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 UMAS, a cost-neutral uncertainty quantification method designed specifically for multi-agent LLM architectures.
Academic study finds fact-checking LLM prompts alters affective tone but fails to eliminate demographic disparities in outputs.
Research on LLM harness self-evolution shows capability gaps, where guidance helping strong models degrades weaker ones.
Researchers test dynamic evaluation, letting language models take gradient steps on text they read to lower loss.
A registered study on self-improving agents finds that when task outcomes are noisy, agents can encode chance into future lessons.
New research on synthetic document fine-tuning explores methods to expose alignment-faking behavior in evaluation-aware language models.
Researchers propose using implicit beliefs as scalable process rewards for training LLMs in multi-turn strategic persuasion.
Researchers introduce SkillShift-Bench to evaluate web agent skill learning under continuous environmental shifts.
Researchers examine whether preference elicitation tasks in large language models accurately predict strategic behavior.
An arXiv paper presents non-autoregressive System 1 decision models for high-speed, low-cost routing, gating, and scoring tasks.
New research shows query-time RAG poisoning filters fail when subtle benchmark construction shortcuts are removed from attack data.
Researchers propose Intervention-Certified Amplification, a validity gate for post-training LLM refinement to prevent undesired behaviours.
Researchers introduce the Continual Learning Exam to benchmark how LLM agents improve through accumulated experience over time.
Researchers introduce a benchmark evaluating whether large language models adapt reliance policies rationally when collaborating with imperfect partners.
Researchers introduced SABench to evaluate tool-using AI agents for unintended side effects and unintended actions during task execution.
Researchers propose Entropic Control for Homeostatic Optimization, a closed-loop framework mitigating catastrophic forgetting in LLM fine-tuning.
Research on Best-of-N inference shows large N tests proxy tails, causing reward hacking when tail calibration curves turn downward.
Researchers propose AndroidFlux to evaluate mobile multimodal agents on error recovery rather than clean-slate task success.
An arXiv paper frames LLM inference reasoning effort as a credence good, analyzing the economics of paying for inference-time compute.
Researchers study Ouro-1.4B and Ouro-2.6B looped language models to understand how later loops repair two-hop retrieval.
Researchers propose recommender-system techniques for managing long-term memory selection in multi-session LLM agents.
Researchers propose ThinkBit, a two-bit quantization method aimed at reducing inference costs for reasoning language models.
Researchers propose a reinforcement learning method enabling LLMs to dynamically adjust abstention thresholds based on confidence targets.
Researchers propose causal voice isolators with 20 ms latency to separate overlapping speakers in audio streams.
Researchers propose Composable and Reusable Item Memory to reduce large language model inference costs in personalized recommendation.
Researchers propose Skill$, a framework for recursive self-improvement in language agents by treating optimization procedures as reusable skills.
Researchers show linear probes for content monitoring transfer from base LLMs to distilled models despite activation space rotations.
Researchers propose error-targeted reinforcement learning for correcting terminal agent misbehaviors on long-horizon command shell tasks.
A systematic academic survey of nearly 250 sources evaluates multi-turn conversational methods for LLM-based agents.
Researchers found latent tokens in LLMs act as uninterpretable placeholders that encourage shortcuts rather than faithful reasoning.