Researchers at the AI4TCI Workshop presented a method to protect behavioral privacy in autonomous negotiation agents, using randomized policies to reduce adversarial inference accuracy by 43–50% while…
Researchers introduced Self-Reflective Program Search for Long Context (SRLM), a framework that uses uncertainty-aware self-reflection to improve long-context handling in language models. SRLM leverag…
Researchers introduced Temporal Global Policy Optimization (TGPO), a reinforcement learning algorithm that improves temporal awareness in multimodal large language models for egocentric video understa…
Researchers introduced a training-free diagnostic framework to evaluate on-policy distillation for reasoning models, finding that distillation guidance aligns better with ideal gradients on incorrect …
Researchers at an undisclosed institution discovered that safety alignment in large language models can be bypassed by manipulating a single neuron, without training or prompt engineering, across seve…
Researchers propose a Cross-Referential Rewriter (CRR) caption framework to improve Text-to-Sounding-Video (T2SV) generation by addressing text conditioning bottlenecks and cross-modal fusion challeng…
Researchers propose LensVLM, an inference framework that enables Vision Language Models to scan compressed images and selectively expand only relevant ones, achieving up to 10.1× effective compression…
Researchers from UCLA, UT Austin, and Lambda, Inc. introduced MT-EditFlow, a reinforcement learning framework for multi-turn image editing that uses flow matching to optimize reward signals. The frame…
Researchers introduced Weblica, a framework for creating scalable and reproducible web environments for training visual web agents. By using HTTP-level caching and LLM-based environment synthesis, the…
Researchers at Google and Apple found that continuous diffusion spoken language models (SLMs) exhibit scaling laws similar to autoregressive models, with validation loss and phoneme Jensen-Shannon div…
Researchers at Apple introduced Annotator Policy Models (APMs), interpretable models that learn annotators' internal safety policies from labeling behavior alone, enabling diagnosis of disagreement so…
A study by researchers including Aneesh Pappu and James Zou found that self-organizing multi-agent LLM teams fail to match the performance of their best individual expert, with losses up to 41.1% on M…
Researchers from Apple propose amortized maximum inner product search (MIPS) using neural networks to predict optimal keys, significantly improving efficiency on the BEIR benchmark. The method trains …
Apple researchers discovered that a single parameter, termed a 'super weight,' can disproportionately determine an LLM's behavior, with removal causing catastrophic quality loss. The finding enables e…
A new study finds that panels of nine large language model judges effectively provide only about two independent votes' worth of information, because the models make correlated errors on the same item…
Apple is presenting new research at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026, taking place in person at the Colorado Convention Center in Denver from June 3 to Ju…
Researchers have introduced VSAS-Bench, a new benchmark for evaluating visual streaming assistant models that process continuous video frames in real time. The framework provides over 18,000 temporall…
Researchers at BalCapRL introduced a balanced reinforcement learning framework for multimodal large language model image captioning that jointly optimizes utility-aware correctness, reference coverage…
Apple hosted a two-day Workshop on Privacy-Preserving Machine Learning & AI in early 2026, bringing together its researchers and the broader academic community to discuss advances in private learning,…
Researchers at Duke University introduced Reward-Variance Policy Optimization (RVPO), a risk-sensitive alignment method that penalizes inter-reward variance to prevent language models from neglecting …