ConCISE: Confidence-guided Compression in Step-by-step Efficient Reasoning

Large Reasoning Models (LRMs) perform strongly in complex reasoning tasks via Chain-of-Thought (CoT) prompting, but often suffer from verbose outputs caused by redundant content, increasing computational overhead, and degrading user experience. Existing compression methods either operate post-hoc pruning, risking disruption to reasoning coherence, or rely on sampling-based selection, which fails to intervene effectively during generation. In this work, we introduce a confidence-guided perspective to explain the emergence of redundant reflection in LRMs, identifying two key patterns: Confidence Deficit, where the model reconsiders correct steps due to low internal confidence, and Termination Delay, where reasoning continues even after reaching a confident answer. Based on this analysis, we propose ConCISE (Confidence-guided Compression In Step-by-step Efficient Reasoning), a framework that simplifies reasoning chains by reinforcing the model's confidence during inference, thus preventing the generation of redundant reflection steps. It integrates Confidence Injection to stabilize intermediate steps and Early Stopping to terminate reasoning when confidence is sufficient. Extensive experiments demonstrate that fine-tuning LRMs on ConCISE-generated data yields significantly shorter outputs, reducing length by up to approximately 50% under SimPO, while maintaining high task accuracy. ConCISE consistently outperforms existing baselines across multiple reasoning benchmarks.

A Dual-Space Framework for General Knowledge Distillation of Large Language Models

Knowledge distillation (KD) is a promising solution to compress large language models (LLMs) by transferring their knowledge to smaller models. During this process, white-box KD methods usually minimize the distance between the output distributions of the teacher model and the student model to transfer more information. However, we reveal that the current white-box KD framework exhibits two limitations: a) bridging probability distributions from different output spaces will limit the similarity between the teacher model and the student model; b) this framework cannot be applied to LLMs with different vocabularies. One of the root causes for these limitations is that the distributions from the teacher and the student for KD are output by different prediction heads, which yield distributions in different output spaces and dimensions. Therefore, in this paper, we propose a dual-space knowledge distillation (DSKD) framework that unifies the prediction heads of the teacher and the student models for KD. Specifically, we first introduce two projectors with ideal initialization to project the teacher/student hidden states into the student/teacher representation spaces. After this, the hidden states from different models can share the same head and unify the output spaces of the distributions. Furthermore, we develop an exact token alignment (ETA) algorithm to align the same tokens in two differently-tokenized sequences. Based on the above, our DSKD framework is a general KD framework that supports both off-policy and on-policy KD, and KD between any two LLMs regardless of their vocabularies. Extensive experiments on instruction-following, mathematical reasoning, and code generation benchmarks show that DSKD significantly outperforms existing methods based on the current white-box KD framework and surpasses other cross-tokenizer KD methods for LLMs with different vocabularies.

Deep Reasoning Translation via Reinforcement Learning

Recently, deep reasoning LLMs (e.g., OpenAI o1/o3 and DeepSeek-R1) have shown promising performance in various complex tasks. Free translation is an important and interesting task in the multilingual world, which requires going beyond word-for-word translation and taking cultural differences into account. This task is still under-explored in deep reasoning LLMs. In this paper, we introduce DeepTrans, a deep reasoning translation model that learns free translation via reinforcement learning. Specifically, we carefully build a reward model with pre-defined scoring criteria on both the translation results and the thought process. Given the source sentences, the reward model teaches the deep translation model how to think and free-translate them during reinforcement learning. In this way, training DeepTrans does not need any labeled translations, avoiding the human-intensive annotation or resource-intensive data synthesis. Experimental results show the effectiveness of DeepTrans. Using Qwen2.5-7B as the backbone, DeepTrans improves performance by 16.3% in literature translation, and outperforms strong deep reasoning baselines as well as baselines that are fine-tuned with synthesized data. Moreover, we summarize the failures and interesting findings during our RL exploration. We hope this work could inspire other researchers in free translation.

D2C: Unlocking the Potential of Continuous Autoregressive Image Generation with Discrete Tokens

In the domain of image generation, latent-based generative models occupy a dominant status; however, these models rely heavily on image tokenizer. To meet modeling requirements, autoregressive models possessing the characteristics of scalability and flexibility embrace a discrete-valued tokenizer, but face the challenge of poor image generation quality. In contrast, diffusion models take advantage of the continuous-valued tokenizer to achieve better generation quality but are subject to low efficiency and complexity. The existing hybrid models are mainly to compensate for information loss and simplify the diffusion learning process. The potential of merging discrete-valued and continuous-valued tokens in the field of image generation has not yet been explored. In this paper, we propose D2C, a novel two-stage method to enhance model generation capacity. In the first stage, the discrete-valued tokens representing coarse-grained image features are sampled by employing a small discrete-valued generator. Then in the second stage, the continuous-valued tokens representing fine-grained image features are learned conditioned on the discrete token sequence. In addition, we design two kinds of fusion modules for seamless interaction. On the ImageNet-256 benchmark, extensive experiment results validate that our model achieves superior performance compared with several continuous-valued and discrete-valued generative models on the class-conditional image generation tasks.

LLaVE: Large Language and Vision Embedding Models with Hardness-Weighted Contrastive Learning

Universal multimodal embedding models play a critical role in tasks such as interleaved image-text retrieval, multimodal RAG, and multimodal clustering. However, our empirical results indicate that existing LMM-based embedding models trained with the standard InfoNCE loss exhibit a high degree of overlap in similarity distribution between positive and negative pairs, making it challenging to distinguish hard negative pairs effectively. To deal with this issue, we propose a simple yet effective framework that dynamically improves the embedding model's representation learning for negative pairs based on their discriminative difficulty. Within this framework, we train a series of models, named LLaVE, and evaluate them on the MMEB benchmark, which covers 4 meta-tasks and 36 datasets. Experimental results show that LLaVE establishes stronger baselines that achieve state-of-the-art (SOTA) performance while demonstrating strong scalability and efficiency. Specifically, LLaVE-2B surpasses the previous SOTA 7B models, while LLaVE-7B achieves a further performance improvement of 6.2 points. Although LLaVE is trained on image-text data, it can generalize to text-video retrieval tasks in a zero-shot manner and achieve strong performance, demonstrating its remarkable potential for transfer to other embedding tasks.

DBudgetKV: Dynamic Budget in KV Cache Compression for Ensuring Optimal Performance

To alleviate memory burden during inference of large language models (LLMs), numerous studies have focused on compressing the KV cache by exploring aspects such as attention sparsity. However, these techniques often require a pre-defined cache budget; as the optimal budget varies with different input lengths and task types, it limits their practical deployment accepting open-domain instructions. To address this limitation, we propose a new KV cache compression objective: to always ensure the full-cache performance regardless of specific inputs, while maximizing KV cache pruning as much as possible. To achieve this goal, we introduce a novel KV cache compression method dubbed DBudgetKV, which features an attention-based metric to signal when the remaining KV cache is unlikely to match the full-cache performance, then halting the pruning process. Empirical evaluation spanning diverse context lengths, task types, and model sizes suggests that our method achieves lossless KV pruning effectively and robustly, exceeding 25% compression ratio on average. Furthermore, our method is easy to integrate within LLM inference, not only optimizing memory space, but also showing reduced inference time compared to existing methods.

Warmup-Distill: Bridge the Distribution Mismatch between Teacher and Student before Knowledge Distillation

The widespread deployment of Large Language Models (LLMs) is hindered by the high computational demands, making knowledge distillation (KD) crucial for developing compact smaller ones. However, the conventional KD methods endure the distribution mismatch issue between the teacher and student models, leading to the poor performance of distillation. For instance, the widely-used KL-based methods suffer the mode-averaging and mode-collapsing problems, since the mismatched probabitliy distribution between both models. Previous studies mainly optimize this issue via different distance calculations towards the distribution of both models. Unfortunately, the distribution mismatch issue still exists in the early stage of the distillation. Hence, to reduce the impact of distribution mismatch, we propose a simple yet efficient method, named Warmup-Distill, which aligns the distillation of the student to that of the teacher in advance of distillation. Specifically, we first detect the distribution of the student model in practical scenarios with its internal knowledge, and then modify the knowledge with low probability via the teacher as the checker. Consequently, Warmup-Distill aligns the internal student's knowledge to that of the teacher, which expands the distribution of the student with the teacher's, and assists the student model to learn better in the subsequent distillation. Experiments on the seven benchmarks demonstrate that Warmup-Distill could provide a warmup student more suitable for distillation, which outperforms the vanilla student by as least +0.4 averaged score among all benchmarks. Noteably, with the assistance of Warmup-Distill, the distillation on the math task could yield a further improvement, at most +1.9% accuracy.

LongDPO: Unlock Better Long-form Generation Abilities for LLMs via Critique-augmented Stepwise Information

Long-form generation is crucial for academic writing papers and repo-level code generation. Despite this, current models, including GPT-4o, still exhibit unsatisfactory performance. Existing methods that utilize preference learning with outcome supervision often fail to provide detailed feedback for extended contexts. This shortcoming can lead to content that does not fully satisfy query requirements, resulting in issues like length deviations, and diminished quality. In this paper, we propose enhancing long-form generation by incorporating process supervision. We employ Monte Carlo Tree Search to gather stepwise preference pairs, utilizing a global memory pool to maintain consistency. To address the issue of suboptimal candidate selection, we integrate external critiques to refine and improve the quality of the preference pairs. Finally, we apply step-level DPO using the collected stepwise preference pairs. Experimental results show that our method improves length and quality on long-form generation benchmarks, with almost lossless performance on general benchmarks across various model backbones.

DeepRAG: Thinking to Retrieval Step by Step for Large Language Models

Large Language Models (LLMs) have shown remarkable potential in reasoning while they still suffer from severe factual hallucinations due to timeliness, accuracy, and coverage of parametric knowledge. Meanwhile, integrating reasoning with retrieval-augmented generation (RAG) remains challenging due to ineffective task decomposition and redundant retrieval, which can introduce noise and degrade response quality. In this paper, we propose DeepRAG, a framework that models retrieval-augmented reasoning as a Markov Decision Process (MDP), enabling strategic and adaptive retrieval. By iteratively decomposing queries, DeepRAG dynamically determines whether to retrieve external knowledge or rely on parametric reasoning at each step. Experiments show that DeepRAG improves retrieval efficiency while improving answer accuracy by 21.99%, demonstrating its effectiveness in optimizing retrieval-augmented reasoning.

Personalized Language Model Learning on Text Data Without User Identifiers

In many practical natural language applications, user data are highly sensitive, requiring anonymous uploads of text data from mobile devices to the cloud without user identifiers. However, the absence of user identifiers restricts the ability of cloud-based language models to provide personalized services, which are essential for catering to diverse user needs. The trivial method of replacing an explicit user identifier with a static user embedding as model input still compromises data anonymization. In this work, we propose to let each mobile device maintain a user-specific distribution to dynamically generate user embeddings, thereby breaking the one-to-one mapping between an embedding and a specific user. We further theoretically demonstrate that to prevent the cloud from tracking users via uploaded embeddings, the local distributions of different users should either be derived from a linearly dependent space to avoid identifiability or be close to each other to prevent accurate attribution. Evaluation on both public and industrial datasets using different language models reveals a remarkable improvement in accuracy from incorporating anonymous user embeddings, while preserving real-time inference requirement.