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 benchmarked chunking strategies on document revisions to measure re-embedding costs and retrieval quality for changing data.
Academic research demonstrates automated safety classifiers exhibit dialectal and cultural bias, mislabeling minority language as toxic.
Researchers examine limitations in pairwise evaluation for reward models, showing prompt-shared components go untested.
Researchers propose a matched protocol to test whether giving language models tools makes benchmark comparisons more stable across rewordings.
An arXiv study demonstrates that LLM-as-a-judge prompting techniques like Chain-of-Thought lack robust transfer across languages and judges.
Researchers analyzed LLM multi-cultural value selection biases using World Values Survey probes across implicit and explicit country identity cues.
Research on memory-aware reward modeling explores how LLMs handle user continuity and personalization preferences across sessions.
Researchers identify 'State Integrity Collapse' in LLMs, where models retain explicit instructions in memory but fail to apply them.
Researchers find Qwen2.5-7B's same-person entity matching accuracy drops from 35.1% in Latin script to 8.5% in Devanagari under Marathi prompts.
Researchers propose an agentic auditing method using budgeted APIs to debug multilingual web corpora and compile executable tests.
An arXiv survey maps 118 RAG defense papers by measured endpoint, highlighting fragmented metrics and inconsistent threat models.
Researchers introduced Agent Mosaic, demonstrating how LLM agents aggregate non-sensitive documents to disclose confidential conclusions.
Researchers introduced Steward, a multi-agent LLM system using a typed core to enforce safeguards during autonomous investment decisions.
An arXiv paper demonstrates that hybrid retrieval using RRF can harm accuracy on specific individual queries compared to single retrievers.
Researchers introduce the SC-matrix to distinguish genuine competence from sycophancy in multilingual language models under pressure.
Research on reasoning models shows the word wait is often inherited from training data rather than newly discovered through reinforcement learning.
Researchers introduced SCOPE, a neurosymbolic method compiling natural-language policies into predicate programs for agent safety evaluation.
Researchers introduced AfterDeleteBench to audit how LLM agents retain deleted memory data through derived states and summaries.
Researchers propose a test-time reasoning optimization method to prune candidate solutions, balancing quality and compute costs.
Research paper examines how mathematical reasoning capabilities transfer across languages in English-centric large language models.
Researchers examine open-source LLM capabilities for online influence campaigns using unmodified checkpoints.
Researchers propose soft-prompt tuning to automate LLM evaluation and overcome formatting penalties in base models.
Researchers propose A-B-D to infer human-like behavioral traits in LLM agents, addressing gaps in self-reports and LLM-judge ratings.
A new paper demonstrates scaling hard-negative reranker training to improve document retrieval quality in large candidate pools.
Research paper arXiv:2610.25123 explores dependency-tethered memory challenges for tool-using LLM agents when underlying premises change.
Researchers propose GFlowPO, a probabilistic prompt optimization framework designed to improve sample efficiency for language model prompts.
Researchers introduced adaptive inference routing using early decoding entropy to determine when chain-of-thought reasoning is cost-effective.
Researchers propose ASR_STEP, a cost-aware metric penalizing repeated model calls during iterative LLM jailbreak attempts.
Researchers propose Specter, an inference system combining quantization, speculative decoding, and expert offloading for MoE models.
Research shows LLM agents using persistent memory trigger a self-reinforcing loop where users disclose more personal data over time.