Search signals, briefings, company results, benchmarks and glossary terms.
Search signals, briefings, company results, benchmarks and glossary terms.
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.
LLM agents exhibit "temporal blindness," failing to account for real-world time elapsed between actions, leading to suboptimal tool use decisions.
Research explored mitigating LLM biases from spurious social contexts using direct preference optimization, focusing on high-stakes decision-making.
Research finds spoken language models (SLMs) lose instructed speaking styles (emotion, accent, volume) over multi-turn conversations.
Researchers propose Prefill-Only Pruning (POP) for LLMs/VLMs to reduce inference costs by targeting prefill stage without accuracy loss.
Research proposes CAP-TTA, a test-time adaptation framework, to debias LLMs during inference by updating LoRA weights for high-bias prompts.
Research introduces ReasonScaffold, a human-AI co-annotation protocol exposing LLM explanations while withholding labels to reduce human annotation variability.
Research introduces 'feedback adaptation' for RAG, evaluating how effectively corrective user feedback propagates through the system.
Research explores using formal methods to generate verifiable, targeted explanations for deep neural networks, aiming for mathematical guarantees.
Research introduces Calibrate-Then-Delegate (CTD), a model-cascade approach for LLM safety monitoring that uses a cheaper model to screen and delegates hard cases to an expert, optimizing for cost and accuracy.
Research models cyber attack surfaces as a queue, integrating AI's impact on vulnerability discovery, exploitation, and patching dynamics.
Researchers demonstrated a small adapter can correct suppressed factual log-probabilities in alignment-tuned LLMs like Qwen3, leveraging hidden states.
Research paper explores LLMs' ability to detect methodological flaws, specifically data leakage, in machine learning studies.
Research paper proposes BitFlipScope, a method to localize and recover from bit-flip corruptions in LLMs, addressing hardware-induced silent data corruption.
Research paper systematically benchmarks optimizers for MLPs in tabular deep learning, finding potential alternatives to AdamW.
Research proposes Interrogative Uncertainty Quantification (IUQ) for long-form LLM generation, addressing challenges beyond short, constrained outputs.
Research identifies explicit step size conditions for zeroth-order (ZO) optimization, improving stability for black-box and memory-efficient model tuning.
ConfLayers proposes an adaptive confidence-based layer skipping method for self-speculative decoding to accelerate LLM inference.
Researchers propose SAGE, a memory-efficient LLM optimizer addressing AdamW's memory bottleneck and the embedding layer dilemma for large model training.
Research paper proposes 'De-Anonymization at Scale' (DAS), an LLM-based method to attribute authorship among tens of thousands of anonymous texts.
Research presents a bit-accurate modeling framework for GPU matrix multiply-accumulate units, revealing undocumented numerical behaviors and discrepancies.
Research introduces MinShap, a modified Shapley value approach for feature selection in machine learning models, addressing non-linear, dependent features.
Research introduces a novel metric-agnostic learning-to-rank method using boosting and rank approximation, moving beyond single-metric optimization.
Research proposes Atropos, an agent architecture improving cost-benefit of LLM-based agents using early termination and model hotswap.
Research finds common fairness metrics often disagree, challenging current single-metric approaches for assessing ML fairness in high-stakes applications.
Research details a black-box adversarial attack method to force LLM routers to select higher-cost, high-capability models.
Research investigates if reinforcement learning expands LLM agent capabilities for tool use or merely improves reliability, introducing PASS@(k,T) metric.
Research paper argues static AI value alignment methods are insufficient for robust alignment given model scaling, distributional shift, and autonomy.
Research introduces Deep Neural Lesion (DNL), a method to catastrophically disrupt DNNs by flipping few parameter bits, data-free and optimization-free.
Research evaluated multilingual text embeddings for hate speech detection in Lithuanian, Russian, and English, optimizing model choices.
Research analyzed reasoning dynamics in 18 Vision-Language Models (VLMs), tracking Chain-of-Thought confidence and modality reliance.
Research identifies 'prolepsis' in small transformers: early, uncorrectable commitment to decisions via task-specific attention heads.
Researchers propose TempusBench, a new evaluation framework for time-series foundation models (TSFMs) to standardize performance benchmarking.
Research finds LLMs used as judges in AI evaluation are susceptible to 'stakes signaling,' affecting verdicts based on perceived downstream impact.
New research proposes Dual-Path Graph Filtering, a graph neural network (GNN) method for fraud detection, addressing relation camouflage in fraud graphs.
Research exposes high per-instance inconsistency in LLM-as-judge frameworks for NLG evaluation, with 33-67% of documents showing transitivity violations.
Research presents an explainable Graph Neural Network (GNN) framework, ST-GAT, for interbank contagion surveillance using U.S. FDIC data.
Research identifies 'attention sink' phenomenon in GPT-2, where the first token receives disproportionately high attention due to specific model interactions.
Research paper empirically evaluates NVIDIA L4 GPU performance against T4 for deep learning inference, focusing on parallelism and architectural improvements.
Research formalizes synthetic data augmentation, identifying a bias-variance trade-off from modifying training distributions, crucial for financial ML data scarcity.
Research paper proposes a model-free method using quantile curves to quantify the 'sim-to-real' gap in generative AI models used for simulation.
Research indicates that optimizing AI interventions solely for predictive accuracy can lead to suboptimal outcomes when service capacity is limited.
Research tested LLM juries against expert panels for scoring medical diagnoses in real-world hospital cases, showing strong correlation.
Research identifies and proposes a solution for the "reward-generation gap" in Direct Alignment Algorithms (DAAs) like DPO and SimPO.
Research analyzes the architecture of 'Claude Code,' an agentic coding tool that executes shell commands and edits files, comparing it to OpenClaw.
Research proposes DPQuant, a method combining differential privacy with dynamic quantization to accelerate neural network training while protecting user data.
PolyBench is a new multimodal benchmark for LLM forecasting and trading on live prediction market data, coupling market snapshots with qualitative news.
Anthropic's Claude Opus 4.7 is reportedly a marginal improvement over 4.6, maintaining its position as a leading frontier model.
AI Snake Oil introduces Project CRUX for open-world evaluations of frontier AI on complex, multi-step tasks, addressing current benchmark limitations.
Alibaba's Qwen3.6-35B-A3B quantized model running locally produced a better image than Claude Opus 4.7 for a specific prompt.
Meta developed an AI agent platform to automate finding and fixing performance issues, optimizing infrastructure capacity and freeing engineers.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion