Researcher at Microsoft working across diffusion language models, EHR clinical reasoning, and LLM-based simulation
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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 metrics are not yet available in the underlying bibliometric data. This axis would measure total publication volume and annual production rate, which here cannot be scored or estimated from the three reviewed catalog papers alone.
Impact—
Impact data such as citation counts is missing from the current snapshot. This axis would measure citations, citation velocity, and downstream uptake of the work; no figure can be quoted, so it is left unscored rather than estimated.
Focus0/100
The focus score is 0, the lowest on this radar, reflecting the raw field distribution: an even 0.5/0.5 split between natural language processing (diffusion language models) and machine learning for healthcare. The reviewed catalog adds a third strand in AI in education (StudentSim). This is broad, application-driven work across LLM architectures, clinical modeling, and education simulators rather than deep concentration in a single subfield.
Collaboration55/100
The collaboration score of 55 is computed from 15 distinct coauthors in the source data. The frequent-coauthor list spans two visible clusters: language-model researchers (Xiaodong Liu, Jianfeng Gao, Michel Galley, Chenglong Wang, Ke Yang) and healthcare-ML researchers (Sheng Zhang, Tristan Naumann, Hoifung Poon, Yuxin Xiao), consistent with the even 0.5/0.5 field split. All three catalog papers are middle-author, non-corresponding contributions, suggesting a team-embedded working style rather than solo leadership.
Activity—
Recent activity metrics are unavailable in this snapshot. This axis would measure publication cadence over recent years and ongoing research momentum; the 2026 arXiv postings hint at current work, but no scored activity data exists yet.
Reproducibility78/100
The reproducibility score of 78 is the strongest signal in this profile, averaged over 2 papers with reproducibility assessments. Within the catalog, the diffusion-language-model register-token paper is rated high readiness, while the EHR reinforcement-learning and StudentSim papers are rated medium — an above-average but not uniform commitment to releasable code, benchmarks, and honest reporting of limitations.
Research summary
Chandan Singh is a researcher at Microsoft (United States) whose reviewed work spans two roughly equal strands: natural language processing — with an emphasis on diffusion language models — and machine learning for healthcare, each accounting for 0.5 of the recorded field distribution. The Evolve catalog contains three 2026 arXiv papers, all with middle-author positions and none as corresponding author. The first, on register tokens for bounded-state reasoning in diffusion language models, carries reasoning across chunk resets and beats text-based carry-over on all 12 benchmarks, though no-carry SFT still wins on math; Evolve characterizes it as an honest method paper and rates its reproducibility readiness high. The second applies reinforcement learning (GRPO/DAPO) with time-aware rewards to EHR foundation models for clinical reasoning; Evolve judges the gains real but modest, single-dataset, with thin statistical rigor. The third, StudentSim, trains per-student LLM-based student simulators that beat GPT-5.4 and Maia2 on fidelity and responsiveness, with a solid benchmark but no validation on real students. Radar scores place reproducibility at 78 (averaged over 2 assessed papers) as the strongest axis, collaboration at 55 (15 distinct coauthors), and focus at 0, reflecting the even two-field split. Citation, output, and activity metrics are not yet available in the underlying data, so this profile should be read as an early-stage, catalog-scoped picture.
Collaboration pattern
With 15 distinct coauthors and a collaboration score of 55, the reviewed work is clearly team-based. The frequent-coauthor list splits along the two field strands: language-model-side coauthors including Xiaodong Liu, Jianfeng Gao, Michel Galley, Chenglong Wang, and Ke Yang, and healthcare-ML-side coauthors including Sheng Zhang, Tristan Naumann, Hoifung Poon, and Yuxin Xiao. All three catalog papers list middle-author, non-corresponding roles, consistent with contributing specific components inside larger research groups.
Research keywords
Highlights
Register tokens give diffusion LMs persistent reasoning state
Time-aware RL fine-tuning for EHR foundation models
Per-student LLM simulators beat frontier baselines on fidelity
Timeline
- 2026Releases the register-token method for bounded-state reasoning in diffusion language models on arXiv (middle-author position; high reproducibility readiness).
- 2026Publishes reinforcement learning over patient trajectories for clinical reasoning in EHR foundation models, using GRPO/DAPO with time-aware rewards.
- 2026Publishes StudentSim, training pooled-and-specialized LLM student simulators that beat GPT-5.4 and Maia2 on fidelity and responsiveness.
- 2026All three catalog papers appear within a single year as middle-author, non-corresponding arXiv preprints.
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.
- This profile is built from only three Evolve-reviewed arXiv papers, all dated 2026; the researcher's broader publication record is outside the current data scope.
- Output, impact, and activity radar axes lack sufficient underlying data and are left unscored rather than estimated.
- The reproducibility score of 78 averages only 2 assessed papers, so it should be read as an early, narrow signal.
- No ORCID or author-disambiguation metadata is available; this profile is name-resolved from public bibliographic data, and all reviewed papers list middle-author, non-corresponding roles.