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.
Research suggests smaller language models with task-aware retrieval can achieve strong performance in scientific knowledge discovery, challenging the 'bigger is better' paradigm.
Meta-Tool explores few-shot tool adaptation for small language models (Llama-3.2-3B-Instruct) using hypernetwork-based LoRA vs. prompting.
Research proposes a method for LLMs to predict full conditional probability distributions from text, using quantile tokens and neighbor context.
KOCO-BENCH evaluates LLM performance on domain-specific software development tasks, focusing on how models learn and apply new domain knowledge.
Research identifies language bias in multilingual RAG rerankers, favoring English and query language, leading to performance gaps.
Research traces how LLMs recall relational knowledge, identifying latent representations supporting linear relation classification and which relation types are easier.
Research identifies two distinct failure modes in LLM 2-bit quantization: signal degradation and computation collapse, impacting efficient deployment.
Research annotated 10,600 persuader turns in 1,017 charitable donation dialogues with 41 strategies to link persuasion tactics to donation outcomes.
Research on LLM summarization of life narratives shows LLMs can introduce positionality and bias, challenging qualitative analysis use cases.
Research finds frontier LLMs exhibit culturally uneven urban perception, biasing descriptions and judgments even with neutral prompts.
Research analyzed 668 ChatGPT logs to quantify the risk of LLMs inferring user personality traits from chat history, identifying privacy risks.
A Dutch academic hospital piloted an EHR-integrated LLM for discharge summaries, generating 379 drafts with high adoption among clinicians.
Research identifies stereotype-related activations within GPT-2 Small and Llama 3.2 neural networks, exploring individual neurons and attention heads.
Research proposes framework to quantify how LLMs express unwarranted confidence, decoupling rhetorical intensity from actual epistemic grounding.
Research identifies 'hallucination neurons' in LLMs that predict factual errors and shows they generalize across knowledge domains.
New research proposes "dual-layer guidance" for self-describing structured data to mitigate LLM's "Lost-in-the-Middle" positional bias in knowledge retrieval.
OpenAI launched a bug bounty program for GPT-5.5 Bio, challenging red teamers to find universal jailbreaks for biosafety risks, offering up to $25k.
Shopify CTO details aggressive AI integration, projecting 2026 usage explosion, leveraging Anthropic Opus 4.6 with unlimited tokens.
OpenAI offers ChatGPT for Clinicians at no cost to verified U.S. physicians, nurse practitioners, and pharmacists for clinical, documentation, and research use.
Google DeepMind introduced Decoupled DiLoCo, a new method for distributed AI training designed to improve resiliency and efficiency in large-scale model development.
OpenAI detailed using WebSockets and caching to optimize API response times for agentic workflows, specifically for its Codex agent loop.
Researchers propose unsupervised method for calibrating LLM confidence from a single generation, addressing deployment reliability challenges.
Research quantifies error introduced by convex relaxations in neural network verification, impacting soundness for improved performance.
New research proposes a multiclass calibration method for ML models using Jensen-Shannon distance, aiming for stronger calibration.
Research proposes Stochastic Attention, an inference-time modification for scientific foundation models to improve calibrated predictive uncertainty.
Research proposes a generative mitigation method using VAEs to purify adversarially perturbed inputs in multi-modal embeddings, addressing 'adversarial illusions'.
FedProxy is a new federated fine-tuning method for LLMs designed to protect IP, ensure privacy, and improve performance on heterogeneous data using proxy SLMs.
Research paper proposes a method for continual machine unlearning, addressing knowledge erosion and forgetting reversal in AI systems.
Research demonstrates LLMs exhibit local linearity, enabling activation steering via model-based linear optimal control for more effective inference-time alignment.
Research demonstrates a side-channel attack, TrEEStealer, capable of extracting Decision Tree models by observing enclave memory access patterns.
Research introduces ARES, an adaptive red-teaming system addressing systemic weaknesses in RLHF by identifying and repairing both LLM and reward model failures.
Research indicates LLM-based scientific agents produce results without adhering to traditional epistemic norms of scientific reasoning.
Research audits LLM fairness in tabular prediction augmented by casenotes for housing placement, finding multi-class classification error disparities.
Sherpa.ai proposes a privacy-preserving entity alignment (PPEA) method for Vertical Federated Learning (VFL) with noisy identifiers, avoiding intersection disclosure.
Research proposes RESFL, a framework for federated learning that balances privacy, fairness, and utility by integrating uncertainty quantification.
Research paper proposes graph data augmentation with contrastive learning to improve graph neural network (GNN) robustness to covariate distribution shifts.
Research identifies a remote Rowhammer attack vector against Federated Learning clients leveraging adversarial observations and sparse gradient updates.
Researchers introduced a new millisecond-resolution network dataset for training time series foundation models, addressing gaps in high-frequency data.
Researchers propose Hydra Ensembles, a method to create efficient, uncertainty-aware transformer ensembles by pruning attention heads and using grouped multi-head attention.
Research identifies critical prompt injection vulnerabilities in LLM-based agentic systems, extending attack surfaces through external tool integrations.
Researchers propose Cross-Learning Score (CLS) to quantify dataset similarity using both input features and label information, improving on feature-only methods.
Research proposes a method for efficient autoregressive inference in transformer probabilistic models, improving joint distribution estimation from set-based models.
Research introduces TROJail, a trajectory-level optimization method for multi-turn LLM jailbreaks, improving on turn-level attack strategies.
Research proposes dynamic LLM safety monitoring, adapting computational cost based on input risk to optimize resource use and detection accuracy.
Research introduces HardNet++, a method to enforce hard nonlinear constraints in neural network outputs during inference, addressing a critical safety gap.
Research paper proposes PREF-XAI, a method for generating personalized, preference-based rule explanations for black-box ML models, moving beyond model-centric XAI.
Research identifies 'distillation traps' (tail noise, off-policy instability, teacher-student gap) that degrade smaller LLM performance during knowledge distillation.
Research identifies 'recognition bottleneck' and 'weakest link effect' as key failure modes in LLM multi-hop reasoning, proposing MFAI as a diagnostic.
Researchers propose ZC-Swish, a new activation function that stabilizes deep batch normalization-free networks, crucial for micro-batch and federated learning.
Research claims Graph Neural Networks (GNNs) do not outperform simpler models for Bitcoin fraud detection under rigorous, leakage-free evaluation.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion