grounding

IBM researchers formalize LLM consistency as scoped semantic invariance over knowledge-graph-defined semantic domains, showing inconsistencies cluster in stable hotspot subgraphs and that selective, layered KG grounding improves cross-model agreement and reliability.

François Chollet argues that intelligence is rarely the bottleneck to solving impactful problems — grounded, iterated contact with reality is — comparing ungrounded theorizing to an intelligent system hallucinating due to lack of sensory input, as echoed by Elon Musk.

GroundingPI 与 XPeng、北大、港大合作发布了3个开源模型：4B基础模型（检测、指代、OCR等多能力）、基于扩散并行解码的快速grounding模型GroundAnything，以及自回归VLM版本的高精度grounding模型。

论文提出 MULTIFINGROUND 多语言金融基准，揭示语言模型虽能跨语言识别财务概念（准确率 97.3-99.4%），却在 26.6-41.9% 的情况下无法完整恢复对应的数值证据，凸显识别与证据落地之间的差距。

OmniCompass 提出了一个面向全模态视频定位与理解的整体基准，涵盖视觉-语言与音视频两条评测轨道，包含时间定位、空间定位与问答三类任务，并对50个多模态大模型进行了系统评估。研究发现开源模型在音视频理解上仍落后于闭源模型，精确的时间与空间定位对领先模型而言仍具挑战性。

Maxime Rivest 推荐使用开放、免费且快速的科学检索方式来为 AI agent 的回答提供事实依据，认为这种做法明显更好且应更早普及。

AI4Fire 评估六大核心大模型在五项野火管理任务上的零样本表现，涵盖裸输入与任务特定 grounding 两种设置，发现 grounding 仅在补充信息直接包含答案时有效（如 SQL 工具将准确率从≤16%提升至≥88%），简单规则常胜过模型，且多个公开数据集存在标签缺陷可能虚高评分。

提出“隐喻电路”假说，解释 Vision Transformer 如何在缺乏直接指称证据时通过具体锚点概念（如 fire）间接接地抽象概念（如 angry），并利用 Transcoders 对 CLIP 和 DINO 的电路追踪与因果干预进行了验证。

Cloudflare launched the Web Search API in beta, allowing AI agents and applications to search the internet and ground responses in live information. It supports three providers (Ceramic.ai, Exa, Linkup) with Zero Data Retention, runs through AI Gateway, and is accessible via REST API or Workers AI binding.

LlamaIndex 发布 Extract v2.5，一个面向结构化文档提取的智能体架构，在不涨价的前提下将 ExtractBench 上各层级精度与 grounding 分数大幅提升，并显著优于前沿模型的价格/性能比。

LlamaIndex introduces Extract v2.5, a set of frontier-agent document extraction models (Cost Effective, Agentic, Agentic Plus) that now claim SOTA price-performance on ExtractBench, outperforming Opus 5.5 and GPT-6 Sol at 30%-4x lower cost, with major accuracy gains on long lists, multi-page records and scanned forms, plus new Advanced Citations and Structural Reasoning features available on LlamaParse.

GroundingPI：一个 4B 参数的视觉定位基础模型，以量化坐标生成点与框，在 34 个定位基准上以 73.68% 平均成绩超越 GPT-6 Astra 等 44 个基线，并在机器人操作与自动驾驶下游任务中显著提升性能。

SpatialCORE is a post-training framework that uses a model's confidence in its generated bounding-box grounding as a learning signal for spatial reasoning in large vision-language models, achieving state-of-the-art results among open-source models on spatial reasoning benchmarks.

ReGround is a large-scale dataset for grounding reviewer comments in multimodal evidence from scientific papers, revealing the challenges and importance of integrating text, tables, and figures in retrieval tasks.

This paper introduces a graph-grounded harness for vision-language models to improve accuracy in answering topology questions about Piping and Instrumentation Diagrams by recovering evidence graphs from images.

This paper audits scene-level confabulation in LLM-generated autobiography against a documented ground-truth corpus, finding a 96.7% verification-failure rate and contributing a reusable audit instrument and a grounding remedy.

Introduces a rubric-based reinforcement learning method to improve visual faithfulness in vision-language models by decomposing responses into atomic propositions and scoring them on visual faithfulness, reasoning consistency, and instruction following.

ExtractBench is a new benchmark for document extraction that strictly scores grounding accuracy, showing VLMs and coding agents lack evidence while LlamaExtract Agentic Plus leads with high page-level and word-level performance.

This paper studies the trade-off between grounding and coverage in long-form hallucination reinforcement learning, proposing rubric-based rewards to represent required and optional information for questions. A soft combination of grounding, rubric coverage, and relevance yields the best balance between support and richness.

LiquidAI releases LFM2.5-VL-3B, a 3B multimodal model for on-device deployment with improved OCR, grounding, and efficient inference, available in multiple formats including GGUF, ONNX, and MLX.