Dartmouth researcher linking representation geometry and LLM reasoning — from Platonic alignment to latent steering
3
papers here
2
reviewed by Evolve
1
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—
No publication-volume metrics are available for this profile yet, so output is unscored. This axis would measure how many papers the researcher produces over time and how consistently.
Impact—
Citation counts and downstream influence are not yet available, so impact is unscored. This axis would measure how often and how widely the researcher's papers are cited and built upon.
Focus0/100
The focus score of 0 reflects an even split in the underlying field distribution: 0.5 in natural language processing (LLM reasoning) and 0.5 in machine learning (representation learning). The 2026 catalog mirrors this — a first-author representation-geometry paper, a latent-steering reasoning paper, and an LLM-agent world-model study. The work is balanced across two threads rather than concentrated in one.
Collaboration61/100
The collaboration score is 61, drawn from 20 distinct coauthors in the current snapshot, 12 of whom recur — including Yujin Tang, Hengyuan Zhang, Zunhai Su, and Jing Xiong. Authorship positions vary across the 2026 papers: first author on the displacement-geometry study and middle positions on the latent-steering and LLM-agent papers, consistent with team-based LLM research.
Activity—
Recent-activity signals such as publication cadence are not available yet, so activity is unscored. This axis would measure how regularly the researcher publishes and sustains research momentum.
Reproducibility78/100
Reproducibility averages 78 across the 2 papers Evolve has rated (reproMean 78, reproPapers 2). The displacement-geometry paper is rated high readiness, while the SPS latent-steering paper is medium; Evolve's verdicts flag whitening and data overlap in the former and thin statistical rigor in the latter as points reviewers should verify.
Research summary
Chenming Shang is a researcher at Dartmouth College working across representation learning and LLM reasoning. In a 2026 first-author arXiv paper, 'Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities,' Shang argues that models share displacement directions rather than representational positions, drawing on evidence from 44 encoders and reframing the debate around the Platonic Representation Hypothesis. A second 2026 paper introduces SPS, a training-free, state-conditioned latent-steering method with progress guidance that delivers consistent Pass@4 gains on Qwen3 models from 1.7B to 14B across 9 benchmarks. A third studies how world models and policies compose in LLM agents through a joint spectral and behavioral account. Evolve's radar shows a collaboration score of 61 with 20 distinct coauthors (12 recurring), a focus score of 0 reflecting an even 0.5/0.5 split between LLM reasoning and representation learning, and a reproducibility score of 78 averaged over 2 rated papers. Output, impact, and activity metrics are not yet available, making this an early-stage profile built from a 2026 arXiv snapshot.
Collaboration pattern
The snapshot records 20 distinct coauthors, with 12 recurring — including Yujin Tang, Hengyuan Zhang, Zunhai Su, Xiao Liang, Hui Shen, Jing Xiong, and Kailai Yang — supporting a collaboration score of 61. Shang takes first-author responsibility on the representation-geometry work and middle positions on the reasoning and agent papers, a pattern typical of team-based LLM research.
Research keywords
Highlights
First-author reframing of the Platonic Representation Hypothesis
SPS: training-free latent steering for more efficient reasoning exploration
How world models and policies compose in LLM agents
A 2026 co-authored study offers a joint spectral and behavioral account of how world models and policies compose in LLM agents. It sits at the intersection of Shang's two main research threads and has not yet received an Evolve verdict, keywords, or reproducibility rating.
Timeline
- 2026Publishes a first-author study using displacement geometry to capture Platonic shared reality across 44 encoders, reframing the Platonic Representation Hypothesis debate.
- 2026Co-authors SPS, a training-free state-conditioned latent-steering method delivering consistent Pass@4 gains on Qwen3 1.7B–14B across 9 benchmarks.
- 2026Co-authors a joint spectral and behavioral account of how world models and policies compose in LLM agents.
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.
- Bibliometric indicators are computed from public bibliometric data (OpenAlex-based); no citation counts or h-index figures are available in this snapshot.
- Coverage is limited to a 2026 arXiv snapshot with three cataloged papers; output, impact, and activity axes lack sufficient data and are unscored.
- The profile is name-resolved from public bibliographic data (ORCID 0009-0004-8945-2165); ambiguity with other researchers sharing the name cannot be fully excluded.
- One cataloged paper (world models and policies in LLM agents) has not yet been reviewed by Evolve, so it carries no verdict, keywords, or reproducibility rating.