Skip to main content
Search

University of California, Santa Cruz

8 related articles

Subscribe to updates
  1. Spectral bias lives in the training kernel, not the wavenumber basis: dense operator learners spread error almost evenly across modes

    Training five operator models on the one-dimensional discrete Poisson problem shows that error is not learned wavenumber by wavenumber but is retained along small-eigenvalue directions of the kernel that governs the training update. Sine-aligned spectral and linear Fourier neural operators give each wavenumber its own rate, so error collects at high wavenumbers; dense networks have eigendirections close to random in the sine basis, so error is nearly equal at every wavenumber and relative error grows only because the true solution spectrum decays. Removing the minibatch cross terms leaves the GELU dense network with under 0.1% of squared error below ηλ≈8×10⁻⁵ under either forcing.

  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. arXiv

    A 27B model learns to propose language-model upgrades: 200 real training runs beat GPT-6-sol on held-out tasks

    Deciding which change is worth trying when a language model improves a language model has depended on costly training experiments. LM2LM uses a frozen LLM as a research world model that predicts a proposal's gain from real experimental records in context, turning those predictions into reinforcement-learning rewards, and periodically executes the current policy's top-rated proposals to write measured outcomes back into the records. Across eight autoregressive and diffusion language-model development environments, Qwen3.8-27B trained this way outperforms GPT-6-sol on held-out tasks, even when GPT-6-sol is given 300 real experimental records in context, at about 33% lower data-generation cost.

  4. arXiv

    PEV, which adds Gaussian noise to prompt embeddings, reaches a 100% jailbreak rate on JailbreakBench across six open-weight LLMs

    The work introduces the Perturbed Embedding Vector (PEV) jailbreak: adding independent Gaussian noise to a prompt's embedding and resampling it repeatedly, with no gradient computation and no weight modification, achieves a 100% jailbreak rate on all 100 JailbreakBench prompts across six open-weight LLMs, with the first successful jailbreak typically arriving within one minute and at lower compute cost than the compared prior methods.

  5. arXiv

    AEGIS couples Mars atmospheric physics to a differentiable Dinosaur core, yielding stable ten-Mars-year integrations that reproduce the seasonal CO2 cycle

    The authors present AEGIS, a differentiable Mars climate model that couples Mars radiation, surface, CO2 frost, and regolith physics to the Dinosaur spectral dynamical core; at T21/L12 it completes stable ten-Mars-year integrations, reproduces the seasonal CO2 cycle while conserving the total atmosphere–frost CO2 inventory to very small drift, produces surface pressure that follows MOLA topography, and yields automatic-differentiation gradients that agree with finite differences for physical calibration and neural-closure training.

  6. arXiv

    Bringing survey small-area estimation into AI evaluation: PP-S and PP-TS improve point and interval estimates on a benchmark and on deployed traffic, while DB-CV picks as well as an independent validation sample from one sample

    The work frames disaggregated AI evaluation as finite-population survey sampling, proposes prediction-powered smoothing (PP-S), a Bayesian model fit to each domain's prediction-powered estimate, plus a taxonomy extension (PP-TS) that borrows strength along a nested reporting hierarchy, and derives an approximately unbiased design-based cross-validation score (DB-CV); on the Open LLM Leaderboard (9,324 questions, 34 task types) and PRISM deployed traffic (68,371 rated responses, 21 LLMs by 3 conversation types), where every outcome is observed, the smoothed estimators beat direct estimators in point and interval estimation with near-nominal 95% coverage, and DB-CV selects as well as an independent validation sample at the same budget while estimating the chosen estimator's error far more ac

  7. arXiv

    ACLArena quantifies catastrophic forgetting across a four-stage post-training pipeline and proposes a LoRA expert mixture that lifts AIME26 from 10.21 toward independently trained experts

    The work builds ACLArena, a four-stage sequential post-training pipeline spanning math, search, e-commerce, and instruction following; diagnoses forgetting and transfer at both the model level and the token level; systematically compares multi-teacher on-policy distillation (MMOPD), self-distilled fine-tuning (SDFT), and model merging; and proposes Mixture of Low-Rank Experts (MLE), which first consolidates multi-domain trajectories into a shared backbone via SDFT, then freezes the backbone and trains one RL-optimized LoRA expert per stage routed by environment context, improving both in-domain and out-of-domain performance on four reasoning and agentic tasks.

  8. arXiv

    Spinverse inverts face permeabilities through a differentiable Bloch-Torrey simulator to reconstruct diverse microstructural interfaces on synthetic meshes

    Spinverse presents a permeability-aware microstructure reconstruction method: on a fixed tetrahedral grid it treats each interior face permeability as a learnable parameter, optimizes face permeabilities by backpropagating a signal-matching loss through a fully differentiable Bloch-Torrey simulator, and recovers an interface by thresholding the learned permeability field; across a collection of synthetic voxel meshes it reconstructs diverse geometries and shows that sequence scheduling and regularization are critical to avoid outline-only solutions while improving boundary accuracy and structural validity.

You’re all caught up