Jubin Chheda

—

Unclaimed · AI-structured profileORCID OpenAlex

LLM researcher focused on hypothesis revision — dialogue clarification failures and social reasoning in LLM agents

papers here

reviewed by Evolve

first-author

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 underlying publication-count data for this axis is unavailable in the current snapshot. Once gathered, it would measure research output volume — number of papers and preprints, and output pace over time.

Impact—

Not yet scored: citation-based evidence is absent from the supplied data. This axis would measure scholarly influence — citations, citing works, and downstream uptake of the two catalogued papers.

Focus0/100

The focus score is 0, and the field distribution shows why: output divides evenly at 0.5/0.5 between natural language processing (LLM dialogue evaluation) and artificial intelligence (LLM agents). Both catalogued 2026 papers attack the same failure mode — models that will not revise an earlier commitment — from a dialogue side and an agent side. That reads as a coherent two-track agenda, but with only two papers there is no dominant field yet.

Collaboration46/100

The collaboration dimension scores 46. The raw network contains 9 distinct co-authors across the 2 catalogued papers, with Jianzhe Lin, Xiaolin Li, and Fei Wang the most repeated names. Jubin Chheda holds the last-author slot on both papers while being designated corresponding author on neither — a senior-position-without-correspondence pattern. The moderate score matches a small, stable core team rather than a broad collaborative web.

Activity—

Not yet scored: the snapshot supplies no multi-year publication history — only two 2026-dated papers. This axis would measure research tempo, such as papers per year and recency of output, once a longer record exists.

Reproducibility38/100

Reproducibility scores 38, the mean readiness rating across 2 reviewed papers (reproMean 38, reproPapers 2). The dialogue paper rates low — Evolve flags judge circularity and no statistical tests; the ReAdapt agent paper rates medium, called a credible proof-of-concept on a single model without ablations or significance testing. Both reviews point to the same gap: interesting effects, under-powered evaluation.

Research summary

Jubin Chheda is an early-stage profile built from public bibliographic data, spanning natural language processing and artificial intelligence in an even 0.5/0.5 split between LLM dialogue evaluation and LLM agents. The Evolve catalog contains two 2026 arXiv papers, both listing Chheda in the last-author position and neither as corresponding author. One thesis unifies them: language models form commitments early and struggle to revise them. "Clarification Is Not Correction: LLMs Fail to Let Go" diagnoses how models treat later user clarification as more context rather than correction, leaving an old hypothesis contaminating the dialogue state; Evolve finds the diagnosis compelling but flags judge circularity and the absence of statistical tests. "When LLM Agents Fail to Read the Room" introduces ReAdapt, where explicit structured social-state updates lift relational accuracy from 37% to 51% and from 69% to 77% — a credible single-model proof of concept, per Evolve, but without ablations or significance testing. Reproducibility readiness rates low and medium respectively, averaging 38 across 2 papers. Collaboration is moderate (score 46; 9 distinct co-authors, with Jianzhe Lin, Xiaolin Li, and Fei Wang recurring). Output, impact, and activity metrics are not yet available.

Collaboration pattern

The raw record shows 9 distinct co-authors across the two catalogued papers, and the frequent-coauthor list is topped by a recurring trio — Jianzhe Lin, Xiaolin Li, and Fei Wang — each appearing twice. Jubin Chheda sits in the last-author position on both papers without corresponding-author designation, a senior-slot contribution to team-led work. The collaboration dimension scores 46, a moderate value matching a small but stable core-plus-rotating network.

Research keywords

LLM dialogue evaluationMulti-turn conversationClarification handlingUncertainty in dialogueDialogue state managementOld-hypothesis contaminationHypothesis revisionLLM agentsReAct-style agentsSocial reasoningAgent benchmarksStructured state tracking

Highlights

Clarification is treated as context, not correction
The 2026 dialogue-evaluation paper claims LLMs commit to an answer early and then treat later user clarification as additional context rather than as a correction, so the old hypothesis keeps contaminating the dialogue state. Evolve calls the diagnosis compelling but weakens it on judge circularity and the absence of statistical tests; reproducibility readiness is rated low.
ReAdapt: explicit social-state updates lift agent relational accuracy
The 2026 agents paper proposes ReAdapt, which performs explicit structured social-state updates inside a ReAct-style loop; Evolve records claimed relational-accuracy gains from 37% to 51% and from 69% to 77%. It is judged a credible proof-of-concept, but on a single model and without ablations or statistical tests; reproducibility readiness is medium.
One thesis across two papers: models fail to let go of stale beliefs
Both catalogued 2026 papers target the same failure mode from two angles — dialogue-state contamination when users correct themselves, and stale social state in LLM agents — positioning the author at the seam between dialogue evaluation and agent design, with each field carrying 0.5 of the catalog distribution.

Timeline

2026First appearance in the Evolve catalog: last-author arXiv paper "Clarification Is Not Correction: LLMs Fail to Let Go" in natural language processing (LLM dialogue evaluation).
2026Also dated 2026, last-author arXiv paper "When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning" in artificial intelligence (LLM agents) is catalogued, with Evolve recording claimed accuracy gains of 37→51% and 69→77%.
2026Evolve reproducibility reviews rate the two papers low and medium respectively, with a mean of 38 across 2 papers, citing missing statistical tests and single-model evaluation.

Papers in the Evolve catalog

2026-09-21
Clarification Is Not Correction: LLMs Fail to Let Go
2026-09-21
When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning

Frequent co-authors

Jianzhe Lin×2 Xiaolin Li×2 Fei Wang×2 Robert Douglas×1 Rajeshkumar Golani×1 Yunda Liu×1 Jianzhe Lin×1 Xiaolin Li×1 Fei Wang×1

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.

Scores are computed from OpenAlex-derived bibliometric data only; no institutional affiliation or country is present in the snapshot.
Output, impact, and activity axes are unscored because the underlying counts are absent from the supplied data.
Only two catalogued papers, both 2026 arXiv preprints, underlie this profile, so all scores rest on a very small base.
The reproducibility score (38) averages Evolve review assessments of 2 papers; it is not an independent replication audit.