Skip to main content
Search

University of Wisconsin–Madison

24 related articles

Subscribe to updates
  1. bioRxiv

    A single-cell simulator with a dial for batch strength: scDeepSim varies batch effects and trajectory geometry on demand

    Benchmarking single-cell methods has been limited by costly annotation of real data and by datasets that capture only one observed condition. scDeepSim encodes expression into latent subvectors assigned to cell type, batch and residual variation, learns their distribution with latent diffusion, and at generation time modifies only the target subvector, imposing batch effects or constructing branching trajectories at user-set strengths. On scIB human pancreas data at intervention strength α=1, the structured model reached a mean cell-type silhouette width of 0.519 versus 0.432 for a plain variational autoencoder.

  2. arXiv

    CyberArena puts AI agents in live attack-and-defense combat, and two frontier models pull ahead of the field

    CyberArena turns cybersecurity evaluation from static targets into multi-round combat: two agents in isolated sandboxes attack each other's services while patching their own, with objectives rotating every round. Across 40 challenges and 466 games, GPT-5.6 Sol and Claude Opus 5 led 12 frontier agents, and on 30 challenges all three compared models solved solo, round-three capture rates fell to 52%, 37% and 35%.

  3. Retrieval-augmented generation gets three separately optimized agents: TRIO splits query expansion, retrieval judging and answering

    Retrieval-augmented generation (RAG), which fetches external material before a language model answers, often falters on weak retrieval and fabricated content. TRIO splits that pipeline into three agents, for query expansion, for judging whether retrieval is sufficient, and for answering, and lets each improve iteratively under its own reward signal: the expansion agent is rewarded by gains in answer quality, the judge agent is supervised by unanimity-voted sufficiency labels, and the answer agent trains with dynamic matching-based rewards plus curriculum learning. Experiments across multiple benchmark datasets report TRIO as more effective than state-of-the-art approaches, and the code is public.

  4. bioRxiv

    Pooling guides hides off-target bias: majority-agreement weighting pulls Perturb-seq target estimates near zero

    Perturb-seq often pools several guides per target as replicates, but a guide's off-target effects can bias the target-level estimate. PETRO assumes a strict majority of guides are valid for each outcome and uses robust weighting to downweight disagreeing guides, recovering the on-target perturbation effect without removing guides or using sequence annotations. In RPE-1 CRISPRa target-outcome pairs whose longest seed match was 13–18 bp, mean pooled bias was 0.087 and PETRO's residual bias was 0.005; in simulation, pooling stayed near 0.075 as cells increased while PETRO fell toward zero.

  5. arXiv

    Sparse autoencoder features do not simply add up: pairwise terms lift coalition rank fidelity from 0.5874 to 0.7790

    Sparse autoencoders split model activations into sparse, linearly decoded features, but combining those features does not make their downstream effects add. The work introduces SAE Additivity Residuals (SAR), which measure the behavioral variation left unexplained above a chosen interaction order; across 18 model and behavior settings, learned SAE coordinates show lower first-order residuals than sum-preserving random rotations of the same subspace, yet the coordinates are not fully additive, and pairwise terms improve held-out coalition ordering in all 18 settings. On 16 published steering directions in Llama-3.1 8B, pairwise modeling raises rank fidelity from 0.5874 to 0.7790 and cuts ordering error from 0.1444 to 0.0606.

  6. medRxiv

    A radiology decision model that never writes text: open weights lift report-error detection from 72.1% to 94.7%

    The open-weight radiology decision model RadKev-27B raised task-averaged accuracy across fifteen radiology decision tasks from 75.9% to 80.8%, and report-error detection from 72.1% to 94.7%, while answering at roughly 1/132 the latency of a reasoning LLM.

  7. arXiv

    Chest X-ray reports stop being one long paragraph: align images to disease labels, then let an LLM rewrite them as Q&A

    Chest X-ray report generation moves to question-answer output: predicted disease labels align image and text representations, a medical knowledge graph supplies retrieved evidence, and a fine-tuned LLM rewrites the result, outperforming prior methods on most metrics on MIMIC-CXR and IU X-Ray.

  8. arXiv

    Installing Natural Values Transfers Only a Small Fraction of the Known–Unknown Abstention Contrast: Auditing LLM Features for Sufficiency Versus Use

    The work proposes an empirical contract that evaluates LLM internal features at values observed on natural inputs through installation, removal, and downstream rescue: installation measures how far a feature suffices for a behavior, while removal and rescue measure how much the model uses it. Applied to published entity-recognition latents, dense known–unknown directions, and a subject–verb agreement feature set, sufficiency and use strengths separate sharply, showing that tracking a concept and steering a behavior do not by themselves show that the model uses a feature.

  9. arXiv

    Pre-trained models with a light inference harness cover more solutions on agentic tasks, while post-training imposes a measurable sharpening tax

    The work finds that pre-trained LLMs with a light inference harness, despite far lower single-shot accuracy (pass@1), often surpass their post-trained counterparts in solution coverage (pass@K) given sufficient test-time budget; it proposes the Sharpening Tax metric to quantify post-training's loss in test-time scalability, shows across 14 base/post-trained model pairs from four families and three agentic benchmarks (42 cases) that the tax is prevalent, estimable from a few rollouts, and correlates with other metrics, and presents PTGS, a difficulty-adaptive Bayesian sampler that reduces the tax.

  10. arXiv

    Percentile requests as generation targets: method realizes 1,422 of 1,440 requests (98.75%) on highD

    The study defines a generated scenario's adversity as its percentile in the conditional distribution of future risk given the observed history, learns a reference risk distribution mapping each requested percentile to a physical risk target, and uses a percentile-conditioned joint diffusion model with sampling-time risk guidance to generate multi-agent futures; on highD with minimum ego-surrounding-vehicle post-encroachment time (PET) as the surrogate, the method realizes 1,422 of 1,440 requests (98.75%) within a 0.05 percentile tolerance, with mean percentile error 0.00673 and PET-target error 0.00991 seconds.