FOGO detects and resolves gradient interference via spectral orthogonalization and compact codebook memory to prevent dominant directions from suppressing rare updates, improving convergence and retention across continual and standard training.
MATO achieves training-free multi-objective LLM alignment via test-time optimization of discovered rewards and adaptive weights during decoding, improving steerability and Pareto performance.
RAIL introduces a CHC-based benchmark evaluating LALMs across five auditory cognitive abilities, revealing highly uneven performance among 26 state-of-the-art models.
IndustryCode is a multi-domain, multi-language benchmark of 579 industrial coding sub-problems; Claude 3.5 Opus reaches 68.1% sub-problem and 42.5% main-problem accuracy.
RankE co-evolves discrete text-to-image policy and decoder via alternating optimization to eliminate latent covariate shift, improving both FID and CLIP scores.
Existing domain unlearning overfits to seen classes; this paper proposes open-vocabulary domain unlearning via Fisher-masked parameter editing and targeted manifold scattering to erase domains across unseen classes with few shots.
OneVision-Encoder applies codec-aligned sparsity to video, processing only high-entropy regions to outperform dense backbones with fewer tokens. It achieves 4.1% higher video accuracy than Qwen3-ViT across 16 benchmarks.
BSO recasts safety alignment as density ratio matching via Bregman divergence minimization, yielding a single-stage loss that improves the safety-helpfulness trade-off without auxiliary models.