Chao Zhang

Research profile

Research summary

Chao Zhang appears in the Evolve catalog through two 2026 arXiv preprints that share a common stance: open, honestly reported LLM systems research. The first, Rufus-Air, an open LLM post-training recipe, lists Zhang as a middle author. It documents an eight-stage post-training pipeline applied to GLM-4.5-Air-Base that reportedly outperforms the official GLM-4.5-Air release; Evolve praises its unusual candor while cautioning that the stage-order rationale is untested. The second, Forecast-Dojo, lists Zhang as last author. It introduces replayable environments for benchmarking and training LLM forecasting agents, and reports a clean result: research tools helped all twelve evaluated models, yet every model still underperformed prediction markets — a finding Evolve considers credible and partly attributable to the system's architecture. Both papers carry favorable repro-readiness ratings (medium and high), reinforcing an infrastructure-and-evaluation orientation. The collaboration record behind these papers is substantial: 28 distinct coauthors and a collaboration score of 67/100, with twelve recurring collaborators including Rongzhi Zhang, Yuan He, Liqin Ye, Haorui Wang, and Sudheer Chava. No citation counts, venue history, affiliation, or cumulative output data are available in this snapshot, so the output, impact, focus, activity, and reproducibility dimensions remain deliberately unscored, and every claim here is grounded in the two reviewed papers alone.

Research keywords

Highlights

• Open eight-stage post-training recipe that beats an official release

  On Rufus-Air (2026, middle authorship), the team post-trains GLM-4.5-Air-Base through an open eight-stage recipe and reports results surpassing the official GLM-4.5-Air release; Evolve credits the work's unusual candor but notes the stage-order rationale is untested.

• Replayable environments expose the limits of LLM forecasting agents

  Forecast-Dojo (2026, last authorship) builds replayable environments for benchmarking and training LLM forecasting agents; across all twelve evaluated models, research tools helped, yet every model still trailed prediction markets — a verdict Evolve finds credible and partly architectural.

• A reproducibility-first working style

  Both reviewed papers carry positive repro-readiness ratings — high for Forecast-Dojo and medium for Rufus-Air — and both release open recipes, environments, and candid negative results, making open infrastructure the through-line of this early-stage record.

Timeline

1. 2026Co-authored (middle position) 'Rufus-Air: An Open LLM Post-Training Recipe', an eight-stage open recipe applied to GLM-4.5-Air-Base that reportedly beats the official GLM-4.5-Air release.
2. 2026Served as last author on 'Forecast-Dojo', replayable environments for benchmarking and training LLM forecasting agents, rated high for repro-readiness.
3. 2026Both papers entered the Evolve catalog, with verdicts praising candid reporting and credible head-to-head findings.
4. 2026The bibliometric record attributes 28 distinct coauthors to this name in the same period, yielding a collaboration score of 67/100.

Papers in the Evolve catalog

• 2026-09-24

  Rufus-Air: An Open LLM Post-Training Recipe

• 2026-09-24

  Forecast-Dojo: Replayable Environments for Benchmarking and Training LLM Forecasting Agents

Frequent co-authors