Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Choice2Seq benchmark reveals LLMs fail to construct correct answers sequentially despite selecting them correctly, measuring Sequence Composition Rate across arithmetic and retrieval tasks.
Efficient exploration strategies for Gaussian process Bayesian optimization with inference regret guarantees for noisy evaluation environments.
SLKR framework performs zero-shot text classification via label-centric retrieval-augmented generation, separating offline LLM knowledge construction from lightweight online inference.
datasette-ports 0.4 released with PID detection support for Datasette instances.
ProCARE automates real-world clinical studies using LLM agents with profile-grounded evidence contracts to handle dataset-specific healthcare semantics and temporal constraints.
EventLens uses reinforcement learning to improve multimodal LLM video understanding of temporal event structures, boundaries, and multi-event relations.
DeepThink proposes knowledge-addressable recurrent language modeling framework with typed Knowledge Units, activation selection, and modular expert design.
Interactive 4D volumetric liquid forecasting task predicts spatiotemporal fluid dynamics conditioned on solid object motion for robotics and digital twin applications.
Training-free latent communication between LLMs via direct KV cache sharing eliminates text bottlenecks and prefill latency in multi-agent coordination tasks.
Fisher information matrix analysis on CIFAR-10 vision models reveals how top eigenvectors decouple generalization from memorization under controlled noise.
Persona Evolution method with task-integrity constraints reveals LLM agent robustness weaknesses under diverse realistic user interactions beyond standardized simulators.
WildTrace evaluates training data detection in post-trained LLMs at realistic scale, addressing benchmark contamination and copyright concerns.
Theoretical study of equilibrium learning in stochastic games with unknown nonlinear dynamics using Gaussian processes.
BARTER reduces dynamic sparse training cost using standard dense kernels via structured parameter reallocation across layers.
Unravel converts single images to interactive 3D scenes with movable objects using agentic layer-by-layer decomposition.
Pollard provides framework-independent agent control via governed semantic execution trees with unified governance, replay, and resource tracking.
Hashrope unifies four LLM context management mechanisms (edits, branching, KV-cache reuse, segment encoding) via persistent weight-balanced rope.
Extends distributional regression to arbitrary response spaces using ReLU networks and event-based probability indexing.
Phantom Merge identifies entity-evidence binding failures in LLM agents and proposes two-stage framework for claim-level remediation.
Integrates LLMs with transportation planning via differentiable user equilibrium layer for natural-language network optimization.
Token-level Modality Alignment reveals omni-modal LLMs perceive but fail to express non-textual information (knows-but-cannot-say gap).
Active Probing Benchmark evaluates LLM multi-turn reasoning via generalized puzzle-solving with partial information and concept identification tasks.
XForecast KDD 2026 benchmark combines historical stock prices and financial news for 4-week ahead forecasting using multimodal FinTexTS dataset.
PID-Gated Continual Prompt Adaptation uses control theory to schedule when noisy feedback should update LLM agent instructions in non-stationary environments.
Localized benchmark evaluates 12 LLMs (proprietary and open-source) on Hong Kong fire safety code interpretation for AEC industry compliance.
Passive imitation framework localizes agent failures, extracts repairs via skill patches, and verifies full-task success across compositional tool-use tasks.
ScaleMAS uses substrate-centric coordination to address scaling failures in LLM multi-agent systems by centralizing state on shared medium instead of agent workflows.
FramePack-U training-free frame packing reduces O(n) memory scaling for hour-length video understanding in VLMs via constant-memory compression.
Study measures LLM multi-agent collaboration robustness under realistic unreliable colleagues who provide incomplete/incorrect/delayed responses.
Regional photovoltaic generation field forecasting improves ultra-short-term solar prediction by modeling spatial fleet dynamics instead of discrete time series.