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 RIMS, a method for preference optimization in RAG for small-scale LLMs, enhancing robustness against noisy retrieval.
Research finds open-weight LLMs sometimes commit to an answer and then generate reasoning, even if it contradicts the premise, showing 'pre-commitment'.
Research finds advanced LLMs predict next word similarly to human brain responses measured by EEG, exploring human-like language processing.
MSCE is a training-free framework enabling LLM agents to convert experience into reusable skills and policies, not just memory retrieval.
OpenLanguageModel (OLM) is a new open-source PyTorch library designed for building and pretraining small language models with transparent architecture.
Research introduces SpecLA, a speculative decoding method for linear-attention models to improve autoregressive decoding efficiency by verifying multiple tokens.
Research explores if LLM arithmetic failures on equivalent problem formulations stem from distinct internal circuits or shared circuit activation states.
Researchers propose Scientific Feasibility Control (SFC), a conformal prediction framework to provide statistical guarantees for scientific reasoning validity in LLM outputs.
JOR-Bench introduces Japanese language benchmarks for evaluating LLMs on 1,319 operations research problems (LP, MIP, NLP).
Research compares cascaded vs. joint multi-task modeling for hierarchical offensive language detection, analyzing accuracy, parameters, and latency.
Research explores schema-constrained, document-level event argument extraction using lightweight LLM fine-tuning to improve accuracy and consistency.
Research introduces Group Entropy-Controlled Policy Optimization for LLMs to manage exploration-exploitation on heterogeneous tasks.
Researchers propose Trace-Based On-Policy Distillation (TOPD), a new teacher-supervised framework for reasoning-oriented post-training of diffusion LLMs.
Research describes a human-in-the-loop AI agent that drafts translational impact summaries, reducing staff time from 15 hours to 1.
Research addresses model collapse in iterative instruction tuning when training LLMs on synthetic data, proposing methods to ensure performance improvement.
Researchers introduced BLAD, a historically contextualized, multilingual dataset of 1,484 Bangladeshi legal acts from 1799 to 2025.
Researchers introduced KyrgyzLLM-Bench, a new benchmark for evaluating LLMs in Kyrgyz, addressing limitations of translated English datasets.
Research introduces the QQ equality as an audit criterion to detect question-order effects in LLMs, characterizing which mechanisms satisfy it.
Research identifies consistent n-gram patterns in LLM-generated text, distinguishing it from human writing and indicating a restricted semantic range.
TalTech developed a system using fine-tuned open-weight speech LLMs for direct SOAP note generation from doctor-patient conversations without transcription.
EvolvingWorld is a new framework and benchmark for co-evolving AI agents and world models in interactive literary simulations, focusing on long-horizon character and world progression.
Research suggests repairing all missing modalities in multimodal sentiment analysis is not always optimal; some samples perform better with modality subsets.
Research proposes "Debate-on-Graph" to improve LLM reasoning by addressing noise and errors in knowledge graphs for QA tasks.
Research finds clinical safety evaluations of LLMs in English do not transfer to other languages (e.g., Hausa), especially for smaller models.
Research explores Bayesian data mixing and empirical X-risk minimization to improve AI-text detection, addressing out-of-distribution performance.
Research finds LLM code generation performance varies significantly based on prompt personas, with model-dependent effects observed.
Research demonstrates multilingual sentence embeddings can effectively replace translation for Linguistic-Integrated Reliability Auditing across 11 assessment items.
New research proposes Token-Level Off-Policy Learning (TOPL) to improve faithful generation of LLMs by training models to distinguish correct tokens.
Research decomposes the net gain from test-time collaboration methods (e.g., self-consistency, best-of-N, verifiers) for LLMs into measurable factors.
Research investigates how shared subword vocabularies in multilingual LLMs handle cross-lingual homographs and false friends, identifying limitations.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion