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 identifies 'concept neurons' in LLMs representing psychological constructs like the Big Five, enabling analysis of their formation and relation to output.
Research proposes Disco-RAG, a discourse-aware RAG strategy to capture structural cues and synthesize knowledge from dispersed evidence across documents.
Research finds natural language rules in coding agents improve performance only when structured as 'guardrails' (forbidden actions) over 'guidance' (suggested actions).
Research identifies 'ChatInject,' a novel indirect prompt injection vector abusing LLM agent chat templates to execute malicious instructions.
Research claims 261x cost reduction for maritime intelligence via LLMs as one-time teachers for specialized Small Language Models (SLMs).
Research suggests sustained interaction with conversational AI systems may contribute to delusional experiences in a subset of users, beyond individual vulnerability.
Research finds substantial increase of AI-generated content in news articles, particularly in local and college media, using advanced AI-text detectors.
LLMs show unreliable counterfactual reasoning in policy evaluation, performing worse on non-intuitive economic and social science findings.
CLSGen, a dual-head fine-tuning framework, aims to provide joint probabilistic classification and verbalized explanations from LLMs.
Research proposes a 'dual-path runtime integrity game' to detect RAG extraction attacks, a security vulnerability where LLMs leak proprietary data.
Research evaluates demographic and linguistic biases in omnimodal (text, image, audio, video) language models across identity, demographics, and activity.
Research paper proposes a framework for evaluating small open LLMs for medical QA, focusing on response consistency and safety in misinformation-prone environments.
FinTrace benchmark introduces trajectory-level evaluation for LLM tool-calling in long-horizon financial tasks, addressing limitations of call-level metrics.
Research finds large vision-language models (LVLMs) exhibit cross-cultural stereotypes, including religious, national, and socioeconomic biases.
Researchers demonstrated LingoLoop, an attack trapping MLLMs in endless loops via linguistic context, exhausting computational resources during inference.
Research proposes Knowledge Composition Sampling (KCS) to diversify multi-hop question generation, integrating more complex knowledge for robust QA.
Research introduces AttnTrace, a method for contextual attribution in long-context LLMs to detect prompt injection and knowledge corruption.
Research quantifies discrepancies between LLM outputs and their self-generated explanations, showing feature importances often differ.
Research investigates if LLMs' epistemic markers (e.g., "fairly confident") accurately reflect their intrinsic uncertainty.
Research proposes "Relational Probing," replacing standard LLM heads with a relation head to directly induce relational graphs for financial prediction from text.
New benchmark, SecureVibeBench, evaluates code agent security by comparing vulnerability introduction to human developer patterns, aiming for realistic assessment.
Research details a new adversarial attack, 'Attention-Guided Visual Jailbreaking,' that blinds Large Vision-Language Models to safety instructions.
StyleBench evaluates five reasoning styles in LLMs, analyzing trade-offs between structured reasoning benefits and computational/control costs.
Research identifies 'salami slicing' multi-turn jailbreaks as persistent LLM security vulnerabilities, bypassing safety controls gradually.
Gemma-3-4b-it encodes nationality-discriminative information in hidden states when generating academic text conditioned by British and Chinese personas.
Research identifies 'LLM Nepotism,' a bias where LLMs favor content expressing trust in AI, impacting fairness in AI-assisted evaluations.
A new Chinese benchmark, C-ReD, evaluates AI-generated text detection using real-world prompts, addressing current limitations in Chinese corpora.
Research proposes latent probing to improve RAG faithfulness, moving beyond black-box interventions to better leverage provided context.
Research paper proposes information density and feedback quality as fundamental limits to ML progress, explaining code generation's success.
Research proposes a self-calibrating method for LLMs using test-time discriminative distillation to mitigate systematic overconfidence without labeled data or high inference cost.
Research identifies significant, unmeasured hidden variance in LLM evaluation pipelines due to prompt rephrasing, judge models, and temperature, leading to unreliable rankings.
Research introduces BEHEMOTH benchmark for heterogeneous memory extraction in LLM-based assistants across 18 datasets, spanning personalization, problem-solving, and agentic tasks.
New benchmark, The Amazing Agent Race (AAR), challenges LLM agents with complex, non-linear tool-use tasks (DAGs), finding existing agents struggle.
Audio Flamingo Next, an open-source audio-language model, improves accuracy across diverse audio understanding tasks including speech, sound, and music.
Research finds LLMs exhibit formulaic discourse patterns in multi-turn empathic dialogues, despite high single-turn empathy ratings.
Research identifies 'resource consumption threats' in LLMs causing excessive generation, impacting efficiency, service availability, and cost.
Research explores using web-scale unlabelled data and LLM-based synthetic annotations to improve multilingual hate speech detection.
Research claims current LLM alignment evaluation is flawed; detection of harmful concepts is distinct from policy-based refusal mechanisms, using Chinese models as case study.
Research identifies motivations and mechanisms behind LLM-generated fake news to improve detection methods against information integrity threats.
New benchmark, METER, evaluates LLM contextual causal reasoning across all three causal ladder levels in a unified context setting.
LLMs exhibit "structural alignment bias" causing them to invoke irrelevant tools, impacting tool-use reliability and potential hallucinations.
Research explores model scheduling for masked diffusion LMs (MDLMs) to accelerate inference by replacing full-sequence denoising passes with a smaller model.
Research evaluates large language models' effectiveness in generating multilingual synthetic data for training smaller models, highlighting capability gaps in non-English languages.
Research evaluates LLM robustness for clinical numerical reasoning beyond simple arithmetic, finding limitations in handling patient measurements in clinical notes.
Research proposes DiaFORGE, a disambiguation-centric finetuning pipeline to improve enterprise tool-calling LLMs facing duplicate tools or underspecified arguments.
Research introduces PODS, a method for down-sampling LLM rollouts in RLVR to address compute and memory asymmetry in policy updates.
Single LLM agents can outperform multi-agent systems in multi-hop reasoning when computational budgets for "thinking tokens" are normalized, based on arXiv research.
Research introduces PICon, a multi-turn interrogation framework to evaluate consistency and factual accuracy of LLM-based persona agents.
New post-training quantization method, SEPTQ, claims improved LLM compression for reduced computational and storage costs without retraining.
Research identifies 'Incomplete Learning Phenomenon' in LLM supervised fine-tuning, where models fail to reproduce training data.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion