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 reveals mechanistic insights into how backdoor attacks operate in LLMs, specifically identifying trigger formation and processing within attention heads.
Research demonstrates a method for systematically identifying rare, critical safety failures in large language models by efficiently sampling diverse responses.
Research identifies 'truncation blind spot' in LLM decoding (top-k, nucleus sampling), systematically excluding human-like, low-probability token choices.
Research indicates LLMs implicitly encode problem difficulty in their internal representations, detectable via linear probing.
Research proposes 'retromorphic testing with hierarchical verification' to detect hallucinations in RAG by judging faithfulness against retrieved context.
Researchers introduced Unsupervised Multimodal Clustering (UMC) for semantics discovery in multimodal utterances without pre-existing labels.
Research introduces MAESTRO, a method for structured pruning of Mixture-of-Experts (MoE) models to reduce memory footprint and improve deployment efficiency.
Researchers developed a low-resource Maltese OCR system using synthetic data, Tesseract 5 fine-tuning, and multi-stream arbitration for improved accuracy.
Research introduces Prismatic Synthesis, a gradient-based method for generating diverse training data to improve LLM generalization in reasoning tasks.
Research introduces "Mediator" for memory-efficient LLM merging, reducing parameter conflicts and improving performance over traditional averaging.
Research argues current LLM multilingualism is incidental, leading to unequal, brittle, and opaque behavior across languages due to training on uneven corpora.
Research shows a 146M-parameter small language model can detect LLM behavioral issues like sycophancy and confabulation more effectively than human raters.
Research explores whether interpreting model internals can predict out-of-distribution (OOD) behavior on unseen data, using synthetic tasks.
Research describes DisarmRAG, a retriever-centric poisoning attack disabling LLM self-correction in Retrieval-Augmented Generation systems.
MMR-V, a new benchmark, addresses MLLM limitations in multi-frame evidence location and multimodal reasoning for video data.
Research shows Vision Language Models (VLMs) can 'see' non-existent visual illusions, highlighting limitations in their perceptual reasoning.
Research introduces a three-stage calibration framework to evaluate how post-training methods (SFT, RL, OPD) reshape LLM confidence during Chain-of-Thought reasoning.
Researchers introduced LEGO Co-builder, a hybrid benchmark for fine-grained vision-language models focused on multimodal assembly instructions.
Research introduces Sequential Tool Attack Chaining (STAC), a multi-turn attack framework exploiting LLM agent tool use to chain harmless tools for harmful ops.
Octopus v4 research introduces a 'Graph of Language Models' to orchestrate multiple niche LLMs, aiming for more efficient and capable systems.
New research proposes a method for learning human value systems to improve alignment in generative AI, addressing a gap in current approaches.
Research introduces C-MTP, a direct supervision method for Continuous Chain-of-Thought (CoT) models, offering faster training than prior indirect methods.
CLARE, a self-evolving AI agent, is proposed to resolve 'intent asymmetry' in 3D tool orchestration by clarifying vague user instructions.
Research explores cross-modal unlearning in Vision-Language Models (VLMs) across three architectures to assess if unlearning transfers across modalities.
Research quantifies uncertainty in LLM benchmark rankings (e.g., MMLU) using rank confidence intervals and modified hypothesis tests.
Research investigates multimodal LLMs' capability for optical coherence tomography (OCT) imaging analysis beyond basic classification, focusing on clinical reasoning.
Research introduces CIGPO, a new reinforcement learning method for multi-turn LLM agents, improving stability in evidence-reading tasks compared to GRPO.
Research explores using Large Language Models with evolutionary algorithms to automate and improve feature generation for machine learning pipelines.
ColGraphRAG introduces late-interaction evidence retrieval for multimodal graph-grounded question answering, improving accuracy for graph-linked images.
Researchers propose SynH-Rank, a code search system that incorporates non-functional qualities like execution speed and memory usage into reranking.
RAIL Guard introduces a closed-loop system for evaluating and iteratively remediating unsafe LLM agent outputs, moving beyond binary blocking.
New research introduces DRNOISE, a benchmark for evaluating deep research agents' ability to maintain sound evidential standards in environments with misleading documents.
Research explores expressing and editing preference model inferences in natural language to address opacity and improve interpretability.
DS@GT developed a hybrid multi-agent LLM system with structured algorithmic guidance for conversational depression screening at eRisk 2026.
Research demonstrates that removing refusal behaviors from LLMs via 'abliteration' causes off-target effects on unrelated decision-making tasks.
Research proposes a semantic map to bridge language model token probabilities to interpretable, declared state uncertainty for professional decisions.
Research evaluates if off-the-shelf LLMs can replicate the diverse functionalities and methodologies of traditional privacy policy analysis tools.
Research explores hyperbolic geometry in AI, finding placement at the loss layer, not trainable adapters, prevents training collapse (HySAT).
Research introduces Otap, a new evaluation metric using optimal transport for LLM agent trajectories, assessing planning, tool use, and execution beyond binary success.
Research introduces WAR, a workload-aware rollout system, to accelerate synchronous agentic reinforcement learning by optimizing decoding and scheduling.
Research identifies 'Parametric Temporal Conflict' in open-weight LLMs, where newer facts are present but outdated ones are preferred without specific prompting.
Research explores "Thinking in Video," a paradigm where video generative models simulate and reason about real-world dynamics, but its promise remains unverified.
Research paper proposes a continuous geometric framework, modeling Transformer architecture components as an integro-differential equation on a semantic fiber bundle.
SlotGuard is a proposed local transcript boundary solution to prevent LLM agents from leaking sensitive data like API keys or file paths into provider-bound transcripts.
Research addresses temporal leakage in legal precedent retrieval, evaluating citation context methods against a temporally-fenced benchmark.
Research identifies 'self-state attacks' on self-hosted AI agents where agents are compromised via their own memory/config files.
PoLoRA introduces a preconditioned, orthogonalized optimizer for LoRA fine-tuning, aiming to improve efficiency and performance over standard Adam.
Research explores rhetorical patterns beyond directive verdicts for AI-assisted information evaluation, suggesting debate-style interactions improve critical assessment.
Research explores LLMs' capability for structure-based drug design, benchmarking performance under spatial constraints for molecule generation.
WorldCupArena is a new dynamic benchmark on arXiv for evaluating LLMs and deep-research agents on real-time football match forecasting.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion