arXiv subjects

Conghui Zhu

Publications and source records attributed to Conghui Zhu.

At least 19 recordsLinked to original sources

arXiv2026

SPAE: Spectrally Guided Autoencoder for Pretrained Visual Latents

Latents from vision foundation models (VFMs) are semantically rich and well suited for visual understanding. Recent representation autoencoder methods such as RAE have shown that they can provide promising latent spaces for image generation. However, VFM latents remain difficult to model directly: DiT-generated latents exhibit spectral mismatch with encoder latents, especially in high-frequency components. Our channel-wise spectral analysis further reveals that these high-frequency components are diffusely distributed across latent channels and entangled with semantic information, making the latent space difficult for DiT to model. To address these challenges, we propose SPAE, latent adaptation framework for generation. Specifically, SPAE employs a compact bottleneck to distill stable semantic information while suppressing high-frequency components, thereby improving the alignment between DiT-generated latents and encoder latents. In addition, we apply a channel-wise masking strategy to promote the decoupling of semantic information and high-frequency details across bottleneck channels. Experiments show that SPAE achieves a favorable balance among visual understanding, generation quality, and reconstruction fidelity.

cs.CV↗

arXiv2026

BiCAA: Bidirectional Credit Assignment for Search-Augmented Agent

Multi-step search is a fundamental capability for search agents, enabling them to iteratively acquire, refine, and integrate external evidence for complex reasoning QA. However, vanilla GRPO allocates rewards exclusively based on the model's final outputs, yielding outcome-only supervision with no supervisory signals for intermediate reasoning steps. Such sparse supervision easily causes training instability and redundant search behaviors on multi-step search tasks. To mitigate this limitation, we adopt process reward to deliver stepwise supervision signals. For this process reward, we propose two complementary criteria to judge each search step: whether the step yields new evidence to facilitate problem solving, and whether it forms an efficient, pivotal intermediate decision within the overall reasoning trajectory. Building on this insight, we propose BiCAA: a bidirectional credit assignment framework that delivers dense, distinguishing process rewards for search-augmented agents. BiCAA builds bidirectional process rewards by fusing two complementary signals: forward solvability gain and hindsight success criticality. The former quantifies step-wise improvements in answer plausibility, while the latter evaluates each step's necessity for final success via hindsight outcome-based criticality scoring. We modulate and aggregate the two signals and then fuse them with the outcome reward. Experiments on search-augmented QA benchmarks show that BiCAA stabilizes policy optimization, reduces redundant search behavior, and achieves competitive performance.

cs.CL↗

arXiv2026

Long-form RewardBench: Evaluating Reward Models for Long-form Generation

The widespread adoption of reinforcement learning-based alignment highlights the growing importance of reward models. Various benchmarks have been built to evaluate reward models in various domains and scenarios. However, a significant gap remains in assessing reward models for long-form generation, despite its critical role in real-world applications. To bridge this, we introduce Long-form RewardBench, the first reward modeling testbed specifically designed for long-form generation. Our benchmark encompasses five key subtasks: QA, RAG, Chat, Writing, and Reasoning. We collected instruction and preference data through a meticulously designed multi-stage data collection process, and conducted extensive experiments on 20+ mainstream reward models, including both classifiers and generative models. Our findings reveal that current models still lack long-form reward modeling capabilities. Furthermore, we designed a novel Long-form Needle-in-a-Haystack Test, which revealed a correlation between reward modeling performance and the error's position within a response, as well as the overall response length, with distinct characteristics observed between classification and generative models. Finally, we demonstrate that classifiers exhibit better generalizability compared to generative models trained on the same data. As the first benchmark for long-form reward modeling, this work aims to offer a robust platform for visualizing progress in this crucial area.

cs.CL↗

arXiv2026

Toward Robust LLM-Based Judges: Taxonomic Bias Evaluation and Debiasing Optimization

Large language model (LLM)-based judges are widely adopted for automated evaluation and reward modeling, yet their judgments are often affected by judgment biases. Accurately evaluating these biases is essential for ensuring the reliability of LLM-based judges. However, existing studies typically investigate limited biases under a single judge formulation, either generative or discriminative, lacking a comprehensive evaluation. To bridge this gap, we propose JudgeBiasBench, a benchmark for systematically quantifying biases in LLM-based judges. JudgeBiasBench defines a taxonomy of judgment biases across 4 dimensions, and constructs bias-augmented evaluation instances through a controlled bias injection pipeline, covering 12 representative bias types. We conduct extensive experiments across both generative and discriminative judges, revealing that current judges exhibit significant and diverse bias patterns that often compromise the reliability of automated evaluation. To mitigate judgment bias, we propose bias-aware training that explicitly incorporates bias-related attributes into the training process, encouraging judges to disentangle task-relevant quality from bias-correlated cues. By adopting reinforcement learning for generative judges and contrastive learning for discriminative judges, our methods effectively reduce judgment biases while largely preserving general evaluation capability.

cs.CL↗

arXiv2025

Video2Layout: Recall and Reconstruct Metric-Grounded Cognitive Map for Spatial Reasoning

Spatial intelligence is a critical frontier for Multimodal Large Language Models (MLLMs), empowering them to comprehend the physical world. Drawing inspiration from human perception mechanisms, prior studies attempt to construct a spatial understanding via grid-based cognitive maps. However, current grid-based map methods rely on discretized representations, which limit the model's ability in fine-grained spatial reasoning. To overcome this limitation, we propose Video2Layout, a framework for reconstructing metric-grounded spatial layouts from video. The framework uses continuous object boundary coordinates to enable quantitative spatial computation, which effectively reduces ambiguity in natural language descriptions of spatial relationships. Specifically, our method comprises two stages. First, in supervised fine-tuning stage, we construct a high-quality dataset from the AI2THOR simulator, which enables the model to learn the mapping from visual inputs to precise boundary coordinates. Subsequently, a reinforcement fine-tuning stage enhances the model's real-world generalization capabilities. Based on the above framework, we investigate factors that affect cognitive map accuracy and quantify its relationship with task performance. Evaluated on mainstream spatial reasoning benchmarks, our model, V2LO-7B, achieves an average improvement of 3.24\% over the model trained on grid maps, validating the superiority of our method.

cs.CV↗

arXiv2025

From Perception to Reasoning: Deep Thinking Empowers Multimodal Large Language Models

With the remarkable success of Multimodal Large Language Models (MLLMs) in perception tasks, enhancing their complex reasoning capabilities has emerged as a critical research focus. Existing models still suffer from challenges such as opaque reasoning paths and insufficient generalization ability. Chain-of-Thought (CoT) reasoning, which has demonstrated significant efficacy in language models by enhancing reasoning transparency and output interpretability, holds promise for improving model reasoning capabilities when extended to the multimodal domain. This paper provides a systematic review centered on "Multimodal Chain-of-Thought" (MCoT). First, it analyzes the background and theoretical motivations for its inception from the perspectives of technical evolution and task demands. Then, it introduces mainstream MCoT methods from three aspects: CoT paradigms, the post-training stage, and the inference stage, while also analyzing their underlying mechanisms. Furthermore, the paper summarizes existing evaluation benchmarks and metrics, and discusses the application scenarios of MCoT. Finally, it analyzes the challenges currently facing MCoT and provides an outlook on its future research directions.

cs.CL↗

arXiv2025

Speculative Decoding Meets Quantization: Compatibility Evaluation and Hierarchical Framework Design

Speculative decoding and quantization effectively accelerate memory-bound inference of large language models. Speculative decoding mitigates the memory bandwidth bottleneck by verifying multiple tokens within a single forward pass, which increases computational effort. Quantization achieves this optimization by compressing weights and activations into lower bit-widths and also reduces computations via low-bit matrix multiplications. To further leverage their strengths, we investigate the integration of these two techniques. Surprisingly, experiments applying the advanced speculative decoding method EAGLE-2 to various quantized models reveal that the memory benefits from 4-bit weight quantization are diminished by the computational load from speculative decoding. Specifically, verifying a tree-style draft incurs significantly more time overhead than a single-token forward pass on 4-bit weight quantized models. This finding led to our new speculative decoding design: a hierarchical framework that employs a small model as an intermediate stage to turn tree-style drafts into sequence drafts, leveraging the memory access benefits of the target quantized model. Experimental results show that our hierarchical approach achieves a 2.78$\times$ speedup across various tasks for the 4-bit weight Llama-3-70B model on an A100 GPU, outperforming EAGLE-2 by 1.31$\times$. Code available at https://github.com/AI9Stars/SpecMQuant.

cs.CL↗

arXiv2025

Lost in Benchmarks? Rethinking Large Language Model Benchmarking with Item Response Theory

The evaluation of large language models (LLMs) via benchmarks is widespread, yet inconsistencies between different leaderboards and poor separability among top models raise concerns about their ability to accurately reflect authentic model capabilities. This paper provides a critical analysis of benchmark effectiveness, examining mainstream prominent LLM benchmarks using results from diverse models. We first propose Pseudo-Siamese Network for Item Response Theory (PSN-IRT), an enhanced Item Response Theory framework that incorporates a rich set of item parameters within an IRT-grounded architecture. PSN-IRT can be utilized for accurate and reliable estimations of item characteristics and model abilities. Based on PSN-IRT, we conduct extensive analysis on 11 LLM benchmarks comprising 41,871 items, revealing significant and varied shortcomings in their measurement quality. Furthermore, we demonstrate that leveraging PSN-IRT is able to construct smaller benchmarks while maintaining stronger alignment with human preference.

cs.CL↗

arXiv2025

MuSC: Improving Complex Instruction Following with Multi-granularity Self-Contrastive Training

Complex instruction-following with elaborate constraints is imperative for Large Language Models (LLMs). While existing methods have constructed data for complex instruction alignment, they all rely on a more advanced model, especially GPT-4, limiting their application. In this paper, we propose a Multi-granularity Self-Contrastive Training (MuSC) framework, to improve the complex instruction alignment without relying on a stronger model. Our method is conducted on both coarse and fine granularity. On coarse-granularity, we construct constraint-aware preference data based on instruction decomposition and recombination. On fine-granularity, we perform token-aware preference optimization with dynamic token-level supervision. Our method is evaluated on open-sourced models, and experiment results show our method achieves significant improvement on both complex and general instruction-following benchmarks, surpassing previous self-alignment methods.

cs.CL↗

arXiv2025

DuplexMamba: Enhancing Real-time Speech Conversations with Duplex and Streaming Capabilities

Real-time speech conversation is essential for natural and efficient human-machine interactions, requiring duplex and streaming capabilities. Traditional Transformer-based conversational chatbots operate in a turn-based manner and exhibit quadratic computational complexity that grows as the input size increases. In this paper, we propose DuplexMamba, a Mamba-based end-to-end multimodal duplex model for speech-to-text conversation. DuplexMamba enables simultaneous input processing and output generation, dynamically adjusting to support real-time streaming. Specifically, we develop a Mamba-based speech encoder and adapt it with a Mamba-based language model. Furthermore, we introduce a novel duplex decoding strategy that enables DuplexMamba to process input and generate output simultaneously. Experimental results demonstrate that DuplexMamba successfully implements duplex and streaming capabilities while achieving performance comparable to several recently developed Transformer-based models in automatic speech recognition (ASR) tasks and voice assistant benchmark evaluations. Our code and model are released.

cs.CL↗

arXiv2025

SEO: Stochastic Experience Optimization for Large Language Models

Large Language Models (LLMs) can benefit from useful experiences to improve their performance on specific tasks. However, finding helpful experiences for different LLMs is not obvious, since it is unclear what experiences suit specific LLMs. Previous studies intended to automatically find useful experiences using LLMs, while it is difficult to ensure the effectiveness of the obtained experience. In this paper, we propose Stochastic Experience Optimization (SEO), an iterative approach that finds optimized model-specific experience without modifying model parameters through experience update in natural language. In SEO, we propose a stochastic validation method to ensure the update direction of experience, avoiding unavailing updates. Experimental results on three tasks for three LLMs demonstrate that experiences optimized by SEO can achieve consistently improved performance. Further analysis indicates that SEO-optimized experience can generalize to out-of-distribution data, boosting the performance of LLMs on similar tasks.

cs.CL↗

arXiv2024

Mitigating the Bias of Large Language Model Evaluation

Recently, there has been a trend of evaluating the Large Language Model (LLM) quality in the flavor of LLM-as-a-Judge, namely leveraging another LLM to evaluate the current output quality. However, existing judges are proven to be biased, namely they would favor answers which present better superficial quality (such as verbosity, fluency) while ignoring the instruction following ability. In this work, we propose systematic research about the bias of LLM-as-a-Judge. Specifically, for closed-source judge models, we apply calibration to mitigate the significance of superficial quality, both on probability level and prompt level. For open-source judge models, we propose to mitigate the bias by contrastive training, with curated negative samples that deviate from instruction but present better superficial quality. We apply our methods on the bias evaluation benchmark, and experiment results show our methods mitigate the bias by a large margin while maintaining a satisfactory evaluation accuracy.

cs.CL↗

arXiv2024

LoRA-drop: Efficient LoRA Parameter Pruning based on Output Evaluation

Low-Rank Adaptation (LoRA) is currently the most commonly used Parameter-efficient fine-tuning (PEFT) method, it introduces auxiliary parameters for each layer to fine-tune the pre-trained model under limited computing resources. However, it still faces resource consumption challenges during training when scaling up to larger models. Most previous studies have tackled this issue by using pruning techniques, which involve removing LoRA parameters deemed unimportant. Nonetheless, these efforts only analyze LoRA parameter features to evaluate their importance, such as parameter count, size, and gradient. In fact, the output of LoRA (product of LoRA parameter and hidden state), directly impacts the final results. Preliminary experiments indicate that a fraction of LoRA elements possesses significantly high output values, substantially influencing the layer output. Motivated by the observation, we propose LoRA-drop. Concretely, LoRA-drop evaluates the importance of LoRA based on the LoRA output. Then we retain LoRA for important layers and the other layers share the same LoRA. We conduct abundant experiments with models of different scales on NLU and NLG tasks. Results demonstrate that LoRA-drop can achieve performance comparable to full fine-tuning and LoRA, while retaining 50\% of the LoRA parameters on average.

cs.LG↗

arXiv2022

CLMLF:A Contrastive Learning and Multi-Layer Fusion Method for Multimodal Sentiment Detection

Compared with unimodal data, multimodal data can provide more features to help the model analyze the sentiment of data. Previous research works rarely consider token-level feature fusion, and few works explore learning the common features related to sentiment in multimodal data to help the model fuse multimodal features. In this paper, we propose a Contrastive Learning and Multi-Layer Fusion (CLMLF) method for multimodal sentiment detection. Specifically, we first encode text and image to obtain hidden representations, and then use a multi-layer fusion module to align and fuse the token-level features of text and image. In addition to the sentiment analysis task, we also designed two contrastive learning tasks, label based contrastive learning and data based contrastive learning tasks, which will help the model learn common features related to sentiment in multimodal data. Extensive experiments conducted on three publicly available multimodal datasets demonstrate the effectiveness of our approach for multimodal sentiment detection compared with existing methods. The codes are available for use at https://github.com/Link-Li/CLMLF

cs.CL↗

arXiv2021

A Neural Conversation Generation Model via Equivalent Shared Memory Investigation

Conversation generation as a challenging task in Natural Language Generation (NLG) has been increasingly attracting attention over the last years. A number of recent works adopted sequence-to-sequence structures along with external knowledge, which successfully enhanced the quality of generated conversations. Nevertheless, few works utilized the knowledge extracted from similar conversations for utterance generation. Taking conversations in customer service and court debate domains as examples, it is evident that essential entities/phrases, as well as their associated logic and inter-relationships can be extracted and borrowed from similar conversation instances. Such information could provide useful signals for improving conversation generation. In this paper, we propose a novel reading and memory framework called Deep Reading Memory Network (DRMN) which is capable of remembering useful information of similar conversations for improving utterance generation. We apply our model to two large-scale conversation datasets of justice and e-commerce fields. Experiments prove that the proposed model outperforms the state-of-the-art approaches.

cs.CL↗

arXiv2020

AI-lead Court Debate Case Investigation

The multi-role judicial debate composed of the plaintiff, defendant, and judge is an important part of the judicial trial. Different from other types of dialogue, questions are raised by the judge, The plaintiff, plaintiff's agent defendant, and defendant's agent would be to debating so that the trial can proceed in an orderly manner. Question generation is an important task in Natural Language Generation. In the judicial trial, it can help the judge raise efficient questions so that the judge has a clearer understanding of the case. In this work, we propose an innovative end-to-end question generation model-Trial Brain Model (TBM) to build a Trial Brain, it can generate the questions the judge wants to ask through the historical dialogue between the plaintiff and the defendant. Unlike prior efforts in natural language generation, our model can learn the judge's questioning intention through predefined knowledge. We do experiments on real-world datasets, the experimental results show that our model can provide a more accurate question in the multi-role court debate scene.

cs.CL↗

arXiv2020

Cross Copy Network for Dialogue Generation

In the past few years, audiences from different fields witness the achievements of sequence-to-sequence models (e.g., LSTM+attention, Pointer Generator Networks, and Transformer) to enhance dialogue content generation. While content fluency and accuracy often serve as the major indicators for model training, dialogue logics, carrying critical information for some particular domains, are often ignored. Take customer service and court debate dialogue as examples, compatible logics can be observed across different dialogue instances, and this information can provide vital evidence for utterance generation. In this paper, we propose a novel network architecture - Cross Copy Networks(CCN) to explore the current dialog context and similar dialogue instances' logical structure simultaneously. Experiments with two tasks, court debate and customer service content generation, proved that the proposed algorithm is superior to existing state-of-art content generation models.

cs.CL↗

arXiv2020

Reliable Evaluations for Natural Language Inference based on a Unified Cross-dataset Benchmark

Recent studies show that crowd-sourced Natural Language Inference (NLI) datasets may suffer from significant biases like annotation artifacts. Models utilizing these superficial clues gain mirage advantages on the in-domain testing set, which makes the evaluation results over-estimated. The lack of trustworthy evaluation settings and benchmarks stalls the progress of NLI research. In this paper, we propose to assess a model's trustworthy generalization performance with cross-datasets evaluation. We present a new unified cross-datasets benchmark with 14 NLI datasets, and re-evaluate 9 widely-used neural network-based NLI models as well as 5 recently proposed debiasing methods for annotation artifacts. Our proposed evaluation scheme and experimental baselines could provide a basis to inspire future reliable NLI research.

cs.CL↗