Search signals, briefings, benchmarks and glossary terms.
Search signals, briefings, 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.
INS-ActBench, a new benchmark, evaluates LLMs on professional actuarial tasks requiring auditable, context-grounded, and tool-executable decisions.
Research identifies a 'blind spot' in AI agent long-term memory systems where retrievers fail to link implicit knowledge to queries.
Research finds AI models GPT-5.6 Sol, Gemini 3.6 Flash, Perplexity Sonar Pro, and Grok 4.5 show bias in naming individuals based on citation type.
Researchers developed a framework to measure how LLM political bias shifts dynamically based on prompt context rather than remaining static.
Researchers introduced LibMoE, a unified framework for benchmarking Mixture of Experts (MoE) architectures in large language models due to high training/evaluation costs.
New research explores if transformers can dynamically adapt their reasoning strategies (latent algorithm routing) based on input data characteristics.
Research explores integrating RAG with locally deployed LLMs for regulatory knowledge management, focusing on epistemic reliability without high-end GPUs.
Research explores pointer-augmented autoregressive generation for patent claims, addressing hierarchical constraints in structured text with LLMs.
An academic practitioner guide detailing the technical architecture, GPU systems, and orchestration layers required to build agentic AI systems.
New research identifies a critical flaw in LLM context attribution methods: they fail to distinguish between information retrieved from input context and knowledge encoded in model weights, producing unreliable scores.
Researchers created GEMCo, a human-written proxy dataset of 86 German e-mail counseling conversations, validated against 124 real conversations.
Research evaluates closed-loop validation-repair for clinical LLMs in healthcare to achieve structured output schema compliance (ICD-10, CPT, HL7 FHIR).
Research isolates performance factors in language model-based entity matching, distinguishing architecture from model backbone and size effects.
Researchers introduce CodexGraph, a framework bridging LLMs and code repositories using graph databases to improve repository-level tasks.
New research introduces LENS, a protocol for evaluating how well machine unlearning algorithms suppress disinformation-aligned narratives in large language models.
Research reveals reward models (RMs) misallocate memorization to easy preferences, memorize dataset-specific shortcuts, and overgeneralize heuristics.
Research indicates no optimal language set exists for multilingual instruction tuning, challenging the assumption that linguistically diverse sets universally yield better models.
Research finds behavioral detection of unfaithful Chain-of-Thought (CoT) explanations fails when LLM answers are incorrect, hindering oversight.
Research identifies that offensive language detection models degrade across datasets and languages, proposing a framework to diagnose and optimize cross-domain generalization.
Research introduces XIH-Bench, a new benchmark to evaluate instruction hierarchy compliance in multilingual LLMs, finding language impacts adherence.
Research shows frontier LLMs use 'invisible reasoning' with semantically irrelevant filler tokens to improve performance on synthetic tasks.
Research finds Activation Oracles (AOs), LMs explaining other models' internals, can develop concept-specific blind spots due to training data.
Research indicates LLM debate patterns differ across languages, with Chinese models less prone to repeating arguments compared to other languages tested.
Research paper proposes a novel method for dynamic evaluation of multimodal automated fact-checking systems to avoid contamination from outdated claims.
New research proposes AutoThinkSQL, a framework that dynamically decides when to apply reasoning chains for Text-to-SQL queries, reducing inference overhead.
Research applies a two-stage LLM pipeline (Gemini 2.5 Pro, Gemini 2.5 Flash) to detect documentation inconsistencies in 3,000 health records.
Research explores why Joint-Embedding Predictive Architectures (JEPAs) are not standard for text encoders, citing a mismatch with language's conditional structure.
LA-RL introduces a label-aware self-reflection method for reinforcement learning in information extraction, improving structured output correction.
Research finds fine-tuning small models with legal context improves their accuracy on legal Q&A, even with retrieved law.
Research evaluates LLM understanding and generation of novel Chinese xiehouyu riddles using MCQs and free-form explanation generation to test reasoning.
Research indicates that diagrams, particularly Euler diagrams, can improve LLM syllogistic reasoning performance compared to natural language or logical notation.
Researchers introduced IndicTalk, a large persona-based multilingual conversational corpus for Indic languages, addressing scarcity of high-quality code-mixed dialogue data.
Research explores training-free methods to personalize language model toxicity sensitivity during inference, moving beyond global alignment for subjective contexts.
GAND is a new dataset for benchmarking gender bias in machine translation, focusing on gender-ambiguous scenarios and contrastive attribution for explainability.
Research compares BERT-based models and LLMs for Named Entity Recognition in Marathi, a low-resource language, finding uncertain effectiveness for LLMs.
Research introduces a method for multi-hop question answering by jointly evolving graph and text memories, enabling coordination of relational and textual evidence.
Researchers propose SINT-Flow, an LLM-based framework with five operators for fully automated, end-to-end schema integration across diverse input tables.
Research finds visually grounding a bilingual LSTM with images and captions in English and Spanish improves semantic understanding within and across languages.
New research proposes "Flick," a few-label text classification method using K-Aware Intermediate Learning to reduce reliance on extensive labeled data, especially for low-resource languages.
New research proposes a method for evidence attribution in visual document understanding without needing explicit coordinate or region labels.
Research introduces Multi-Task GRPO to improve LLM reasoning reliability across diverse tasks, addressing imbalances in standard multi-task RL optimization.
New research proposes attention-guided strategies for contrastive decoding (e.g., DoLa) in LLMs to improve factuality beyond vocabulary divergence.
Researchers developed ADAGE, a language-agnostic pipeline combining native-speaker curation with LLM-assisted generation for culturally-grounded analogical reasoning benchmarks.
Academic study demonstrates LLMs internally represent accurate token counts even when generating incorrect numerical outputs.
Researchers introduce ProvenanceGuard, a metric to detect cross-source misattribution in MCP-based agentic LLM workflows.
Research compares LLM-based vs. lexicon-based sentiment analysis for detecting tail-risk signals from Reddit data on meme stocks like GME and AMC.
Research identifies two distinct failure points in compressed short-text generation: information loss in the codec or weak codes from the latent generator.
Researchers developed formally verified IEEE-754 FP32 and BF16 arithmetic for ARCH HDL, a language intended for LLM generation, ensuring mathematical correctness.
Research details a simple language normalization method for cross-lingual speaker verification, improving performance on the TidyVoice 2026 Challenge.
Researchers propose Zhijing, a framework for measuring and integrating social intelligence in LLMs, including the SoMBench benchmark.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion