Selected work

Representative Papers

Anticipate & Act: Integrating LLMs and Classical Planning for Efficient Task Execution in Household Environments†

May 13, 2024IEEE International Conference on Robotics and Automation

To address low multi-task execution efficiency of assistive agents in domestic environments, this paper proposes an LLM-driven joint task planning framework. It leverages large language models (LLMs) with few-shot prompting to achieve zero-shot high-level task anticipation, then uniformly encodes the anticipated multi-task set as a PDDL goal for classical planning—specifically, the FF Planner—to generate a synergistically optimized, fine-grained action sequence. This work establishes the first seamless integration of LLM-based task anticipation with symbolic classical planning, enabling cross-task action coordination without any training data. Evaluated in the VirtualHome simulation environment, the framework reduces task completion time by 31% compared to serial single-task execution baselines, demonstrating its effectiveness in action reuse, temporal optimization, and resource coordination.

7 citations1 influentialRead paper

Just KIDDIN: Knowledge Infusion and Distillation for Detection of INdecent Memes

Nov 19, 2024arXiv.org

This paper addresses the challenge of toxicity detection in multimodal hate memes—composite images with overlaid text. To this end, we propose a neuro-symbolic framework integrating knowledge distillation and explicit commonsense injection. Methodologically: (i) cross-modal knowledge distillation from a large vision-language model (LVLM) captures implicit toxic semantics; (ii) a ConceptNet subgraph is constructed and embedded into a multimodal alignment space to explicitly model commonsense-driven toxic associations between image and text; (iii) a relation-aware reasoning module orchestrates synergistic interaction between the two components. To our knowledge, this is the first work to jointly leverage LVLM-based distillation and structured knowledge graph infusion for hate speech detection. Evaluated on two benchmark datasets for hateful meme detection, our framework achieves absolute improvements of 1.1%, 7.0%, and 35.0% in AU-ROC, F1-score, and Recall, respectively—substantially outperforming existing state-of-the-art methods.

1 citationsRead paper

How Do VLMs Behave When Blind or Misled? Behavioral Evaluation of VLMs on Scientific Figures

Aug 13, 2026

This work addresses a critical gap in the evaluation of vision-language models (VLMs), which typically emphasize perceptual and reasoning accuracy while neglecting behavioral reliability under missing or misleading visual evidence. To this end, we introduce SciFigBench—a challenging diagnostic benchmark for scientific figure understanding comprising over 34,000 test samples—designed with image perturbations, adversarial probes, and selective blurring, complemented by human-annotated labels and multidimensional metrics including MQM scores and reasoning accuracy. We further propose the A-R-I framework to systematically assess whether models Acknowledge insufficient evidence, Resist misleading cues, and Infer cautiously under uncertainty. Empirical results reveal that while GPT-5.2 achieves high accuracy, it frequently hallucinates; in contrast, Gemini 3.1 Pro demonstrates comparable performance with markedly higher reliability, explicitly acknowledging uncertainty in 71% of cases and attaining a resistance-to-misleading score of 0.91.

0 citationsRead paper

MapRoute++: Surrogate-Guided Semantic Routing for Visual Concept Unlearning

Aug 13, 2026

This work addresses the challenge of precisely removing specific visual concepts while preserving both unrelated and semantically proximate concepts in concept erasure tasks. Building upon the Stable Diffusion v1.4 framework, the authors propose a method that integrates a semantic routing mechanism with a proxy-guided strategy. By incorporating task-specific training objectives, enhanced concept representations, and a dynamic mapper selection scheme, the approach achieves fine-grained and high-fidelity concept forgetting. Evaluated on the official Genμ 2.0 Challenge benchmark, the method outperforms the current state-of-the-art baseline by an average of 12.1% in terms of the ERR metric, demonstrating consistent and significant improvements across all five concept categories.

0 citationsRead paper

Cultivar: A Contrastive and Locale-Oriented Translation Benchmark for Investigating Contamination and Localisation Robustness

Aug 10, 2026

This study addresses the limitations of current machine translation evaluation practices, which are predominantly English-centric, overlook regional cultural differences, and are susceptible to data contamination, thereby failing to assess model robustness on localized content. To remedy this, the authors propose a source-language contrastive evaluation paradigm and introduce Cultivar—a benchmark derived from a localized subset of FLORES—that compares model performance on localized versus non-localized translations to detect data contamination and evaluate regional adaptability. This framework extends the unit of evaluation from language pairs to localized content, systematically uncovering performance disparities across regional contexts. Experiments on 32 open-source models reveal insufficient robustness in specialized translation systems, evidence of overfitting to FLORES in some cases, and a consistent bias favoring U.S.-localized content.

0 citationsRead paper

Recent publications

Latest Papers

How Do VLMs Behave When Blind or Misled? Behavioral Evaluation of VLMs on Scientific Figures

Aug 13, 2026

This work addresses a critical gap in the evaluation of vision-language models (VLMs), which typically emphasize perceptual and reasoning accuracy while neglecting behavioral reliability under missing or misleading visual evidence. To this end, we introduce SciFigBench—a challenging diagnostic benchmark for scientific figure understanding comprising over 34,000 test samples—designed with image perturbations, adversarial probes, and selective blurring, complemented by human-annotated labels and multidimensional metrics including MQM scores and reasoning accuracy. We further propose the A-R-I framework to systematically assess whether models Acknowledge insufficient evidence, Resist misleading cues, and Infer cautiously under uncertainty. Empirical results reveal that while GPT-5.2 achieves high accuracy, it frequently hallucinates; in contrast, Gemini 3.1 Pro demonstrates comparable performance with markedly higher reliability, explicitly acknowledging uncertainty in 71% of cases and attaining a resistance-to-misleading score of 0.91.

0 citationsRead paper

MapRoute++: Surrogate-Guided Semantic Routing for Visual Concept Unlearning

Aug 13, 2026

This work addresses the challenge of precisely removing specific visual concepts while preserving both unrelated and semantically proximate concepts in concept erasure tasks. Building upon the Stable Diffusion v1.4 framework, the authors propose a method that integrates a semantic routing mechanism with a proxy-guided strategy. By incorporating task-specific training objectives, enhanced concept representations, and a dynamic mapper selection scheme, the approach achieves fine-grained and high-fidelity concept forgetting. Evaluated on the official Genμ 2.0 Challenge benchmark, the method outperforms the current state-of-the-art baseline by an average of 12.1% in terms of the ERR metric, demonstrating consistent and significant improvements across all five concept categories.

0 citationsRead paper

Cultivar: A Contrastive and Locale-Oriented Translation Benchmark for Investigating Contamination and Localisation Robustness

Aug 10, 2026

This study addresses the limitations of current machine translation evaluation practices, which are predominantly English-centric, overlook regional cultural differences, and are susceptible to data contamination, thereby failing to assess model robustness on localized content. To remedy this, the authors propose a source-language contrastive evaluation paradigm and introduce Cultivar—a benchmark derived from a localized subset of FLORES—that compares model performance on localized versus non-localized translations to detect data contamination and evaluate regional adaptability. This framework extends the unit of evaluation from language pairs to localized content, systematically uncovering performance disparities across regional contexts. Experiments on 32 open-source models reveal insufficient robustness in specialized translation systems, evidence of overfitting to FLORES in some cases, and a consistent bias favoring U.S.-localized content.

0 citationsRead paper

SurakshaEval: An Indic Safety Benchmark for Multilingual LLMs

Aug 07, 2026

This study addresses the critical gap in safety evaluations of large language models, which have predominantly focused on English and Western contexts while overlooking the cultural diversity and local sensitivities of multilingual regions such as India. The work proposes the first systematically constructed, human-authored safety benchmark encompassing ten major Indian languages alongside English, integrating both general and region-specific prompts. It employs a fine-grained safety taxonomy and a context-aware evaluation protocol. Comprehensive assessments of prominent multilingual models reveal pervasive issues in Indian languages—particularly when written in native scripts—including excessive refusal rates, failure to detect implicit biases, and inadequate understanding of regional context. These findings underscore the urgent need for localized safety evaluation frameworks beyond English-centric paradigms.

0 citationsRead paper

Resourced Authority A Mechanism-Design Model for Participatory Governance of Deployed AI Agents

Aug 06, 2026

This work addresses the challenge of implementing continuous, effective, and self-enforcing participatory governance over deployed AI agents to prevent behavioral drift from human interests. It proposes a mechanism-design-based governance framework that uniquely leverages computational resources as a core governance instrument. By introducing a dedicated governance currency and market mechanisms, the framework enables empirically verified human stakeholders to express preferences through sequential games. These preferences are aggregated via a dual-threshold rule into a binary authorization signal, which is then mapped to signed computational permits bounded by a safety cap, thereby enabling hardware-level self-execution of decisions. The framework formally characterizes the class of governable agents, identifies “agent manipulation of voters” as a critical vulnerability, and establishes a new paradigm for “safe AI.”

0 citationsRead paper