Tag
本文提出一个物理接地的多模态智能体框架，用于磁约束聚变装置中性束注入离子源的非平稳自适应调控，在EXL-50U平台455次放电的时序回放中验证了符号化物理量优于波形图像，并在监督运行下实现86.80%的束提取率。
本文提出'solver-sufficient symbolization'概念，从推理算法视角定义符号的计算有用性，并开发反例引导的符号发明算法，用于混合离散-连续机器人规划，在仿真与真实机器人任务中降低求解成本同时保持解质量与可解释性。
SAT-RL is a reinforcement learning framework that trains LLMs to interact with symbolic solvers like Z3 through multi-turn tool execution, combining synthetic data generation, supervised warm-up on solver-augmented traces, and solver-grounded RL with tool-aware rewards to improve logical reasoning on in-domain and out-of-domain tasks.
The paper introduces SAE-IDF, which uses sparse autoencoders and rarity-weighted features as a shared coordinate system to match symbolic units across heterogeneous systems, outperforming dense LLM baselines on 12 tasks while remaining human-interpretable.
本文提出RoleWorld，一种符号化世界建模方法，通过可复用的功能角色抽象（functional role abstractions）实现LLM智能体在不同环境间的知识迁移，在MARS基准上展示了跨环境的稳健泛化能力。
Philip Zucker introduces 'Refinement E-Graphs', an extension to e-graphs that adds a privileged <= relation alongside equality, enabling unidirectional refinement rewrites useful for compiler optimizations like don't-care logic in digital circuits.
This paper presents a neurosymbolic router that uses grammatical inference (L* algorithm) to learn a DFA-based dispatcher, sending structured queries to deterministic symbolic solvers and reserving small language models for open-ended problems on edge hardware. Evaluated on a Raspberry Pi 4B, the router achieves 98.3% overall accuracy and runs 8.8x faster than Program-of-Thought baselines.
This paper presents a framework that augments Large Language Models with geometric vision parsing and symbolic solving to match state-of-the-art multimodal models on complex geometry problems, using a new benchmark from 2025 Chinese Zhongkao exams for evaluation.
This paper introduces Certification-Driven Reinforcement Learning (CDRL), a framework that leverages symbolic reasoning to generate reusable constraints for improving reinforcement learning in combinatorial search spaces, demonstrated in neutrino flavor model discovery with higher valid model rates and efficiency.
This paper introduces the RAIL principles (Reasoning, Assurances, Interfacing, Learning) for designing neurosymbolic AI systems that integrate machine learning with symbolic reasoning, arguing this approach is crucial for reliable, efficient, and trustworthy AI.
Proposes SymCA, an interpretable column annotation framework using LLMs to materialize annotation as a global-to-local symbolic decision process, achieving significant accuracy improvements over baselines.
This paper proposes that AI memory consolidation should recombine knowledge across domains (like dreaming) rather than merely replaying experiences, and demonstrates that cross-domain consolidation improves performance in both neural (LoRA fine-tuning) and symbolic systems.
This paper introduces Inductive Logic Programming to extract symbolic representations of RL policies and proposes novel explainability metrics (activation rate, feature coverage, syntactic and semantic distance) for objective evaluation in single- and multi-agent settings.
GATS introduces a Graph-Augmented Tree Search with a layered world model (symbolic, learned, generative) to eliminate LLM calls during planning, achieving 100% success on synthetic tasks and stress tests, outperforming LATS and ReAct.
This paper presents a novel fuzzy-logic-based extension of Answer Set Programming (ASP) that integrates numerical information with qualitative reasoning using membership functions, demonstrated through a travel recommendation case study.
Introduces RusFinChain, the first Russian-language symbolic benchmark for verifiable chain-of-thought reasoning in finance, spanning 17 domains with 5,280 parameterized examples and enhanced evaluation metrics including fuzzy numeric alignment.
RuleChef is a framework that uses LLMs to generate human-editable, executable rules for NLP tasks, iteratively improving them based on examples and human feedback, resulting in fast, deterministic, and inspectable rule systems.
The paper argues that data-driven machine learning systems, including GPT-5, cannot achieve symbolic-level logical reasoning through scaling alone, due to inherent limitations in distinguishing logical structures from statistical regularities.
This paper presents Process-Verified Reinforcement Learning, using the Lean proof assistant as a process oracle to provide fine-grained tactic-level feedback during training, improving theorem proving performance.
This paper introduces RecurrReason, a difficulty-controlled benchmark of four symbolic logic puzzles to evaluate multi-step reasoning in sequence models. Fine-tuning experiments on T5 and GPT-2 show that architecture determines success more than scale, and that pre-training transfer depends on local transition structure.