Skip to main content
Search

Washington University in St. Louis

9 related articles

Subscribe to updates
  1. arXiv

    Diffusion OOD detection stops trusting reconstruction alone: DUO adds a feature-sensitivity readout and reaches 0.897 average AUROC on 12 pairs

    Diffusion-model out-of-distribution detection often relies on reconstruction error, yet OOD inputs can be denoised just as well or better, so those scores overlap or even invert. DUO adds a second readout on a frozen diffusion model: the sensitivity spectrum of intermediate denoising features to small input perturbations, with the feature layer, noise levels and two-channel fusion calibrated on in-distribution data alone. Across 12 image pairs DUO reaches 0.897 average AUROC versus 0.838 for the strongest baseline, EigenScore, at 0.28 seconds per sample.

  2. arXiv

    No weight updates at inference: steering neurons per query lifts seven models from 3B to 32B

    Instead of a fixed intervention chosen in advance, Q-NeuS reads each query's own internal activations to decide which feed-forward neurons to amplify, improving every one of seven models from 3B to 32B parameters across three benchmarks.

  3. arXiv

    PAC-Private Diffusion Adaptation Perturbs the Learned Component Once, Beating DP-SGD on Subject Identity and Generation Quality at Equal Reconstruction Privacy

    The work proposes a PAC-private diffusion model adaptation that first learns a compact data-dependent component via LoRA or Textual Inversion and then calibrates anisotropic Gaussian noise from the covariance of repeated mechanism outputs, perturbing the learned component only once after optimization to avoid privacy composition across gradient updates; on few-shot concept personalization and full-dataset image synthesis it better preserves subject identity, generation quality, and downstream classification accuracy than DP while achieving the same reconstruction privacy.

  4. arXiv

    SRFT lets LLM agents learn from their own failures, cutting AgentDojo attack success from 7.59% to 1.26%

    The work proposes Self-Reflection Fine-Tuning (SRFT): it injects attack instructions into clean expert trajectories, samples the agent's own potentially failing actions under the attacked context, and has an expert model generate structured self-reflection reasoning that contrasts unsafe and optimal actions, which then supervises fine-tuning; instantiated as SR-Agent on Llama-3.1-8B-Instruct and Qwen3-8B, it reduces AgentDojo attack success rate from 7.59% to 1.26% and from 16.97% to 1.05% respectively, and keeps a low attack success rate under the RL-Hammer adaptive attack.

  5. arXiv

    MILO co-evolves agent harnesses and their search strategy, reaching 86.1% on Terminal-Bench 2.1 above the official leaderboard top and tightening three mathematical bounds

    MILO (Meta-evolutionary Island Orchestration) co-evolves an agent harness together with the strategy that discovers it: a hierarchical island-based lineage memory treats rejected mutations as negative evidence, per-island mutator agents rewrite complete harnesses, and an orchestrator adapts the search through lineage grafting, speciation, mutator reassignment, and curriculum revision; across Terminal-Bench 2.1, PaperBench, and DeepSWE it improves resolution over its initial harness by 12.0%, 28.3%, and 10.3% with Opus 4.8 (versus best prior-search gains of 4.5%, 18.3%, and 0%), reaches 86.1±2.0% on Terminal-Bench 2.1 (official leaderboard top 83.8±2.

  6. Cancer

    CIBMTR real-world data: about one-third of multiple myeloma patients receiving cilta-cel are frail, with 12-month PFS of 62.7% and OS of 72.8% and doubled ICANS risk

    Using Center for International Blood and Marrow Transplant Research registry data, this study evaluated 541 patients with relapsed/refractory multiple myeloma who received commercial cilta-cel between March 2022 and December 2023 and found that 183 (33.8%) were frail by an adapted simplified frailty score; overall response rates were comparable (82.8% vs. 88.5%), but frail patients had worse 12-month progression-free survival (62.7% vs. 75.9%) and overall survival (72.8% vs. 90.4%) and more any-grade ICANS (32.2% vs. 17.6%), with frailty independently predicting worse PFS (HR 1.67), OS (HR 2.46), and any-grade ICANS (OR 2.01) on multivariable analysis.

  7. A 40-day measurement of 55,393 trending queries finds Google AI Overviews activate on 13.7% of searches (64.7% for question-form queries) and that 11.0% of 98,020 atomic claims are unsupported by their cited sources

    Using a Puppeteer crawler on AWS Lambda, the authors ran a 40-day longitudinal measurement (March 13–April 21, 2026) over 55,393 trending Google queries across 19 topical categories and found that AI Overviews activate on 13.7% of queries overall (64.7% for question-form queries), that cited domains are more credible than co-displayed first-page results yet 29.8% do not appear on the first page at all, that 11.0% of 98,020 atomic claims are unsupported by the cited pages with omission as the dominant failure mode, and that at least 50.6% of cited pages carry display advertising.

  8. arXiv

    BioGait-VLM reaches 68.1% accuracy on an 8-class gait benchmark and is chosen as the better model in 69.2% of blinded expert cases

    The work proposes BioGait-VLM, a tri-modal (RGB vision, language, explicit biomechanics) framework built on a frozen InternVL3.5-1B, combining a Temporal Evidence Distillation branch with a Biomechanical Tokenization branch that projects 3D skeleton kinematics into language-aligned semantic tokens; on a subject-disjoint 8-class benchmark of 1,181 clips formed by merging public GAVD with a newly collected 30-patient degenerative cervical myelopathy (DCM) cohort, it reaches 68.1% accuracy and 52.9% macro F1 (98.1% F1 on DCM, 80.4% on Abnormal gait), and in a blinded review of 52 held-out clips four DCM-domain experts selected it as superior in 69.2% of cases (36/52, p<0.001).

  9. arXiv

    Regularizing the Self-Improvement of Agent Harnesses: How RRSI Makes Recursive Improvement Transfer

    The work introduces RRSI, which keeps an agent harness (prompts, control flow, tooling, memory and context management) fully editable while regularizing both the proposal and the selection of edits, gaining up to 14.1 points on the split it evolves against and up to 4.7 points on five out-of-distribution benchmarks across eight benchmarks in coding, agentic workspace and engineering design, while running on 30% fewer policy tokens than unregularized evolution.

You’re all caught up