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.
MEMCoder research introduces a multi-dimensional evolving memory system for LLMs to improve code generation using private enterprise libraries.
Research identifies 'Persona Collapse' in LLMs, where distinct agents converge into homogeneous behavior, limiting diversity in multi-agent simulations.
Research proposes MEG-RAG, a new metric and methodology to quantify multimodal evidence grounding in Retrieval-Augmented Generation systems.
Research indicates users can effectively post-edit LLM-generated text to infuse personal style, addressing a key adoption barrier for personalized content.
Research from arXiv highlights advanced image generation models creating photorealistic, search-grounded synthetic visual evidence, increasing real-world risk.
Research finds LLMs adopting specific personas exhibit gender bias in narratives, with personality cues interacting with gender stereotypes across languages.
Research evaluates LLM prompting strategies for cross-lingual text simplification (CLTS) between English and French, addressing both translation and linguistic complexity.
Research explores chunk filtering strategies (semantic, topic, named-entity) to reduce redundancy in RAG indexed corpora while preserving retrieval quality.
Research introduces CanMT, a new dataset and evaluation framework for assessing culture-aware machine translation performance of LLMs, highlighting current gaps.
New research introduces AIPsy-Affect, a keyword-free stimulus battery to improve mechanistic interpretability of emotion in LLMs by avoiding lexical confounding.
Research explores methods for quantifying uncertainty in Large Language Model (LLM) function calls to prevent incorrect, irreversible actions.
Research compares domain fine-tuning against RAG for small LLMs in medical question answering, holding variables fixed at 4B-parameter scale.
A research paper describes a deployed system for customer support automation using LLMs, leveraging copilot feedback and UI interaction traces.
MTRouter proposes a method for cost-aware multi-turn LLM routing, selecting models from a pool to optimize cost within a budget using history-model joint embeddings.
Research proposes T, a new test-based framework for evaluating semantic correctness of LLM-generated formal proofs, moving beyond lexical overlap.
Research identifies layer-wise feature vulnerabilities in Gemma-2-2B, demonstrating that internal mechanisms, not just prompts, drive jailbreak success.
Research finds AI safety benchmark results are highly sensitive to the configuration of LLM judges, specifically model and prompt choices.
Research proposes using a small language model (SLM) to resolve semantic ambiguity in large language model (LLM) prompts, improving task performance.
Researchers propose a general-purpose automated red teaming model to identify vulnerabilities unique to specific LLMs beyond content safety benchmarks.
Research proposes a zero-shot prompting method for automatic readability assessment using 10 open-source LLMs and provides a comprehensive evaluation.
Research proposes LinguDistill, a method to recover degraded linguistic abilities in vision-language models (VLMs) caused by cross-modal adaptation.
Research systematically compares various graph-based RAG methods for LLMs, evaluating their impact on factual accuracy and interpretability.
Research finds that irrelevant audio, including silence and noise, reduces accuracy and increases volatility in Large Audio-Language Models (LALMs) on text reasoning tasks.
Researchers introduced Game-Time Benchmark to evaluate Spoken Language Models' (SLMs) capacity for temporal dynamics in real-time speech.
Research identifies 'temporal scope stability' as a new challenge for multi-turn language models, assessing their ability to maintain context over time.
New arXiv paper proposes a divergence-based method for weighting and averaging probabilistic predictions from various statistical and ML models.
DenoGrad proposes a gradient-based framework to iteratively correct noisy tabular and time-series data using a pretrained neural network.
NVIDIA's CuTile, a Python abstraction for GPU kernel development, evaluated across Hopper and Blackwell GPUs for efficiency against cuBLAS, Triton.
Research introduces True Thinking Score (TTS) to quantify causal contribution of each step in LLM Chain-of-Thought (CoT) reasoning.
Research formalizes comparison of fine-tuning (FT) vs. in-context learning (ICL) in LLMs to determine proficiency and inductive biases.
Research explores approximating high-dimensional uniform random rotations using structured Hadamard rotations to reduce computational cost.
Research investigates active learning algorithms' effectiveness for text annotation, accounting for real-world noisy, fallible crowd-sourced labels.
MOCA introduces a transformer-based modular framework for causal inference, improving stability for complex, non-linear observational data.
LLM-based mental health support agents show clinical harm in 33% of simulated cases; only 16% of interventions are clinically tested.
Research evaluates LLaMA 3.2 and Mistral for local bug detection in Python, focusing on privacy-sensitive environments over cloud LLMs.
Research proposes Hindsight Preference Optimization (HPO) to train language models for financial time series advisory, using retrospective outcome data.
Research explores three techniques for vector quantization-based model weight compression, improving efficiency and end-to-end training.
Researchers propose FLEXI-Haz, a deep neural network for survival data with a partially linear structure, combining interpretability with complex time-covariate interactions.
Research details methods to scale Mixture-of-Experts (MoE) LLM inference by optimizing expert load balancing and token routing across multi-node setups.
Research finds large language models used as 'silicon samples' systematically reduce heterogeneity in philosophical opinions compared to human panels.
KARL is a new reinforcement learning framework designed to reduce LLM hallucinations by enabling models to abstain from answering questions beyond their knowledge boundaries.
Research finds that LLMs undergoing continual fine-tuning can experience a collapse in uncertainty reliability (conformal coverage) before accuracy degrades.
Audio2Tool introduces a new 30,000-query dataset to benchmark Speech Language Models' (SpeechLMs) tool-calling capabilities across diverse domains.
Research introduces 'Stochastic KV Routing' to reduce LLM Key-Value cache memory footprint by adaptive depth-wise cache sharing.
Rabtriever proposes an efficient rationale-based retrieval method using independent query/document encoding and distilled generative rerankers.
New research proposes Coverage-Based Calibration, a Post-Training Quantization method using weighted set cover to activate outlier channels for improved LLM compression.
Research outlines a layered security framework for agentic AI systems, addressing persistent memory, tool invocation, and multi-agent coordination.
Research proposes adaptive quantization and differential privacy for Federated Learning, addressing communication bottlenecks and privacy in non-IID data.
Research paper details an LLM-powered pipeline for automated data extraction and structuring from scientific literature, exemplified with concrete materials.
Research identifies new security risks in multi-agent AI systems due to architectural decisions, separate from individual agent robustness.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion