ConceptGuard: Benchmarking Context-Sensitive Unlearning in Large Language Models
本文提出ConceptGuard基准，用于评估大型语言模型在概念层面的上下文敏感性删除能力，以解决现有方法未能有效区分有害与有益知识的问题。
0 citationsRead paper
Selected work
Representative Papers
L3Cube-IndicQuest v2: A Large-Scale Multilingual Benchmark for Evaluating Factual Knowledge of Large Language Models Across Indic Languages
This study addresses the absence of evaluation benchmarks for indigenous Indian factual knowledge in large language models by constructing a multilingual dataset comprising 69,000 question-answer pairs across 20 languages. We propose a hybrid construction strategy integrating context-aware generation, semantic deduplication, and human verification to ensure data quality and scalability, while validating the consistency of multiple evaluation protocols. Experimental results demonstrate that proprietary models significantly outperform open-source alternatives, with Gemma-4-31B comprehensively surpassing Sarvam-30B. Furthermore, model rankings remain highly consistent across diverse evaluation protocols. These findings establish a reliable paradigm for assessing regional knowledge capabilities in LLMs, highlighting both current performance disparities and the robustness of the proposed benchmarking framework for low-resource linguistic contexts.
0 citationsRead paper
Server-side Anti-cheat in FPS games for Aimbot detection using Deep learning and Machine learning
This study addresses the rampant proliferation of aimbot cheats in first-person shooter (FPS) games by proposing YAACS, a deep learning–based server-side detection system. YAACS models sequential player behavioral data—including aiming velocity, firing frequency, and target distance—and innovatively introduces contextualized temporal sequences into anti-cheat mechanisms to effectively discriminate between legitimate and cheating behaviors. The method employs a stacked LSTM architecture followed by dense layers to process 128-tick action sequences and integrates a parser and middleware for seamless deployment within game servers. Experimental results demonstrate that YAACS achieves a classification accuracy of 88.6% with a false positive rate of only 0.97%, substantially outperforming a decision tree baseline (2.68% false positives), thereby enhancing both competitive fairness and player experience without compromising detection precision.
0 citationsRead paper
MuonAll: Muon Variant for Efficient Finetuning of Large Language Models
Existing Muon optimizers only adapt a subset of model parameters, limiting their efficiency in fine-tuning large language models (LLMs). To address this, we propose MuonAll—the first Muon-based optimizer that uniformly incorporates *all* model parameters. It employs a systematic 2D matrix reparameterization technique to enable efficient gradient updates, significantly broadening applicability to open-source pre-trained language models with up to one billion parameters. The method supports distributed training and maintains compatibility with mainstream deep learning frameworks. Extensive experiments across model scales (100M–1B) and standard benchmarks (GLUE, Evaluating LLMs) demonstrate that MuonAll achieves performance on par with AdamW, validating its viability as a general-purpose drop-in replacement optimizer. The implementation is publicly available.
0 citationsRead paper
Opinion Mining Based Entity Ranking using Fuzzy Logic Algorithmic Approach
Existing opinion mining approaches struggle to jointly optimize fine-grained sentiment classification and entity ranking. To address this, we propose a fuzzy logic–based method for aspect-level opinion mining and entity ranking. First, we employ fuzzy inference to perform fine-grained classification of aspect–sentiment pairs in user reviews, overcoming the limitations of traditional binary or discrete sentiment modeling. Second, the resulting sentiment intensities are integrated into an entity ranking model to enable sentiment-aware, precise ranking. This work is the first to systematically apply fuzzy logic to aspect-level sentiment classification and tightly couple it with an entity ranking framework, thereby resolving the long-standing challenge of co-optimizing granularity refinement and ranking performance. Experiments on e-commerce and social media review datasets demonstrate significant improvements in ranking accuracy—averaging +12.7%—establishing a novel paradigm for fine-grained opinion mining.
0 citationsRead paper
Recent publications
Latest Papers
ConceptGuard: Benchmarking Context-Sensitive Unlearning in Large Language Models
本文提出ConceptGuard基准，用于评估大型语言模型在概念层面的上下文敏感性删除能力，以解决现有方法未能有效区分有害与有益知识的问题。
0 citationsRead paper
L3Cube-IndicQuest v2: A Large-Scale Multilingual Benchmark for Evaluating Factual Knowledge of Large Language Models Across Indic Languages
This study addresses the absence of evaluation benchmarks for indigenous Indian factual knowledge in large language models by constructing a multilingual dataset comprising 69,000 question-answer pairs across 20 languages. We propose a hybrid construction strategy integrating context-aware generation, semantic deduplication, and human verification to ensure data quality and scalability, while validating the consistency of multiple evaluation protocols. Experimental results demonstrate that proprietary models significantly outperform open-source alternatives, with Gemma-4-31B comprehensively surpassing Sarvam-30B. Furthermore, model rankings remain highly consistent across diverse evaluation protocols. These findings establish a reliable paradigm for assessing regional knowledge capabilities in LLMs, highlighting both current performance disparities and the robustness of the proposed benchmarking framework for low-resource linguistic contexts.
0 citationsRead paper
Server-side Anti-cheat in FPS games for Aimbot detection using Deep learning and Machine learning
This study addresses the rampant proliferation of aimbot cheats in first-person shooter (FPS) games by proposing YAACS, a deep learning–based server-side detection system. YAACS models sequential player behavioral data—including aiming velocity, firing frequency, and target distance—and innovatively introduces contextualized temporal sequences into anti-cheat mechanisms to effectively discriminate between legitimate and cheating behaviors. The method employs a stacked LSTM architecture followed by dense layers to process 128-tick action sequences and integrates a parser and middleware for seamless deployment within game servers. Experimental results demonstrate that YAACS achieves a classification accuracy of 88.6% with a false positive rate of only 0.97%, substantially outperforming a decision tree baseline (2.68% false positives), thereby enhancing both competitive fairness and player experience without compromising detection precision.
0 citationsRead paper
MuonAll: Muon Variant for Efficient Finetuning of Large Language Models
Existing Muon optimizers only adapt a subset of model parameters, limiting their efficiency in fine-tuning large language models (LLMs). To address this, we propose MuonAll—the first Muon-based optimizer that uniformly incorporates *all* model parameters. It employs a systematic 2D matrix reparameterization technique to enable efficient gradient updates, significantly broadening applicability to open-source pre-trained language models with up to one billion parameters. The method supports distributed training and maintains compatibility with mainstream deep learning frameworks. Extensive experiments across model scales (100M–1B) and standard benchmarks (GLUE, Evaluating LLMs) demonstrate that MuonAll achieves performance on par with AdamW, validating its viability as a general-purpose drop-in replacement optimizer. The implementation is publicly available.
0 citationsRead paper
Opinion Mining Based Entity Ranking using Fuzzy Logic Algorithmic Approach
Existing opinion mining approaches struggle to jointly optimize fine-grained sentiment classification and entity ranking. To address this, we propose a fuzzy logic–based method for aspect-level opinion mining and entity ranking. First, we employ fuzzy inference to perform fine-grained classification of aspect–sentiment pairs in user reviews, overcoming the limitations of traditional binary or discrete sentiment modeling. Second, the resulting sentiment intensities are integrated into an entity ranking model to enable sentiment-aware, precise ranking. This work is the first to systematically apply fuzzy logic to aspect-level sentiment classification and tightly couple it with an entity ranking framework, thereby resolving the long-standing challenge of co-optimizing granularity refinement and ranking performance. Experiments on e-commerce and social media review datasets demonstrate significant improvements in ranking accuracy—averaging +12.7%—establishing a novel paradigm for fine-grained opinion mining.
0 citationsRead paper