NUS-based researcher linking LLM reasoning and mechanistic interpretability in the RL-training era
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 snapshot contains no publication-count or productivity data for this researcher beyond the three Evolve-reviewed papers. This axis would measure total output volume and its distribution over venues and years.
Impact—
Not yet scored: no citation counts, h-index, or venue-level influence data are available in this snapshot. This axis would measure how widely the researcher's papers are cited and their influence on subsequent LLM reasoning and interpretability research.
Focus0/100
The focus score of 0 reflects an evenly split field distribution: the three catalogued papers sit one-third each in NLP (LLM reasoning) and two machine-learning labels centred on mechanistic interpretability. In practice the two ML labels describe one coherent interpretability agenda, so apparent breadth partly reflects labeling rather than genuine topic scatter; a 3-paper catalog is too small for a settled read.
Collaboration46/100
The collaboration score of 46 rests on 9 distinct coauthors across the reviewed work. All three 2026 papers place Liu in the last author position, consistent with a senior or supervisory role, and none is flagged corresponding. The recurring pool — Xufeng Duan, Arash Lagzian, Srinivas Anumasa, Lingheng Du, Yiming Tang, Harshvardhan Saini, Samyak Jha and Huaming Chen — suggests small rotating teams rather than one long-running group.
Activity—
Not yet scored: the snapshot lacks a publication-date series across recent years, so publishing tempo cannot be assessed. This axis would measure output cadence and recent productivity; here it is represented only by three papers all dated 2026.
Reproducibility43/100
Reproducibility scores 43 as the mean readiness across 3 reviewed papers. MIRAGE rates low: the review notes its 'consistently outperforms CoT' claim is contradicted by its own tables. The two interpretability papers rate medium — the Fixed-SAE synthetic benchmark is never evaluated, and Capsule Lens shows near-chance visual specificity with single-run cases. Code and data availability are not confirmed in this snapshot.
Research summary
Dianbo Liu is a researcher at the National University of Singapore whose reviewed work sits at the intersection of large language model reasoning and mechanistic interpretability. The Evolve catalog holds three 2026 arXiv papers, in all of which Liu is the last author — a position usually associated with senior or supervisory responsibility — and none is flagged as corresponding author. The field distribution is nearly even: one paper on NLP-side LLM reasoning (MIRAGE) and two on mechanistic interpretability. MIRAGE trains a reinforcement-learning selector that routes queries over 20 reasoning perspectives; the reviewed verdict finds the gains promising but notes that the claim of consistently outperforming chain-of-thought is contradicted by the paper's own tables. The interpretability line asks what reinforcement learning actually changes inside a model: a shared frozen sparse autoencoder tracks representation drift during RLVR/GRPO training and finds that short RL mainly boosts formatting features, while Capsule Lens fits closed-form capsules to concept regions to locate and track concept geometry. Collaboration is moderate, with 9 distinct coauthors and a recurring pool of eight frequent collaborators. Reproducibility readiness averages 43 across the three papers, limited mainly by unverified or partially evaluated benchmarks.
Collaboration pattern
Across all three reviewed 2026 papers Liu appears as last author (none flagged corresponding), working with a pool of 9 distinct coauthors. Recurring names include Xufeng Duan, Arash Lagzian, Srinivas Anumasa, Lingheng Du, Yiming Tang, Harshvardhan Saini, Samyak Jha and Huaming Chen. The pattern suggests small, project-based teams assembled around an interpretability and LLM-reasoning agenda rather than a single long-running lab line.
Research keywords
Highlights
RL-guided perspective routing for creative LLM reasoning
A fixed sparse-autoencoder window on what RL changes in an LLM
Capsule Lens: closed-form tracking of concept geometry
Timeline
- 2026Last-authored arXiv paper MIRAGE on reinforcement-learning-guided multi-perspective LLM reasoning (reproducibility readiness: low).
- 2026Fixed-SAE Track paper: a mechanistic-interpretability study of what RLVR/GRPO training changes inside an LLM (readiness: medium).
- 2026Capsule Lens paper: closed-form capsule fitting to locate and track concept geometry in model representations (readiness: medium).
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.
- Radar scores and coauthor counts come from public bibliometric data (OpenAlex-style sources); citation, output and activity metrics are missing from this snapshot.
- The reviewed catalog covers only 3 arXiv papers, all dated 2026, and may not represent the full publication record.
- The profile is name-resolved from public bibliographic data (ORCID 0000-0002-3042-9161); 'Dianbo Liu' may overlap with other researchers of the same name.
- No corresponding-author papers appear in the catalog; the senior-author reading is inferred from last author position only.