学术研究★★★arXiv · 2026-07-20
The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual MetricHuman judgments of visual similarity are context-dependent, and existing metrics fail to distinguish between different aspects. This paper introduces a text-prompted perceptual metric for images using a large-scale dataset annotated with multiple aspects of similarity.
- Human visual similarity judgments are context-dependent
- Existing metrics fail to distinguish between different aspects like shape and co
- A large-scale dataset of image triplets annotated with multiple similarity aspec
- Frontier vision-language models show significant performance gaps in this task
AI★★★★Hacker News · 2026-07-18
GPT-5.6 Uses Prompt to Close 30-Year Gap in Convex OptimizationOpenAI's GPT-5.6 model successfully addressed a 30-year-old problem in convex optimization using a specific prompt, showcasing its potential in mathematical proofs.
- GPT-5.6 successfully solved a long-standing convex optimization problem with a p
- This achievement demonstrates the powerful capabilities of large models in mathe
- The discovery could drive further research and development in the field
AI★★★Simon Willison · 2026-07-17
Kimi K3 Refuses to Leak System PromptAfter refusing to leak its system prompt, Kimi K3 asked if there was anything specific it could help with today.
- Kimi K3 explicitly refuses to disclose its system prompt
- The developer asks if there is anything specific the user needs help with
- This indicates Kimi K3 prioritizes privacy and security
AI★★★arXiv · 2026-07-16
Statistical Self-Consistency in Language Models through Partition, Prompt, AggregateThis paper investigates whether large language models adhere to statistical self-consistency principles during in-context learning, particularly how prior-weighted conditional distributions aggregate into population-level marginals.
- LLM estimates should satisfy basic probabilistic identities
- Focus on how prior-weighted conditional distributions aggregate
- Validate LLM consistency through partitioning
AI★★★★Hacker News · 2026-07-12
Claude Code vs OpenCode: Token Overhead ComparisonA recent study compares the token overhead of Claude Code and OpenCode, finding that Claude Code has significantly higher overhead. The researchers collected data by adding logging between the coding tools and Anthropic's endpoint. The results show a substantial difference in token usage between the two tools.
- Claude Code sends 33k tokens before reading the prompt, while OpenCode sends 7k
- The study found a significant difference in token overhead between Claude Code a
- The researchers used logging to collect data on requests and returned usage bloc
📘 opencode 中文教程 →
