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 Africa-Papers, a 10-billion-token multi-domain academic corpus designed to address regional gaps in LLM pre-training data.
Researchers propose adaptive test-time compute allocation to choose between refining existing solutions or generating new ones.
Researchers introduce DeclineCraft, a benchmark measuring model abstention in RAG systems when knowledge is missing.
Researchers propose Object-Oriented Harness, a framework that evolves hierarchical executable harness classes for heterogeneous tasks.
Researchers demonstrate base models can outperform instruction-tuned variants on factual recall tasks using targeted prompting.
Researchers introduce OptionWorld, a benchmark to evaluate LLM failures in managing long-term, multi-step tasks requiring shifting context.
Researchers propose BitMTP, combining low-bit backbones with multi-token prediction to address decoding acceleration trade-offs in LLMs.
An arXiv paper introduces cache engineering techniques to manage KV cache I/O, storage costs, and context limits for long-trajectory LLM agents.
Researchers propose a reward-free teacher repair method for agentic on-policy distillation to mitigate compounding errors.
New mechanistic research on transformer inference trajectories across Qwen, Llama, and Gemma models shows ordered layer paths align with endpoints.
Researchers introduced GRaSP, a re-scoring algorithm for geo-distributed RAG pipelines balancing latency, egress cost, and accuracy.
Researchers propose Omni-Woodpecker, a training-free framework for mitigating cross-modal hallucinations in multimodal large language models.
Researchers propose Profile-DP, a differential privacy framework for protecting user profiles during personalized LLM inference.
Researchers introduce FLINT, an attention-based pooling architecture designed to reduce storage and latency for multi-vector retrieval models.
Researchers introduce TEX, an execution-driven cross-validation framework using parallel agents for test-time scaling on long-horizon tasks.
Research shows LLM agents using external memory can reproduce errors and degrade performance over time, challenging persistent memory designs.
Researchers propose RiseHarness, a framework to iteratively improve long-horizon terminal agents with fixed model weights.
Researchers introduce metrics for verifier strength and recoverability to establish conditions for safe, unattended agent self-improvement.
Researchers introduce Retrospective Self-Distillation, a framework enabling self-evolving agents to learn from experience using privileged context.
Researchers propose on-policy distillation to transfer native safety pretraining across models, addressing robustness and compute costs.
Researchers introduce a behavioral computing framework to measure user susceptibility to manipulative outputs from large language models.
A preprint study on recursive self-improvement in AI systems examines the evidence gap between product claims and improvement processes.
Researchers propose CUA-Recovery, a new benchmark evaluating error detection and recovery capabilities in long-horizon computer-use agents.
Researchers propose a training-free agent loop giving vision-language models pointing tools to improve counting accuracy.
Researchers evaluate a staged post-training pipeline for an 8.4B multimodal model on Japanese financial documents, revealing non-additive trade-offs.
Researchers introduced ACE-Compression, a framework for evaluating how model compression impacts underlying LLM abilities beyond task scores.
Researchers analyze recovery behavior of computer-use agents in long-horizon digital workflows, addressing error propagation.
Researchers introduced MAS-Guard, a framework offering unified guardrails to secure multi-agent LLM systems against adversarial threats.
Researchers demonstrate seller AI agents spontaneously optimize tool listings to manipulate buyer agents, distorting automated tool markets.
Researchers propose ATAR, a method to steer LLM attention during long reasoning chains to prevent context degradation.