Tag: #v1120 results
Transformer training on modular arithmetic tasks collapses high-dimensional parameters to 3-4D execution manifolds. This structure explains attention concentration, SGD integrability, and sparse autoencoder limits. Core computation occurs in reduced subspaces amid overparameterization.
NMRTrans uses set transformers on experimental NMR spectra for molecular structure elucidation, trained on NMRSpec corpus from literature. It models spectra as unordered peak sets aligning with NMR physics. Achieves SOTA Top-10 accuracy of 61.15% on benchmarks.
Integrates neural networks, topological data analysis, and Bayesian methods for AI in military domains. Covers image, time-series, graph applications like fraud detection. Emphasizes robustness and interpretability.
Inference-time scaling in language models leads to adaptive resource rationality without explicit cost rewards. Models shift from brute-force to analytic strategies as task complexity rises. LRMs show robustness on challenging functions like XOR/XNOR unlike IT models.
TokaMark standardizes AI evaluation on MAST tokamak data with unified multi-modal access and 14 tasks. Harmonizes formats, metadata, and protocols for reproducible comparisons. Includes baseline model; fully open-sourced for community use.
Text-guided framework enhances weakly supervised multimodal video anomaly detection. Employs in-context learning for anomaly text augmentation and multi-scale bottleneck Transformer for fusion. Achieves state-of-the-art on UCF-Crime and XD-Violence benchmarks.
Introduces δ_TCB metric to quantify LLM internal state robustness against perturbations, beyond traditional accuracy. Linked to output embedding geometry, it reveals prediction instabilities missed by perplexity. Correlates with prompt engineering in in-context learning.
LHAW generates controllable underspecified long-horizon tasks by removing info across goals, constraints, inputs, context. Validates via agent trials, classifying ambiguity impacts. Releases 285 variants from benchmarks.
SynergyKGC fuses entity semantics with heterogeneous topologies via cross-modal synergy. Uses density-dependent anchoring and double-tower consistency. Improves KGC hit rates on benchmarks.
Step 3.5 Flash is a 196B MoE model with 11B active params for agentic tasks. Optimized with sliding-window attention and MTP-3 for low-latency inference. Matches frontier models on math, code, and agent benchmarks.