Carla P. Gomes: ML researcher in LLM mechanistic interpretability and efficient language model architectures
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) 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 output score is available yet: the snapshot lacks full publication counts and yearly productivity data. This axis would measure total and per-year publication volume across venues, which cannot be inferred from two reviewed 2026 papers alone.
Impact—
Impact cannot be scored yet: citation counts, h-index, and venue-level influence are absent from the input data. This axis would measure citations to her work and its uptake by other researchers, which requires a larger bibliographic record.
Focus0/100
Focus scores 0 on the 0-100 scale, consistent with reviewed output split exactly 50/50 between two subfields — LLM mechanistic interpretability and efficient language model architectures — rather than concentrated in one line. Both sit inside machine learning and share an interest in how, and at what parameter cost, large language models reason, so the low score reads as breadth across two active threads rather than a scattered agenda, on evidence from only two papers.
Collaboration44/100
Collaboration scores 44 with 8 distinct co-authors across the two reviewed papers, a mid-range figure consistent with small, focused teams. She is last author on the interpretability study and a middle author on MoRE, and recurring names include Kilian Q. Weinberger, Justin Lovelace, Zhongdi Qu, and Eric S. Qiu, indicating stable working groups around both research threads.
Activity—
No activity score is available: the data contain no publication dates beyond 2026 and no yearly output series. This axis would measure recent publication cadence and ongoing research engagement, which a two-paper snapshot from a single year cannot support.
Reproducibility100/100
Reproducibility is the strongest measured axis, with a mean score of 100 across 2 reviewed papers, both rated high reproducibility readiness. The interpretability study backs its four-stage pipeline with controlled, multi-method evidence — activation patching, linear probing, direct logit attribution — while MoRE reports small but consistent perplexity and downstream gains up to 1.15B parameters, making both bodies of claims directly checkable.
Research summary
Carla P. Gomes is an early-stage Evolve profile, name-resolved from public bibliographic data and currently grounded in two 2026 arXiv papers reviewed by Evolve, both rated high reproducibility readiness. Her reviewed work splits exactly 50/50 across two machine learning subfields: LLM mechanistic interpretability and efficient language model architectures. On the interpretability study she is last author; the paper proposes a four-stage decomposition of how large language models solve math word problems and localizes distractor-driven fragility to Operation Planning heads, using activation patching, linear probing, and direct logit attribution on GSM8K. The efficiency paper, MoRE: Mixture of Reused Experts, shares MoE expert pools across adjacent layers and adds depth embeddings, yielding small but consistent perplexity and downstream gains at up to 1.15B parameters. Collaboration metrics show 8 distinct co-authors and a mid-range collaboration score of 44, with Kilian Q. Weinberger, Justin Lovelace, and Zhongdi Qu among recurring collaborators. Reproducibility is the strongest measured axis, with a mean score of 100 across both reviewed papers, reflecting controlled, multi-method evidence and directly verifiable claims. Output, impact, and activity cannot yet be scored because the snapshot contains no citation counts, publication history, or yearly productivity series; these limitations are recorded in the profile caveats rather than estimated.
Collaboration pattern
Across the two reviewed 2026 papers she works in small teams — 8 distinct co-authors in total, collaboration score 44. She is last author on the mechanistic interpretability study and middle author on MoRE. Recurring co-authors include Kilian Q. Weinberger, Justin Lovelace, Christian Belardi, Zhongdi Qu, Eric S. Qiu, Utku Umur Acikalin, and Arjun B. Mulchandani, suggesting two stable working groups: one around LLM interpretability, one around efficient model architectures.
Research keywords
Highlights
Four-stage internal pipeline for LLM word-problem solving
Distractor fragility localized to Operation Planning heads
MoRE reuses expert pools across adjacent layers
Timeline
- 2026Last-author paper on a four-stage decomposition of LLM math reasoning appears on arXiv, entering the catalog in the LLM mechanistic interpretability field.
- 2026Co-authors MoRE: Mixture of Reused Experts, an efficient architecture paper on shared expert pools, in the middle of the author list and rated high reproducibility readiness.
- 2026Both 2026 papers are reviewed by Evolve with a reproducibility mean of 100, making reproducibility the strongest measured dimension of this profile.
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.
- Institution, country, and ORCID are absent from the input data; the profile is name-resolved from public bibliographic data, so identity should be verified before use.
- Radar scores are computed from a limited bibliometric snapshot covering only the two reviewed 2026 arXiv papers; output, impact, and activity axes lack underlying data.
- Citation counts, h-index, venue list, and full publication history are not present in the input data and are therefore not quoted anywhere in this profile.
- Timeline coverage is limited to 2026, the only publication year present in the reviewed record.