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 introduced StealthRL, a reinforcement learning framework designed to bypass AI-text detectors using adaptive paraphrasing.
Researchers propose VLMShield and the MAFE framework to defend vision-language models against malicious prompt attacks during visual integration.
Researchers benchmarked tabular foundation models against gradient boosting and logistic regression for credit risk default prediction.
Researchers introduce O'Prior to improve synthetic pretraining distributions for tabular foundation model robustness.
Researchers introduced Conformal Elo Estimation to calibrate LLM-as-a-judge rankings and quantify uncertainty against human annotations.
Researchers introduced AU-Harness, an open-source toolkit designed for efficient, standardized evaluation of large audio language models.
An empirical study benchmarks 16 deep learning forecasting architectures on 15 years of futures data for risk-adjusted returns.
Researchers argue generative model watermarking should be reframed as a monitoring primitive rather than solely a security tool.
Researchers propose a state-space encoding method for linear-time context compression to reduce transformer inference latency.
Researchers propose TraceQuery to reduce diffusion language model inference costs by selectively recomputing token representations.
A new arXiv paper introduces Recursive Semantic Divergence to measure non-deterministic behavioral drift in LLM agents across runs.
Researchers introduced SLICE-Mem, a lightweight memory framework designed to reduce LLM agent memory admission and extraction costs.
Researchers propose parametric session states to recover retrieval intent in long-running assistants without replaying interaction history.
Researchers introduced WhoWins, a benchmark testing how LLM agents handle conflicting instructions from system prompts, tools, and memory.
A new independent audit evaluates low-cost, frozen classification models for binary content moderation at scale.
Researchers propose FlexA-Mem, a budget-adaptive memory architecture designed to handle varying retrieval and step-tracking needs for LLM agents.
Researchers propose a method to transfer representational signatures for detecting LLM reliability issues without domain-specific data.
Researchers introduce a French surface-variant benchmark measuring how typos and missing diacritics affect LLM accuracy and reliability.
Researchers demonstrate cross-model representational alignment across five independently trained models using activations alone without weight merging.
Researchers propose a learnable retriever architecture to improve graph-based retrieval-augmented generation and structured knowledge grounding.
A new reinforcement learning technique addresses reward hacking in rubric-based models by balancing recall with a precision objective.
Researchers propose joint training strategies to balance single-pass and looped inference performance in parameter-efficient transformers.
Researchers introduced SEDATE, a method investigating whether LLMs share cross-lingual neural mechanisms for aggressive or unsafe outputs.
Researchers propose Trajectory2Gym to improve tool-use agent supervision via compositional coverage and decision grounding.
A new meta-study categorizes hallucination detection benchmarks across provenance, typology, and granularity to improve model evaluation.
Researchers introduced RSI Arena, evaluating ten frontier models acting as training agents for autonomous model post-training.
Researchers release a systematic study on how instruction design, few-shot prompting, and backbone models affect LLM-based translation evaluation.
Researchers introduced TriageRAG, a self-training method for adapting retrieval-augmented language models using unlabelled corpora.
Researchers introduce IntentDelta, a human-constructed benchmark for evaluating multi-turn LLM agent adaptability under evolving user intent.
Researchers propose a global citizenship values alignment method to address hidden safety risks in persona-conditioned LLM prompts.