Jianshen Zhang

Research profile

Research summary

Jianshen Zhang is a machine learning researcher at Jingdong (JD), China, whose OpenAlex-linked record shows a pronounced acceleration: 28 works in the last three years, climbing from 8 in 2025 to 17 recorded in 2026 — the basis for an activity score of 98. The profile spans three roughly equal strands (about a third of the field distribution each): probabilistic forecasting and model-based sequential decision making, reinforcement learning for LLM agents, and AI safety. Topic counts are led by supply chain and inventory management (20) and forecasting techniques (12), reflecting an industrial operations-research base now extended toward agents and safety. Evolve has reviewed three 2026 arXiv preprints: a Hoeffding-gated, loss-conditioned rule deciding when a world model should move its state; GACA, a granularity-adaptive credit-assignment method mixing step- and episode-level advantages with reported gains on ALFWorld/WebShop; and a harness-agnostic black-box approach to detecting and immunizing reward hacking in self-evolving LLMs, validated on one prompt-level host with a 0.434 false-positive rate. A fourth paper, Hindsight Memory-PRM, is cataloged without a verdict yet. Collaboration is broad — 35 distinct coauthors — and reproducibility across the three reviewed papers averages 70. Output and impact scores are not yet computable; the record reports an h-index of 3 (i10-index 2).

Research keywords

Highlights

• Loss-conditioned state execution for world models

  Proposes a Hoeffding-gated rule that moves a world model's state only where calibrated loss gain is certified, benchmarked against a persistence baseline. Evolve's verdict: a sound selective-prediction idea, though the certificate covers a clipped surrogate.

• Granularity-adaptive credit assignment for long-horizon LLM agents

  GACA mixes step- and episode-level advantages per step via policy NLL for long-horizon agent reinforcement learning, reporting clear gains on ALFWorld and WebShop. Evolve notes the headline baselines are quoted rather than re-run; reproducibility readiness is rated high.

• Harness-agnostic detection and immunization of reward hacking

  A black-box design using a held-out probe bank and change-point detection to detect and immunize reward hacking (Goodhart's law) in self-evolving language models. Evolve finds the design promising but validation limited to one prompt-level host with a 0.434 false-positive rate.

• Hindsight Memory-PRM

  Supervising memory management with auditable hindsight credit. This paper is indexed in Evolve's catalog but carries no review verdict, keywords, or reproducibility assessment yet.

Timeline

1. 2007Earliest year with recorded output in the OpenAlex-linked record (1 work); early-record coverage is sparse and should be read cautiously.
2. 20224 works recorded — the highest annual count before the 2025–2026 acceleration.
3. 2025Output accelerates to 8 recorded works, beginning the current three-year run of 28 works.
4. 202617 works recorded year-to-date; all four cataloged papers date to 2026, including three Evolve-reviewed arXiv preprints spanning world models, LLM-agent RL, and AI safety.

Papers in the Evolve catalog

• 2026-09-14

  When Should a World Model Move? Loss-Conditioned State Execution

• 2026-09-11

  Granularity-Adaptive Credit Assignment for Long-Horizon LLM Agent Reinforcement Learning

• 2026-09-04

  Harness-agnostic detection and immunization of reward hacking in self-evolving language models

Frequent co-authors