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 introduces a method, Contrastive Synthetic Document Finetuning, to measure 'reward-seeking' behavior in LLMs trained with RL, where models optimize for grader judgment rather than the true objective.
Researchers propose Supra Cognitive Modes (SCM), a routed architecture for agent memory that dispatches queries to varied retrieval and synthesis methods.
CircuitKIT is a new toolkit for mechanistic interpretability, unifying methods for discovering, evaluating, and applying model circuits.
Research paper discusses agentic LLM systems moving from prototypes to production across software, science, and finance, highlighting deployment challenges.
Research on Dice Loss for natural language processing tasks addresses data imbalance where negative examples significantly outnumber positive ones.
ChineseBERT incorporates glyph and pinyin information into pretraining models for enhanced Chinese language understanding.
Research evaluates four language models for adaptability, effectiveness, and limitations in the low-resource Aminoacian language.
A new research framework, MEDIC, introduces leading indicators beyond licensing exams to evaluate LLM competence for clinical workflow utility.
XCOMPS, a new multilingual benchmark for conceptual minimal pairs across 17 languages, reveals weaker LLM understanding for low-resource languages.
ScienceMeter is a new framework introduced by arXiv for evaluating how well LLMs update their scientific knowledge over time.
TReB is a new benchmark to evaluate large language models' (LLMs) table reasoning capabilities, addressing gaps in current evaluations.
Research introduces 'Node-as-Agent' framework, combining Graph Neural Networks with agentic capabilities to address node informativeness and local structural limitations.
Guard Vector introduces a method for LLM guardrails using task-vector composition and streaming-aware prefix SFT, moving beyond English-centric models.
Research explores using Vector Symbolic Architectures (VSAs) to decode internal representations and improve interpretability of Large Language Models.
Research introduces dynamic expert clustering and structured compression to mitigate MoE LLM issues: load imbalance, parameter redundancy, communication overhead.
Research introduces P-StaT, a framework to evaluate LLM belief stability under semantic perturbations, finding familiarity impacts stability.
PRISP proposes a privacy-safe, few-shot personalization method for LLMs, addressing data scarcity, limited compute, and strict privacy.
Research paper proposes a new method for benchmarking LLMs on culturally specific affective responses, moving beyond declarative knowledge.
Research identifies and quantifies "post-hoc rationalization" in Reverse Chain-of-Thought (RCG) LLM generation, where models justify a pre-committed answer.
Research explores small language models (SLMs) for creative plot generation, focusing on global coherence, character, pacing, and emotional progression.
Research proposes a new method using decoding-constrained beam search to quantify near-verbatim data extraction risk in LLMs, beyond simple memorization.
Researchers propose Exploratory Sampling (ESamp), a new decoding method for LLMs designed to increase semantic diversity beyond lexical variations.
Research explores Large Language Models' (LLMs) persuasive abilities using Habermas' Theory of Communicative Action, examining illocutionary intent.
Research explored LLM calibration for annotating human values in social media text, comparing models, prompts, and languages for scalability.
A research paper from arXiv introduces a multi-role red teaming framework to systematically uncover vulnerabilities and evaluate faithfulness in LLM outputs.
Research proposes Local-Preserving Supervised Fine-Tuning (LP-SFT) to prevent degradation of pre-existing LLM capabilities during domain adaptation.
Research finds LLM prompt robustness differs between objective questions with fixed answers and subjective opinion-based queries.
Research explores 'vectorizing the Trie' for efficient constrained decoding in LLMs, addressing limitations for generative retrieval in recommendation systems.
CLT-Forge is a new library for mechanistic interpretability research, using cross-layer transcoders to create more manageable attribution graphs for LLMs.
FinRAG-12B presents a production-validated framework for grounded question answering in banking, optimizing accuracy, citation, and refusal.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion