Jing Zhang

Research profile

Research summary

Jing Zhang is an early-stage researcher at The University of Texas at Arlington whose cataloged work spans two complementary strands of machine learning. The first strand is empirical reinforcement learning: as first author of a 2026 arXiv paper, Zhang proposes RSIQL, a method that improves offline goal-conditioned RL by selectively adding reward at states of estimated progress, addressing sparse rewards and long-horizon credit assignment in an implicit Q-learning (IQL) setting; Evolve's review credits careful analysis and strong AntMaze gains while flagging stylized theory and the absence of a policy-invariance guarantee, with medium reproduction readiness. The second strand is analytical: a 2026 co-authored survey recasts complex problem solving in large language models as sequence estimation and process control, drawing on statistical decision theory, Bayes filters, and uncertainty quantification; Evolve finds it clear and candid but notes its diagnostic hypotheses remain untested, and rates reproduction readiness low. OpenAlex bibliometrics confirm the profile is in its early phase: h-index and i10-index are both recorded as 0, only 1 work is attributed in the last three years, and per-year counts show a single 2026 entry. Collaboration is the strongest signal, with 37 distinct coauthors and 12 recurring collaborators. Field distribution splits evenly between machine learning/statistics and reinforcement learning, and assigned topics include topic modeling, NLP techniques, and AI-based problem solving and planning. Because 'Jing Zhang' is a common name, this profile is name-resolved from public bibliographic data.

Research keywords

Highlights

• Selective reward stimulation lifts offline goal-conditioned RL

  As first author of a 2026 arXiv paper, Zhang introduces RSIQL, which adds reward at states of estimated progress in offline goal-conditioned RL. Evolve's review credits careful analysis and strong AntMaze gains, targeting sparse rewards and long-horizon credit assignment, while noting stylized theory and no policy-invariance guarantee; reproduction readiness is rated medium.

• Statistical control framework for LLM complex problem solving

  In a 2026 survey where Zhang holds a middle author position, complex problem solving in large language models is recast as sequence estimation and process control, building on statistical decision theory, Bayes filters, and uncertainty quantification. Evolve finds the survey clear and candid, but its diagnostic hypotheses remain untested and reproduction readiness is rated low.

• A method-plus-survey arc bridging RL and LLM analysis

  The two 2026 papers split evenly across reinforcement learning and machine learning/statistics (field distribution 0.5 each), pairing a concrete offline RL method with a diagnostic LLM survey; together with assigned topics such as topic modeling and AI-based problem solving and planning, they sketch a bridge between empirical control and the analysis of model behavior.

Timeline

1. 2026Released a first-authored arXiv paper in reinforcement learning, 'Improving Offline Goal-Conditioned Reinforcement Learning via Selective Reward Stimulation,' proposing RSIQL with strong AntMaze gains.
2. 2026Co-authored an arXiv survey in machine learning/statistics as a middle-position author, 'Complex Problem Solving in Large Language Models: A Statistical Control Survey and Diagnostic Framework.'
3. 2026OpenAlex per-year counts record the single attributed work in 2026, the only year with output in the recent three-year window.

Papers in the Evolve catalog

• 2026-09-18first author

  Improving Offline Goal-Conditioned Reinforcement Learning via Selective Reward Stimulation

• 2026-09-17

  Complex Problem Solving in Large Language Models: A Statistical Control Survey and Diagnostic Framework

Frequent co-authors