Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 22, 2026São Paulo, Brazil

Local-Geometry Signals of Capability Emergence During Portuguese Grammar Acquisition in a Small Language Model

Elvis Sikora, Carolina Oliveira, José Pedro Brito · Team DI #0003

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Local-Geometry Signals of Capability Emergence During Portuguese Grammar Acquisition in a Small Language Model

Open in a new tab (opens in new tab)
Share

This work trains a small AI model on Portuguese and finds that an internal structure measure, the LLC, jumps right when the model learns grammar, even though the usual loss curve shows nothing. Control experiments confirm the signal appears only when real language structure is being learned. The promise is a way to spot a new ability as it forms, before it shows up in behavior. The most exciting next step is to test whether safety rules learned in one language survive further training in another, using the LLC as an early warning.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. The main results rely on a few samples from a single model training run, and the described trend does not seem abrupt enough to warrant much confidence in their results.

    Still, the write-up was exceptionally clear, the set-up was really clean such that I trust the results they state, and it seems to build upon and slot well within the surrounding literature. It's the best article I've reviewed so far in my batch for this hackathon

  2. The demonstration that the Local Learning Coefficient exhibits a "change point" aligned with the emergence of grammatical competence in Portuguese (while the loss curve remains smooth) is a clean, well-controlled result, and the matched controls with scrambled Portuguese and matched English are methodologically sound. The AI safety implication is relevant in terms of if this signal scales, it could complement behavioral evaluations by flagging capability formation before it is expressed in behavior. To consolidate the work, the most important next steps are replicating with a second independent seed, adding formal statistical tests of the LLC-behavior alignment, and exploring whether the same pattern appears in larger models.

    The future work question about whether safety-like constraints learned in one language are weaken by continued training in another is one of the most important questions the paper leaves open, and would be a highly valuable contribution if pursued.

    Read full reviewShow less

  3. This is a clean developmental-monitoring result in a tiny controlled language setting. I would frame the novelty narrowly, around the continued-pretraining + Portuguese grammar + controls.

    Also make the repo more reproducible, right now a fresh clone does not contain the checkpoints/corpora needed to rerun.

    Other than that, excellent work!

Cite this project

@misc{sikora2026localgeometry,
  title = {{Local-Geometry Signals of Capability Emergence During Portuguese Grammar Acquisition in a Small Language Model}},
  author = {Elvis Sikora and Carolina Oliveira and José Pedro Brito},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/localgeometry-signals-of-capability-emergence-during-portuguese-grammar-acquisition-in-a-small-language-model-b24y}},
  url = {https://apartresearch.com/sprints/projects/localgeometry-signals-of-capability-emergence-during-portuguese-grammar-acquisition-in-a-small-language-model-b24y}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026