiv preprint on supervising memory management with auditable hindsight credit. It is not yet reviewed by Evolve, so no verdict, keywords or reproducibility rating is available; the slug entry is pending.

Timeline

2026OpenAlex records 17 works in the year, with 31 distinct co-authors across the recent three-year window.
2026Granularity-Adaptive Credit Assignment for Long-Horizon LLM Agent Reinforcement Learning posted to arXiv, with Bin Chong as a middle author.
2026Harness-agnostic detection and immunization of reward hacking in self-evolving language models posted to arXiv, with Bin Chong as last author.
2026Hindsight Memory-PRM: Supervising Memory Management with Auditable Hindsight Credit posted to arXiv, with Bin Chong as last author.

Papers in the Evolve catalog

Frequent co-authors

Jianshen Zhang×3 Shang Luo×2 Yingguang Yang×2 Yulin Huang×2 Congjing Ran×2 Haoxuan Jia×2 Yongzhi Qi×2 Taoran Liang×1 Yang Liu×1 Rongrong Zhang×1 Yingzong Min×1 Yongzhi Qi×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.

All bibliometric figures come from an OpenAlex snapshot; h-index and i10-index are unpopulated (0), so output and impact scores are missing.
All three catalog papers are 2026 arXiv preprints; no peer-reviewed venue information is present in the data.
No ORCID is provided; identity is name-resolved from OpenAlex, so homonym risk cannot be excluded.
Only two papers carry Evolve review verdicts; Hindsight Memory-PRM is unreviewed, with no keywords or reproducibility rating.