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.
Pyramid MoA proposes a probabilistic, hierarchical Mixture-of-Agents architecture to optimize LLM inference cost by escalating queries only when necessary.
Research paper introduces 'Text-to-Big SQL' benchmark to evaluate LLM agents generating SQL for large-scale data processing workflows.
Research finds LLMs struggle to infer complex causal relationships from real-world, unsimplified text, despite prior claims based on synthetic data.
Research models how increasing AI agent choices in economic games (bargaining, negotiation, persuasion) alters strategic market interactions.
Research suggests interpreting LLM reasoning requires analyzing multiple chains-of-thought, not just single samples, by resampling subsequent text.
Research introduces BadGraph, a backdoor attack method targeting latent diffusion models for text-guided graph generation.
Researchers propose Proximal Supervised Fine-Tuning (PSFT), a method inspired by RL's TRPO/PPO, to mitigate catastrophic forgetting in LLMs.
Research paper introduces G-TRACE, a region-aware framework for quantifying the carbon emissions of Generative AI training and inference.
Researchers propose a framework for localized adversarial anonymization using small-scale models to address privacy risks with remote LLM APIs.
OpenAI extends its 'Trusted Access for Cyber' program, making an early version of GPT-5.4-Cyber available to vetted cybersecurity organizations.
A speculative timeline by Joe Reis outlines a progression toward autonomous AI models through 2028, focusing on AI's ability to 'think for itself.'
Import AI 453 discusses AI agents, MirrorCode, and a philosophical debate on gradual disempowerment, likening AI to historical paradigm shifts.
Cloudflare integrates OpenAI's GPT-5.4 and Codex into its Agent Cloud, allowing enterprises to develop and deploy AI agents securely.
Research proposes a 'Many-Tier Instruction Hierarchy' for LLM agents to resolve conflicting instructions from diverse sources, improving safety and reliability.
Research proposes Anchored Sliding Window (ASW) framework to improve robustness and imperceptibility in LLM-based linguistic steganography.
SiMing-Bench evaluates MLLMs for procedural correctness in clinical skill videos, tracking continuous interactions and state updates, moving beyond event recognition.
Research finds supervised fine-tuning (SFT) can decorrelate LLM confidence scores from output quality, impairing uncertainty quantification.
Research finds Vision-Language Models (VLMs) encode visual evidence accurately but fail to arbitrate conflicting visual-linguistic information.
Research identifies new fake news generation strategies using LLMs to embed subtle inaccuracies in credible narratives, challenging binary detection.
Research investigates how different quality aspects of preference data (generator-level, output-level) impact reasoning gains in LLMs using DPO/KTO.
Research surveys reasons for multilingual model performance disparities, examining intrinsic linguistic difficulty vs. model design choices like tokenization and data exposure.
New research proposes Subsentence-level Policy Optimization (SSPO), an RLVR algorithm designed to improve LLM reasoning stability and reduce high-variance tokens.
New benchmark, CONDESION-BENCH, evaluates LLMs in conditional decision-making with compositional action spaces, moving beyond static action sets.
Research paper explores credit assignment in RL for LLMs, addressing challenges in distributing rewards across long reasoning chains and multi-turn agentic actions.
Researchers demonstrated an exploit against diffusion-based language models (dLLMs) by re-masking early-stage refusal tokens, bypassing safety alignment.
Researchers propose a distillation and RL method, 'Multi-head Twig', to accelerate large Vision-Language Models by pruning visual tokens.
Research finds LLMs underperform smaller, graph-based architectures for supervised relation extraction in complex linguistic graphs.
Research models how AI-generated text entering public datasets creates 'model drift' from original distributions and 'selection' for common outputs.
Research proposes VisionFoundry, a method using targeted synthetic images from keywords to improve VLM visual perception tasks like spatial understanding.
Research finds LLMs overstate attitudinal influence and ignore network effects when simulating human susceptibility to misinformation.
Research proposes 'Verbalized Assumptions' framework to elicit and control LLM sycophancy by making implicit user assumptions explicit.
A new academic benchmark, TaxPraBen, evaluates LLMs specifically for Chinese tax practice, highlighting gaps in specialized, legally regulated domains.
Quantization (Q5_K_M) alters Llama-3-8B's self-assessment (metacognition) differently across knowledge domains, not uniformly degrading it.
Research paper details data exfiltration risk through indirect prompt injection in LLM agents using web search tools and RAG with sensitive corporate data.
Research paper introduces MuTSE, a human-in-the-loop tool for comparative evaluation of LLM-generated text simplifications across prompts and architectures.
Research identifies OCR bottlenecks in VLM architectures (Qwen3-VL, Phi-4, InternVL3.5) by analyzing activation differences with text-inpainted images.
New research proposes facet-level diagnostics for RAG to trace evidence uncertainty and hallucination, improving evaluation beyond answer-level.
Research investigates if LLMs homogenize academic writing, analyzing native language identification trends in papers across pre-NN, pre-LLM, and post-LLM eras.
New research proposes two improved multi-bit generative watermarking schemes for LLMs, outperforming prior work under worst-case false-alarm constraints.
LG AI Research released EXAONE 4.5, an open-weight vision language model integrating a visual encoder for multimodal pretraining on document-centric data.
Research proposes Temperature-Controlled Verdict Aggregation (TCVA) to align LLM evaluations with human assessments by adapting strictness to application domains.
Research introduces Litmus (Re)Agent, a benchmark and agentic system for predictive evaluation of multilingual model performance on unseen tasks and languages.
Research proposes BERT-as-a-Judge for LLM evaluation, claiming it's a robust alternative to lexical methods for reference-based assessment.
VerifAI, an open-source expert system for biomedical Q&A, integrates RAG with a novel post-hoc claim verification mechanism using NLI.
Research proposes LOM-action, an event-driven ontology simulation framework to ground LLM-based agent decisions in specific business scenarios for auditable AI.
Research proposes Hierarchical Alignment to enforce instruction priorities in LLMs, resolving common conflicts from varied sources like system policies and user requests.
Research proposes Evidential Transformation Network (ETN) to add post-hoc uncertainty estimation to existing pretrained models without retraining.
Research paper provides theoretical guarantees for OPTQ/GPTQ, a post-training quantization (PTQ) method for LLMs, addressing previous lack of rigor.
Research identifies Semantic Intent Fragmentation (SIF), an attack where benign subtasks from an LLM orchestrator jointly violate policy, bypassing current safety.
Researchers propose H-MRS, a novel algorithm for learning Directed Acyclic Graphs (DAGs) from observational data with positive-valued variables like asset prices, addressing multiplicative dynamics.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion