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 propose an active preference acquisition method for web agents to learn user preferences across repeated tasks.
Researchers propose a decision-theoretic framework for evaluating memory management in language agents across operational states.
A new pre-print paper examines inference challenges in nonstationary multi-armed bandit algorithms for adaptive experiments.
Researchers introduced LTEM, a markdown-based table format designed to preserve complex headers and merged cells for LLM processing.
Researchers propose training agentic reinforcement learning through diverse harnesses to improve generalization across tool use and context.
Research shows CLIP dual encoders assign higher similarity scores when false details are added to accurate text prompts.
Research demonstrates that human facial attractiveness biases image generation preference judgments and reward models.
Researchers introduced LoWAgent, a benchmark for evaluating web agents across low-resource multilingual websites.
A controlled study on speculative decoding examines the trade-offs between acceptance length and latency across training and inference.
Researchers propose a reversible pseudo-anonymization method to protect PII during interactions with remote large language models.
Researchers propose Interaction-as-a-Judge, using computer-use agents to interactively evaluate code and applications built by AI agents.
A systematic study shows that LLM-based novelty evaluators perform poorly when scoring generated ideas instead of human papers.
Researchers introduced Agents' Last Exam (ALE), a benchmark evaluating AI agents on real, economically valuable professional workflows.
Researchers propose ranking prior alignment to improve credit risk modeling performance in cold-start scenarios with scarce labeled data.
Researchers examine whether safety alignment is preserved across recurrent inference depths in looped language models.
Researchers introduced WorkForge, a framework designed to synthesize scalable, verifiable environments for long-horizon AI work agents.
Researchers propose VETTA, a method for coordinating turn- and token-level credit assignment in multi-turn LLM agents.
Researchers propose JevAdvBench to evaluate the robustness of reinforcement learning models making automated calibrated decisions.
Researchers released a dataset of real scam and spam call conversations collected using an active voice-agent honeypot.
Researchers propose JevLite, a small language model approach using typed decisions for real-time scam-call screening.
An empirical study benchmarks self-supervised mask-and-recover pretraining on 14 tabular classification tasks under label scarcity.
Analysis of 10,211 scam and spam calls collected via an AI voice-agent honeypot reveals operational patterns of phone scammers.
Study shows step-level AI agent credit signals (LLM-judge, logprobs, confidence) perform no better than chance against causal replay.
New research evaluates LLMs on substantive legal citation accuracy beyond basic hallucination checks, targeting unsupported propositions.
Researchers propose CoinRAG, a framework optimizing KV cache reuse in long-context RAG by focusing on fine-grained information nuggets.
Researchers introduce FinVerse, a new benchmarking framework designed to evaluate financial time-series forecasting foundation models.
Researchers analyzed LLM evaluations of non-native Japanese writing, finding models mimic human biases regarding fluency, status, and solidarity.
Researchers show that KV cache compression in reasoning LLMs can preserve final-answer accuracy while degrading supporting rationales.
Research demonstrates that monitoring agentic chain-of-thought (CoT) reasoning is easily bypassed by adversarial rewriting of reasoning steps.
Research shows knowledge distillation from larger models like Gemma-2-9B to smaller models improves context accuracy but worsens bias calibration.