Tag
This paper argues that LLMs' pretrained world knowledge is accessible for temporal event prediction when the interface is properly designed, introducing DescribeEvent (~150K events, 7 datasets) and showing that natural-language event descriptions plus symbolic labels let capable LLMs outperform specialized baselines.
TerraVis 提出了一个基于 MLLM 工作流的框架，用于评估文生图模型生成图像的世界一致性（如畸形结构、物理上不可能的交互、空间关系不一致等），并定义了涵盖对象、交互和场景层级的 18 类一致性违规分类体系。实验表明该框架与人类判断相关性最强，且常规指标表现优秀的模型仍可能存在大量世界一致性缺陷。
EVOKE is a post-training method that elicits world knowledge already present in pretrained LLM agents by contrasting alternative goals at fixed states, training policies to rank candidate actions by their consequences rather than contextual habits. It achieves state-of-the-art results across ALFWorld, WebShop, and search-based QA, and transfers well to unseen environments without needing world models or inference-time planning.
The author discusses the need for AI models with better world knowledge, leveraging N-gram technology to fit more knowledge into smaller models, and questions why development focuses more on coding capabilities than broader world knowledge.
The author advises against deleting older AI models due to their valuable world knowledge and clarity, sharing a personal experience where an older model like DeepSeekV3.2 performed best for a specific task.
The article suggests using a smaller 4B LLM with Kiwix skill and local Wikipedia to avoid hallucinations about world knowledge, instead of relying on larger models.
A user observes that Gemini 3.1 possesses strong world knowledge but tends to output lazy responses, not fully utilizing its capabilities.
According to the DeepSeek V4 technical report's evaluation of mainstream LLMs, Gemini 3.1 Pro is considered to have the strongest world knowledge, but users generally find it hard to use because the model does not proactively use search tools.
A new paper introduces an outcome-based reward that quantifies how self-generated world knowledge boosts task success, enabling agents to improve without external guidance at inference.
This paper proposes a method to train LLM agents with intrinsic meta-evolution capabilities, enabling spontaneous self-improvement without external rewards at inference time. Applied to Qwen3-30B and Seed-OSS-36B, the approach yields a 20% performance boost on web navigation benchmarks, with a 14B model outperforming Gemini-2.5-Flash.