Maël
Trémouille.

The question

Knowledge-editing methods can rewrite a fact inside an LLM, but the edit often fails to transfer across languages. Rather than benchmark the failure, the project locates its mechanism: which keys, residual-stream positions and attention heads carry the edited fact, and why the multilingual read-out misses them.

My role

I own the experimental side. I built the multilingual medical evaluation sets, reproduced and extended AlphaEdit, MEMIT, INCLINE, CAA and ITI on LLaMA, Qwen, Mistral and BLOOMZ, and located where cross-lingual access breaks by patching keys, residual streams and attention under causal controls and ablations — with checkpoint audits so every number can be traced back to a run. Work conducted towards an ICLR 2027 submission.

The problem

A risk analyst states an indicator — the probability of a cyberattack on a critical service, say — and needs to know whether the outside world agrees. The system reads the documents, extracts the comparable facts, and returns a verdict that a human can audit down to the sentence it came from.

My role

I developed the backend, the agents and the deployment, and owned the comparison-and-decision component: two named agents that turn documents into structured queries, check semantic and numerical compatibility with the claim, and produce a verdict with traceable citations. Multi-step orchestration in LangGraph across several models, a FastAPI backend and installable CLI, an offline test suite and reproducible benchmarks to measure verdict quality, Docker Compose deployment behind Nginx.

The catch

On dense transaction data, a naive last-price baseline is very hard to beat, and any feature that peeks at the current trade will beat it for the wrong reason. The value of the project is in the evaluation, not the model.

My role

I owned data integrity and evaluation. I found and fixed seven data leaks in the feature engineering — five where a current value stood in for a lagged one, two where training statistics were reused on validation — then rebuilt the evaluation as a seven-fold walk-forward backtest over 14-day windows, from 547k to 4.6M training rows, and added SHAP analysis of market and order-book features. With honest numbers, the team's LightGBM model with order-book features cut MAE by 11.7% against the naive baseline on high-yield bonds, winning five folds out of seven, and reduced RMSE by a factor of 2.6 on the tails.

The finding

Bigger is not better here. A cleaned TF-IDF with logistic regression reaches 84.2% test accuracy; a fine-tuned 110M-parameter CamemBERT reaches 71.0%, and a frozen sentence encoder 66.8%. A temporal split — train on early elections, test on later ones — drops every model by ten points and shows how much of the signal is period-specific vocabulary.

My role

Solo project. I built the pipeline end to end: corpus cleaning, three model families, five-fold cross-validation with bootstrap confidence intervals, the temporal split that exposed the drift, an error analysis, and t-SNE / UMAP maps of inter-party discourse. The cleaning step alone lifted temporal accuracy by 2.7 points and made the features interpretable.

The problem

Known positives, and an unlabeled remainder that mixes positives and negatives. Treating every unknown as negative biases the classifier; and without true negatives, recall in the usual sense does not exist.

My role

I reviewed the PU literature and implemented the methods that mattered: SCAR and SAR propensity estimation with an EM-style algorithm, the non-negative risk estimator nnPU, probability-gap relabelling, and Kernel Mean Matching solved as a quadratic program. Since recall could not be computed, I designed a recall-on-masked-positives metric to compare them, tuned with TPE, and shipped the result as an internal Python scoring package. I presented the method to the group's scientific committee, in front of 100 people.

The setup

Train against the worst-case perturbation an attacker can find, then ask what capacity buys you: a 4.3M-parameter CNN, ResNet-18 and a 36M-parameter WideResNet-28-10 on CIFAR-10 and GTSRB, attacked with FGSM, PGD, MIM and a universal L0 patch.

My role

Solo project. I re-implemented Madry's min–max training in PyTorch as a configurable pipeline with checkpoints and CSV reports, ran the full benchmark, and measured the trade-off. Adversarial training cut the universal patch's success rate from 89.7% to 14.8% on the CNN and from 71.3% to 27.1% on ResNet-18; lowering the share of adversarial examples during training recovered up to 5.9 points of clean accuracy at little or no cost in robustness.

The idea

When the objective is noisy, the Gaussian-process variance near observed points is wrong for acquisition; re-interpolating it gives the optimizer an honest picture of what it still does not know.

My role

I implemented the variance re-interpolation for Bayesian optimization in the public SMT codebase, with the optimization code, unit tests and documentation required to merge into a library used by engineering teams.