Two independent research projects, one repo. They share nothing but a Python environment and a working style: every claim is measured, every discarded run is recorded, and nothing is written down that the experiments have not earned.
| project | question | state |
|---|---|---|
tabagent/ |
Can an LLM agent repair the frozen synthetic prior of a tabular foundation model (TabICL), after the checkpoint has shipped? | Negative result at small scale. Every arm loses to leaving the checkpoint alone. A full-scale run was retracted — a revoked key had turned the agent into random search, silently. The live follow-up is repairing the context instead of the prior, and it is running now. |
laion_data/ |
Does Doerig et al. (Nature Machine Intelligence 2025), "High-level visual representations in the human brain are aligned with large language models", reproduce — and does it generalise off COCO? | Reproduction in progress. Figures 1 and 3 reproduce; one published claim retracted after audit; the paper's Fig. 4 activations turn out to be irreproducible from the authors' own code. |
TabICL/TabPFN are pre-trained on a prior over random SCMs, frozen at publication. The agent is shown only anonymised statistics of downstream tasks, revises the prior's generating config and its generator code, and LoRA-adapts the released checkpoint to each candidate. Control: random search over the same knobs, same GPU seconds.
A C2ST says the prior is perfectly distinguishable from real tables (AUC 1.000). The agent closes 30% of that gap — and the model gets worse. The unrealism is domain randomisation; narrowing it destroys the coverage the model depends on.
The retraction is worth reading too: every silent fallback in that codebase turned out to be a
lie generator. A revoked API key, an exception handler that returned "", and a caller that
read the silence as "the agent had no proposal" produced a full-scale table that reported random
search under the name agent.
→ tabagent/RESULTS.md (findings) · tabagent/NEXT.md (handoff)
A reproduction of Doerig et al. 2025 on NSD, plus a generalisation test on LAION-fMRI, off the COCO distribution the paper's captions come from.
→ laion_data/reports/master_report.md (what has been
verified, and what has not) · laion_data/RE_VISION_PLAN.md (the plan, in Korean)
tabagent/LLM_에이전트_기반_Tabular_Prior_생성기_문헌조사.md— literature review behind tabagent.
Data and secrets, on purpose. NSD betas are DUA-gated and ~315 GB; COCO derivatives are
regenerable from laion_data/src/; model caches, logs, vendored clones and .env are
gitignored. Generated paper artifacts (tabagent/paper/numbers.tex, main.pdf) are gitignored
too, so a stale number can never be mistaken for a result.