Institution profile

Ant Group

Industry researchasia · cn
Official website
Research library1,032linked papers
Opportunities0open roles
Selected work

Representative Papers

Hide and Seek in Noise Labels: Noise-Robust Collaborative Active Learning with LLMs-Powered Assistance

Apr 03, 2025Annual Meeting of the Association for Computational Linguistics

To address the challenge of accurately identifying and correcting mislabeled samples in learning with noisy labels, this paper proposes NoiseAL—a novel framework that achieves fine-grained separation of clean and noisy samples via dual lightweight model co-prediction and a dynamically adjusted confidence threshold. It further introduces an LLM-driven active labeling mechanism for semantic-level correction of noisy labels. Innovatively, we establish a hierarchical collaborative learning paradigm for noisy data and design subset-specific multi-objective optimization (employing CE, GCE, and SCE losses) tailored to varying sample quality. Extensive experiments on both synthetic and real-world noisy benchmarks demonstrate that NoiseAL improves noise robustness by 12.7% over state-of-the-art methods and reduces human annotation cost by over 40%, thereby overcoming the coarse-grained partitioning limitation inherent in conventional label-noise learning approaches.

5 citationsRead paper

Unify Graph Learning with Text: Unleashing LLM Potentials for Session Search

May 13, 2024The Web Conference

Existing conversational search methods inadequately model graph-structured interactions and fail to fuse fine-grained textual semantics with structural information. Method: We propose the Symbolic Graph Ranker (SGR), which (i) formalizes dialogue history as a conversational graph and symbolically encodes it—via an interpretable, grammar-based syntax—into LLM-processable textual sequences; (ii) introduces multi-granularity self-supervised objectives (link prediction, node generation, contrastive generation) to explicitly guide LLMs in learning graph topology; and (iii) bypasses reliance on generic document embeddings to jointly model lexical semantics and graph structure. Contribution/Results: SGR achieves significant improvements over state-of-the-art methods on the AOL and Tiangong-ST benchmarks, demonstrating both the effectiveness and generalizability of graph-text joint modeling for conversational search.

4 citationsRead paper

E2LLM: Encoder Elongated Large Language Models for Long-Context Understanding and Reasoning

Sep 10, 2024arXiv.org

Addressing the “impossible triangle” in long-context modeling for large language models—where high performance, computational efficiency, and compatibility with pretrained models are mutually exclusive—this paper proposes an encoder-extension architecture. It freezes a pretrained text encoder (e.g., BERT or CLIP) to compress long inputs into soft prompts, then introduces a learnable adapter coupled with dual-objective adaptation training: reconstruction loss and long-context instruction fine-tuning. This enables decoder-only LLMs to model extended contexts efficiently without modifying the backbone architecture. The method is fully plug-and-play compatible with existing decoder-only models. Evaluated on multi-turn dialogue and long-document summarization, it outperforms state-of-the-art long-context approaches, achieving a 2.3× inference speedup and 37% reduction in GPU memory consumption—demonstrating strong performance, high efficiency, and seamless integration.

3 citationsRead paper

CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences

Mar 16, 2025

Existing uniform layer-wise KV cache eviction schemes fail to accommodate inter-layer variations in attention patterns and temporal importance decay, leading to memory explosion in long-context reasoning for large language models. Method: We propose a hierarchical preference-driven cascaded adaptive cache eviction mechanism: (1) a novel layer-preference-aware global cache allocation paradigm modeled as dynamic “cake-cutting”-style resource partitioning; (2) a dynamic weighted eviction scoring metric integrating spatiotemporal attention analysis; and (3) a cascaded memory-constrained management framework enabling deep synergy with FlashAttention-2. Results: On LongBench and NeedleBench, our method retains full model performance using only 3.2% of the original KV cache, reduces decoding latency by over 10× for 128K-context inputs, and significantly outperforms state-of-the-art methods under stringent memory constraints.

2 citations1 influentialRead paper

Interpretable All-Type Audio Deepfake Detection with Audio LLMs via Frequency-Time Reinforcement Learning

Jan 06, 2026arXiv.org

This work addresses the challenge of generalizing audio deepfake detection across diverse audio types—including speech, environmental sounds, singing, and music—where existing methods struggle to balance performance and interpretability. The authors propose a two-stage training framework based on Audio Large Language Models (ALLMs). First, they construct interpretable supervision signals via a frequency-time structured chain-of-thought (CoT) with automatic annotation. Subsequently, they perform reinforcement fine-tuning by integrating supervised fine-tuning (SFT) with a novel Frequency-Time Grouped Relative Policy Optimization (FT-GRPO). The resulting model achieves state-of-the-art performance across all audio forgery detection tasks while generating human-interpretable reasoning grounded in frequency-time features, effectively mitigating reward hacking and hallucination issues.

2 citationsRead paper
Recent publications

Latest Papers