Recent advances in Multimodal Large Language Models (MLLMs) have shown promising results in integrating diverse modalities such as texts and images. MLLMs are heavily influenced by modality bias, often relying on language while under-utilizing other modalities like visual inputs. This position paper argues that MLLMs are deeply affected by modality bias. Firstly, we diagnose the current state of modality bias, highlighting its manifestations across various tasks. Secondly, we propose a systematic research road-map related to modality bias in MLLMs. Thirdly, we identify key factors of modality bias in MLLMs and offer actionable suggestions for future research to mitigate it. To substantiate these findings, we conduct experiments that demonstrate the influence of each factor: 1. Data Characteristics: Language data is compact and abstract, while visual data is redundant and complex, creating an inherent imbalance in learning dynamics. 2. Imbalanced Backbone Capabilities: The dominance of pretrained language models in MLLMs leads to overreliance on language and neglect of visual information. 3. Training Objectives: Current objectives often fail to promote balanced cross-modal alignment, resulting in shortcut learning biased toward language. These findings highlight the need for balanced training strategies and model architectures to better integrate multiple modalities in MLLMs. We call for interdisciplinary efforts to tackle these challenges and drive innovation in MLLM research. Our work provides a fresh perspective on modality bias in MLLMs and offers insights for developing more robust and generalizable multimodal systems-advancing progress toward Artificial General Intelligence.

Inspired by the power of large language models (LLMs), our research adapts them to quantum federated learning (QFL) to boost efficiency and performance. We propose a federated fine-tuning method that distills an LLM within QFL, allowing each client to locally adapt the model to its own data while preserving privacy and reducing unnecessary global updates. The fine-tuned LLM also acts as a reinforcement agent, optimizing QFL by adjusting optimizer steps, cutting down communication rounds, and intelligently selecting clients. Experiments show significant efficiency gains. We pioneer a synergy between LLM and QFL, offering: i) practical efficiency: Reduced communication costs and faster convergence. ii) theoretical rigor: Provable guarantees for adaptive federated optimization. iii) scalability: PEFT methods (LoRA, QLoRA) enable deployment on resource-constrained quantum devices. Code implementation is available here 1.

Climate-Eval is a comprehensive benchmark designed to evaluate natural language processing models across a broad range of tasks related to climate change. Climate-Eval aggregates existing datasets along with a newly developed news classification dataset, created specifically for this release. This results in a benchmark of 25 tasks based on 13 datasets, covering key aspects of climate discourse, including text classification, question answering, and information extraction. Our benchmark provides a standardized evaluation suite for systematically assessing the performance of large language models (LLMs) on these tasks. Additionally, we conduct an extensive evaluation of open-source LLMs (ranging from 2B to 70B parameters) in both zero-shot and few-shot settings, analyzing their strengths and limitations in the domain of climate change.

Chain-of-thought (CoT) distillation allows a large language model (LLM) to guide a small language model (SLM) in reasoning tasks. Existing methods train the SLM to learn the long rationale in one iteration, resulting in two issues: 1) Long rationales lead to a large token-level batch size during training, making gradients of core reasoning tokens (i.e., the token will directly affect the correctness of subsequent reasoning) over-smoothed as they contribute a tiny fraction of the rationale. As a result, the SLM converges to sharp minima where it fails to grasp the reasoning logic. 2) The response is slow, as the SLM must generate a long rationale before reaching the answer. Therefore, we propose chunk-wise training (CWT), which uses a heuristic search to divide the rationale into internal semantically coherent chunks and focuses SLM on learning from only one chunk per iteration. In this way, CWT naturally isolates non-reasoning chunks that do not involve the core reasoning token (e.g., summary and transitional chunks) from the SLM learning for reasoning chunks, making the fraction of the core reasoning token increase in the corresponding iteration. Based on CWT, skip-thinking training (STT) is proposed. STT makes the SLM automatically skip non-reasoning medium chunks to reach the answer, improving reasoning speed while maintaining accuracy. We validate our approach on a variety of SLMs and multiple reasoning tasks.

Large language models (LLMs) suffer from high inference latency due to the auto-regressive decoding process. Speculative decoding accelerates inference by generating multiple draft tokens using a lightweight model and verifying them in parallel. However, existing verification methods rely heavily on distributional consistency while overlooking semantic correctness, thereby limiting the potential speedup of speculative decoding. While some methods employ additional models for relaxed verification of draft tokens, they often fail to generalize effectively to more diverse or open-domain settings. In this work, we propose Reflective Verification, a training-free and semantics-aware approach that achieves a better trade-off between correctness and efficiency. Specifically, we leverage the inherent reflective capacity of LLMs to semantically assess the correctness of draft tokens in parallel during verification. Using prompt-based probing, we obtain both the original and reflective distributions of draft tokens in a single forward pass. The fusion of these distributions enables semantic-level verification of draft tokens that incorporates both consistency and correctness. Experiments across multiple domain benchmarks and model scales demonstrate that our method significantly increases the acceptance length of draft tokens without compromising model performance. Furthermore, we find that the proposed Reflective Verification is orthogonal to existing statistical verification methods, and their combination yields additional 5$\sim$15\% improvements in decoding speed.

Recently, significant progress has been made in developing reasoning-capable Large Language Models (LLMs) through long Chain-of-Thought (CoT) techniques. However, this long-CoT reasoning process imposes substantial memory overhead due to the large Key-Value (KV) Cache memory overhead. Post-training KV Cache quantization has emerged as a promising compression technique and has been extensively studied in short-context scenarios. However, directly applying existing methods to long-CoT LLMs causes significant performance degradation due to the following two reasons: (1) Large cumulative error: Existing methods fail to adequately leverage available memory, and they directly quantize the KV Cache during each decoding step, leading to large cumulative quantization error. (2) Short-context calibration: Due to Rotary Positional Embedding (RoPE), the use of short-context data during calibration fails to account for the distribution of less frequent channels in the Key Cache, resulting in performance loss. We propose Progressive Mixed-Precision KV Cache Quantization (PM-KVQ) for long-CoT LLMs to address the above issues in two folds: (1) To reduce cumulative error, we design a progressive quantization strategy to gradually lower the bit-width of KV Cache in each block. Then, we propose block-wise memory allocation to assign a higher bit-width to more sensitive transformer blocks. (2) To increase the calibration length without additional overhead, we propose a new calibration strategy with positional interpolation that leverages short calibration data with positional interpolation to approximate the data distribution of long-context data. Extensive experiments on 7B-70B long-CoT LLMs show that PM-KVQ improves reasoning benchmark performance by up to 8% over SOTA baselines under the same memory budget. Our code is available at https://github.com/thu-nics/PM-KVQ.

Large Language Models (LLMs) demonstrate impressive general capabilities but often struggle with step-by-step reasoning, especially in complex applications such as games. While retrieval-augmented methods like GraphRAG attempt to bridge this gap through cross-document extraction and indexing, their fragmented entity-relation graphs and overly dense local connectivity hinder the construction of coherent reasoning. In this paper, we propose a novel framework based on Goal-Oriented Graphs (GoGs), where each node represents a goal and its associated attributes, and edges encode logical dependencies between goals. This structure enables explicit retrieval of reasoning paths by first identifying high-level goals and recursively retrieving their subgoals, forming coherent reasoning chains to guide LLM prompting. Our method significantly enhances the reasoning ability of LLMs in game-playing tasks, as demonstrated by extensive experiments on the Minecraft testbed, outperforming GraphRAG and other baselines.

Causal attention has become a foundational mechanism in autoregressive vision-language models (VLMs), unifying textual and visual inputs under a single generative framework. However, existing causal mask-based strategies are inherited from large language models (LLMs) where they are tailored for text-only decoding, and their adaptation to vision tokens is insufficiently addressed in the prefill stage. Strictly masking future positions for vision queries introduces overly rigid constraints, which hinder the model's ability to leverage future context that often contains essential semantic cues for accurate inference. In this work, we empirically investigate how different causal masking strategies affect vision-language inference and then propose a family of future-aware attentions tailored for this setting. We first empirically analyze the effect of previewing future tokens for vision queries and demonstrate that rigid masking undermines the model's capacity to capture useful contextual semantic representations. Based on these findings, we propose a lightweight attention family that aggregates future visual context into past representations via pooling, effectively preserving the autoregressive structure while enhancing cross-token dependencies. We evaluate a range of causal masks across diverse vision-language inference settings and show that selectively compressing future semantic context into past representations benefits the inference.

Multimodal large language models (MLLMs) have made significant progress in document understanding. However, the information-dense nature of document images still poses challenges, as most queries depend on only a few relevant regions, with the rest being redundant. Existing one-pass MLLMs process entire document images without considering query relevance, often failing to focus on critical regions and producing unfaithful responses. Inspired by the human coarse-to-fine reading pattern, we introduce Doc-CoB (Chain-of-Box), a simple-yet-effective mechanism that integrates human-style visual reasoning into MLLM without modifying its architecture. Our method allows the model to autonomously select the set of regions (boxes) most relevant to the query, and then focus attention on them for further understanding. We first design a fully automatic pipeline, integrating a commercial MLLM with a layout analyzer, to generate 249k training samples with intermediate visual reasoning supervision. Then we incorporate two enabling tasks that improve box identification and box-query reasoning, which together enhance document understanding. Extensive experiments on seven benchmarks with four popular models show that Doc-CoB significantly improves performance, demonstrating its effectiveness and wide applicability. All code, data, and models will be released publicly.

The development of large language models (LLMs) has provided new tools for research in supply chain management (SCM). In this paper, we introduce a retrieval-augmented generation (RAG) framework that dynamically integrates external knowledge into the inference process, and develop a domain-specialized SCM LLM, which demonstrates expert-level competence by passing standardized SCM examinations and beer game tests. We further employ the use of LLMs to conduct horizontal and vertical supply chain games, in order to analyze competition and cooperation within supply chains. Our experiments show that RAG significantly improves performance on SCM tasks. Moreover, game-theoretic analysis reveals that the LLM can reproduce insights from the classical SCM literature, while also uncovering novel behaviors and offering fresh perspectives on phenomena such as the bullwhip effect. This paper opens the door for exploring cooperation and competition for complex supply chain network through the lens of LLMs.