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 released an open bilingual benchmark and engine for reversible document pseudonymisation before sending data to external LLMs.
Researchers analyze GEPA prompt optimization instability, showing merge operator performance varies significantly between GPT and Qwen models.
Academic evaluation shows dedicated AI agent memory systems often underperform naive RAG systems across standard benchmarks.
Research across 150 models reveals identically trained LLMs with different random seeds fail on entirely different jailbreak attacks.
Researchers propose T2D, a target-to-draft model pruning and adaptation method to improve inference efficiency in speculative decoding.
An arXiv paper proposes training-free semantic trees to structure context for long-context models rather than relying on raw retrieval.
Researchers introduce DeltaTau-Bench, an executable benchmark evaluating agent performance across evolving and dynamic tool ecosystems.
Researchers introduce PerspectivesQA to evaluate how LLM answers vary across languages on subjective and contested queries.
Research demonstrates RAG systems are vulnerable to stance poisoning, where factually correct but one-sided context skews model outputs.
Researchers introduced STEP, a training-free method using step-by-step probability scoring to estimate LLM prediction accuracy.
Researchers propose a synthetic pipeline for early tool calling in latency-sensitive spoken dialogue agents.
Researchers introduce DecisionTraceBench, a new speech state-tracking benchmark evaluating multi-party conversational information evolution.
Researchers introduce a multi-level evaluation framework testing whether language models can apply theory of mind to verifiable tasks.
Researchers propose DynEdit, an entropy-guided sequential method for editing long-sequence knowledge in large language models.
Researchers examine how prompt position and attribution influence whether language models abandon correct answers under conflicting advice.
Researchers propose semantic atoms to streamline text analysis workloads within LLM-based agent data analytics pipelines.
Researchers examine conformal factuality in agentic RAG where language-model policies adaptively issue follow-up retrieval queries.
Researchers introduce RubriX, an encoder-based framework for evaluating rubric levels that reduces inference costs for LLM scoring.
Researchers introduced NumTrace, a benchmark to audit numerical claim provenance and financial semantics in financial analyst agents.
Researchers propose SVD-Attention, a post-hoc method to decompose and score per-head singular modes in transformer self-attention blocks.
Researchers identify Conflict of Skills (CoS), a vulnerability where composition of safe LLM agent tools causes policy violations.
Researchers propose QERA, a question-driven evidence reasoning architecture addressing frontier LLM failures on complex real-life context tasks.
Researchers introduced CalibraRL, a reinforcement learning method using query difficulty and confidence to reduce RAG hallucinations.
Research on arXiv examines linguistic bias in LLM-based peer review as automated evaluation tools see wider academic adoption.
An independent safety evaluation assesses DeepSeek V4.1 Flash-1017 for loss of control and harmful manipulation risks in open-weight AI.
A new arXiv paper develops a unified statistical framework for evaluating LLM self-confidence and LLM-as-a-judge approaches.
Researchers propose SEMA-Auto, an autonomous multi-turn adversarial attack method for LLMs bridging static and interactive jailbreaks.
An arXiv paper identifies claim smoothing in LLM research agents, where diverse or conflicting sources are collapsed into overconfident summaries.
Research on coding agents indicates providing whole files outperforms compression and retrieval techniques when fixing code issues.
Researchers introduced weak-to-strong on-policy distillation to mitigate strategic agent collusion in simulated insider trading environments.