{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-28T06:20:51.496Z","headline":"ABBEL：通过信念瓶颈提升LLM长程交互中的摘要学习效率","description":"BAIR提出ABBEL框架，将长程交互中的摘要生成隔离为自然语言信念状态，并通过信念评分进行监督。在CollabBench协作编程测试中，基于重建的信念评分将全上下文模型的性能差距缩小约50%，训练步数减少50%。该方法解决了递归摘要（如Cursor Composer 2.5采用的压缩技术）在有限训练数据下性能下降的问题。","url":"https://www.aioga.com/news/cms2p1wyq037aro3fl2in7qni/","mainEntityOfPage":"https://www.aioga.com/news/cms2p1wyq037aro3fl2in7qni/","datePublished":"2026-07-26T09:00:00.000Z","dateModified":"2026-07-26T09:00:00.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["http://bair.berkeley.edu/blog/2026/07/26/abbel","https://aihot.virxact.com/items/cms2p1wyq037aro3fl2in7qni"],"canonicalUrl":"https://www.aioga.com/news/cms2p1wyq037aro3fl2in7qni/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：BAIR提出ABBEL框架，将长程交互中的摘要生成隔离为自然语言信念状态，并通过信念评分进行监督。 Aioga 将其归入「论文研究」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cms2p1wyq037aro3fl2in7qni/","dateCreated":"2026-07-26T09:00:00.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"BAIR：Berkeley AI Research Blog source article","url":"http://bair.berkeley.edu/blog/2026/07/26/abbel","datePublished":"2026-07-26T09:00:00.000Z","provider":{"@type":"Organization","name":"BAIR：Berkeley AI Research Blog","url":"http://bair.berkeley.edu/blog/2026/07/26/abbel"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cms2p1wyq037aro3fl2in7qni","datePublished":"2026-07-26T09:00:00.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cms2p1wyq037aro3fl2in7qni"}}],"aggregationSource":"BAIR：Berkeley AI Research Blog","originalPublisher":{"name":"BAIR：Berkeley AI Research Blog","url":"http://bair.berkeley.edu/blog/2026/07/26/abbel"},"article":{"id":"cms2p1wyq037aro3fl2in7qni","slug":"cms2p1wyq037aro3fl2in7qni","url":"https://www.aioga.com/news/cms2p1wyq037aro3fl2in7qni/","title":"ABBEL：通过信念瓶颈提升LLM长程交互中的摘要学习效率","title_en":"Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction","summary":"BAIR提出ABBEL框架，将长程交互中的摘要生成隔离为自然语言信念状态，并通过信念评分进行监督。在CollabBench协作编程测试中，基于重建的信念评分将全上下文模型的性能差距缩小约50%，训练步数减少50%。该方法解决了递归摘要（如Cursor Composer 2.5采用的压缩技术）在有限训练数据下性能下降的问题。","source":"BAIR：Berkeley AI Research Blog","sourceUrl":"http://bair.berkeley.edu/blog/2026/07/26/abbel","aiHotUrl":"https://aihot.virxact.com/items/cms2p1wyq037aro3fl2in7qni","publishedAt":"2026-07-26T09:00:00.000Z","category":"论文研究","score":51,"selected":false,"articleBody":["--> 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 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. We address this with ABBEL：https://arxiv.org/abs/2512.20111: a framework that isolates and supervises the information content of summaries in the form of natural-language belief states.","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. For example, Cursor’s latest model composer 2.5 uses compaction during training for improved performance (Cassano et al., 2026：https://cursor.com/blog/self-summarization). Alongside composer, Grandcode (DeepReinforce et al., 2026：https://arxiv.org/abs/2604.02721), the first system to consistently beat all human competitors in online coding competitions, despite using one of the newest efficient attention models (Qwen 3.5-397B), 1 ：#fn:efficientattn still found it necessary to employ context summarization.","But compaction has a problem. Despite seemingly low performance gaps in benchmarks, model servers like Cursor continue to recommend that users avoid compaction with their coding assistants in the middle of a task (Heule et al., 2026：https://cursor.com/blog/continually-improving-agent-harness).","To understand why, see below the performance over RL fine-tuning of a Context summary model compared to full context models in Combination Lock, a Wordle-like game that allows up to 16 guesses. 2 ：#fn:wordlehard Though both model types improve over the course of training, the summary model never closes the gap.","--> Fig. 1: Average attempts to guess the target word on Combination Lock over RL fine-tuning (lower is better). Context-summary policies improve with training but do not close the gap to full-context policies.","Making models self-summarize while completing a task increases the complexity of the learning problem. While this could typically be addressed by training with more data, the performance degradation observed in real world interactive settings likely arises from the difficulty we have in creating and using human simulators effectively to generate high quality training environments (Lin et al., 2025：https://jessylin.com/2025/07/10/user-simulators-1/, Tomlin et al., 2025：https://nickatomlin.github.io/blog/user-simulators-2.html). Thus, the better you can learn to summarize on the limited and messy multiturn interaction trajectories you can collect, the better off your model will be for downstream users.","--> Fig. 2: Autoencoder-inspired belief grading. The model encodes prior belief, action and observation (b t , a t , o t ) into posterior belief b t+1 and is rewarded for how well select information from the history can be reconstructed from that belief.","To address poor learning efficiency, we isolate the summary generation task. Drawing inspiration from recursive Bayesian estimation, we formulate summaries as belief states, which we periodically prompt the model to update based on new information. 3 ：#fn:videoabbelslow","We then extract and supervise the contents of the belief states (Fig. 2, Belief Grading). Belief grading can be thought of as adding an auxiliary RL task, using heuristics designed to capture what makes a good belief as the reward. An example heuristic for coding could be shorter is better, but closer to being able to reconstruct the git diff is also better, so balancing these would yield a good belief. In domains where good heuristics are hard to define, we propose a general autoencoding-inspired grading function, which treats the current language model π θ as both encoder and decoder of information from the history, and the belief states as the codes. We grade each belief b t+1 by how well it can be used by the current model π θ to reconstruct the most recent observation o t :","--> Eq. 1: Reconstruction grading objective. Here b t+1 is the updated belief, o t the latest observation, a t the action just taken, b t the prior belief, p I the task prompt, and π θ the current model. Higher grades reward beliefs that retain information needed to decode the latest observation.","We demonstrate the utility of belief grading in our motivating domain of human-driven assistive coding, with the CollabBench environment from Sweet-RL (Zhou et al., 2025：https://arxiv.org/pdf/2503.15478).","--> Fig. 4: CollabBench collaborative coding environment. The agent asks clarifying questions, then submits a function scored against hidden unit tests.","We see that with the general reconstruction-based belief grading function we reduce the performance gap from full context models by about 50%, and train in 50% fewer steps compared to training models to summarize without belief grading (no BG). After training, ABBEL still uses significantly less memory than the full context setting, as measured by the peak context token length (Peak Tokens).","Additionally, in CombinationLock, we demonstrate that ABBEL with a belief grader which leverages domain knowledge (by computing useful statistics over the history and checking that they can be reconstructed from the belief state), enables even higher learning efficiency than full context (FULL CTX) models.","--> Fig. 6: Average attempts to guess the target word on Combination Lock (lower is better). With domain-knowledge belief grading, ABBEL approaches or exceeds FULL CTX in this setting; without belief grading, learning is slower.","In a third environment, multi-objective question answering (from MEM1 Zhang et al., 2025：https://arxiv.org/pdf/2512.24601, a recent work which performed end-to-end optimization in a modified version of typical recursive summarization), we demonstrate the utility of isolating belief states from reasoning, by showing that a Peak Belief length Penalty (more details in paper) significantly reduces memory usage with minimal performance degradation, unlike is commonly observed when penalizing reasoning lengths (Arora et al., 2025：https://proceedings.neurips.cc/paper_files/paper/2025/file/579b5b84311e122584dedab1d3b7613f-Paper-Conference.pdf).","--> Fig. 7: Exact-match score and peak memory versus number of objectives in multi-objective QA. ABBEL with a peak belief penalty (PBP) maintains comparable performance while using less memory than MEM1 and ABBEL without PBP in this evaluation.","Many more possibilities are enabled through using explicit belief states as information bottlenecks for multi-step interaction. You could reward actions based on their effect on the belief state to guide exploration, transmit the explicit belief states for better communication between agents, or even improve user controllability by directly modifying the memories on which the agents’ decisions are based.","Some forms of information, e.g., what a person looks like, are not represented well by text alone. A continuously learning system will also have to capture such information. Additionally, if we want a system to learn to communicate in a brand new language or to play a brand new game better than any person in the world, the skills accumulated over the lifetime of conversations or games must be stored in a very compressed form, essentially taking on the role of the weights of the model itself.","More powerful systems will likely utilize a combination of multiple forms of memory, where the contents of the context may correspond to working memory while other approaches are used for short and long-term memory. How to instantiate these other forms of memory, for instance via test-time training, adapter memories, continuous context memories, or some combination thereof, presents an exciting challenge.","If abbel was inspiring for your future work, please cite us with this! And here is some advice：https://jakob-bjorner.github.io/abbel-advice.html for doing similar research!","With newer models the number of tokens till 50% compute spend is on attention gets much larger than 25K. Interleaving linear attention alternatives with full attention as is done with gpt-oss and DeepSeekv4, results in massive flops reductions for the attention computation. For example with DeepSeekv4-Pro (1.6T A49B) it requires nearly 450 thousand tokens to reach the 50% tradeoff point. Grandcode uses Qwen-3.5-397B-A17B a model which hits 50% FLOPs for attention at ~150 thousand tokens. ↩：#fnref:efficientattn","This setting is technically solvable with much more computationally effective tools, but serves as a flexible test bed to study properties of recursive summarization. Bertsimas et al., 2022：https://wordle-page.s3.amazonaws.com/assets/Wordle_Paper_Final.pdf, showed that an exact solution for the wordle game instantiated with the original vocabulary of the javascript game can be found with dynamic programming, but evidently the general formulation of wordle as a guessing game on K letters with L attempts and some dictionary of valid words and correct words D is NP hard to determine the minimal number of moves required. ↩：#fnref:wordlehard","In practice there is an O(N/K) overhead cost for summary. N is the total number of actions. K is the number of actions till summarization is triggered. This is necessarily true for any summary approach. For ease of illustration this gif uses K = 1. In our experiments, to put more emphasis on summarization weaknesses we also use K=1. In practice overhead is small as K can be chosen to be near the efficient hardware limit. ↩：#fnref:videoabbelslow"],"articleImages":[{"sourceUrl":"https://bair.berkeley.edu/static/blog/abbel/frames/frame_00.png","alt":"ABBEL overview animation frame","afterParagraph":8,"url":"/media/articles/cms2p1wyq037aro3fl2in7qni/e7fa3a3f049a8c7c.png"}],"mediaStatus":"ok","articleBodyZh":["--> ABBEL 相较于传统递归总结的概述。信念代替完整的交互历史作为代理的工作上下文，而信念评分通过监督每个信念状态的内容来提升性能。","随着任务范围的扩大，LLM 的上下文无法无限扩展。自我总结可以生成简洁、可解释的上下文，但代价是显著的性能损失，特别是在高质量数据稀缺的人类辅助领域，例如协作代码生成。我们使用 ABBEL 解决这一问题：https://arxiv.org/abs/2512.20111：一个框架，以自然语言信念状态的形式隔离和监督总结的信息内容。","为了使语言模型能够有效地协助处理日益复杂的任务，如软件开发，它们必须能够与我们进行数百甚至数千步的交互。对于如此长的任务，在上下文中保留整个交互历史是不现实的。到目前为止使用的启发式方法是生成总结，有时称为上下文压缩。例如，Cursor 的最新模型 Composer 2.5 在训练期间使用压缩以提高性能（Cassano 等, 2026：https://cursor.com/blog/self-summarization）。与 Composer 一起，Grandcode（DeepReinforce 等, 2026：https://arxiv.org/abs/2604.02721），这是首次在在线编码竞赛中持续击败所有人类竞争者的系统，尽管它使用了最新的高效注意力模型之一（Qwen 3.5-397B），仍发现有必要使用上下文总结。","但压缩存在问题。尽管在基准测试中表现差距似乎很小，模型服务商如 Cursor 仍建议用户在任务进行中避免对代码助手进行上下文压缩（Heule 等, 2026：https://cursor.com/blog/continually-improving-agent-harness）。","要理解原因，请参见下图对 Context 总结模型与完整上下文模型在 Combination Lock（一种类似 Wordle 的游戏，允许最多 16 次猜测）上 RL 微调的性能比较。虽然两种模型类型在训练过程中都有提升，但总结模型从未缩小差距。","--> 图 1：在 RL 微调过程中，在 Combination Lock 上猜测目标词的平均尝试次数（越低越好）。上下文摘要策略随着训练而改进，但仍未达到全上下文策略的水平。","在完成任务的同时让模型自我总结会增加学习问题的复杂性。虽然通常可以通过更多数据的训练来解决，但在现实交互环境中观察到的性能下降，很可能源于我们在有效创建和使用人类模拟器以生成高质量训练环境方面的困难（Lin et al., 2025：https://jessylin.com/2025/07/10/user-simulators-1/, Tomlin et al., 2025：https://nickatomlin.github.io/blog/user-simulators-2.html）。因此，你在有限且混乱的多轮交互轨迹上学习总结的能力越强，你的模型对下游用户就会越有利。","--> 图 2：受自编码器启发的信念评分。模型将先验信念、动作和观察（b t、a t、o t）编码为后验信念 b t+1，并根据从该信念中重构历史中所选信息的效果进行奖励。","为了解决学习效率低的问题，我们将摘要生成任务隔离开来。借鉴递归贝叶斯估计的思想，我们将摘要公式化为信念状态，并定期提示模型根据新信息更新信念。 3：#fn:videoabbelslow","然后我们提取并监督信念状态的内容（图 2，信念评分）。信念评分可以被理解为添加一个辅助的 RL 任务，使用旨在捕捉良好信念特征的启发式作为奖励。一种编码的示例启发式是越短越好，但越接近能够重构 git diff 也越好，因此平衡这些便会产生一个良好的信念。在很难定义良好启发式的领域，我们提出了一种通用的、自编码器启发的评分函数，将当前语言模型 π θ 视为历史信息的编码器和解码器，将信念状态视为代码。我们通过当前模型 π θ 能多好地重构最近的观察 o t 来对每个信念 b t+1 进行评分：","--> 式 1：重建评分目标。这里 b t+1 是更新后的信念，o t 是最新观察，a t 是刚采取的动作，b t 是先前信念，p I 是任务提示，π θ 是当前模型。较高的分数奖励能够保留解码最新观察所需信息的信念。","我们在以人为驱动的辅助编程这一激励领域中展示了信念评分的实用性，使用的是 Sweet-RL 的 CollabBench 环境（Zhou 等, 2025：https://arxiv.org/pdf/2503.15478）。","--> 图 4：CollabBench 协作编程环境。智能体提出澄清性问题，然后提交一个函数，该函数会根据隐藏的单元测试进行评分。","我们看到，通过通用重建型信念评分函数，我们将与全上下文模型的性能差距减少了约 50%，同时相比训练模型总结而不使用信念评分（无 BG），训练步骤减少了 50%。训练后，ABBE L 的内存使用量仍显著低于全上下文设定，这通过峰值上下文令牌长度（Peak Tokens）进行衡量。","此外，在 CombinationLock 中，我们演示了 ABBEL 使用借助领域知识的信念评分器（通过对历史数据计算有用统计并检查是否可以从信念状态重建这些统计）时，实现了比全上下文（FULL CTX）模型更高的学习效率。","--> 图 6：在 Combination Lock 上猜测目标单词的平均尝试次数（数值越低越好）。在有领域知识的信念评分下，ABBE L 在这一设定中接近甚至超过 FULL CTX；无信念评分时，学习速度更慢。","在第三个环境中，多目标问答（来自 MEM1 Zhang 等, 2025：https://arxiv.org/pdf/2512.24601，一项近期工作，在典型递归总结的修改版本中进行了端到端优化），我们展示了将信念状态与推理隔离的实用性，通过显示 Peak Belief 长度惩罚（具体细节见论文）在最小性能下降的情况下显著减少内存使用，而不像通常在惩罚推理长度时观察到的情况（Arora 等, 2025：https://proceedings.neurips.cc/paper_files/paper/2025/file/579b5b84311e122584dedab1d3b7613f-Paper-Conference.pdf）。","--> 图 7：多目标问答中，精确匹配分数和峰值内存随目标数量的变化。在本次评估中，带峰值信念惩罚（PBP）的 ABBEL 保持了可比性能，同时比 MEM1 和不带 PBP 的 ABBEL 使用更少内存。","通过将显式信念状态用作多步交互的信息瓶颈，可以实现更多可能性。你可以根据动作对信念状态的影响来奖励动作，以指导探索；传递显式信念状态以改善智能体之间的沟通；甚至可以通过直接修改智能体决策所依赖的记忆来提高用户的可控性。","某些类型的信息，例如一个人的长相，仅靠文本很难很好地表示。一个持续学习的系统也必须捕捉这些信息。此外，如果我们希望系统学习使用全新的语言进行交流，或者在全新的游戏中表现优于任何人，那么在对话或游戏过程中积累的技能必须以非常压缩的形式存储，本质上承担模型自身权重的作用。","更强大的系统可能会利用多种记忆形式组合，其中上下文的内容可能对应于工作记忆，而其他方法用于短期和长期记忆。如何实例化这些其他形式的记忆，例如通过测试时训练、适配器记忆、连续上下文记忆或其组合，是一个令人兴奋的挑战。","如果 ABBEL 对你的未来工作有启发，请引用我们的工作！这里有一些建议：https://jakob-bjorner.github.io/abbel-advice.html，用于进行类似研究!","在新模型中，达到 50% 计算量在注意力上的分配所需的标记数量远大于 25K。像 GPT-OSS 和 DeepSeekv4 那样在全注意力中交错线性注意力替代方案，可以大幅减少注意力计算的 FLOPs。例如，对于 DeepSeekv4-Pro（1.6T A49B），达到 50% 权衡点几乎需要 45 万个标记。Grandcode 使用 Qwen-3.5-397B-A17B，这是一个在大约 15 万标记处达到注意力 50% FLOPs 的模型。 ↩：#fnref:efficientattn","这个设置在技术上可以用更高效的计算工具来解决，但它作为一个灵活的测试平台，用于研究递归总结的特性。Bertsimas 等人，2022：https://wordle-page.s3.amazonaws.com/assets/Wordle_Paper_Final.pdf，显示对于以 JavaScript 游戏原始词汇表实例化的 Wordle 游戏，可以通过动态规划找到精确解，但显然，将 Wordle 一般化为使用 K 个字母进行猜测的游戏，并且有 L 次尝试和某个有效词典以及正确词 D 的情况下，要确定所需的最少步数是 NP 难的。↩：#fnref:wordlehard","在实践中，总结有一个 O(N/K) 的额外开销。N 是动作的总数。K 是触发总结前的动作数。对于任何总结方法，这都是必然的。为了便于说明，这个 gif 使用 K = 1。在我们的实验中，为了更强调总结的弱点，我们也使用 K = 1。在实践中开销很小，因为 K 可以选择接近高效硬件的限制。↩：#fnref:videoabbelslow"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：BAIR提出ABBEL框架，将长程交互中的摘要生成隔离为自然语言信念状态，并通过信念评分进行监督。 Aioga 将其归入「论文研究」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：模型与研究类动态需要结合能力边界、开放方式、成本、可用性和真实任务表现判断，单项指标领先不等于已经形成稳定采用。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文档、实际可用性、价格变化、开发者反馈和竞品回应。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-07-28T06:29:10.489Z","sourceHash":"ac3f57dff0bf46f3","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["论文研究","BAIR：Berkeley AI Research Blog"],"translations":{"zh-CN":{"title":"ABBEL：通过信念瓶颈提升LLM长程交互中的摘要学习效率","summary":"BAIR提出ABBEL框架，将长程交互中的摘要生成隔离为自然语言信念状态，并通过信念评分进行监督。在CollabBench协作编程测试中，基于重建的信念评分将全上下文模型的性能差距缩小约50%，训练步数减少50%。该方法解决了递归摘要（如Cursor Composer 2.5采用的压缩技术）在有限训练数据下性能下降的问题。","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL：通过信念瓶颈提升LLM长程交互中的摘要学习效率 - Aioga AI资讯","description":"BAIR提出ABBEL框架，将长程交互中的摘要生成隔离为自然语言信念状态，并通过信念评分进行监督。在CollabBench协作编程测试中，基于重建的信念评分将全上下文模型的性能差距缩小约50%，训练步数减少50%。该方法解决了递归摘要（如Cursor Composer 2.5采用的压缩技术）在有限训练数据下性能下降的问题。","url":"https://www.aioga.com/news/cms2p1wyq037aro3fl2in7qni/"},"en":{"title":"ABBEL: Improving Summary Learning Efficiency in Long-Range Interactions of LLMs Through Belief Bottlenecks","summary":"BAIR proposed the ABBEL framework, which isolates summary generation in long-range interactions into natural language belief states and supervises it through belief scoring. In the CollabBench collaborative programming test, using reconstructed belief scores reduced the performance gap of full-context models by about 50% and cut the number of training steps by 50%. This method addresses the performance degradation of recursive summarization (such as the compression technique used by Cursor Composer 2.5) under limited training data.","category":"Research","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Improving Summary Learning Efficiency in Long-Range Interactions of LLMs Through Belief Bottlenecks - Aioga AI News","description":"BAIR proposed the ABBEL framework, which isolates summary generation in long-range interactions into natural language belief states and supervises it through belief scoring. In the...","url":"https://www.aioga.com/en/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:03:33.118Z"},"ja":{"title":"ABBEL：信念のボトルネックを通じてLLMの長距離対話における要約学習効率を向上させる","summary":"BAIRはABBELフレームワークを提案し、長期的な相互作用における要約生成を自然言語の信念状態として分離し、信念スコアによって監督します。CollabBenchの協調プログラミングテストでは、再構築された信念スコアに基づき、全コンテキストモデルの性能差を約50％縮小し、学習ステップ数を50％削減しました。この方法は、限定された学習データ下で再帰的要約（例えばCursor Composer 2.5が採用している圧縮技術）の性能低下問題を解決します。","category":"論文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL：信念のボトルネックを通じてLLMの長距離対話における要約学習効率を向上させる - Aioga AIニュース","description":"BAIRはABBELフレームワークを提案し、長期的な相互作用における要約生成を自然言語の信念状態として分離し、信念スコアによって監督します。CollabBenchの協調プログラミングテストでは、再構築された信念スコアに基づき、全コンテキストモデルの性能差を約50％縮小し、学習ステップ数を50％削減しました。この方法は、限定された学習データ下で再帰的要約（例え...","url":"https://www.aioga.com/ja/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:03:43.541Z"},"ko":{"title":"ABBEL: 신념 병목을 통해 LLM 장기 상호작용에서 요약 학습 효율 향상","summary":"BAIR는 ABBEL 프레임워크를 제안하여 장기 상호작용에서 요약 생성을 자연어 신념 상태로 분리하고, 신념 점수를 통해 감독한다. CollabBench 협업 프로그래밍 테스트에서, 재구성된 신념 점수 기반은 전체 컨텍스트 모델의 성능 격차를 약 50% 줄이고, 학습 단계 수를 50% 감소시켰다. 이 방법은 제한된 학습 데이터에서 재귀 요약(예: Cursor Composer 2.5가 채택한 압축 기술) 성능 저하 문제를 해결한다.","category":"연구","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: 신념 병목을 통해 LLM 장기 상호작용에서 요약 학습 효율 향상 - Aioga AI 뉴스","description":"BAIR는 ABBEL 프레임워크를 제안하여 장기 상호작용에서 요약 생성을 자연어 신념 상태로 분리하고, 신념 점수를 통해 감독한다. CollabBench 협업 프로그래밍 테스트에서, 재구성된 신념 점수 기반은 전체 컨텍스트 모델의 성능 격차를 약 50% 줄이고, 학습 단계 수를 50% 감소시켰다. 이 방법은 제한된 학습...","url":"https://www.aioga.com/ko/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:04:30.644Z"},"es":{"title":"ABBEL: Mejorar la eficiencia del aprendizaje de resúmenes en interacciones a largo plazo de LLM mediante el cuello de botella de la creencia","summary":"BAIR propuso el marco ABBEL, que aísla la generación de resúmenes en interacciones a largo plazo en estados de creencia en lenguaje natural y supervisa mediante una puntuación de creencia. En la prueba de programación colaborativa CollabBench, la puntuación de creencia basada en la reconstrucción redujo aproximadamente un 50% la brecha de rendimiento del modelo de contexto completo y disminuyó los pasos de entrenamiento en un 50%. Este método resolvió el problema del descenso de rendimiento de los resúmenes recursivos (como la técnica de compresión utilizada por Cursor Composer 2.5) con datos de entrenamiento limitados.","category":"Investigación","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Mejorar la eficiencia del aprendizaje de resúmenes en interacciones a largo plazo de LLM mediante el cuello de botella de la creencia - Aioga Noticias de IA","description":"BAIR propuso el marco ABBEL, que aísla la generación de resúmenes en interacciones a largo plazo en estados de creencia en lenguaje natural y supervisa mediante una puntuación de c...","url":"https://www.aioga.com/es/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:04:23.424Z"},"fr":{"title":"ABBEL : Améliorer l'efficacité de l'apprentissage du résumé dans les interactions à long terme du LLM grâce aux goulots d'étranglement de la croyance","summary":"BAIR a proposé le cadre ABBEL, qui isole la génération de résumés dans les interactions à longue portée sous forme d'états de croyance en langage naturel, et supervise via un score de croyance. Dans les tests de programmation collaborative CollabBench, le score de croyance basé sur la reconstruction a réduit d'environ 50 % l'écart de performance avec le modèle contextuel complet, tout en diminuant de 50 % le nombre de pas d'entraînement. Cette méthode résout le problème de la baisse de performance des résumés récursifs (comme la technique de compression adoptée par Cursor Composer 2.5) avec des données d'entraînement limitées.","category":"Recherche","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL : Améliorer l'efficacité de l'apprentissage du résumé dans les interactions à long terme du LLM grâce aux goulots d'étranglement de la croyance - Aioga Actualités IA","description":"BAIR a proposé le cadre ABBEL, qui isole la génération de résumés dans les interactions à longue portée sous forme d'états de croyance en langage naturel, et supervise via un score...","url":"https://www.aioga.com/fr/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:05:14.743Z"},"de":{"title":"ABBEL: Verbesserung der Zusammenfassungslern-Effizienz in langfristigen Interaktionen von LLM durch Glaubens-Flaschenhals","summary":"BAIR schlug das ABBEL-Framework vor, das die Zusammenfassungsgenerierung in Ferninteraktionen in einen natürlichen Sprach-Glaubenszustand isoliert und durch Glaubensbewertung überwacht. Im CollabBench-Kollaborationsprogrammierungstest verringerte die auf rekonstruierten Glaubensbewertungen basierende Leistungslücke des Full-Context-Modells um etwa 50 % und die Anzahl der Training Schritte um 50 %. Diese Methode löste das Problem des Leistungsabfalls rekursiver Zusammenfassungen (wie die von Cursor Composer 2.5 verwendete Kompressionstechnologie) bei begrenzten Trainingsdaten.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Verbesserung der Zusammenfassungslern-Effizienz in langfristigen Interaktionen von LLM durch Glaubens-Flaschenhals - Aioga KI-News","description":"BAIR schlug das ABBEL-Framework vor, das die Zusammenfassungsgenerierung in Ferninteraktionen in einen natürlichen Sprach-Glaubenszustand isoliert und durch Glaubensbewertung überw...","url":"https://www.aioga.com/de/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:05:11.630Z"},"pt-BR":{"title":"ABBEL: Melhorando a eficiência da aprendizagem de resumos em interações de longo alcance de LLM por meio do gargalo de crença","summary":"BAIR propôs o framework ABBEL, isolando a geração de resumos em interações de longo alcance para um estado de crença em linguagem natural, e supervisionando por meio da pontuação de crenças. No teste de programação colaborativa CollabBench, a pontuação de crenças reconstruída reduziu a diferença de desempenho entre os modelos com contexto completo em cerca de 50%, diminuindo o número de etapas de treinamento em 50%. Este método resolveu o problema da queda de desempenho em resumos recursivos (como a técnica de compressão usada pelo Cursor Composer 2.5) com dados de treinamento limitados.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Melhorando a eficiência da aprendizagem de resumos em interações de longo alcance de LLM por meio do gargalo de crença - Aioga Notícias de IA","description":"BAIR propôs o framework ABBEL, isolando a geração de resumos em interações de longo alcance para um estado de crença em linguagem natural, e supervisionando por meio da pontuação d...","url":"https://www.aioga.com/pt-BR/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:05:54.199Z"},"ru":{"title":"ABBEL: Повышение эффективности обучения резюме в долгосрочном взаимодействии LLM через узкие места веры","summary":"BAIR предложил рамочную структуру ABBEL, которая изолирует генерацию резюме в долгосрочном взаимодействии в состояние убеждений на естественном языке и осуществляет наблюдение через оценку убеждений. В тесте совместного программирования CollabBench оценка убеждений на основе реконструкции сократила разрыв в производительности модели с полным контекстом примерно на 50% и уменьшила количество шагов обучения на 50%. Этот метод решает проблему снижения производительности рекурсивного резюмирования (например, используемой в Cursor Composer 2.5 технологии сжатия) при ограниченных обучающих данных.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Повышение эффективности обучения резюме в долгосрочном взаимодействии LLM через узкие места веры - Aioga Новости ИИ","description":"BAIR предложил рамочную структуру ABBEL, которая изолирует генерацию резюме в долгосрочном взаимодействии в состояние убеждений на естественном языке и осуществляет наблюдение чере...","url":"https://www.aioga.com/ru/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:05:58.004Z"},"ar":{"title":"ABBEL: تحسين كفاءة تعلم الملخصات في التفاعلات الطويلة المدى لنماذج اللغة الكبيرة من خلال عنق الزجاجة الإيماني","summary":"قدم BAIR إطار عمل ABBEL، حيث يفصل توليد الملخص في التفاعلات طويلة المدى إلى حالة اعتقاد طبيعية اللغة، ويشرف عليها من خلال تقييم الاعتقاد. في اختبار برمجة التعاون CollabBench، قلصت درجات الاعتقاد المستندة إلى إعادة البناء فجوة أداء نموذج السياق الكامل بحوالي 50٪، مع تقليل عدد خطوات التدريب بنسبة 50٪. تحل هذه الطريقة مشكلة انخفاض الأداء في الملخصات التكرارية (مثل تقنية الضغط المستخدمة في Cursor Composer 2.5) عند وجود بيانات تدريب محدودة.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: تحسين كفاءة تعلم الملخصات في التفاعلات الطويلة المدى لنماذج اللغة الكبيرة من خلال عنق الزجاجة الإيماني - Aioga أخبار الذكاء الاصطناعي","description":"قدم BAIR إطار عمل ABBEL، حيث يفصل توليد الملخص في التفاعلات طويلة المدى إلى حالة اعتقاد طبيعية اللغة، ويشرف عليها من خلال تقييم الاعتقاد. في اختبار برمجة التعاون CollabBench، قلصت...","url":"https://www.aioga.com/ar/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:06:41.499Z"},"hi":{"title":"ABBEL：विश्वास बॉटलनेक के माध्यम से LLM लंबी दूरी की इंटरैक्शन में सारांश सीखने की दक्षता बढ़ाना","summary":"BAIR ने ABBEL ढांचे का प्रस्ताव रखा, जो लंबी अवधि की इंटरैक्शन में सारांश उत्पादन को प्राकृतिक भाषा विश्वास स्थिति में अलग करता है, और विश्वास स्कोरिंग के माध्यम से निगरानी करता है। CollabBench सहयोगी प्रोग्रामिंग परीक्षण में, पुनर्निर्मित विश्वास स्कोर के आधार पर संपूर्ण संदर्भ मॉडल के प्रदर्शन अंतर को लगभग 50% तक घटा दिया गया और प्रशिक्षण चरणों की संख्या 50% तक कम हो गई। यह तरीका पुनरावर्ती सारांश (जैसे Cursor Composer 2.5 द्वारा अपनाई गई संपीड़न तकनीक) की सीमित प्रशिक्षण डेटा के तहत प्रदर्शन गिरने की समस्या का समाधान करता है।","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL：विश्वास बॉटलनेक के माध्यम से LLM लंबी दूरी की इंटरैक्शन में सारांश सीखने की दक्षता बढ़ाना - Aioga AI समाचार","description":"BAIR ने ABBEL ढांचे का प्रस्ताव रखा, जो लंबी अवधि की इंटरैक्शन में सारांश उत्पादन को प्राकृतिक भाषा विश्वास स्थिति में अलग करता है, और विश्वास स्कोरिंग के माध्यम से निगरानी करता है...","url":"https://www.aioga.com/hi/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:06:45.329Z"},"it":{"title":"ABBEL: Migliorare l'efficienza dell'apprendimento del sommario nelle interazioni a lungo termine degli LLM tramite colli di bottiglia di credenza","summary":"BAIR ha proposto il framework ABBEL, isolando la generazione di riassunti nelle interazioni a lungo termine nello stato di convinzione del linguaggio naturale e supervisionandola tramite punteggi di convinzione. Nei test di programmazione collaborativa CollabBench, i punteggi di convinzione basati sulla ricostruzione hanno ridotto di circa il 50% il divario di prestazioni dei modelli a contesto completo, riducendo del 50% anche il numero di passi di addestramento. Questo metodo ha risolto il problema del calo di prestazioni dei riassunti ricorsivi (come la tecnica di compressione adottata da Cursor Composer 2.5) con dati di addestramento limitati.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Migliorare l'efficienza dell'apprendimento del sommario nelle interazioni a lungo termine degli LLM tramite colli di bottiglia di credenza - Aioga Notizie IA","description":"BAIR ha proposto il framework ABBEL, isolando la generazione di riassunti nelle interazioni a lungo termine nello stato di convinzione del linguaggio naturale e supervisionandola t...","url":"https://www.aioga.com/it/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:07:29.467Z"},"nl":{"title":"ABBEL: Het verbeteren van de samenvattingsleerefficiëntie in langeafstandsinteracties van LLM's door geloofsbottlenecks","summary":"BAIR stelde het ABBEL-framework voor, waarbij het genereren van samenvattingen in lange-afstandsinteracties wordt gescheiden in natuurlijke taalgeloofstoestanden en wordt gecontroleerd door middel van geloofsscores. In de CollabBench-samenwerkingsprogrammeertest verkleint op reconstructie gebaseerde geloofsscore het prestatieverschil van het volledige contextmodel met ongeveer 50% en wordt het aantal trainingsstappen met 50% verminderd. Deze methode lost het probleem op van prestatievermindering bij recursieve samenvattingen (zoals de compressietechniek gebruikt door Cursor Composer 2.5) bij beperkte trainingsgegevens.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Het verbeteren van de samenvattingsleerefficiëntie in langeafstandsinteracties van LLM's door geloofsbottlenecks - Aioga AI-nieuws","description":"BAIR stelde het ABBEL-framework voor, waarbij het genereren van samenvattingen in lange-afstandsinteracties wordt gescheiden in natuurlijke taalgeloofstoestanden en wordt gecontrol...","url":"https://www.aioga.com/nl/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:07:31.306Z"},"tr":{"title":"ABBEL: İnanç dar boğazları aracılığıyla LLM uzun menzilli etkileşimlerinde özet öğrenme verimliliğini artırma","summary":"BAIR, uzun menzilli etkileşimlerde özetleme işlemini doğal dil inanç durumlarına izole eden ABBEL çerçevesini öneriyor ve inanç puanlaması yoluyla denetim sağlıyor. CollabBench işbirlikçi programlama testlerinde, yeniden yapılandırılmış inanç puanlamasına dayalı olarak tüm bağlamlı modelin performans farkını yaklaşık %50 oranında azaltıyor ve eğitim adımlarını %50 azaltıyor. Bu yöntem, sınırlı eğitim verileri altında performans düşüşüne yol açan özyinelemeli özetleme (Cursor Composer 2.5'te kullanılan sıkıştırma teknolojisi gibi) sorununu çözüyor.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: İnanç dar boğazları aracılığıyla LLM uzun menzilli etkileşimlerinde özet öğrenme verimliliğini artırma - Aioga AI Haberleri","description":"BAIR, uzun menzilli etkileşimlerde özetleme işlemini doğal dil inanç durumlarına izole eden ABBEL çerçevesini öneriyor ve inanç puanlaması yoluyla denetim sağlıyor. CollabBench işb...","url":"https://www.aioga.com/tr/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:08:16.796Z"},"vi":{"title":"ABBEL: Nâng cao hiệu quả học tóm tắt trong tương tác dài hạn của LLM thông qua nút thắt niềm tin","summary":"BAIR đã đề xuất khung ABBEL, tách việc tạo tóm tắt trong tương tác dài thành trạng thái niềm tin ngôn ngữ tự nhiên và thực hiện giám sát thông qua đánh giá niềm tin. Trong thử nghiệm lập trình hợp tác CollabBench, dựa trên điểm niềm tin tái tạo đã thu hẹp khoảng cách hiệu suất của mô hình toàn bộ ngữ cảnh khoảng 50%, đồng thời giảm 50% số bước huấn luyện. Phương pháp này giải quyết vấn đề suy giảm hiệu suất của tóm tắt đệ quy (như kỹ thuật nén được sử dụng trong Cursor Composer 2.5) khi dữ liệu huấn luyện hạn chế.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Nâng cao hiệu quả học tóm tắt trong tương tác dài hạn của LLM thông qua nút thắt niềm tin - Tin tức AI Aioga","description":"BAIR đã đề xuất khung ABBEL, tách việc tạo tóm tắt trong tương tác dài thành trạng thái niềm tin ngôn ngữ tự nhiên và thực hiện giám sát thông qua đánh giá niềm tin. Trong thử nghi...","url":"https://www.aioga.com/vi/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:08:15.997Z"},"id":{"title":"ABBEL: Meningkatkan Efisiensi Pembelajaran Ringkasan dalam Interaksi Jangka Panjang LLM melalui Bottleneck Keyakinan","summary":"BAIR mengajukan kerangka ABBEL, yang memisahkan pembuatan ringkasan dalam interaksi jangka panjang menjadi keadaan keyakinan bahasa alami, dan melakukan pengawasan melalui penilaian keyakinan. Dalam pengujian pemrograman kolaboratif CollabBench, penilaian keyakinan berbasis rekonstruksi memperkecil selisih kinerja model konteks penuh sekitar 50%, dengan jumlah langkah pelatihan berkurang 50%. Metode ini mengatasi masalah penurunan kinerja pada ringkasan rekursif (seperti teknik kompresi yang digunakan oleh Cursor Composer 2.5) ketika data pelatihan terbatas.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Meningkatkan Efisiensi Pembelajaran Ringkasan dalam Interaksi Jangka Panjang LLM melalui Bottleneck Keyakinan - Berita AI Aioga","description":"BAIR mengajukan kerangka ABBEL, yang memisahkan pembuatan ringkasan dalam interaksi jangka panjang menjadi keadaan keyakinan bahasa alami, dan melakukan pengawasan melalui penilaia...","url":"https://www.aioga.com/id/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:08:57.878Z"},"th":{"title":"ABBEL: ยกระดับประสิทธิภาพการเรียนรู้สรุปในการโต้ตอบระยะยาวของ LLM ผ่านคอขวดความเชื่อ","summary":"BAIR นำเสนอกรอบงาน ABBEL ซึ่งแยกการสร้างสรุประยะยาวออกเป็นสถานะความเชื่อในภาษาธรรมชาติ และทำการควบคุมโดยใช้คะแนนความเชื่อ ในการทดสอบการเขียนโปรแกรมเชิงความร่วมมือ CollabBench คะแนนความเชื่อที่สร้างขึ้นใหม่ช่วยลดช่องว่างประสิทธิภาพของโมเดลบริบททั้งหมดได้ประมาณ 50% และลดจำนวนก้าวการฝึกลง 50% วิธีนี้แก้ปัญหาการลดประสิทธิภาพของการสรุปแบบวนซ้ำ (เช่น เทคนิคการบีบอัดที่ใช้ใน Cursor Composer 2.5) ภายใต้ข้อมูลการฝึกที่จำกัด","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: ยกระดับประสิทธิภาพการเรียนรู้สรุปในการโต้ตอบระยะยาวของ LLM ผ่านคอขวดความเชื่อ - ข่าว AI Aioga","description":"BAIR นำเสนอกรอบงาน ABBEL ซึ่งแยกการสร้างสรุประยะยาวออกเป็นสถานะความเชื่อในภาษาธรรมชาติ และทำการควบคุมโดยใช้คะแนนความเชื่อ ในการทดสอบการเขียนโปรแกรมเชิงความร่วมมือ CollabBench คะแนน...","url":"https://www.aioga.com/th/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:09:05.958Z"},"pl":{"title":"ABBEL: Poprawa efektywności uczenia się streszczeń w długodystansowej interakcji LLM poprzez wąskie gardło wiary","summary":"BAIR zaproponowało ramy ABBEL, oddzielając generowanie streszczeń w długoterminowych interakcjach poprzez stan przekonań w języku naturalnym i nadzorując je za pomocą oceny przekonań. W teście współpracy przy programowaniu CollabBench, oparta na odbudowanej ocenie przekonań metoda zmniejszyła różnicę w wydajności pełnego modelu kontekstowego o około 50% i skróciła liczbę kroków treningowych o 50%. Ta metoda rozwiązuje problem spadku wydajności rekurencyjnych streszczeń (takich jak technika kompresji stosowana w Cursor Composer 2.5) przy ograniczonej ilości danych treningowych.","category":"论文研究","source":"BAIR：Berkeley AI Research Blog","aggregationSource":"BAIR：Berkeley AI Research Blog","pageTitle":"ABBEL: Poprawa efektywności uczenia się streszczeń w długodystansowej interakcji LLM poprzez wąskie gardło wiary - Aioga Wiadomości AI","description":"BAIR zaproponowało ramy ABBEL, oddzielając generowanie streszczeń w długoterminowych interakcjach poprzez stan przekonań w języku naturalnym i nadzorując je za pomocą oceny przekon...","url":"https://www.aioga.com/pl/news/cms2p1wyq037aro3fl2in7qni/","contentTranslated":true,"sourceHash":"1a284f79577d801b","translatedAt":"2026-07-27T04:09:58.210Z"}}}}