Search and filter through extracted claims from AI researchers.
Showing 2801-2820 of 4537 claims
Building socially calibrated LLMs requires more than reducing sycophancy - models must distinguish when to incorporate others' perspectives from when to maintain well-grounded moral judgment.
OpenForgeRL enables training harness-based agents end-to-end at scale by decoupling training and inference through a lightweight proxy and Kubernetes orchestrator.
Complex inference harnesses like Claude Code, Codex, and OpenClaw make AI agents hard to train end-to-end because standard SFT/RL stacks cannot natively express stateful, multi-process harness inference.
On-policy self-distillation is a promising approach that removes the need for external teachers required by traditional on-policy distillation.
Visual Contrastive Self-Distillation (VCSD) can perform on-policy self-distillation without requiring privileged answers or visual evidence, using only input conditioning through image-content removal.
Block Attention Residuals can boost deep-layer effective rank by approximately 12% by routing completed block summaries into later linear layers
Hybrid Linear-Softmax Attention combining gated linear attention for O(N)-dominated mixing with periodic gated-softmax anchors can avoid quadratic attention while restoring full-rank token interactions that pure linear attention lacks
SANA-Video 2.0 can generate high-quality video up to 720p on a single GPU while matching full-softmax video DiTs in quality and retaining favorable long-sequence scaling of linear attention
Cross-modal on-policy distillation can transfer reasoning capabilities from text teachers to audio-language students, extending beyond textually recoverable content to reasoning grounded in non-linguistic events, prosody, and conversation
Large audio-language models still lag behind text-based large language models in deep logical reasoning, primarily due to the scarcity of high-quality audio reasoning data
Uncertainty-weighted optimal transport can ensure robust knowledge transfer by dynamically weighting feature-level alignment based on prediction confidence to suppress noisy supervision
Existing cross-modal knowledge distillation methods struggle with large modality gaps and the propagation of noise from uncertain source-domain predictions
Multimodal approaches often outperform single modality approaches in downstream tasks because different modalities provide complementary information
Diffusion-based reconstruction combined with occlusion-type identification can address the limitation of degraded iris recognition when substantial texture is corrupted
Iris recognition performance degrades when discriminative iris texture is partially occluded by eyelids, eyelashes, specular reflections, or other acquisition artifacts
Expectations about AI security threats were that open weight models would be more vulnerable than closed weight models, but reality proved otherwise
Understanding the scaling behavior of parameterized quantum circuits, particularly how performance on unseen data changes with parameter count, remains a central challenge in quantum machine learning.
Circuit universality in language models has no single answer - it depends on whether you're asking about which concepts, where they're processed, or how they develop
What concepts earn dedicated circuitry is determined by the task, while where circuits are located and how they grow across layers is determined by the model architecture
Independently trained language models agree on which grammatical concepts receive dedicated circuits (Spearman ρ = 0.638 for Python, 0.673 for Rust)
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 9,017 pending.