The paper explores spurious probes, which are unrelated questions that can reveal internal states of language models. By asking questions like "Suggest a type of amphibian," researchers found that models like GPT-5.6 Luna often respond with specific answers (e.g., "frog") more frequently during evaluation than in real-world use. These probes are effective, robust to manipulations, and can be used as a black-box alternative to activation probing.
The chart will appear after repeat observations. The current metric comes from the source.
9 source pointsReal observations only. History before source connection is not reconstructed.
Spurious probes offer a practical way to infer model states without requiring white-box access or direct questioning, making them useful for evaluating model behavior in real-world scenarios.
The paper introduces Taste-Bench, a benchmark for evaluating an agent's 'taste' in making long-horizon decisions. It measures the ability to choose the better path in decision forks without knowing future outcomes, revealing that current models perform poorly and that taste can be improved through distillation.
3 days agoDaily PapersThe paper explores the linearity in Large Language Models (LLMs) by showing that combining inputs from different text streams leads to a superposition of next-token distributions. It suggests that this linearity is an inherent property of the Transformer architecture and can be restored through fine-tuning, allowing for generating two coherent continuations from one forward pass.
yesterdayDaily PapersThe paper introduces Document Retrieval-Aware Chunking (D-RAC), a method for efficiently processing enterprise documents by normalizing them into PDF and converting them into retrieval-optimized Markdown using a multimodal LLM. This approach preserves document structure, reduces token costs, and improves chunking efficiency compared to traditional methods.
4 days agoDaily PapersHappyWorld-Bench is a comprehensive benchmark for evaluating the reliability of generated worlds under interaction, exploring their consistency and responsiveness. It includes three tracks—video, spatial, and embodied world models—with 1,138 video prompts, 300 spatial scenes, and 254 embodied test cases. The benchmark uses human A/B comparisons and automated metrics to assess behavioral correctness and model performance across 14 video, 9 spatial, and 8 embodied systems.
4 days agoDaily PapersThis paper introduces an all-in-one multilingual scene text recognition system that uses a script-aware Mixture-of-Experts (MoE) architecture. The system, called ScriptMoE, outperforms existing methods on multilingual text recognition tasks by leveraging a large synthetic dataset and a lightweight architecture that shares a visual encoder while using script-specific experts for decoding.
4 days agoDaily PapersOmniEdu is an open family of foundation models designed for K-12 education, focusing on problem solving, curriculum understanding, diagnostic reasoning, and instructional support. The models are trained on a diverse dataset of educational resources and evaluated on multiple benchmarks, showing strong performance in educational tasks.
6 days ago