ICML 2025 Papers

State Space Models (SSMs) are gaining attention as an efficient alternative to Transformers due to their constant memory complexity and comparable performance. Yet, deploying large-scale SSMs on cloud-based services or resource-constrained devices faces challenges. To address this, quantizing SSMs using low bit-width data types is proposed to reduce model size and leverage hardware acceleration. Given that SSMs are sensitive to quantization errors, recent advancements focus on quantizing a specific model or bit-width to improve their efficiency while maintaining performance. However, different bit-width configurations, such as W4A8 for cloud service throughput and W4A16 for improving question-answering on personal devices, are necessary for specific scenarios. To this end, we present Quamba2, compatible with \textbf{W8A8}, \textbf{W4A8}, and \textbf{W4A16} for both \textbf{Mamba} and \textbf{Mamba2}, addressing the rising demand for SSM deployment across various platforms. We propose an offline approach to quantize inputs of a linear recurrence in 8-bit by sorting and clustering for $x$, combined with a per-state-group quantization for $B$ and $C$. To ensure compute-invariance in the SSM output, we offline rearrange weights according to the clustering sequence. The experiments show Quamba2-8B outperforms several state-of-the-art SSMs quantization methods and delivers 1.3$\times$ and 3$\times$ speedup in the pre-filling and generation stages and 4$\times$ memory reduction with only a $1.6$% accuracy drop on average. The code and quantized models will be released at:

Decoding with autoregressive language models traditionally occurs sequentially, generating one token after another. Recent attempts to introduce parallelism require a pre-determined structure in the generated content to implement parallel generation, such as by pattern-matching on bullet points. In this work, we present a new technique to automate parallel generation by dynamically exploiting the semantic independence of generation outputs to implement asynchronous decoding. We introduce an annotation language Pasta-Lang for language models to initiate asynchronous decoding at inference time. We also develop an accompanying Pasta-Lang interpreter that performs on-the-fly asynchronous decoding, effectively implementing parallel generation and speeding up inference. We present an instruction-finetuning dataset with Pasta-Lang-annotated responses for teaching LLMs to annotate semantic independence with Pasta-Lang as well as the methodology for creating the dataset. Our evaluation shows using the interpreter with a Pasta-Lang-equipped model achieves significant speedup while maintaining the same generation quality.

Extending the context length of Language Models (LMs) by improving Rotary Position Embedding (RoPE) has become a trend. While prior works mainly address RoPE's limitations within attention, this paper uncovers the adverse effects on length generalization from nearly all parts of LMs. Using *Discrete Signal Processing* theory, we show that RoPE enables periodic attention by implicitly achieving *Non-Uniform Discrete Fourier Transform*. However, this periodicity is undermined by the spectrum damage caused by: 1) linear layers and activation functions outside of attention; 2) insufficiently trained frequency components brought by time-domain truncation. Building on our observations, we propose ***Fourier Position Embedding (FoPE)***, which enhances attention's frequency-domain properties to improve both its periodic extension and length generalization. FoPE constructs *Fourier Series* and zero-outs the destructive frequency components, increasing model robustness against the spectrum damage. Experiments across various model scales and benchmarks show that, within varying context windows, FoPE maintains a more stable performance compared to other baselines. Several analyses and ablations bring further support to our method and theoretical modeling.

Post-training quantization is a key technique for reducing the memory and inference latency of large language models by quantizing weights and activations without requiring retraining. However, existing methods either (1) fail to account for the varying importance of hidden features to the end loss or, when incorporating end loss, (2) neglect the critical interactions between model weights. To address these limitations, we propose GuidedQuant, a novel quantization approach that integrates gradient information from the end loss into the quantization objective while preserving cross-weight dependencies within output channels. GuidedQuant consistently boosts the performance of state-of-the-art quantization methods across weight-only scalar, weight-only vector, and weight-and-activation quantization. Additionally, we introduce a novel non-uniform scalar quantization algorithm, which is guaranteed to monotonically decrease the quantization objective value, and outperforms existing methods in this category. We release the code at https://github.com/snu-mllab/GuidedQuant.

Large Language Models (LLMs) are vulnerable to backdoor attacks that manipulate outputs via hidden triggers. Existing defense methods—designed for vision/text classification tasks—fail for text generation. We propose *Internal Consistency Regularization (CROW)*, a defense leveraging the observation that backdoored models exhibit unstable layer-wise hidden representations when triggered, while clean models show smooth transitions. CROW enforces consistency across layers via adversarial perturbations and regularization during finetuning, neutralizing backdoors without requiring clean reference models or trigger knowledge—only a small clean dataset. Experiments across Llama-2 (7B, 13B), CodeLlama (7B, 13B), and Mistral-7B demonstrate CROW’s effectiveness: it achieves significant reductions in attack success rates across diverse backdoor strategies (sentiment steering, targeted refusal, code injection) while preserving generative performance. CROW’s architecture-agnostic design enables practical deployment.

We introduce ConfPO, a method for preference learning in Large Language Models (LLMs) that identifies and optimizes preference-critical tokens based solely on the training policy's confidence, without requiring any auxiliary models or compute. Unlike prior Direct Alignment Algorithms (DAAs) such as Direct Preference Optimization (DPO), which uniformly adjust all token probabilities regardless of their relevance to preference, ConfPO focuses optimization on the most impactful tokens. This targeted approach improves alignment quality while mitigating overoptimization (i.e., reward hacking) by using the KL divergence budget more efficiently. In contrast to recent token-level methods that rely on credit-assignment models or AI annotators, raising concerns about scalability and reliability, ConfPO is simple, lightweight, and model-free. Experimental results on challenging alignment benchmarks, including AlpacaEval 2 and Arena-Hard, demonstrate that ConfPO consistently outperforms uniform DAAs across various LLMs, delivering better alignment with zero additional computational overhead.

Pretraining large language models (LLMs) is resource-intensive, often requiring months of training time even with high-end GPU clusters. There are two approaches of mitigating such computational demands: reusing smaller models to train larger ones (upcycling), and training computationally efficient models like mixture-of-experts (MoE). In this paper, we study the upcycling of LLMs to MoE models, of which the scaling behavior remains underexplored. Through extensive experiments, we identify empirical scaling laws that describe how performance depends on dataset size and model configuration. Particularly, we show that, while scaling these factors improves performance, there is a novel interaction term between the dense and upcycled training dataset that limits the efficiency of upcycling at large computational budgets. Based on these findings, we provide guidance to scale upcycling, and establish conditions under which upcycling outperforms from-scratch trainings within budget constraints.

Decoder-only models generate tokens autoregressively by caching key/value vectors, but as the cache grows, inference becomes memory-bounded. To address this challenge, we introduce CLOVER (Cross-Layer Orthogonal Vectors) pruning, a novel approach that treats pairs of components of the attention mechanism as low-rank decompositions. CLOVER applies Singular Value Decomposition (SVD) to the Q-K and V-O pairs within each attention head. The resulting singular values, in turn, guide pruning and further serve as trainable parameters for efficient fine-tuning, ultimately enabling the model to recover its performance to the level before pruning.After pruning and fine-tuning, these values are reintegrated into the model without increasing its parameter count. Visualizations across various models show that CLOVER effectively removes linear redundancies within attention heads, greatly improving pruning efficiency. For example, pruning 70\% of the Q-K head dimension in GPT-2 XL results in a perplexity comparable to that of pruning just 8\% using vanilla pruning. The combination of CLOVER and TransMLA achieves a speedup of up to 11.1$\times$ over LLaMA-2-7B.

The rapid development of Multimodal Large Language Models (MLLMs), such as GPT-4, marks a significant step toward artificial general intelligence. Existing methods typically align vision encoders with LLMs via supervised fine-tuning (SFT), but this often deteriorates their ability to handle multiple languages as training progresses. We empirically observe that imbalanced SFT datasets, largely English-centric, degrade performance on non-English languages due to the failure in multilingual token alignment. To address this, we propose Parrot, a novel approach that leverages textual guidance for visual token alignment at the language level. Parrot conditions visual tokens on diverse language inputs and uses Mixture-of-Experts (MoE) to align multilingual tokens. By computing cross-attention between initial visual features and textual embeddings, we select the most relevant experts, converting visual tokens into language-specific representations. Additionally, we introduce the Massive Multilingual Multimodal Benchmark (MMMB), a new benchmark comprising 6 languages, 15 categories, and 12,000 questions, to assess multilingual capabilities. Parrot achieves state-of-the-art performance on both the multilingual benchmarks and a wide range of multimodal tasks. Code and dataset are available at: \url{https://github.com/AIDC-AI/Parrot}.

We present a perception in reflection paradigm designed to transcend the limitations of current large vision-language models (LVLMs), which are expected yet often fail to achieve perfect perception initially. Specifically, we propose Reflective Perception (RePer), a dual-model reflection mechanism that systematically alternates between policy and critic models, enables iterative refinement of visual perception. This framework is powered by Reflective Perceptual Learning (RPL), which reinforces intrinsic reflective capabilities through a methodically constructed visual reflection dataset and reflective unlikelihood training Comprehensive experimental evaluation demonstrates RePer's quantifiable improvements in image understanding, captioning precision, and hallucination reduction. Notably, RePer achieves strong alignment between model attention patterns and human visual focus, while RPL optimizes fine-grained and free-form preference alignment. These advancements establish perception in reflection as a robust paradigm for future multimodal agents, particularly in tasks requiring complex reasoning and multi-step manipulation. Project Page: [https://weiyana.github.io/Perception-in-Reflection](https://weiyana.github.io/Perception-in-Reflection)

While recent code-specific large language models (LLMs) have greatly enhanced their code generation capabilities, the safety of these models remains under-explored, posing potential risks as insecure code generated by these models may introduce vulnerabilities into real-world systems. Existing methods collect security-focused datasets from real-world vulnerabilities for instruction tuning in order to mitigate such issues. However, they are largely constrained by the data sparsity of vulnerable code, and have limited applicability in the multi-stage post-training workflows of modern LLMs. In this paper, we propose ProSec, a novel proactive security alignment approach designed to align code LLMs with secure coding practices. ProSec systematically exposes the vulnerabilities in a code LLM by synthesizing vulnerability-inducing coding scenarios from Common Weakness Enumerations (CWEs) and generates fixes to vulnerable code snippets, allowing the model to learn secure practices through preference learning objectives. The scenarios synthesized by ProSec trigger 25× more vulnerable code than a normal instruction-tuning dataset, resulting in a security-focused alignment dataset 7× larger than the previous work. Experiments show that models trained with ProSec are 25.2% to 35.4% more secure compared to previous work without degrading models' utility.

We introduce Lexico, a novel KV cache compression method that leverages sparse coding with a universal dictionary. Our key finding is that key-value cache in modern LLMs can be accurately approximated using sparse linear combination from a small, input-agnostic dictionary of ~4k atoms, enabling efficient compression across different input prompts, tasks and models. Using orthogonal matching pursuit for sparse approximation, Lexico achieves flexible compression ratios through direct sparsity control. On GSM8K, across multiple model families (Mistral, Llama 3, Qwen2.5), Lexico maintains 90-95% of the original performance while using only 15-25% of the full KV-cache memory, outperforming both quantization and token eviction methods. Notably, Lexico remains effective in low memory regimes where 2-bit quantization fails, achieving up to 1.7x better compression on LongBench and GSM8K while maintaining high accuracy.

In-context learning with large language models (LLMs) delivers strong few-shot performance by choosing few-shot demonstrations from the entire training dataset. However, previous few-shot in-context learning methods, which calculate similarity scores for choosing demonstrations, incur high computational costs by repeatedly retrieving large-scale datasets for each query. This is due to their failure to recognize that not all demonstrations are equally informative, and many less informative demonstrations can be inferred from a core set of highly informative ones. To this end, we propose FEEDER (FEw yet Essential Demonstration prE-selectoR), a novel \emph{pre-selection} framework that identifies a core subset of demonstrations containing the most informative examples. This subset, referred to as the FEEDER set, consists of demonstrations that capture both the ''sufficiency'' and ''necessity'' information to infer the entire dataset. Notice that FEEDER is selected before the few-shot in-context learning, enabling more efficient few-shot demonstrations choosing in a smaller set. To identify FEEDER, we propose a novel effective tree based algorithm. Once selected, it can replace the original dataset, leading to improved efficiency and prediction accuracy in few-shot in-context learning. Additionally, FEEDER also benefit fine-tuning LLMs, we propose a bi-level optimization method enabling more efficient training without sacrificing performance when datasets become smaller. Our experiments are on 6 text classification datasets, 1 reasoning dataset, and 1 semantic-parsing dataset, across 6 LLMs (ranging from 335M to 7B parameters), demonstrate that: (i) In few-shot inference, FEEDER achieves superior (or comparable) performance while utilizing only half the input training data. (ii) In fine-tuning, FEEDER significantly boosts the performance of LLMs.

Large language models are pre-trained on ever-growing token budgets under the assumption that better pre-training performance translates to improved downstream models. In this work, we challenge this assumption and show that extended pre-training can make models harder to fine-tune, leading to degraded final performance. We term this phenomenon \textbf{catastrophic overtraining}. For example, the instruction-tuned OLMo-1B model pre-trained on 3T tokens leads to over 2\% worse performance on multiple standard LLM benchmarks than its 2.3T token counterpart. Through controlled experiments and theoretical analysis, we show that catastrophic overtraining arises from a systematic increase in the broad sensitivity of pre-trained parameters to modifications, including but not limited to fine-tuning. Our findings call for a critical reassessment of pre-training design that considers the downstream adaptability of the model.

The rapidly increasing size of large language models (LLMs) presents significant challenges in memory usage and computational costs. Quantizing both weights and activations can address these issues, with hardware-supported fine-grained scaling emerging as a promising solution to mitigate outliers. However, existing methods struggle to capture nuanced block data distributions. We propose BlockDialect, a block-wise fine-grained mixed format technique that assigns a per-block optimal number format from a formatbook for better data representation. Additionally, we introduce DialectFP4, a formatbook of FP4 variants (akin to dialects) that adapt to diverse data distributions. To leverage this efficiently, we propose a two-stage approach for online DialectFP4 activation quantization. Importantly, DialectFP4 ensures energy efficiency by selecting representable values as scaled integers compatible with low-precision integer arithmetic. BlockDialect achieves 10.78% (7.48%) accuracy gain on the LLaMA3-8B (LLaMA2-7B) model compared to MXFP4 format with lower bit usage per data, while being only 5.45% (2.69%) below full precision even when quantizing full-path matrix multiplication. Focusing on how to represent over how to scale, our work presents a promising path for energy-efficient LLM inference.

The sample efficiency of Bayesian optimization algorithms depends on carefully crafted acquisition functions (AFs) guiding the sequential collection of function evaluations. The best-performing AFs can vary significantly across optimization problems, often requiring ad-hoc and problem-specific choices. This work tackles the challenge of designing novel AFs that perform well across a variety of experimental settings. Based on FunSearch, a recent work using Large Language Models (LLMs) for discovery in mathematical sciences, we propose FunBO, an LLM-based method that can be used to learn new AFs written in computer code by leveraging access to a number of evaluations for a limited set of objective functions. We provide the analytic expression of all discovered AFs and evaluate them on various global optimization benchmarks and hyperparameter optimization tasks. We show how FunBO identifies AFs that generalize well both in and out of the training distribution of functions, thus outperforming established general-purpose AFs and achieving competitive performance against AFs that are customized to specific function types and are learned via transfer-learning algorithms.

Recent studies have shown that supervised fine-tuning of LLMs on a small number of high-quality datasets can yield strong reasoning capabilities. However, full fine-tuning (Full FT), while powerful, is computationally expensive and susceptible to overfitting and catastrophic forgetting, particularly when data is limited. Sparse fine-tuning, which previously achieved notable success by updating only a small subset of model parameters, offers a promising trade-off between efficiency and effectiveness. Yet, it has lagged behind in the LLM era due to the difficulty of identifying parameters truly critical for reasoning. In this work, we state that weights with the largest magnitude after low-rank approximation are critical weights for fine-tuning, which we call *Principal Weights*. Surprisingly, while magnitude-based sparse fine-tuning performs poorly as a baseline on LLM fine-tuning, it becomes highly effective after rank reduction. These insights motivate our method: **L**ow-rank **I**nformed Sparse **F**ine-**T**uning ($\texttt{LIFT}$). $\texttt{LIFT}$ only updates the top 5% *Principal Weights* throughout training and consistently achieves better performance on reasoning tasks than Full FT, while maintaining memory efficiency on par with popular parameter-efficient fine-tuning methods. In addition to strong performance on target domains such as arithmetic reasoning, $\texttt{LIFT}$ also retains up to 20% more source-domain knowledge, compared to Full FT and LoRA. Our code is available at: https://github.com/zihanghliu/LIFT.

Offline preference alignment for language models such as Direct Preference Optimization (DPO) is favored for its effectiveness and simplicity, eliminating the need for costly reinforcement learning. Various offline algorithms have been developed for different data settings, yet they lack a unified understanding. In this study, we introduce Pior-Informed Preference Alignment (PIPA), a unified, RL-free probabilistic framework that formulates language model preference alignment as a Maximum Likelihood Estimation (MLE) problem with prior constraints. This method effectively accommodates both paired and unpaired data, as well as answer and step-level annotations. We illustrate that DPO and KTO are special cases with different prior constraints within our framework. By integrating different types of prior information, we developed two variations of PIPA: PIPA-M and PIPA-N. Both algorithms demonstrate a $3\sim10\%$ performance enhancement on the GSM8K and MATH benchmarks across all configurations, achieving these gains without additional training or computational costs compared to existing algorithms.

Large Language Models (LLMs) often excel in specific domains but fall short in others due to the limitations of their training. Thus, enabling LLMs to solve problems collaboratively by integrating their complementary knowledge promises to improve their performance across domains. To realize this potential, we introduce a novel Collaborative Speculative Decoding (CoSD) algorithm that enables efficient LLM knowledge fusion at test time without requiring additional model training. CoSD employs a draft model to generate initial sequences and an easy-to-learn rule or decision tree to decide when to invoke an assistant model to improve these drafts. CoSD not only enhances knowledge fusion but also improves inference efficiency, is transferable across domains, and offers greater explainability. Experimental results demonstrate that CoSD improves accuracy by up to 10% across benchmarks compared to existing methods, providing a scalable and effective solution for LLM-based applications. Our code has been released at https://github.com/ATP-1010/CoSD.