Experience-Driven Dynamic Exits for LLMs with Reinforcement Learning

Large Language Models suffer from slow autoregressive inference. While self-speculative decoding accelerates this process, its efficiency is hampered by static configurations like fixed exit layers and speculation lengths. We reframe this optimization as a \textbf{Markov Decision Process} and propose \textbf{LEDE}, a framework that uses offline reinforcement learning. LEDE learns a policy to dynamically select the optimal exit layer and speculation length based on the local context of the generated sequence at each step, balancing computational cost and draft quality. Comprehensive evaluations on Llama-2 and Llama-3 models show LEDE achieves up to a $2.0\times$$\sim$$2.7\times$ speedup over autoregressive decoding and and provides an additional 17\% speedup over the static speculative baselines.

The Bridge-Garden Dilemma in LLM Distillation: Why Mixing Hard and Soft Labels Works

Knowledge distillation (KD) transfers knowledge from a large teacher model to a smaller student. In language modeling, the student is trained either on tokens sampled from the teacher (hard labels) or the teacher's full next-token distribution (soft labels). Despite soft labels appear strictly richer, we find that mixing hard and soft labels consistently yields better results. Crucially, we show that this gain cannot be explained by closer teacher matching during training. Instead, it comes from reduced exposure bias, the mismatch between training and inference distributions. To explain this phenomenon, we introduce the Bridge-Garden Decomposition theory, which categorizes generation steps into two types: Bridges, where the next token must be exact, and Gardens, where it can be flexible. We show that hard-only KD excels in Bridges by avoiding risky deviations, while soft-only KD preserves diversity in Gardens. A hybrid strategy handles both cases and, as a result, reduces exposure bias across the sequence. Guided by this theory, we develop a family of Bridge-Garden hybrid supervision methods that adaptively balance hard and soft labels. Across a primary suite of seven teacher-student pairs (including Qwen, Llama, Gemma, and DeepSeek) and benchmarks in reasoning and coding, our approach outperforms divergence-based and on-policy KD baselines while reducing training cost by 9.7x, enabling efficient model compression. Code is available at https://github.com/ghwang-s/bridge_garden_hybrid_kd_release.

Mind the Way You Select Negative Texts: Pursuing the Distance Consistency in OOD Detection with VLMs

Out-of-distribution (OOD) detection seeks to identify samples from unknown classes, a critical capability for deploying machine learning models in open-world scenarios. Recent research has demonstrated that Vision-Language Models (VLMs) can effectively leverage their multi-modal representations for OOD detection. However, current methods often incorporate intra-modal distance during OOD detection, such as comparing negative texts with ID labels or comparing test images with image proxies. This design paradigm creates an inherent inconsistency against the inter-modal distance that CLIP-like VLMs are optimized for, potentially leading to suboptimal performance. To address this limitation, we propose InterNeg, a simple yet effective framework that systematically utilizes consistent inter-modal distance enhancement from textual and visual perspectives. From the textual perspective, we devise an inter-modal criterion for selecting negative texts. From the visual perspective, we dynamically identify high-confidence OOD images and invert them into the textual space, generating extra negative text embeddings guided by inter-modal distance. Extensive experiments across multiple benchmarks demonstrate the superiority of our approach. Notably, our InterNeg achieves state-of-the-art performance compared to existing works, with a 3.47\% reduction in FPR95 on the large-scale ImageNet benchmark and a 5.50\% improvement in AUROC on the challenging Near-OOD benchmark.

GMSA: Enhancing Context Compression via Group Merging and Layer Semantic Alignment

Large language models (LLMs) have achieved impressive performance in a variety of natural language processing (NLP) tasks. However, when applied to long-context scenarios, they face two challenges, i.e., low computational efficiency and much redundant information. This paper introduces GMSA, a context compression framework based on the encoder-decoder architecture, which addresses these challenges by reducing input sequence length and redundant information. Structurally, GMSA has two key components: Group Merging and Layer Semantic Alignment (LSA). Group merging is used to effectively and efficiently extract summary vectors from the original context. Layer semantic alignment, on the other hand, aligns the high-level summary vectors with the low-level primary input semantics, thus bridging the semantic gap between different layers. In the training process, GMSA first learns soft tokens that contain complete semantics through autoencoder training. To furtherly adapt GMSA to downstream tasks, we propose Knowledge Extraction Fine-tuning (KEFT) to extract knowledge from the soft tokens for downstream tasks. We train GMSA by randomly sampling the compression rate for each sample in the dataset. Under this condition, GMSA not only significantly outperforms the traditional compression paradigm in context restoration but also achieves stable and significantly faster convergence with only a few encoder layers. In downstream question-answering (QA) tasks, GMSA can achieve approximately a 2x speedup in end-to-end inference while outperforming both the original input prompts and various state-of-the-art (SOTA) methods by a large margin.

OpenworldAUC: Towards Unified Evaluation and Optimization for Open-world Prompt Tuning

Prompt tuning adapts Vision-Language Models like CLIP to open-world tasks with minimal training costs. In this direction, one typical paradigm evaluates model performance separately on known classes (i.e., base domain) and unseen classes (i.e., new domain). However, real-world scenarios require models to handle inputs without prior domain knowledge. This practical challenge has spurred the development of open-world prompt tuning, which demands a unified evaluation of two stages: 1) detecting whether an input belongs to the base or new domain (P1), and 2) classifying the sample into its correct class (P2). What's more, as domain distributions are generally unknown, a proper metric should be insensitive to varying base/new sample ratios (P3). However, we find that current metrics, including HM, overall accuracy, and AUROC, fail to satisfy these three properties simultaneously. To bridge this gap, we propose OpenworldAUC, a unified metric that jointly assesses detection and classification through pairwise instance comparisons. To optimize OpenworldAUC effectively, we introduce Gated Mixture-of-Prompts (GMoP), which employs domain-specific prompts and a gating mechanism to dynamically balance detection and classification. Theoretical guarantees ensure generalization of GMoP under practical conditions. Experiments on 15 benchmarks in open-world scenarios show GMoP achieves SOTA performance on OpenworldAUC and other metrics. We release the code at https://github.com/huacong/OpenworldAUC

ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $α$-$β$-Divergence

Knowledge Distillation (KD) transfers knowledge from a large teacher model to a smaller student model by minimizing the divergence between their output distributions, typically using forward Kullback-Leibler divergence (FKLD) or reverse KLD (RKLD). It has become an effective training paradigm due to the broader supervision information provided by the teacher distribution compared to one-hot labels. We identify that the core challenge in KD lies in balancing two mode-concentration effects: the \textbf{\textit{Hardness-Concentration}} effect, which refers to focusing on modes with large errors, and the \textbf{\textit{Confidence-Concentration}} effect, which refers to focusing on modes with high student confidence. Through an analysis of how probabilities are reassigned during gradient updates, we observe that these two effects are entangled in FKLD and RKLD, but in extreme forms. Specifically, both are too weak in FKLD, causing the student to fail to concentrate on the target class. In contrast, both are too strong in RKLD, causing the student to overly emphasize the target class while ignoring the broader distributional information from the teacher. To address this imbalance, we propose ABKD, a generic framework with $\alpha$-$\beta$-divergence. Our theoretical results show that ABKD offers a smooth interpolation between FKLD and RKLD, achieving an effective trade-off between these effects. Extensive experiments on 17 language/vision datasets with 12 teacher-student settings confirm its efficacy. The code is available at https://github.com/ghwang-s/abkd.

Focal-SAM: Focal Sharpness-Aware Minimization for Long-Tailed Classification

Real-world datasets often follow a long-tailed distribution, making generalization to tail classes difficult. Recent methods resorted to long-tail variants of Sharpness-Aware Minimization (SAM), such as ImbSAM and CC-SAM, to improve generalization by flattening the loss landscape. However, these attempts face a trade-off between computational efficiency and control over the loss landscape. On the one hand, ImbSAM is efficient but offers only coarse control as it excludes head classes from the SAM process. On the other hand, CC-SAM provides fine-grained control through class-dependent perturbations but at the cost of efficiency due to multiple backpropagations. Seeing this dilemma, we introduce Focal-SAM, which assigns different penalties to class-wise sharpness, achieving fine-grained control without extra backpropagations, thus maintaining efficiency. Furthermore, we theoretically analyze Focal-SAM's generalization ability and derive a sharper generalization bound. Extensive experiments on both traditional and foundation models validate the effectiveness of Focal-SAM.

EDGE: Unknown-aware Multi-label Learning by Energy Distribution Gap Expansion

Multi-label Out-Of-Distribution (OOD) detection aims to discriminate the OOD samples from the multi-label In-Distribution (ID) ones. Compared with its multiclass counterpart, it is crucial to model the joint information among classes. To this end, JointEnergy, which is a representative multi-label OOD inference criterion, summarizes the logits of all the classes. However, we find that JointEnergy can produce an imbalance problem in OOD detection, especially when the model lacks enough discrimination ability. Specifically, we find that the samples only related to minority classes tend to be classified as OOD samples due to the ambiguous energy decision boundary. Besides, imbalanced multi-label learning methods, originally designed for ID ones, would not be suitable for OOD detection scenarios, even producing a serious negative transfer effect. In this paper, we resort to auxiliary outlier exposure (OE) and propose an unknown-aware multi-label learning framework to reshape the uncertainty energy space layout. In this framework, the energy score is separately optimized for tail ID samples and unknown samples, and the energy distribution gap between them is expanded, such that the tail ID samples can have a significantly larger energy score than the OOD ones. What's more, a simple yet effective measure is designed to select more informative OE datasets. Finally, comprehensive experimental results on multiple multi-label and OOD datasets reveal the effectiveness of the proposed method.

Suppress Content Shift: Better Diffusion Features via Off-the-Shelf Generation Techniques

Diffusion models are powerful generative models, and this capability can also be applied to discrimination. The inner activations of a pre-trained diffusion model can serve as features for discriminative tasks, namely, diffusion feature. We discover that diffusion feature has been hindered by a hidden yet universal phenomenon that we call content shift. To be specific, there are content differences between features and the input image, such as the exact shape of a certain object. We locate the cause of content shift as one inherent characteristic of diffusion models, which suggests the broad existence of this phenomenon in diffusion feature. Further empirical study also indicates that its negative impact is not negligible even when content shift is not visually perceivable. Hence, we propose to suppress content shift to enhance the overall quality of diffusion features. Specifically, content shift is related to the information drift during the process of recovering an image from the noisy input, pointing out the possibility of turning off-the-shelf generation techniques into tools for content shift suppression. We further propose a practical guideline named GATE to efficiently evaluate the potential benefit of a technique and provide an implementation of our methodology. Despite the simplicity, the proposed approach has achieved superior results on various tasks and datasets, validating its potential as a generic booster for diffusion features. Our code is available at https://github.com/Darkbblue/diffusion-content-shift.

Not All Diffusion Model Activations Have Been Evaluated as Discriminative Features

Diffusion models are initially designed for image generation. Recent research shows that the internal signals within their backbones, named activations, can also serve as dense features for various discriminative tasks such as semantic segmentation. Given numerous activations, selecting a small yet effective subset poses a fundamental problem. To this end, the early study of this field performs a large-scale quantitative comparison of the discriminative ability of the activations. However, we find that many potential activations have not been evaluated, such as the queries and keys used to compute attention scores. Moreover, recent advancements in diffusion architectures bring many new activations, such as those within embedded ViT modules. Both combined, activation selection remains unresolved but overlooked. To tackle this issue, this paper takes a further step with a much broader range of activations evaluated. Considering the significant increase in activations, a full-scale quantitative comparison is no longer operational. Instead, we seek to understand the properties of these activations, such that the activations that are clearly inferior can be filtered out in advance via simple qualitative evaluation. After careful analysis, we discover three properties universal among diffusion models, enabling this study to go beyond specific models. On top of this, we present effective feature selection solutions for several popular diffusion models. Finally, the experiments across multiple discriminative tasks validate the superiority of our method over the SOTA competitors. Our code is available at https://github.com/Darkbblue/generic-diffusion-feature.