AEGIS: Awareness-Enhanced Guidance for Iterative Safeguard
AEGIS is a new research framework exploring span-guided multilingual detoxification for LLMs across English, Mandarin Chinese, and Korean.
arXiv cs.CL — Computation and Language·21 Jul 2026
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.
AEGIS is a new research framework exploring span-guided multilingual detoxification for LLMs across English, Mandarin Chinese, and Korean.
Research paper proposes C^2KV, a new method for compressed and composable KV cache reuse to improve LLM inference efficiency for long contexts.
Research evaluates SOTA LLMs for citation function classification, achieving new high benchmarks on the ACL-ARC dataset.
Research proposes Evidence-Grounded Terminology Adaptation (EGTA) for simultaneous speech translation, focusing on recovering paper-specific terminology.
ESCUCHA is introduced as the first Spanish speech understanding benchmark to evaluate large audio language models (LALMs) across diverse acoustic conditions.
A new benchmark dataset and method for disambiguating culturally entangled Bangla homographs in low-resource LLMs was introduced on arXiv.
DeLIVeR, a new framework, uses a Planner LLM to decompose complex claims for fact-checking via reinforced knowledge graph exploration.
New research introduces Pancasila-Dilemmas, a dataset of 1,834 questions from Indonesian news to evaluate LLM value alignment with country-specific values.
VEHBench is a new diagnostic benchmark for evaluating LLM performance across different stages of iterative physical engineering design workflows for vibration energy harvesters.
PPL-Factory is a research paper proposing a task-aware and budget-aware method for selecting training data to fine-tune LLMs, improving efficiency.
SWE-Pruner Pro, a new method for LLM coding agents, prunes tool outputs by leveraging the agent's internal relevance representations.
Research introduces a typology for evaluating LLM responses to user expressions of belief, noting linguistic diversity impacts LLM persuasiveness.
DocOCR-Eval proposes a framework for selecting OCR tools and MLLMs for document parsing without requiring ground truth data for evaluation.
ColGraphRAG introduces late-interaction evidence retrieval for multimodal graph-grounded question answering, improving accuracy for graph-linked images.
RAIL Guard introduces a closed-loop system for evaluating and iteratively remediating unsafe LLM agent outputs, moving beyond binary blocking.
Research explores expressing and editing preference model inferences in natural language to address opacity and improve interpretability.
Research introduces CIGPO, a new reinforcement learning method for multi-turn LLM agents, improving stability in evidence-reading tasks compared to GRPO.
Research explores using Large Language Models with evolutionary algorithms to automate and improve feature generation for machine learning pipelines.
New RL policy gradient method aims to improve long-horizon language agent training by better attributing actions to outcomes, reducing optimization variance.
Research quantifies uncertainty in LLM benchmark rankings (e.g., MMLU) using rank confidence intervals and modified hypothesis tests.
CLARE, a self-evolving AI agent, is proposed to resolve 'intent asymmetry' in 3D tool orchestration by clarifying vague user instructions.
Research explores cross-modal unlearning in Vision-Language Models (VLMs) across three architectures to assess if unlearning transfers across modalities.
Research investigates multimodal LLMs' capability for optical coherence tomography (OCT) imaging analysis beyond basic classification, focusing on clinical reasoning.
Researchers propose a dependency-aware framework for automated code generation, aiming to improve logical completeness and integration.
DS@GT developed a hybrid multi-agent LLM system with structured algorithmic guidance for conversational depression screening at eRisk 2026.
Research explores methods to safeguard facial identities against unauthorized manipulation by unified multimodal image editing models using 'cross-branch conflict'.
New research proposes a method for learning human value systems to improve alignment in generative AI, addressing a gap in current approaches.
Research introduces C-MTP, a direct supervision method for Continuous Chain-of-Thought (CoT) models, offering faster training than prior indirect methods.
Research evaluates if off-the-shelf LLMs can replicate the diverse functionalities and methodologies of traditional privacy policy analysis tools.
Research introduces Otap, a new evaluation metric using optimal transport for LLM agent trajectories, assessing planning, tool use, and execution beyond binary success.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion