ICLR 2025 Papers

Multimodal large language models (MLLMs) have recently shown great progress in text-rich image understanding, yet they still struggle with complex, multi-page visually-rich documents. Traditional methods using document parsers for retrieval-augmented generation suffer from performance and efficiency limitations, while directly presenting all pages to MLLMs leads to inefficiencies, especially with lengthy ones. In this work, we present a novel framework named **S**elf-**V**isual **R**etrieval-**A**ugmented **G**eneration (SV-RAG), which can broaden horizons of *any* MLLM to support long-document understanding. We demonstrate that **MLLMs themselves can be an effective multimodal retriever** to fetch relevant pages and then answer user questions based on these pages. SV-RAG is implemented with two specific MLLM adapters, one for evidence page retrieval and the other for question answering. Empirical results show state-of-the-art performance on public benchmarks, demonstrating the effectiveness of SV-RAG.

Large Language Models (LLMs) generate text by sampling the next token from a probability distribution over the vocabulary at each decoding step. Popular sampling methods like top-p (nucleus sampling) often struggle to balance quality and diversity, especially at higher temperatures which lead to incoherent or repetitive outputs. We propose min-p sampling, a dynamic truncation method that adjusts the sampling threshold based on the model's confidence by using the top token's probability as a scaling factor. Our experiments on benchmarks including GPQA, GSM8K, and AlpacaEval Creative Writing show that min-p sampling improves both the quality and diversity of generated text across different model families (Mistral and Llama 3) and model sizes (1B to 123B parameters), especially at higher temperatures. Human evaluations further show a clear preference for min-p sampling, in both text quality and creativity. Min-p sampling has been adopted by popular open-source LLM frameworks, including Hugging Face Transformers, VLLM, and many others, highlighting its considerable impact on improving text generation quality.

Temporal Knowledge Graph(TKG) serves as an efficacious way to store dynamic facts in real-world. Extrapolation reasoning on TKGs, which aims at predicting possible future events, has attracted consistent research interest. Recently, some rule-based methods have been proposed, which are considered more interpretable compared with embedding-based methods. Existing rule-based methods apply rules through path matching or subgraph extraction, which falls short in inference ability and suffers from missing facts in TKGs. Besides, during rule application period, these methods consider the standing of facts as a binary 0 or 1 problem and ignores the validity as well as frequency of historical facts under temporal settings. In this paper, by designing a novel paradigm for rule application, we propose INFER, a neural-symbolic model for TKG extrapolation. With the introduction of Temporal Validity Function, INFER firstly considers the frequency and validity of historical facts and extends the truth value of facts into continuous real number to better adapt for temporal settings. INFER builds Temporal Weight Matrices with a pre-trained static KG embedding model to enhance its inference ability. Moreover, to facilitates potential integration with existing embedding-based methods, INFER adopts a rule projection module which enables it apply rules through conducting matrices operation on GPU. This feature also improves the efficiency of rule application. Experimental results show that INFER achieves state-of-the-art performance on various TKG datasets and significantly outperforms existing rule-based models on our modified, more sparse TKG datasets, which demonstrates the superiority of our model in inference ability.

Visual instruction datasets from various distributors are released at different times and often contain a significant number of semantically redundant text-image pairs, depending on their task compositions (i.e., skills) or reference sources. This redundancy greatly limits the efficient deployment of continually adaptable multimodal large language models, hindering their ability to refine existing skills and acquire new competencies over time. To address this, we reframe the problem of lifelong Instruction Tuning (LiIT) via data selection, where the model automatically selects beneficial samples to learn from earlier and new datasets based on the current state of acquired knowledge in the model. Based on empirical analyses that show that selecting the best data subset using a static importance measure is often ineffective for multi-task datasets with evolving distributions, we propose Adapt-$\infty$, a new multi-way and adaptive data selection approach that dynamically balances sample efficiency and effectiveness during LiIT. We first construct pseudo-skill clusters by grouping gradient-based sample vectors. Next, we select the best-performing data selector for each skill cluster from a pool of selector experts, including our newly proposed scoring function, Image Grounding score. This data selector samples a subset of the most important samples from each skill cluster for training. To prevent the continuous increase in the size of the dataset pool during LIT, which would result in excessive computation, we further introduce a cluster-wise permanent data pruning strategy to remove the most semantically redundant samples from each cluster, keeping computational requirements manageable. We validate the effectiveness and efficiency of Adapt-$\infty$ over a sequence of various multimodal instruction tuning datasets with various tasks, including (Knowledge) VQA, multilingual, grounding, reasoning, language-only, and multi-image comprehension tasks. Training with samples selected by Adapt-$\infty$ alleviates catastrophic forgetting, especially for rare tasks, and promotes forward transfer across the continuum using only a fraction of the original datasets.

Out-of-Distribution (OOD) detection, aiming to distinguish outliers from known categories, has gained prominence in practical scenarios. Recently, the advent of vision-language models (VLM) has heightened interest in enhancing OOD detection for VLM through few-shot tuning. However, existing methods mainly focus on optimizing global prompts, ignoring refined utilization of local information with regard to outliers. Motivated by this, we freeze global prompts and introduce Local-Prompt, a novel coarse-to-fine tuning paradigm to emphasize regional enhancement with local prompts. Our method comprises two integral components: global prompt guided negative augmentation and local prompt enhanced regional regularization. The former utilizes frozen, coarse global prompts as guiding cues to incorporate negative augmentation, thereby leveraging local outlier knowledge. The latter employs trainable local prompts and a regional regularization to capture local information effectively, aiding in outlier identification. We also propose regional-related metric to empower the enrichment of OOD detection. Moreover, since our approach explores enhancing local prompts only, it can be seamlessly integrated with trained global prompts during inference to boost the performance. Comprehensive experiments demonstrate the effectiveness and potential of our method. Notably, our method reduces average FPR95 by 5.17% against state-of-the-art method in 4-shot tuning on challenging ImageNet-1k dataset, even outperforming 16-shot results of previous methods.

Mamba-based architectures have shown to be a promising new direction for deep learning models owing to their competitive performance and sub-quadratic deployment speed. However, current Mamba multi-modal large language models (MLLM) are insufficient in extracting visual features, leading to imbalanced cross-modal alignment between visual and textural latents, negatively impacting performance on multi-modal tasks. In this work, we propose Empowering Multi-modal Mamba with Structural and Hierarchical Alignment (EMMA), which enables the MLLM to extract fine-grained visual information. Specifically, we propose a pixel-wise alignment module to autoregressively optimize the learning and processing of spatial image-level features along with textual tokens, enabling structural alignment at the image level. In addition, to prevent the degradation of visual information during the cross-model alignment process, we propose a multi-scale feature fusion (MFF) module to combine multi-scale visual features from intermediate layers, enabling hierarchical alignment at the feature level. Extensive experiments are conducted across a variety of multi-modal benchmarks. Our model shows lower latency than other Mamba-based MLLMs and is nearly four times faster than transformer-based MLLMs of similar scale during inference. Due to better cross-modal alignment, our model exhibits lower degrees of hallucination and enhanced sensitivity to visual details, which manifests in superior performance across diverse multi-modal benchmarks. Code provided at https://github.com/xingyifei2016/EMMA.

Multimodal Large Language Models (LLMs) are pivotal in revolutionizing customer support and operations by integrating multiple modalities such as text, images, and audio. Federated Prompt Learning (FPL) is a recently proposed approach that combines pre-trained multimodal LLMs such as vision-language models with federated learning to create personalized, privacy-preserving AI systems. However, balancing the competing goals of personalization, generalization, and privacy remains a significant challenge. Over-personalization can lead to overfitting, reducing generalizability, while stringent privacy measures, such as differential privacy, can hinder both personalization and generalization. In this paper, we propose a Differentially Private Federated Prompt Learning (DP-FPL) approach to tackle this challenge by leveraging a low-rank factorization scheme to capture generalization while maintaining a residual term that preserves expressiveness for personalization. To ensure privacy, we introduce a novel method where we apply local differential privacy to the two low-rank components of the local prompt, and global differential privacy to the global prompt. Our approach mitigates the impact of privacy noise on the model performance while balancing the tradeoff between personalization and generalization. Extensive experiments demonstrate the effectiveness of our approach over other benchmarks.

Large language models (LLMs) have driven significant advancements across diverse NLP tasks, with long-context models gaining prominence for handling extended inputs. However, the expanding key-value (KV) cache size required by Transformer architectures intensifies the memory constraints, particularly during the decoding phase, creating a significant bottleneck. Existing sparse attention mechanisms designed to address this bottleneck have two limitations: (1) they often fail to reliably identify the most relevant tokens for attention, and (2) they overlook the spatial coherence of token selection across consecutive Transformer layers, which can lead to performance degradation and substantial overhead in token selection. This paper introduces TidalDecode, a simple yet effective algorithm and system for fast and accurate LLM decoding through position persistent sparse attention. TidalDecode leverages the spatial coherence of tokens selected by existing sparse attention methods and introduces a few token selection layers that perform full attention to identify the tokens with the highest attention scores, while all other layers perform sparse attention with the pre-selected tokens. This design enables TidalDecode to substantially reduce the overhead of token selection for sparse attention without sacrificing the quality of the generated results. Evaluation on a diverse set of LLMs and tasks shows that TidalDecode closely matches the generative performance of full attention methods while reducing the LLM decoding latency by up to $2.1\times$.

Backdoor unalignment attacks against Large Language Models (LLMs) enable the stealthy compromise of safety alignment using a hidden trigger while evading normal safety auditing. These attacks pose significant threats to the applications of LLMs in the real-world Large Language Model as a Service (LLMaaS) setting, where the deployed model is a fully black-box system that can only interact through text. Furthermore, the sample-dependent nature of the attack target exacerbates the threat. Instead of outputting a fixed label, the backdoored LLM follows the semantics of any malicious command with the hidden trigger, significantly expanding the target space. In this paper, we introduce BEAT, a black-box defense that detects triggered samples during inference to deactivate the backdoor. It is motivated by an intriguing observation (dubbed the **probe concatenate effect**), where concatenated triggered samples significantly reduce the refusal rate of the backdoored LLM towards a malicious probe, while non-triggered samples have little effect. Specifically, BEAT identifies whether an input is triggered by measuring the degree of distortion in the output distribution of the probe before and after concatenation with the input. Our method addresses the challenges of sample-dependent targets from an opposite perspective. It captures the impact of the trigger on the refusal signal (which is sample-independent) instead of sample-specific successful attack behaviors. It overcomes black-box access limitations by using multiple sampling to approximate the output distribution. Extensive experiments are conducted on various backdoor attacks and LLMs (including the closed-source GPT-3.5-turbo), verifying the effectiveness and efficiency of our defense. Besides, we also preliminarily verify that BEAT can effectively defend against popular jailbreak attacks, as they can be regarded as "natural backdoors". Our source code is available at https://github.com/clearloveclearlove/BEAT.

In this paper, we investigate Open-Vocabulary 3D Instance Segmentation (OV-3DIS) with free-form language instructions. Earlier works mainly rely on annotated base categories for training which leads to limited generalization to unseen novel categories. To mitigate the poor generalizability to novel categories, recent works generate class-agnostic masks or projecting generalized masks from 2D to 3D, subsequently classifying them with the assistance of 2D foundation model. However, these works often disregard semantic information in the mask generation, leading to sub-optimal performance. Instead, generating generalizable but semantic-aware masks directly from 3D point clouds would result in superior outcomes. To the end, we introduce Segment any 3D Object with LanguagE ($\textbf{SOLE}$), which is a semantic and geometric-aware visual-language learning framework with strong generalizability by generating semantic-related masks directly from 3D point clouds. Specifically, we propose a multimodal fusion network to incorporate multimodal semantics in both backbone and decoder. In addition, to align the 3D segmentation model with various language instructions and enhance the mask quality, we introduce three types of multimodal associations as supervision. Our SOLE outperforms previous methods by a large margin on ScanNetv2, ScanNet200, and Replica benchmarks, and the results are even closed to the fully-supervised counterpart despite the absence of class annotations in the training. Furthermore, extensive qualitative results demonstrate the versatility of our SOLE to language instructions. The code will be made publicly available.

Vision-language (VL) learning requires extensive visual perception capabilities, such as fine-grained object recognition and spatial perception. Recent works typically rely on training huge models on massive datasets to develop these capabilities. As a more efficient alternative, this paper proposes a new framework that Transfers the knowledge from a hub of Vision Experts (ToVE) for efficient VL learning, leveraging pre-trained vision expert models to promote visual perception capability. Specifically, building on a frozen CLIP image encoder that provides vision tokens for image-conditioned language generation, ToVE introduces a hub of multiple vision experts and a token-aware gating network that dynamically routes expert knowledge to vision tokens. In the transfer phase, we propose a "residual knowledge transfer" strategy, which not only preserves the generalizability of the vision tokens but also allows selective detachment of low-contributing experts to improve inference efficiency. Further, we explore to merge these expert knowledge to a single CLIP encoder, creating a knowledge-merged CLIP that produces more informative vision tokens without expert inference during deployment. Experiment results across various VL tasks demonstrate that the proposed ToVE achieves competitive performance with two orders of magnitude fewer training data.

Tracking and segmenting multiple similar objects with distinct or complex parts in long-term videos is particularly challenging due to the ambiguity in identifying target components and the confusion caused by occlusion, background clutter, and changes in appearance or environment over time. In this paper, we propose a robust video object segmentation framework that learns spatial-semantic features and discriminative object queries to address the above issues. Specifically, we construct a spatial-semantic block comprising a semantic embedding component and a spatial dependency modeling part for associating global semantic features and local spatial features, providing a comprehensive target representation. In addition, we develop a masked cross-attention module to generate object queries that focus on the most discriminative parts of target objects during query propagation, alleviating noise accumulation to ensure effective long-term query propagation. The experimental results show that the proposed method sets new state-of-the-art performance on multiple data sets, including the DAVIS2017 test (\textbf{87.8\%}), YoutubeVOS 2019 (\textbf{88.1\%}), MOSE val (\textbf{74.0\%}), and LVOS test (\textbf{73.0\%}), which demonstrate the effectiveness and generalization capacity of the proposed method. We will make all the source code and trained models publicly available.

Speculative decoding (SD) has emerged as a widely used paradigm to accelerate LLM inference without compromising quality. It works by first employing a compact model to draft multiple tokens efficiently and then using the target LLM to verify them in parallel. While this technique has achieved notable speedups, most existing approaches necessitate either additional parameters or extensive training to construct effective draft models, thereby restricting their applicability across different LLMs and tasks. To address this limitation, we explore a novel plug-and-play SD solution with layer-skipping, which skips intermediate layers of the target LLM as the compact draft model. Our analysis reveals that LLMs exhibit great potential for self-acceleration through layer sparsity and the task-specific nature of this sparsity. Building on these insights, we introduce SWIFT, an on-the-fly self-speculative decoding algorithm that adaptively selects intermediate layers of LLMs to skip during inference. SWIFT does not require auxiliary models or additional training, making it a plug-and-play solution for accelerating LLM inference across diverse input data streams. Our extensive experiments across a wide range of models and downstream tasks demonstrate that SWIFT can achieve over a $1.3\times$$\sim$$1.6\times$ speedup while preserving the original distribution of the generated text. We release our code in https://github.com/hemingkx/SWIFT.

Test-time prompt tuning enhances zero-shot generalization of vision-language models but tends to ignore the relatedness among test samples during inference. Online test-time prompt tuning provides a simple way to leverage the information in previous test samples, albeit with the risk of prompt collapse due to error accumulation. To enhance test-time prompt tuning, we propose DynaPrompt, short for dynamic test-time prompt tuning, exploiting relevant data distribution information while reducing error accumulation. Built on an online prompt buffer, DynaPrompt adaptively selects and optimizes the relevant prompts for each test sample during tuning. Specifically, we introduce a dynamic prompt selection strategy based on two metrics: prediction entropy and probability difference. For unseen test data information, we develop dynamic prompt appending, which allows the buffer to append new prompts and delete the inactive ones. By doing so, the prompts are optimized to exploit beneficial information on specific test data, while alleviating error accumulation. Experiments on fourteen datasets demonstrate the effectiveness of dynamic test-time prompt tuning.

Graphical User Interface (GUI) agents hold great potential for automating complex tasks across diverse digital environments, from web applications to desktop software. However, the development of such agents is hindered by the lack of high-quality, multi-step trajectory data required for effective training. Existing approaches rely on expensive and labor-intensive human annotation, making them unsustainable at scale. To address this challenge, we propose AgentTrek, a scalable data synthesis pipeline that generates high-quality web agent trajectories by leveraging web tutorials. Our method automatically gathers tutorial-like texts from the internet, transforms them into task goals with step-by-step instructions, and employs a visual-language model (VLM) agent to simulate their execution in a real digital environment. A VLM-based evaluator ensures the correctness of the generated trajectories. We demonstrate that training GUI agents with these synthesized trajectories significantly improves their grounding and planning performance over the current models. Moreover, our approach is more cost-efficient compared to traditional human annotation methods. This work underscores the potential of guided replay with web tutorials as a viable strategy for large-scale GUI agent training, paving the way for more capable and autonomous digital agents.

As language models grow ever larger, so do their vocabularies. This has shifted the memory footprint of LLMs during training disproportionately to one single layer: the cross-entropy in the loss computation. Cross-entropy builds up a logit matrix with entries for each pair of input tokens and vocabulary items and, for small models, consumes an order of magnitude more memory than the rest of the LLM combined. We propose Cut Cross-Entropy (CCE), a method that computes the cross-entropy loss without materializing the logits for all tokens into global memory. Rather, CCE only computes the logit for the correct token and evaluates the log-sum-exp over all logits on the fly. We implement a custom kernel that performs the matrix multiplications and the log-sum-exp reduction over the vocabulary in flash memory, making global memory consumption for the cross-entropy computation negligible. This has a dramatic effect. Taking the Gemma 2 (2B) model as an example, CCE reduces the memory footprint of the loss computation from 24 GB to 1 MB, and the total training-time memory consumption of the classifier head from 28 GB to 1 GB. To improve the throughput of CCE, we leverage the inherent sparsity of softmax and propose to skip elements of the gradient computation that have a negligible (i.e. below numerical precision) contribution to the gradient. Experiments demonstrate that the dramatic reduction in memory consumption is accomplished without sacrificing training speed or convergence.

Diffusion transformers have demonstrated remarkable performance in visual generation tasks, such as generating realistic images or videos based on textual instructions. However, larger model sizes and multi-frame processing for video generation lead to increased computational and memory costs, posing challenges for practical deployment on edge devices. Post-Training Quantization (PTQ) is an effective method for reducing memory costs and computational complexity. When quantizing diffusion transformers, we find that existing quantization methods face challenges when applied to text-to-image and video tasks. To address these challenges, we begin by systematically analyzing the source of quantization error and conclude with the unique challenges posed by DiT quantization. Accordingly, we design an improved quantization scheme: ViDiT-Q (**V**ideo \& **I**mage **Di**ffusion **T**ransformer **Q**uantization), tailored specifically for DiT models. We validate the effectiveness of ViDiT-Q across a variety of text-to-image and video models, achieving W8A8 and W4A8 with negligible degradation in visual quality and metrics. Additionally, we implement efficient GPU kernels to achieve practical 2-2.5x memory optimization and a 1.4-1.7x end-to-end latency speedup.

Controlled text generation allows for enforcing user-defined constraints on large language model outputs, an increasingly important field as LLMs become more prevalent in everyday life. One common approach uses energy-based decoding, which defines a target distribution through an energy function that combines multiple constraints into a weighted average. However, these methods often struggle to balance fluency with constraint satisfaction, even with extensive tuning of the energy function's coefficients. In this paper, we identify that this suboptimal balance arises from sampling in continuous space rather than the natural discrete space of text tokens. To address this, we propose \emph{Discrete Auto-regressive Biasing}, a controlled decoding algorithm that leverages gradients while operating entirely in the discrete text domain. Specifically, we introduce a new formulation for controlled text generation by defining a joint distribution over the generated sequence and an auxiliary bias sequence. To efficiently sample from this joint distribution, we propose a Langevin-within-Gibbs sampling algorithm using gradient-based discrete MCMC. Our method significantly improves constraint satisfaction while maintaining comparable or better fluency, all with lower computational costs. We demonstrate the advantages of our controlled decoding method on sentiment control, language detoxification, and keyword-guided generation.

Ensembling strategies for Large Language Models (LLMs) have demonstrated significant potential in improving performance across various tasks by combining the strengths of individual models. However, identifying the most effective ensembling method remains an open challenge, as neither maximizing output consistency through self-consistency decoding nor enhancing model diversity via frameworks like "Mixture of Agents" has proven universally optimal. Motivated by this, we propose a unified framework to examine the trade-offs between task performance, model diversity, and output consistency in ensembles. More specifically, we introduce a consistency score that defines a gating mechanism for mixtures of agents and an algorithm for mixture refinement to investigate these trade-offs at the semantic and model levels, respectively. We incorporate our insights into a novel inference-time LLM ensembling strategy called the Dynamic Mixture of Agents (DMoA) and demonstrate that it achieves a new state-of-the-art result in the challenging Big Bench Hard mixed evaluations benchmark. Our analysis reveals that cross-validation bias can enhance performance, contingent on the expertise of the constituent models. We further demonstrate that distinct reasoning tasks—such as arithmetic reasoning, commonsense reasoning, and instruction following—require different model capabilities, leading to inherent task-dependent trade-offs that DMoA balances effectively.