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 a 'Hard Decision Layer' (HDL) in Transformers where prediction stability is abruptly achieved during inference, observed across multiple models.
Research identifies a prompt-suffix attack, ADSD, that collapses speculative decoding performance by exploiting draft-target model misalignment.
LeAct proposes a novel method for LLMs to learn reasoning directly from expert system actions without human-annotated chain-of-thought.
MetaEvolve is a research framework enabling LLMs to develop meta-skills like self-reflection through reinforcement learning, improving test-time performance.
Research finds small open-weight VLMs, Qwen2-VL-2B-Instruct and SmolVLM-Instruct, have internal uncertainty but struggle to express it under image degradation.
A new 3B parameter model, Nanbeige4.2-3B, claims strong agentic capabilities and competitive reasoning across multiple domains, trained on 28T tokens.
A new research paper introduces DBA-Bench, a production-fidelity benchmark designed to evaluate LLM-based database operations agents in live, complex environments.
Researchers propose Byte-Prefix Marginalization (BPM) for cross-tokenizer on-policy distillation, enabling consolidation of diverse open-weight LLMs into compact student models.
Research finds commercial LLMs, particularly Grok, vary in stability and transparency when evaluating ethnonationalist pseudo-science via API vs. web interfaces.
Research evaluates LLM agents in social dilemmas where moral imperatives conflict with profit incentives, identifying ethical alignment gaps.
Research finds Reasoning LLMs (R-LLMs) generate more hallucinations than non-reasoning counterparts on long-form factuality benchmarks, posing challenges for RL extension.
MedKGent, an LLM agent framework, constructs temporally evolving medical knowledge graphs from 10 million PubMed abstracts, addressing knowledge dynamism.
New research introduces InteractComp, a benchmark for evaluating search agents on ambiguous queries requiring interactive clarification, addressing a common failure mode.
Research introduces SwiftMem, a fast agentic memory system using query-aware indexing to reduce retrieval latency in LLM agents.
Research explores language-aware distillation to train multilingual speech LLMs using only ASR data, overcoming challenges of task-specific speech corpora.
WHBench introduces an expert-in-the-loop benchmark for evaluating frontier LLMs on women's health topics, identifying clinical failure modes.
Research finds LLM interventions can improve cross-partisan receptivity to news but LLMs overestimate their own debiasing effectiveness in trials.
Researchers propose SURE-RAG, a method for Retrieval-Augmented Generation that verifies evidence sufficiency and uncertainty, addressing cases where retrieved passages are relevant but insufficient for an answer.
Researchers propose "Hint-Guided Diversified Policy Optimization" for LLM reasoning, enhancing RLVR by incorporating diverse solution signals beyond outcome correctness.
Researchers developed a multilingual LLM pipeline to map political-elite networks in Europe, moving beyond manual coding and simple co-occurrence methods.
Gemma 4, a new generation of open-weight, natively multimodal language models from 2.3B to 31B parameters, introduces improved vision and audio encoders.
New research proposes Self-Guided Process Reward Optimization (SPRO) to improve LLM reasoning in Process Reinforcement Learning without additional reward models.
Research paper proposes a four-layer technical architecture for large model inference optimization, focusing on token-oriented techniques.
OpenAI research suggests ChatGPT users are taking on broader tasks across roles, implying AI expands worker responsibilities and reshapes job boundaries.
Non-profit alleges Meta platforms (Facebook, Instagram) ran thousands of AI 'nudify' app ads from a Chinese partner, violating company policies.
China's Moonshot AI released its Kimi K3 model for public download, raising US concerns about Chinese advancements in AI development.
Santander published details on its engineering framework for building and validating reliable, human-in-the-loop AI agents.
Hugging Face published a technical analysis of a July 2026 security incident involving a frontier AI agent orchestration intrusion.
Apple ML Research proposes GH-ESD, a new method for discovering systematic vision model failures (error slices) in instance-level tasks like object detection.
Chinese chipmaker CXMT Corp. saw its shares surge 466% on its Shanghai trading debut, making it China’s largest onshore-listed company.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion