Xiao Zhou

Research profile

Research summary

Xiao Zhou is a researcher affiliated with Shanghai Jiao Tong University whose reviewed work spans two halves of the NLP landscape: AI evaluation and scientometrics on one side, AI safety on the other — the bibliometric snapshot splits the field distribution exactly 50/50 between them. The Evolve catalog contains two 2026 arXiv papers. The first, BackTrend, lists Xiao Zhou as first author: it evaluates scientific weak-signal prediction through backward reconstruction, using LLM-as-judge, deep research agents, and RAG, and Evolve validated it as a retrospective benchmark while cautioning that low system scores — F1 at or below 10.1% — may partly reflect strict matching and construction bias. The second, on the efficiency–safety dilemma in large reasoning models, lists them in a middle position; it argues that compression's apparent safety gains stem from degraded reasoning rather than alignment — a study Evolve found credible but statistically thin. The collaboration dimension scores 53 with 13 distinct coauthors recorded, and reproducibility scores 55 as the mean across two scored papers, both rated medium in reproducibility readiness. Because 'Xiao Zhou' is a common name, this profile is name-resolved from public bibliographic data. Output, impact, and activity dimensions lack underlying citation and frequency data and are left unscored, so treat this as an early-stage profile on those axes.

Research keywords

Highlights

• BackTrend: a validated benchmark for scientific weak-signal prediction

  As first author, Xiao Zhou proposes evaluating scientific weak-signal prediction via backward reconstruction, with a toolkit spanning LLM-as-judge, deep research agents, and RAG. Evolve validated the retrospective benchmark design but cautions that low system scores — F1 at or below 10.1% — may partly reflect strict matching and construction bias; reproducibility readiness is rated medium.

• Compression's safety gains in large reasoning models are not alignment

  In a middle author position on this 2026 arXiv study, Xiao Zhou contributes to evidence that apparent safety improvements from quantization, pruning, and KV cache compression in large reasoning models stem from degraded reasoning rather than alignment. Evolve judges it a credible empirical study while flagging thin statistics and inconsistencies; reproducibility readiness is rated medium.

• A methodological bridge between evaluation and safety research

  The two reviewed 2026 papers share an evaluation-first mindset: BackTrend turns scientific foresight into a measurable benchmark, while the efficiency–safety study treats safety claims as something to be measured against reasoning quality. Together they sketch a profile oriented toward rigorous measurement of AI systems, spanning scientometrics and model-internal analysis.

Timeline

1. 2026Publishes BackTrend on arXiv as first author, introducing backward-reconstruction evaluation of scientific weak-signal prediction with LLM-as-judge, deep research agents, and RAG in its toolkit.
2. 2026Co-authors (middle position) the arXiv study on the efficiency–safety dilemma in large reasoning models, covering model compression via quantization, pruning, and KV cache compression.
3. 2026Bibliometric snapshot records the year's reviewed output split evenly (50/50) between AI evaluation/scientometrics (NLP) and AI safety/NLP, with 13 distinct coauthors and a reproducibility mean of 55 across two scored papers.

Papers in the Evolve catalog

• 2026-09-21first author

  BackTrend: Evaluating Scientific Weak-Signal Prediction via Backward Reconstruction

• 2026-09-20

  On the Efficiency-Safety Dilemma in Large Reasoning Models

Frequent co-authors