QGPINNs: A Physics-Informed Neural Network Framework for Nonlocal Differential Equations on Quantum Graphs
提出QGPINNs框架，使用神经网络解决量子图上的非局部微分方程问题，通过物理信息和图适应学习策略提高解的准确性和稳定性。
0 citationsRead paper
Selected work
Representative Papers
LLMs Don't Pay for the Jump
This study addresses the deficiency of scientific abductive reasoning in large language models by proposing that this capability relies on the thermodynamic coupling between cognitive error and physical cost, rather than embodied simulation alone. Through formal modeling and correlation analysis of output entropy with causal complexity, we demonstrate that fixed-weight Transformers lack this coupling mechanism, resulting in output entropy that fails to adapt to task difficulty and consequently causes complex reasoning failures. The research confirms that genuine abductive reasoning necessitates a physical cost mechanism that compels error correction. These findings establish a novel theoretical paradigm and define critical missing elements for overcoming current bottlenecks in machine intelligence, shifting the focus from purely computational architectures to thermodynamically grounded cognitive processes.
0 citationsRead paper
Self-Referential Induction Increases Response Instability Relative to Unresolvable and Verifiable Questions in Large Language Models
This study investigates whether subjective experience reports generated by large language models under self-referential prompts exhibit response consistency, and compares their stability against those elicited by intractable philosophical questions and verifiable factual questions. Through 30 independent sampling trials, output instability is quantified using sentence embeddings and mean pairwise cosine similarity. The findings reveal, for the first time quantitatively, that self-referential prompts yield significantly unstable subjective reports (similarity: 0.343 ± 0.047), markedly higher than both philosophical (0.192 ± 0.008) and verifiable questions (0.105 ± 0.058). This suggests that the observed uncertainty stems from a distinct mechanism induced by self-reference rather than general question openness. Experiments were conducted using the Gemini API (temperature = 0.7) with systematic prompt engineering.
0 citationsRead paper
Caching for the Future: Scrub Jay Episodic Memory Principles for Agent Memory Systems
Current memory systems in large language model (LLM) agents struggle to differentiate the temporal relevance of information, leading to retrieval contaminated by outdated content. Inspired by the episodic memory mechanisms of western scrub-jays, this work proposes a novel memory architecture that represents memories as What-Where-When triplets and assigns each memory type a learnable, type-conditioned temporal decay coefficient. Coupled with query-adaptive scoring, the system enables efficient retrieval and O(1)-complexity retrospective updates. This is the first effort to integrate the biologically observed principle of type-dependent temporal decay from episodic memory into artificial agent systems. The authors also introduce the Temporal Generalization Test (TGT) benchmark for evaluation. Experiments demonstrate that the proposed method achieves a +0.108 generalization gain on TGT and outperforms Mem0 and Qwen3-Embedding-4B by 2.66 and 3.09 F1 points, respectively, on the MemoryAgentBench EventQA-64k task.
0 citationsRead paper
Early to Share, Late to Save: Synchronisation-Driven Communication Gating in Bandwidth-Constrained Cooperative VLN
This work addresses the challenge of communication efficiency in collaborative Vision-and-Language Navigation (VLN) under strict bandwidth constraints. The authors propose an efficient communication framework based on a hindsight-gated mechanism, which learns communication decisions through supervised training on failed navigation trajectories. Their analysis reveals and validates a counterintuitive communication paradigm—“early synchronization, late independence”—demonstrating that agents should communicate when they are highly confident rather than uncertain. By integrating lightweight supervised gating, recurrent hidden-state alignment analysis, and GRU-based sequence modeling, the method achieves near-unconstrained communication performance with only three communication rounds, improving communication efficiency by 260% and 320% over random and entropy-based baselines, respectively.
0 citationsRead paper
Recent publications
Latest Papers
QGPINNs: A Physics-Informed Neural Network Framework for Nonlocal Differential Equations on Quantum Graphs
提出QGPINNs框架，使用神经网络解决量子图上的非局部微分方程问题，通过物理信息和图适应学习策略提高解的准确性和稳定性。
0 citationsRead paper
LLMs Don't Pay for the Jump
This study addresses the deficiency of scientific abductive reasoning in large language models by proposing that this capability relies on the thermodynamic coupling between cognitive error and physical cost, rather than embodied simulation alone. Through formal modeling and correlation analysis of output entropy with causal complexity, we demonstrate that fixed-weight Transformers lack this coupling mechanism, resulting in output entropy that fails to adapt to task difficulty and consequently causes complex reasoning failures. The research confirms that genuine abductive reasoning necessitates a physical cost mechanism that compels error correction. These findings establish a novel theoretical paradigm and define critical missing elements for overcoming current bottlenecks in machine intelligence, shifting the focus from purely computational architectures to thermodynamically grounded cognitive processes.
0 citationsRead paper
Self-Referential Induction Increases Response Instability Relative to Unresolvable and Verifiable Questions in Large Language Models
This study investigates whether subjective experience reports generated by large language models under self-referential prompts exhibit response consistency, and compares their stability against those elicited by intractable philosophical questions and verifiable factual questions. Through 30 independent sampling trials, output instability is quantified using sentence embeddings and mean pairwise cosine similarity. The findings reveal, for the first time quantitatively, that self-referential prompts yield significantly unstable subjective reports (similarity: 0.343 ± 0.047), markedly higher than both philosophical (0.192 ± 0.008) and verifiable questions (0.105 ± 0.058). This suggests that the observed uncertainty stems from a distinct mechanism induced by self-reference rather than general question openness. Experiments were conducted using the Gemini API (temperature = 0.7) with systematic prompt engineering.
0 citationsRead paper
Caching for the Future: Scrub Jay Episodic Memory Principles for Agent Memory Systems
Current memory systems in large language model (LLM) agents struggle to differentiate the temporal relevance of information, leading to retrieval contaminated by outdated content. Inspired by the episodic memory mechanisms of western scrub-jays, this work proposes a novel memory architecture that represents memories as What-Where-When triplets and assigns each memory type a learnable, type-conditioned temporal decay coefficient. Coupled with query-adaptive scoring, the system enables efficient retrieval and O(1)-complexity retrospective updates. This is the first effort to integrate the biologically observed principle of type-dependent temporal decay from episodic memory into artificial agent systems. The authors also introduce the Temporal Generalization Test (TGT) benchmark for evaluation. Experiments demonstrate that the proposed method achieves a +0.108 generalization gain on TGT and outperforms Mem0 and Qwen3-Embedding-4B by 2.66 and 3.09 F1 points, respectively, on the MemoryAgentBench EventQA-64k task.
0 citationsRead paper
Early to Share, Late to Save: Synchronisation-Driven Communication Gating in Bandwidth-Constrained Cooperative VLN
This work addresses the challenge of communication efficiency in collaborative Vision-and-Language Navigation (VLN) under strict bandwidth constraints. The authors propose an efficient communication framework based on a hindsight-gated mechanism, which learns communication decisions through supervised training on failed navigation trajectories. Their analysis reveals and validates a counterintuitive communication paradigm—“early synchronization, late independence”—demonstrating that agents should communicate when they are highly confident rather than uncertain. By integrating lightweight supervised gating, recurrent hidden-state alignment analysis, and GRU-based sequence modeling, the method achieves near-unconstrained communication performance with only three communication rounds, improving communication efficiency by 260% and 320% over random and entropy-based baselines, respectively.
0 citationsRead paper