Selected work

Representative Papers

Can Vision-Language Models Assess Proxemic Risk from Egocentric Robot Images?

Aug 12, 2026

This study addresses the challenge of assessing proxemic danger in human environments from a robot’s first-person perspective to enhance embodied navigation safety. We evaluate the performance of three vision-language models—InternVL, Qwen-VL, and SmolVLM—on a four-class proximity risk classification task, systematically comparing multiple prompting strategies and two rounds of QLoRA fine-tuning. We further analyze the relationship between models’ spatial localization capabilities and their risk judgment accuracy. Results show that unmodified models perform near random baseline levels; while overall gains from fine-tuning remain modest, Qwen-VL combined with advanced prompting significantly improves recall in high-risk scenarios. Our findings highlight current limitations of vision-language models in fine-grained proxemic reasoning and spatial grounding, demonstrate that targeted prompting can effectively mitigate model deficiencies, and reveal that correct risk classification does not necessarily rely on accurate spatial attention.

0 citationsRead paper

Trajectory Design and Budgeted Querying for Digital Twin Calibration

Aug 09, 2026

This study addresses the challenge of high interaction data acquisition costs in digital twin calibration by proposing a novel framework that jointly optimizes trajectory generation and parameter querying under limited budgets. The approach explicitly treats trajectory design and query allocation as coupled optimization variables, integrating an incentive-driven reinforcement learning controller, a GRU-based recurrent parameter estimator that incorporates predictive uncertainty, and a budget-aware query strategy. Experimental results demonstrate significant improvements in calibration efficiency and accuracy under data scarcity: on the Pend日晚间 task, only three queries reduce the terminal parameter error to 0.0092 from an uncalibrated baseline of 0.2031; in the Waterworld environment, the online normalized error stabilizes at 4–5%, underscoring the method’s effectiveness in resource-constrained settings.

0 citationsRead paper

Canonical Joint Energy-Based Model on CIFAR-10: failure modes and practical indistinguishability of Predictor-Corrector and SGLD samplers

Aug 05, 2026

This work systematically evaluates whether the Predictor-Corrector (PC) sampler outperforms Stochastic Gradient Langevin Dynamics (SGLD) in standard Joint Energy-Based Model (JEM) training. Using the WideResNet-28-10 architecture, we conduct a fair comparison between the two samplers across image generation, classification, and out-of-distribution (OOD) detection tasks, employing a replay buffer, Fréchet Inception Distance (FID), AUROC, and statistical equivalence testing. Under a fixed noise schedule, PC and SGLD exhibit no statistically significant differences in test accuracy (≈92.9%), FID (difference < 0.5), or AUROC (difference < 0.007), and both share similar failure modes, including late-stage training collapse. To our knowledge, this is the first study to reproduce and directly compare the PC sampler within the standard JEM framework, demonstrating that its practical performance is indistinguishable from SGLD and thereby challenging claims of its superiority.

0 citationsRead paper

Probe, Don't Prompt: A Hidden-State Probe for Metadata Filtering in Multi-Meta-RAG

Jul 04, 2026

Current large language models rely on prompting to extract metadata for multi-hop question answering retrieval, suffering from high computational costs, uncontrollable outputs, and an inability to reject empty queries. This work proposes a deterministic probing method based on shallow hidden states of small-scale open-source language models, which accurately extracts metadata from a predefined vocabulary using mean pooling and a lightweight linear classification head—requiring only a few initial layers of forward propagation. The approach enables rejection of empty queries, enforces strictly constrained outputs, and eliminates the need for external API calls. Evaluated on 2,556 queries, it achieves a set-wise exact match accuracy of 90.9%, substantially outperforming GPT-3.5 (80.9%) and a substring baseline (88.0%), with a 135M-parameter model matching the performance of a 1.5B-parameter counterpart.

0 citationsRead paper

Query-Aware Spreading Activation for Multi-Hop Retrieval over Knowledge Graphs

Jun 29, 2026

This work addresses the limitation of existing multi-hop knowledge graph retrieval methods, which often neglect query semantics during traversal, resulting in low path relevance. The authors propose a query-aware diffusion activation approach that employs a semantic gating mechanism at each step, computed via cosine similarity between entity descriptions and the input question, to efficiently guide traversal within a fixed number of steps. Crucially, the entire retrieval process is encoded into a single Cypher query and executed in one step within Neo4j, eliminating the need for external memory loading. This method uniquely reduces query-aware traversal to fixed-iteration semantic gating entirely inside the graph database. Evaluated on MuSiQue, it achieves 32.80 EM, outperforming the HippoRAG baseline by 5.3 EM and 3.4 F1; ablation studies further demonstrate that semantic gating alone yields F1 improvements of 3.6–7.4 and reduces latency by 1.5–4.9×.

0 citationsRead paper

Recent publications

Latest Papers

Can Vision-Language Models Assess Proxemic Risk from Egocentric Robot Images?

Aug 12, 2026

This study addresses the challenge of assessing proxemic danger in human environments from a robot’s first-person perspective to enhance embodied navigation safety. We evaluate the performance of three vision-language models—InternVL, Qwen-VL, and SmolVLM—on a four-class proximity risk classification task, systematically comparing multiple prompting strategies and two rounds of QLoRA fine-tuning. We further analyze the relationship between models’ spatial localization capabilities and their risk judgment accuracy. Results show that unmodified models perform near random baseline levels; while overall gains from fine-tuning remain modest, Qwen-VL combined with advanced prompting significantly improves recall in high-risk scenarios. Our findings highlight current limitations of vision-language models in fine-grained proxemic reasoning and spatial grounding, demonstrate that targeted prompting can effectively mitigate model deficiencies, and reveal that correct risk classification does not necessarily rely on accurate spatial attention.

0 citationsRead paper

Trajectory Design and Budgeted Querying for Digital Twin Calibration

Aug 09, 2026

This study addresses the challenge of high interaction data acquisition costs in digital twin calibration by proposing a novel framework that jointly optimizes trajectory generation and parameter querying under limited budgets. The approach explicitly treats trajectory design and query allocation as coupled optimization variables, integrating an incentive-driven reinforcement learning controller, a GRU-based recurrent parameter estimator that incorporates predictive uncertainty, and a budget-aware query strategy. Experimental results demonstrate significant improvements in calibration efficiency and accuracy under data scarcity: on the Pend日晚间 task, only three queries reduce the terminal parameter error to 0.0092 from an uncalibrated baseline of 0.2031; in the Waterworld environment, the online normalized error stabilizes at 4–5%, underscoring the method’s effectiveness in resource-constrained settings.

0 citationsRead paper

Canonical Joint Energy-Based Model on CIFAR-10: failure modes and practical indistinguishability of Predictor-Corrector and SGLD samplers

Aug 05, 2026

This work systematically evaluates whether the Predictor-Corrector (PC) sampler outperforms Stochastic Gradient Langevin Dynamics (SGLD) in standard Joint Energy-Based Model (JEM) training. Using the WideResNet-28-10 architecture, we conduct a fair comparison between the two samplers across image generation, classification, and out-of-distribution (OOD) detection tasks, employing a replay buffer, Fréchet Inception Distance (FID), AUROC, and statistical equivalence testing. Under a fixed noise schedule, PC and SGLD exhibit no statistically significant differences in test accuracy (≈92.9%), FID (difference < 0.5), or AUROC (difference < 0.007), and both share similar failure modes, including late-stage training collapse. To our knowledge, this is the first study to reproduce and directly compare the PC sampler within the standard JEM framework, demonstrating that its practical performance is indistinguishable from SGLD and thereby challenging claims of its superiority.

0 citationsRead paper

Probe, Don't Prompt: A Hidden-State Probe for Metadata Filtering in Multi-Meta-RAG

Jul 04, 2026

Current large language models rely on prompting to extract metadata for multi-hop question answering retrieval, suffering from high computational costs, uncontrollable outputs, and an inability to reject empty queries. This work proposes a deterministic probing method based on shallow hidden states of small-scale open-source language models, which accurately extracts metadata from a predefined vocabulary using mean pooling and a lightweight linear classification head—requiring only a few initial layers of forward propagation. The approach enables rejection of empty queries, enforces strictly constrained outputs, and eliminates the need for external API calls. Evaluated on 2,556 queries, it achieves a set-wise exact match accuracy of 90.9%, substantially outperforming GPT-3.5 (80.9%) and a substring baseline (88.0%), with a 135M-parameter model matching the performance of a 1.5B-parameter counterpart.

0 citationsRead paper

Query-Aware Spreading Activation for Multi-Hop Retrieval over Knowledge Graphs

Jun 29, 2026

This work addresses the limitation of existing multi-hop knowledge graph retrieval methods, which often neglect query semantics during traversal, resulting in low path relevance. The authors propose a query-aware diffusion activation approach that employs a semantic gating mechanism at each step, computed via cosine similarity between entity descriptions and the input question, to efficiently guide traversal within a fixed number of steps. Crucially, the entire retrieval process is encoded into a single Cypher query and executed in one step within Neo4j, eliminating the need for external memory loading. This method uniquely reduces query-aware traversal to fixed-iteration semantic gating entirely inside the graph database. Evaluated on MuSiQue, it achieves 32.80 EM, outperforming the HippoRAG baseline by 5.3 EM and 3.4 F1; ablation studies further demonstrate that semantic gating alone yields F1 improvements of 3.6–7.4 and reduces latency by 1.5–4.9×.

0 citationsRead paper