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 DeALOG, a decentralized multi-agent framework utilizing a shared natural-language log for multimodal question answering.
Researchers propose OdysseyArena, a benchmark designed to evaluate LLMs on long-horizon, active, and inductive agent interactions.
Research on arXiv examines how word-level lexical alignment between pre-training corpora and benchmarks affects zero-shot model accuracy.
Researchers introduce OmniBehavior, a benchmark built on real-world data to evaluate large language models on human behavior simulation.
Researchers use sparse autoencoders to investigate causal mechanisms behind uneven emotion recognition performance in large language models.
Research demonstrates first-token AI safety scores vary widely across minor prompt rewordings, introducing reliability risks in guardrails.
Researchers propose LatentRAG, combining latent reasoning and retrieval for efficient multi-step agentic RAG workflows.
Researchers propose ReVision to reduce visual token redundancy in computer-use agents, lowering token costs across long interactions.
Researchers propose VerifyMAS, a hypothesis verification method for attributing failures in LLM-based multi-agent systems.
Researchers released QUEST, a family of open models from 2B to 35B parameters trained on fully synthetic tasks for deep research agents.
A recent arXiv paper proposes a unified context evolution method for LLM agents to retain strategy across multi-step interactive tasks.
Researchers study behavioral signatures of multi-agent LLM teams to determine when process-level coordination control adds value.
Researchers propose RIVO, a lightweight adaptive rerouting method to improve trajectory control in deep research agents.
Researchers propose a method for external selective answering in LLMs using confidence thresholds, avoiding internal abstention training.
Researchers propose P2N, a method reusing previous token deep representations to enhance reasoning without increasing layer counts.
Researchers evaluate Jev as a decision reader for truncated LLM reasoning, examining accuracy impacts when limiting reasoning budgets.
Researchers study geometric relationships in conversation turns and memory texts to predict retrieval-bound failures in AI agents.
Researchers propose a simulation optimization method to expand and refine prompt sets under limited evaluation budgets.
Research evaluates multimodal LLMs including Claude 3.7 and OpenAI o1 on their capability to detect document fraud.
Researchers propose Fast KVzip, a gated KV cache eviction method aiming to reduce LLM inference memory overhead without performance loss.
Researchers propose post-hoc embedding alignment to improve performance in compressed transformer models without full teacher retraining.
Researchers introduce RAPTOR, a controlled study evaluating compact self-supervised learning backbones for audio deepfake detection.
Researchers argue web agents should use a compile-and-execute paradigm instead of ReAct to handle untrusted web content safely.
Researchers introduced FireFly, a pipeline for generating verified tool-call data from real Model Context Protocol servers.
Researchers show social-engineering attacks manipulate autonomous web agents into submitting personally identifiable information (PII).
Researchers introduced TellOnce, a layer designed to make coding agents retain and apply user corrections across future interactions.
Researchers propose CausalMix, applying causal inference to optimize data mixture weights for language model training without retraining.
Researchers introduce SearchJev, a calibrated System-1 model designed to handle rapid search agent decisions separately from reasoning.
A preregistered study on prompt harness search reveals that performance gains often stem from evaluation defects and parser visibility changes.
Researchers propose SoundnessBench to test whether AI agents can judge methodological viability before resource-intensive execution.