Skip to main content
Search

Université Paris-Saclay

12 related articles

Subscribe to updates
  1. bioRxiv

    CellScribe folds cell identity and state into one generative interface: 0.223 macro-F1 across 309 held-out types, ahead of a frozen-embedding probe

    Single-cell annotation has long treated cell identity (lineage) and state (active transcriptional programs) with separate tools, while fixed-output classifiers can only return predefined types. CellScribe feeds frozen single-cell foundation-model embeddings into a pretrained language model through AdaLN adapters, so one shared model generates either an ontology-grounded cell type or a meta-program profile with activity scores under task-specific prompts. On the 309-type, 4.44M-cell study-holdout split of CELLxGENE, the full pipeline reaches 0.223 macro-F1 and 0.484 weighted-F1, above the 0.194 and 0.366 of a supervised probe on frozen scFoundation embeddings; on malignant cells from held-out 3CA studies it reaches 0.344 hallmark accuracy against the probe's 0.227.

  2. H&E slides skip spatial transcriptomics: ALHENA distills cell and neighborhood similarity, reaching 0.341 patient-level signature correlation

    ALHENA distills spatial transcriptomic information into a routine H&E encoder that reads only images at inference. Rather than regressing individual genes, it trains the H&E representation to match the sample-to-sample similarity structure of two frozen transcriptomic models at cell and neighborhood scales. Across 110 cancer-type and signature combinations in TCGA, mean Spearman correlation was 0.341 versus 0.161 for unadapted UNI2-h; on the Thunder pathology benchmark the median change was −0.004.

  3. arXiv

    Synapses That Remember Two Moments: Equilibrium Propagation Moves Its Contrastive Update Inside the Device, Matching Standard MNIST Error

    Equilibrium Propagation (EqProp) updates each synapse by comparing a local quantity at the free and nudged equilibria, which normally requires extra memory or a duplicated network to hold the earlier value. Here that temporal comparison is carried by the synapse's own two-timescale dynamics: a volatile state relaxes toward the product of pre- and postsynaptic activities while a non-volatile state stores the learned parameter, and the volatile state's change is transferred to the non-volatile one through a global gate. On MNIST with a single-hidden-layer network, the phase-based and continual forms reach test errors of 2.01% and 2.24%, against 2.00% and 2.28% for standard EqProp, within one standard deviation.

  4. arXiv

    A CT foundation model built for cancer patients' chest-abdomen-pelvis scans lifts AUC to 0.767 from 0.695

    CAPTOR, a vision-language foundation model pretrained on 25,512 chest-abdomen-pelvis CT-report pairs from cancer-focused exams, outperformed existing CT foundation models in its target setting. Disease classification reached an AUC of 0.767 versus 0.695 for the next-best model on the full chest-abdomen-pelvis subset, and 0.784 versus 0.734 on the oncologic subset. It encodes organ-centered crops of whole exams but makes study-level, not lesion-level, judgments.

  5. Swap one Schelling agent for an LLM and the segregation trajectory holds, but runtime climbs from 2 seconds to 8,917

    In a Mesa implementation of the Schelling segregation model, one agent's symbolic rule for classifying neighbors is replaced by a locally served Qwen3.5 model that classifies through tool calls, while the rest of the dynamics stays fixed. Statistical model checking with MultiVeStA shows the expected ratio of happy agents rising from about 0.65 to about 1.0 within 60 steps in both variants, with point-wise t-tests finding no difference, yet the LLM version takes 8,917 seconds against about 2 seconds. The 0.8b and 2b models show semantic errors, wrong tool-call counts or stuck generation in unit-style checks, while 4b and 9b pass all of them.

  6. arXiv

    Speech LLMs Hear Age and Gender but Misjudge When to Use It: Explicit Labels Lift Personalization and Stereotyping Together

    Eight speech LLMs mostly recognize a speaker's age and gender, yet rarely turn that perception into requested personalization, and adding explicit demographic labels raises personalization while also increasing stereotypical choices.

  7. arXiv

    Set any loudness you want: continuous SPL control hits 83% of the target change in WavLM features

    Vocal intensity conversion no longer relies only on a reference utterance or a few discrete categories; it now takes a continuous target sound pressure level in dB SPL. A converter trained as a conditional cycle-consistent GAN works in the feature space of a frozen encoder/decoder pair without parallel data. With WavLM layer 6, it achieves 83% of the target intensity change, and listeners perceive increasing loudness with the target level.

  8. arXiv

    Site-level trade-off between stochastic and nearest rounding in low-precision transformer inference: SR in the MLP with RN at the head brings perplexity to 1.10x the full-precision reference

    The work extends the PRISM vectorized rounding library to arbitrary virtual precision via a variable-precision stochastic rounding (VPSR) algorithm and, holding the numerical format fixed while varying only the rounding rule at individual operation sites, emulates DistilGPT-2; it supports the site-level trade-off with a probabilistic forward-error bound for linear projections and a second-order decomposition of expected cross-entropy at the softmax, finding that at t=6 significand bits SR in the MLP gives 1.15x the full-precision reference perplexity versus 2.21x for RN, that the ordering reverses at the language-model head, and that a mixed-precision configuration (MLP output at t=6, SR in the MLP, RN at the head) brings perplexity within 1.10x, a 28% reduction over matched-bit RN.

  9. arXiv

    Treating EEG masking geometry as the only variable: 58 pre-trained models point to a moderate spatial radius with short temporal blocks, and expose a JEPA-specific bias-inflation collapse

    The study unifies EEG self-supervised masking strategies into a three-parameter framework of spatial radius, temporal length and mask ratio, trains 58 models under a fixed REVE-Small backbone and corpus across MAE and JEPA, evaluates them with a linear probe on the 12 OpenEEGBench datasets, and finds that both frameworks agree on a moderate spatial radius with short temporal blocks as the optimum while identifying a JEPA-specific bias-inflation collapse at full-channel masking.

  10. arXiv

    IMaX swaps the marginal entropy term of mutual information for a Tsallis α-entropy, lifting accuracy by up to 7.3 points on ESCA and retinal datasets under long-tailed semi-supervised domain generalization

    The work shows that semi-supervised domain generalization methods such as FBCSA and DGWM degrade substantially under long-tailed class distributions, and proposes IMaX: maximizing the mutual information between learned features and latent labels under supervision constraints from labeled samples, while replacing the standard marginal entropy term with a Tsallis α-entropy to relax the class-balance assumption; across two modalities (ESCA histopathology and diabetic retinopathy grading), three SSL frameworks and three SSDG methods, IMaX improves accuracy in all but one setting, with gains up to 7.3 points in the low-label regime (mL=5).