Zhiwei Liu — Salesforce researcher working on LLM-agent tool-use RL and diagnostic evaluation of time-series reasoning
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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—
Not yet scored. This snapshot lacks the publication counts this axis measures — total volume, cadence, and venue spread over time. Evolve therefore reports no estimate here rather than guessing from only two reviewed papers.
Impact—
Not yet scored. Citation counts, citing works, and downstream influence are absent from this snapshot, so the impact axis — which measures how widely a researcher's work is used and built upon — cannot be computed and is left unrated.
Focus0/100
The focus score is 0, reflecting a field distribution split exactly 50/50 between machine learning and natural language processing / LLM evaluation for time-series reasoning. The two reviewed papers sit on opposite sides of that split — reinforcement learning for multi-turn tool use versus LLM benchmarking for time-series — so the profile currently reads as dual-track rather than narrowly specialized. With only two papers sampled, this breadth measure is provisional.
Collaboration61/100
The collaboration score is 61, computed from 20 distinct coauthors in the available records. Twelve recurring collaborators are listed, including Zixiang Chen, Wenting Zhao, Jianguo Zhang, Shelby Heinecke, Silvio Savarese, Huan Wang, and Jie Gong. Both cataloged papers carry middle-position authorship with no corresponding role, consistent with team-based industrial research inside a larger group rather than solo-led projects.
Activity—
Not yet scored. The activity axis measures recent publishing cadence and year-over-year engagement, but this snapshot contains no yearly publication counts beyond the two 2026 preprints reviewed here, so Evolve reports no activity estimate.
Reproducibility55/100
The reproducibility score is 55, the mean readiness rating across the 2 papers Evolve has reviewed, both rated medium. That rating fits the evidence in the verdicts: Critical-State RL rests on four-cell BFCL experiments limited to one benchmark, while TimeLitmus is a solid diagnostic benchmark but relies on rule-based labels over a small pair set. Methods look describable, yet neither paper's evidence base is broad.
Research summary
Zhiwei Liu is a researcher at Salesforce (United States) whose Evolve profile is name-resolved from public bibliographic data. The current catalog contains two 2026 arXiv papers, both with middle-position authorship and no corresponding role, and both reviewed by Evolve. The first, Critical-State RL, proposes a nested-sampling diagnostic — built on nested Monte Carlo sampling and variance decomposition — that identifies which call in a multi-turn tool-use trajectory is trainable, supported by four-cell BFCL experiments; Evolve notes the evidence is confined to a single benchmark. The second, TimeLitmus, is a diagnostic benchmark for event-conditioned time-series prediction, showing that LLM accuracy on such tasks overstates genuine cross-modal understanding and probing explanation faithfulness and shortcut learning; Evolve flags rule-based labels and a small pair set as caveats. A common thread across both papers is a diagnostic stance: rather than only proposing new capability, each work builds instruments that locate where models actually fail. Radar scores give collaboration 61 (from 20 distinct coauthors and a 12-person recurring collaborator list) and reproducibility 55 (mean across 2 papers, both medium readiness), while focus scores 0, matching a 50/50 field split between machine learning and NLP/LLM evaluation for time-series reasoning. Output, impact, and activity lack underlying data and are unscored.
Collaboration pattern
With 20 distinct coauthors on record and a recurring circle of 12 — including Zixiang Chen, Wenting Zhao, Jianguo Zhang, Akshara Prabhakar, Jielin Qiu, Shelby Heinecke, Silvio Savarese, Huan Wang, and Jie Gong — Zhiwei Liu's visible work is team-produced. Both cataloged papers carry middle authorship without corresponding role, so the pattern reads as contributing specialist within larger industry lab groups; the collaboration dimension scores 61.
Research keywords
Highlights
A nested-sampling diagnostic for multi-turn tool RL
TimeLitmus shows accuracy overstates cross-modal skill
A shared diagnostic-first methodology
Across both reviewed papers the same instinct appears: build instruments that locate failure before training. Critical-State RL diagnoses trainable states; TimeLitmus diagnoses whether benchmark success reflects understanding — positioning Zhiwei Liu's visible 2026 output around measurement rather than pure capability claims.
Timeline
- 2026Critical-State RL posted to arXiv, introducing a nested-sampling diagnostic for trainable states in multi-turn tool use.
- 2026TimeLitmus posted to arXiv as a diagnostic benchmark for event-conditioned time-series prediction and explanation faithfulness.
- 2026Both papers appear in the same year, marking a diagnostic-oriented, two-track agenda across LLM-agent RL and time-series evaluation.
Papers in the Evolve catalog
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.
- Radar scores and raw metrics derive from public bibliometric data only; no citation counts, h-index, or venue-level statistics were provided.
- Zhiwei Liu is a common name; this profile is name-resolved from public bibliographic data and may merge or split identities.
- Only 2 papers reviewed by Evolve, both 2026 arXiv preprints with middle authorship; coverage of the researcher's full output is limited.
- Output, impact, and activity dimensions have insufficient underlying data and are unscored in this profile.