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 GARG-AML, a graph-based framework designed to detect smurfing and money laundering patterns in transactional networks.
Researchers introduced Safhire, a hybrid framework using fully homomorphic encryption to enable private cloud ML inference.
Researchers introduced AuditKIT, an evaluation library unifying benchmark accuracy, judge assessment, and agentic trace tracking.
Researchers introduced SAG2, an embedding-free multi-hop retrieval architecture using BM25 and a SQLite database to lower infrastructure overhead.
Research demonstrates that multi-agent LLM systems are vulnerable to prompt injection, with attacks succeeding in up to 61% of cases.
Researchers propose a unified taxonomy and efficient model architecture for PII detection to lower deployment costs on high-volume traffic.
Academic research demonstrates that LLM uncertainty metrics indicate RAG hallucination risk only when retrieved evidence covers the answer.
Researchers propose trajectory-guided steering to mitigate spurious feature bias in clinical large language model classification.
Researchers introduced JUPE, a benchmark showing plausible textual explanations can mislead multimodal LLM judges on visual tasks.
Researchers introduced ToMStories, an evaluation dataset of 6,000 synthetic vignettes testing language model theory-of-mind inferences.
Researchers introduced a mechanistic uncertainty concept to quantify LLM reliability by measuring internal model intervention thresholds.
Researchers propose an evidence-gated bilingual RAG framework to bridge consumer terminology and banking product pages in Arabic and English.
Researchers propose MFM, a multimodal filesystem memory architecture for long-horizon agents to retain visual and spatial details.
Researchers propose BRIA, a method for LLM agents to explore and evaluate large tool libraries independently during execution.
Researchers propose TokenEconomizer, an experience-driven method to control inference token costs for coding agents over sequential tasks.
Researchers propose harness-policy optimization to jointly evolve LLM coding agents and their environment interaction harnesses.
Researchers proposed a joint-embedding pruning method to remove privacy-leaking neurons from fine-tuned LLMs without degrading performance.
An arXiv study reveals that position bias continues to undermine the reliability of retrieval-augmented generation and model evaluations.
Researchers proposed a per-instruction defense mechanism for LLMs processing external data to mitigate indirect prompt injection attacks.
Researchers introduce NLData2Opt to benchmark LLM performance on noisy, real-world optimization problem formulation from text and data.
Researchers propose a competence-based evaluation framework to better measure logical reasoning limitations in large language models.
Researchers propose FederatedBytes, a tokenizer-free, byte-level small language model designed for efficient multimodal federated learning.
Research on ASR transcript normalization shows that removing pauses and punctuation discards pragmatic signals for dialogue act classification.
Researchers find that self-evolving LLM agents can fix failures but struggle to generalize from them without positive anchors.
Researchers propose Prism Attention, a new index sharing mechanism designed to reduce KV cache memory bottlenecks in long-context LLMs.
Research shows 41% of US households shift stated risk tolerance over time, leading LLM advisers to produce inconsistent stock advice.
A paper finds seven open-weight models execute harmful requests via tool calls 74.5% of the time without retrieval safety checks.
Research identifies Safety Loss through Interlingual Processing (SLIP), where multilingual retrieval agents bypass safety guardrails.
Researchers introduced a low-overhead method to measure LLM context reliance and detect hallucinations using the unembedding dark space.
Researchers propose MGE-PO, an entropy-aware credit assignment method that shapes training rewards using rollout generation uncertainty.