Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Researchers propose a Bayesian computational framework, the 2-Step Agent, to model how humans learn from machine learning decision support.
Researchers introduced Lean4Agent, a formal verification framework using the Lean 4 theorem prover to verify multi-step LLM agent trajectories.
Researchers introduce Varest, an image-token compression method using variable-resolution tokenization and relevance pruning.
Researchers propose SkillEval, a cognitive diagnosis framework for cross-model, skill-level evaluation of large language models.
Research introduces an audit framework using counterfactual markets to evaluate whether reinforcement learning LLM agents enforce price collusion.
Researchers examine how large language models encode symbolic legal factors and case-based reasoning patterns in structured text.
A new benchmark evaluates 11 retrieval configurations across 476 long-document questions to compare full-document scanning with traditional indexing.
Researchers propose GRAFT to address sampling limitations in continuous diffusion language models.
Researchers introduced DiagSafe, an LLM safety evaluation framework using prompt-level rubrics instead of binary safe/unsafe labels.
Researchers introduce MOSAIC, a benchmark for evaluating long-term memory and cross-modal reasoning in conversational agents.
Researchers introduced SearchWeaver, a framework for automatic search interface optimization to improve agentic search performance.
Researchers propose Contrastive Attribution Topology Search (CATS) to identify optimal multi-agent system teams and communication pathways.
Researchers introduced Signals and Queries, a lightweight filtering method designed to reduce token costs for user-facing LLM agents.
Researchers propose memory-centric continual learning approaches to help autonomous AI agents learn from long-term deployment experience.
Researchers introduce DynaWeave, a framework for synthesizing dynamic environments to train LLM agents against unexpected external events.
Researchers introduce Omnilingual-MsASR, a multimodal and multilingual benchmark evaluating long-form multi-speaker speech recognition in omni LLMs.
New research challenges the conventional dichotomy of temporal arithmetic and temporal semantics in large language models.
Researchers propose RCSD, a reinforcement and continual learning method to improve draft model acceptance rates in speculative decoding.
Researchers propose PACE, a 4-bit recurrent state quantization method for linear-attention language models to improve memory efficiency.
An academic paper analyzes how model scale and quantization affect social bias and abstention behaviors using the BBQ benchmark.
Researchers propose Uni-Mem, an end-to-end online reinforcement learning approach for unified agentic memory management.
A new arXiv paper analyzes how combining chunking, reranking, and verification in RAG pipelines impacts real performance.
Researchers propose an operation-level measurement framework to bridge foundational and agentic capabilities in large language models.
Researchers introduce a framework to study recursive self-improvement in AI systems operating within continuously changing environments.
Researchers propose MAPS, a macro-adaptive parallel reasoning method to regulate search momentum in retrieval-augmented generation systems.
Research paper MemoryTax evaluates agent memory systems versus raw history search across coding agents and conversation tasks.
Research shows non-native grammatical errors in prompts alter social bias evaluations in large language models.
A new paper evaluates how fine-tuning LLMs for specific value alignment creates unintended side effects on unrelated model capabilities.
A new arXiv preprint proposes graph topology and energy-driven pathfinding to improve RAG contextual consistency in multi-hop QA.
Researchers propose a provenance-preserving hypergraph memory architecture to maintain visual evidence and entity evolution in multimodal agents.