Xixin Wu (CUHK): teaching LLM agents when to clarify and how to gauge their own confidence
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—
Output data is not available yet — this axis would measure publication and preprint volume over time. The profile currently rests on two Evolve-reviewed 2026 arXiv preprints, so the axis is left unscored rather than estimated.
Impact—
Impact data is not available yet — this axis would measure citations and downstream uptake of the work. With only two Evolve-reviewed 2026 preprints and no citation record provided, it is left unscored.
Focus0/100
The focus score is 0, reflecting the raw field distribution: work splits evenly, 0.5/0.5, between two labeled fields — machine learning (GUI agents, uncertainty estimation) and natural language processing (LLM agents, reinforcement learning). In substance the strands are close cousins: both concern LLM-based agents, and both 2026 preprints deal with confidence, ambiguity, or clarifying behaviour, so the low score captures breadth across two field labels more than genuine scatter.
Collaboration22/100
The collaboration score is 22, based on only 2 distinct coauthors across the two reviewed papers, with Yunxiang Li and Helen Meng recurring. Both papers list Xixin Wu in a middle author position without corresponding-author role, so the visible record shows contribution within a small team rather than independent leadership. With just two 2026 preprints behind it, this score should be read as an early-stage footprint, not a mature collaboration pattern.
Activity—
Activity data is not available yet — this axis would measure recent publication cadence and sustained output over time. Two 2026 preprints cannot establish a trajectory, so the axis is left unscored.
Reproducibility55/100
The reproducibility score is 55, computed as the mean over 2 Evolve-reviewed papers, each rated medium readiness. Evolve's verdicts flag concrete limits: CIGAsk reports single-seed results and is simulator-dependent, while PACE's gains are real but modest — 2 of 6 bootstrap confidence intervals cross zero — and scoped to coordinate formats. The mid-range score reflects genuinely checkable methods whose evidence is not yet airtight.
Research summary
Xixin Wu is a researcher at the Chinese University of Hong Kong whose visible bibliometric footprint, as captured in the Evolve catalog, centers on LLM-based agents — specifically the twin questions of when an agent should ask for clarification and how confidently it should act. The profile rests on two Evolve-reviewed 2026 arXiv preprints. The first, CIGAsk, teaches large language models when and how to pose clarifying questions, using counterfactual information gain against frozen references plus an ambiguity bonus within GRPO reinforcement learning; Evolve judges the gains solid but notes single-seed results and simulator dependence. The second, PACE, weights digit-level entropy by place value to estimate GUI click confidence; Evolve finds the gains real but modest, with two of six bootstrap confidence intervals crossing zero and scope limited to coordinate formats. The field distribution splits evenly between machine learning (GUI agents, uncertainty estimation) and natural language processing (LLM agents, reinforcement learning) — a pairing that in practice forms one coherent research thread on calibrated, interactive agent behaviour. On the platform's radar, collaboration scores 22 (2 distinct coauthors; Yunxiang Li and Helen Meng recur, with middle, non-corresponding authorship on both papers), focus scores 0 on the two-field split, and reproducibility averages 55 across the two medium-readiness papers. Output, impact, and activity axes lack sufficient data and are left unscored.
Collaboration pattern
Across the two Evolve-reviewed 2026 preprints, Xixin Wu appears in a middle author position and is not corresponding author on either. The bibliometric record shows only 2 distinct coauthors, with Yunxiang Li and Helen Meng recurring as frequent collaborators — consistent with a small, stable working team. The footprint is therefore small and early-stage: small-team contributions within an existing group rather than an independent, widely-networked agenda, a picture more papers would quickly refine.
Research keywords
Highlights
CIGAsk: teaching LLMs when and how to ask clarifying questions
PACE: place-value-weighted coordinate entropy for GUI click confidence
Timeline
- 2026CIGAsk preprint appears on arXiv, proposing counterfactual information gain with an ambiguity bonus in GRPO to teach LLM agents when and how to clarify.
- 2026PACE preprint appears on arXiv, introducing place-aware coordinate entropy for GUI agent confidence estimation.
- 2026Both papers enter the Evolve review catalog, each rated medium reproducibility readiness (mean reproducibility score 55).
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.
- Dimension scores are computed by Evolve from public bibliometric data; output, impact and activity axes lacked sufficient data and are unscored rather than estimated.
- No ORCID or country metadata was available, so this profile is name-resolved from public bibliographic records and identity disambiguation is limited.
- Only two 2026 arXiv preprints underpin this profile, both with middle, non-corresponding authorship; career facts such as role and biography are not covered.
- The collaboration (22) and focus (0) scores rest on just 2 distinct coauthors and a two-field split — a very small evidence base likely to shift with further publications.