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 analyzed token allocation in multi-agent and self-refinement AI systems operating within a fixed context window.
Researchers systematically evaluated confidence-based small-to-large model routing across three model families, four languages, and three task types.
Academic research on 33,000 trajectories reveals low prompt-injection success in small agents stems from instruction-following failures.
HydraLLM introduces adaptive width slicing for elastic text and vision-language models, enabling deployment across heterogeneous hardware.
Researchers introduce Munin, an open-source framework for self-hosted agentic scientific assistants using local document corpora.
A new research paper on arXiv introduces a progressive evaluation and training suite for time-series understanding using large language models.
Researchers propose SIT-Eval, a situation-based interactive turn-by-turn evaluation method for conversational AI agents across diverse scenarios.
Research on LLM medical verification shows models over-credit separate publications from the same clinical trial during evidence aggregation.
Researchers introduced DLite, a history-free speculative decoding method that prevents draft model performance drops in agentic AI.
Researchers study an elicitation agent that improves LLM-generated formal specifications through interactive clarification.
Researchers propose PRIME, an LLM agent framework designed to improve verifiable, evidence-grounded reasoning against complex organizational rules.
An arXiv study benchmarks Large Reasoning Models against standard non-reasoning models on everyday logical reasoning tasks.
A new arXiv study evaluates privacy leakage versus model capability when fine-tuning models on sensitive synthetic data using canaries.
Researchers propose TAG-Nav, a task-agnostic graph navigation framework providing mobile GUI agents with persistent app navigation memory.
Researchers propose NoLBERT, a news-derived sentence encoder pretrained without hindsight bias to test stock returns.
Researchers introduce EmoShift, a benchmark showing expressive speech synthesis with intra-utterance emotion shifts evades deepfake detectors.
Research demonstrates production LLM agents fail to adapt strategies dynamically in multi-agent collective-risk simulations.
Researchers propose ECHO, an optimization-based jailbreak attack method that uses decoded response harmfulness instead of compliance proxies.
Researchers propose tree-aware posterior distillation to address sparse credit assignment and policy staleness in long-horizon agentic RL.
A recent research paper examines the limits of allocating test-time compute for language models when verifiers are imperfect.
Researchers introduce Self-Reflective Chain-of-Thought (SRCoT), a closed-loop retrieve-reflect-revise pipeline for edge LLMs.
Researchers study interface drift in tool-augmented LLM agents, proposing a training-free repair layer for function-calling reliability.
Researchers proposed ScoutRoute, a lightweight post-hoc routing method for efficient Mixture-of-Experts systems across heterogeneous LLMs.
Researchers identify flaws in Vision-Language Model hallucination benchmarks that misclassify non-assertive mentions as hallucinations.
An arXiv pre-print evaluates Qwen3-TTS zero-shot voice cloning performance across diverse speech domains and conditions.
Researchers introduced RedDarwin, an agentic evolutionary framework designed to dynamically adapt jailbreak skills against LLM safeguards.
Researchers introduced VocalGrad to evaluate acoustic and paralinguistic perception in large audio language models.
Researchers propose a singular spectrum-based compression method for large language models to address limitations in standard SVD techniques.
Researchers propose Atomic Consistency Preference Optimization to reduce LLM factoid hallucinations without relying on stronger models.
Researchers propose Cross-Layer Attention Probing (CLAP), an activation probing technique designed to detect hallucinations in LLMs.