Santiago Aranguri

Manuel Fernández Burda, Santiago Aranguri, Iván Arcuschin Moreno et al.

Protein language models (PLMs) are becoming practical tools for de novo protein design, yet their dual-use potential raises safety concerns. We show that domain adaptation to specific taxonomic groups can elicit toxic protein generation, even when toxicity is not the training objective. To address this, we adapt Logit Diff Amplification (LDA) as an inference-time control mechanism for PLMs. LDA modifies token probabilities by amplifying the logit difference between a baseline model and a toxicity-finetuned model, requiring no retraining. Across four taxonomic groups, LDA consistently reduces predicted toxicity rate (measured via ToxDL2) below the taxon-finetuned baseline while preserving biological plausibility. We evaluate quality using Fréchet ESM Distance and predicted foldability (pLDDT), finding that LDA maintains distributional similarity to natural proteins and structural viability (unlike activation-based steering methods that tend to degrade sequence properties). Our results demonstrate that LDA provides a practical safety knob for protein generators that mitigates elicited toxicity while retaining generative quality.

Santiago Aranguri, Giulio Biroli, Marc Mezard et al.

Recent works have shown that diffusion models can undergo phase transitions, the resolution of which is needed for accurately generating samples. This has motivated the use of different noise schedules, the two most common choices being referred to as variance preserving (VP) and variance exploding (VE). Here we revisit these schedules within the framework of stochastic interpolants. Using the Gaussian Mixture (GM) and Curie-Weiss (CW) data distributions as test case models, we first investigate the effect of the variance of the initial noise distribution and show that VP recovers the low-level feature (the distribution of each mode) but misses the high-level feature (the asymmetry between modes), whereas VE performs oppositely. We also show that this dichotomy, which happens when denoising by a constant amount in each step, can be avoided by using noise schedules specific to VP and VE that allow for the recovery of both high- and low-level features. Finally we show that these schedules yield generative models for the GM and CW model whose probability flow ODE can be discretized using $Θ_d(1)$ steps in dimension $d$ instead of the $Θ_d(\sqrt{d})$ steps required by constant denoising.

Frank Xiao, Santiago Aranguri

We propose activation-based data attribution, a method that traces behavioral changes in post-trained language models to responsible training datapoints. By computing activation-difference vectors for both test prompts and preference pairs and ranking by cosine similarity, we identify datapoints that cause specific behaviors and validate these attributions causally by retraining with modified data. Clustering behavior-datapoint similarity matrices also enables unsupervised discovery of emergent behaviors. Applying this to OLMo 2's production DPO training, we surfaced distractor-triggered compliance: a harmful behavior where the model complies with dangerous requests when benign formatting instructions are appended. Filtering top-ranked datapoints reduces this behavior by 63% while switching their labels achieves 78%. Our method outperforms gradient-based attribution and LLM-judge baselines while being over 10 times cheaper than both. This in-the-wild model organism - emerging from contaminated preference data rather than deliberate injection - provides a realistic benchmark for safety techniques.

Santiago Aranguri, Francesco Insulla

We analyze the training of a two-layer autoencoder used to parameterize a flow-based generative model for sampling from a high-dimensional Gaussian mixture. Previous work shows that the phase where the relative probability between the modes is learned disappears as the dimension goes to infinity without an appropriate time schedule. We introduce a time dilation that solves this problem. This enables us to characterize the learned velocity field, finding a first phase where the probability of each mode is learned and a second phase where the variance of each mode is learned. We find that the autoencoder representing the velocity field learns to simplify by estimating only the parameters relevant to each phase. Turning to real data, we propose a method that, for a given feature, finds intervals of time where training improves accuracy the most on that feature. Since practitioners take a uniform distribution over training times, our method enables more efficient training. We provide preliminary experiments validating this approach.