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 SkipCat, a low-rank compression technique using shared projection and block skipping to reduce LLM memory overhead.
Researchers introduce PlanAhead, a static planner-executor framework evaluating natural language plan representations for LLM web agents.
Researchers introduce HEART-Bench, a new evaluation framework assessing whether LLM agents make memory-grounded, persona-consistent decisions.
Researchers introduced the Agent Planning Benchmark (APB), a diagnostic dataset of 4,209 cases to isolate planning from execution failures.
Researchers introduced PropMe, a framework evaluating large language model propensity to leak training data without adversarial prompting.
Researchers introduced Causal Attribution Pruning, a method to prune LLM weights while maintaining multi-step reasoning capabilities.
Researchers identify cross-user poisoning vulnerabilities in multi-user collaborative language agents due to task boundary failures.
Researchers introduce relaxed on-policy distillation to address token entropy collapse when transferring reasoning to smaller models.
Researchers introduced ImpasseBench, demonstrating that coding agents over-persist, hallucinate success, and fail to stop when blocked.
Researchers introduced Annotrace, an open-source tool for step-level annotation and failure analysis of LLM agent traces.
Researchers investigate mechanistic interpretability, identifying specific attention heads and neurons responsible for network info retrieval in LLMs.
Researchers found exporter precision casting in ternary models can silently negate code changes during model deployment.
Researchers introduced LongSpark, a speculative decoding technique maintaining fixed drafting costs as prompt context scales.
Researchers introduced TwinRouterBench, a benchmark evaluating LLM routers on multi-step agent workflows to reduce inference costs.
Research paper examines how pair difficulty affects LLM-as-a-judge evaluation reliability, questioning standard proxy metrics.
Researchers introduced Forecast-Dojo, a replayable environment using prediction-market data and historical news to benchmark LLM forecasting agents.
Researchers propose SAKI, a score-aware low-rank KV cache compression method to improve attention accuracy in long-context models.
Research identifies a critical security vulnerability in reasoning models where adversarial inputs starve safety monitors of token budget.
Researchers propose TextCloak, using reinforcement learning to inject perturbations into text to prevent unauthorized LLM training.
Research paper introduces AutoWorld, a self-supervised world model for learning multi-agent traffic simulation, improving realism over abstraction-based methods.
Research proposes reusing internal audio language model representations for temporal localization to improve speed and accuracy over token generation.
Research proposes training Attention Graph Neural Networks without labeled data by minimizing finite-volume method residuals for thermo-fluid field prediction.
Research proposes Isotonic Conformal Prediction for reliable uncertainty quantification, aiming for self-calibration and prediction-conditional validity.
Research applies time-lag-aware deep reinforcement learning to flexible job-shop scheduling in prefabricated construction module factories.
ShredBench evaluates Multimodal LLMs on document reconstruction from shredded fragments, a challenging task requiring semantic and visual integration.
IndiaFinBench is a new public benchmark evaluating LLM performance on Indian financial regulatory text, addressing a gap in non-Western financial NLP.
Research finds AI content watermarking efficacy varies significantly across languages, cultural traditions, and demographic groups due to content properties.
Note conversions reveal Amazon holds massive equity ownership in Anthropic alongside venture capital backers.
SK Hynix reported higher third-quarter operating profits driven by strong demand for high-bandwidth memory chips used in AI hardware.
Anthropic updated its Claude for Financial Advisors tool, expanding partner integrations to 22 third-party services.