Search and filter through extracted claims from AI researchers.
Showing 1321-1340 of 2685 claims of type "fact"
LLMs exhibit three systematic biases when simulating human belief updates: overrepresentation of neutral positions, more frequent but smaller belief shifts than humans, and failure to rank comments accurately
Speech deepfake detection systems can improve generalization through training recipe enhancements (multi-source data, attack-balanced sampling, audio augmentation) rather than additional architectural complexity
15.3% of FAIL verdicts in computer-use agent benchmarks are wrong, with 10.7% being evaluator false negatives and 4.7% being broken tasks
For computer-use agents, verification/feedback and planning failures dominate execution/grounding errors
Treating multimodal agent trajectories as provenance-constrained state machines with structured evidence ledgers enables better grounding verification
Pairwise graph formulations are insufficient to capture multi-way cross-modal dependencies in video misinformation detection, whereas hypergraphs offer a suitable representation
HyperClaim framework enables better localized authenticity detection by constructing sparse heterogeneous hypergraphs over query tokens, evidence tokens, and sampled frames
A symmetry-based experimental framework can systematically compare model preferences across representation types by reducing confounding factors
Concept Activation Vectors extracted via Sparse Autoencoders can explain image similarity in a model- and metric-agnostic way
Hierarchical Multilevel Monte Carlo neural critics can achieve the bias of long optimization runs with only logarithmic expected sample cost, resolving the bias-cost bottleneck
Finetuned graph language model retrievers generalize better out of domain than GNN-based retrievers, achieving SOTA on two multi-hop benchmarks
Adaptive Anticipatory Policy Trees improve agent success rate from 0.50 to 0.79 within contested decision windows by pre-computing conditional policy trees during idle periods
QuantWAMs framework can align quantization decisions with calibration context for World Action Models through shared-basis outlier calibration, co-training-objective saliency, and fixed-intervention rollout auditing
PACE framework extends the use of LLMs in finance from what to trade to how to execute, requiring neither explicit market assumptions nor task-specific training
PACE outperforms TWAP, Almgren-Chriss, and learning-based baselines by 0.65 bps on Shenzhen Stock Exchange Level-1 data
Windowed thinning achieves query complexity of O(κ^(1/2)d(d log κ + log 1/ε)) gradient queries for bouncy particle sampler and O(κd^(1/4)(d log κ + log 1/ε)) full-gradient equivalents for Zigzag
Static structured pruning methods for LLMs cause substantial accuracy degradation under aggressive sparsity due to input-agnostic computation allocation
Dynamic sparsity methods improve quality retention by adapting computation to individual inputs, but their practical acceleration under real-world inference scenarios remains challenging
WIDE is the first end-to-end differentiable token-level dynamic width pruning framework designed for both prefill and decode scenarios
Token-level dynamic width pruning enables fine-grained computation allocation by allowing each token to dynamically select attention-head groups and FFN-channel groups at neuron-block-level granularity
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 8,986 pending.