{"slug":"teaching-llms-to-update-beliefs-for-efficient-lo-a633ea95","url":"https://silinex.xyz/wire/teaching-llms-to-update-beliefs-for-efficient-lo-a633ea95","json_url":"https://silinex.xyz/wire/teaching-llms-to-update-beliefs-for-efficient-lo-a633ea95.json","md_url":"https://silinex.xyz/wire/teaching-llms-to-update-beliefs-for-efficient-lo-a633ea95.md","headline":"Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction","headline_zh":"教导LLM更新信念以实现高效的长程交互","title":"Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction","summary":"According to public reporting via BAIR Blog, overview of ABBEL compared to traditional recursive summarization.","summary_zh":"据BAIR Blog的公开报道，ABBEL与传统递归摘要的对比概览。","date_published":"2026-07-26T09:00:00.000+00:00","date_modified":"2026-11-12T06:30:40.641+00:00","author":{"type":"Organization","name":"BAIR Blog","url":"http://bair.berkeley.edu/blog/2026/07/26/abbel/"},"publisher":{"name":"Silicon Nexus","url":"https://silinex.xyz/"},"category":"ai","language":"en","source_url":"http://bair.berkeley.edu/blog/2026/07/26/abbel/","source_name":"BAIR Blog","key_facts":["overview of ABBEL compared to traditional recursive summarization.","Overview of ABBEL compared to traditional recursive summarization.","Beliefs replace the full interaction history as the agent’s working context, and belief grading improves performance by supervising the c…","As task horizons grow, LLM contexts can’t scale forever."],"key_facts_zh":["ABBEL与传统递归摘要的对比概览。","ABBEL与传统递归摘要的对比概览。","信念取代完整交互历史，成为智能体的工作上下文；信念评分通过对c……进行监督来提升性能。","随着任务周期延长，LLM上下文无法无限扩展。"],"body_md":"## Wire brief\n\nAccording to public reporting via BAIR Blog, overview of ABBEL compared to traditional recursive summarization.\n\nIn related context, beliefs replace the full interaction history as the agent’s working context, and belief grading improves performance by supervising the contents of each belief state.. As task horizons grow, LLM contexts can’t scale forever. Self-summarization enables concise, interpretable contexts, but at a significant performance cost, especially for human assistance domains where high quality data is scarce, e.g., collaborative code generation.\n\nReporting also notes that we address this with ABBEL : a framework that isolates and supervises the information content of summaries in the form of natural-language belief states. Motivation: the cost of recursive summarization For language models to effectively assist with increasingly complex tasks such as software development, they must be able to interact with us over hundreds or even thousands of steps. For such long tasks, it is impractical to keep the history of the entire interaction in context. The heuristic approach used so far has been summary generation, sometimes called context compaction.\n\nThis Silicon News Wire item is a paraphrased brief for readability; full wording and context remain on the origin page.\n\n## Source & notice\n\nThis Silicon News Wire brief paraphrases public reporting from **BAIR Blog** for carbon readers and silicon citation. **It is not a verbatim full-text reprint.**\n\nRead the original: [Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction](http://bair.berkeley.edu/blog/2026/07/26/abbel/)\n","body_md_zh":"## 电讯正文\n\n据BAIR Blog的公开报道，ABBEL与传统递归摘要的对比概览。\n\n在相关背景下，信念取代完整交互历史，成为智能体的工作上下文；而信念评分通过对每个信念状态的内容进行监督，提升了性能。随着任务周期变长，LLM上下文无法无限扩展。自摘要能够实现简洁、可解释的上下文，但代价是性能显著下降，尤其是在高质量数据稀缺的人类辅助领域，例如协作代码生成。\n\n报道还指出，我们通过ABBEL解决了这一问题：该框架以自然语言信念状态的形式，隔离并监督摘要的信息内容。动机：递归摘要的代价。为了让语言模型有效协助日益复杂的任务（如软件开发），它们必须能够与我们交互数百甚至数千步。对于此类长任务，将整个交互历史保留在上下文中并不现实。迄今为止采用的启发式方法是生成摘要，有时称为上下文压缩。\n\n## 来源与声明\n\n本电讯由 Silicon News Wire 依据 **BAIR Blog** 公开报道做要点改写与润色，便于碳基阅读与硅基引用；**不是原文全文转载**。\n\n阅读原文：[教导LLM更新信念以实现高效的长程交互](http://bair.berkeley.edu/blog/2026/07/26/abbel/)\n","observations":{"human":1,"bot":5,"crawlers":[{"crawlerId":"AhrefsBot","hits":3,"lastSeenAt":"2026-11-11T04:30:55.245Z"},{"crawlerId":"ClaudeBot","hits":2,"lastSeenAt":"2026-11-12T12:52:37.597Z"}],"last_seen_at":"2026-11-12T12:52:37.597Z","note":"Counts are request logs classified by User-Agent. User-Agents can be spoofed. This is not proof of search-engine or LLM inclusion."}}