Studies when LLMs and agents fail to revise beliefs — from dialogue clarification to agent social reasoning.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-08) 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: the data snapshot lacks any publication-volume signal for this researcher — no total paper count, venue record, or authorship-role distribution — so this axis, which would measure overall research output, cannot be estimated and is left unscored.
Impact—
Not yet scored: no citation counts, downstream uptake, or influence metrics are available in the provided data, so this axis — which would measure how widely the work is read, cited, and built upon by others — cannot be estimated and remains empty.
Focus0/100
The focus score is 0, reflecting an even split in the underlying field distribution: 50% natural language processing (LLM dialogue evaluation) and 50% artificial intelligence (LLM agents). With only two reviewed papers, this reads less like scattered curiosity than a deliberate two-sided attack on one problem — belief-revision failures — examined through dialogue evaluation on one side and agent architectures on the other. More publications would be needed to tell whether this is a stable dual focus or an early pivot.
Collaboration46/100
The collaboration score is 46, built on 9 distinct coauthors across the two reviewed papers. Named frequent collaborators include Jianzhe Lin, Fei Wang, Jubin Chheda, Robert Douglas, Rajeshkumar Golani, and Yunda Liu. On both papers this researcher sits in a middle authorship position and is not the corresponding author, suggesting contribution within a team frame rather than independent leadership so far — a mid-range, team-embedded collaboration pattern.
Activity—
Not yet scored: the data provides no publication-frequency signal — only two 2026 papers, with no multi-year cadence to measure — so this axis, which would track output rhythm and recent research engagement over time, is left unscored rather than guessed.
Reproducibility38/100
The reproducibility score is 38, a mean computed over exactly 2 papers. The split is instructive: the dialogue-evaluation paper is rated low readiness — Evolve flags judge circularity and no statistical tests — while the ReAdapt agent paper reaches medium readiness despite a single model, no ablations, and no statistical testing. The reported gains (37% to 51%, 69% to 77%) are concrete but unverified, making independent replication the obvious next step.
Research summary
This profile is name-resolved from public bibliographic data: 'Xiaolin Li' is a common name, and the evidence base here is two 2026 arXiv preprints reviewed in the Evolve catalog, on both of which this researcher appears as a middle author rather than corresponding author. The two papers form one coherent research thread around how large language models fail to revise their beliefs. The dialogue-evaluation paper argues that LLMs commit to early answers and treat later clarification as additional context rather than correction, leaving stale hypotheses contaminating multi-turn conversations; Evolve's review finds the diagnosis compelling but flags judge circularity and the absence of statistical tests. The agent paper proposes ReAdapt, claiming that explicit structured social-state updates lift LLM-agent relational accuracy from 37% to 51% and from 69% to 77%; the review deems it a credible proof of concept limited by a single model, no ablations, and no statistical testing. Bibliometric signals are modest and thin: a collaboration score of 46 built on 9 distinct coauthors, a focus score of 0 reflecting an even 50/50 split between LLM dialogue evaluation and LLM agents, and a reproducibility score of 38 averaged over the two papers (low and medium readiness). Output, impact, and activity axes are unscored for lack of data.
Collaboration pattern
Across the two reviewed papers, Xiaolin Li works in a team-embedded pattern: 9 distinct coauthors, a collaboration score of 46, middle authorship on both, and no corresponding-author role. Jianzhe Lin and Fei Wang appear repeatedly in the co-author frequency list, alongside Jubin Chheda, Robert Douglas, Rajeshkumar Golani, and Yunda Liu. The pool spans both research halves — dialogue evaluation and LLM agents — consistent with a small core group attacking one problem from two angles.
Research keywords
Highlights
LLMs commit early and treat clarification as context, not correction
ReAdapt: explicit social-state updates lift agent relational accuracy
One thread across two fields: hypothesis revision
The two catalog papers share a single concern — whether language models can let go of a wrong belief — studied from the evaluation side (dialogue clarification) and the intervention side (agent state updates). This even 50/50 split between NLP and AI underlies the researcher's focus score of 0.
Timeline
- 2026Co-authored (middle position, non-corresponding) 'Clarification Is Not Correction: LLMs Fail to Let Go', an arXiv study of old-hypothesis contamination in LLM dialogue.
- 2026Co-authored (middle position, non-corresponding) 'When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning', an arXiv paper on structured social-state updates for LLM agents.
- 2026Both papers entered the Evolve review catalog in the same year, spanning LLM dialogue evaluation and LLM agents in an even 50/50 field split.
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.
- Profile is name-resolved from public bibliographic data; 'Xiaolin Li' is a common name, and no ORCID or institutional affiliation is available in this snapshot.
- Dimension scores are computed by Evolve from a small public bibliometric snapshot; only 2 reviewed papers underlie the collaboration, focus, and reproducibility axes.
- Output, impact, and activity axes are unscored (dimMissing): no citation counts, publication-volume, or multi-year cadence data was provided.
- Both papers are 2026 arXiv preprints with middle, non-corresponding authorship; venue acceptance and career history are unknown.