ICML 2025 Papers

As Large Language Model (LLM) context windows expand, the necessity of Retrieval-Augmented Generation (RAG) for integrating external knowledge is debated. Existing RAG vs. long-context (LC) LLM comparisons are often inconclusive due to benchmark limitations. We introduce LaRA, a novel benchmark with 2326 test cases across four QA tasks and three long context types, for rigorous evaluation. Our analysis of eleven LLMs reveals the optimal choice between RAG and LC depends on a complex interplay of model capabilities, context length, task type, and retrieval characteristics, offering actionable guidelines for practitioners. Our code and dataset is provided at:https://github.com/Alibaba-NLP/LaRA

We introduce "representative generation," extending the theoretical framework for generation proposed by Kleinberg et al. (2024) and formalized by Li et al. (2024), to additionally address diversity and bias concerns in generative models. Our notion requires outputs of a generative model to proportionally represent groups of interest from the training data. We characterize representative uniform and non-uniform generation, introducing the ``group closure dimension'' as a key combinatorial quantity. For representative generation in the limit, we analyze both information-theoretic and computational aspects, demonstrating feasibility for countably infinite hypothesis classes and collections of groups under certain conditions, but proving a negative result for computability using only membership queries. This contrasts with Kleinberg et al.'s (2024) positive results for standard generation in the limit. Our findings provide a rigorous foundation for developing more diverse and representative generative models.

Recently, a plethora of works have proposed inference-time algorithms (e.g. best-of-n), which incorporate verifiers to assist the generation process. Their quality-efficiency trade-offs have been empirically benchmarked on a variety of constrained generation tasks, but the algorithmic design landscape is still largely poorly understood. In this paper, we develop a mathematical framework for reasoning about constrained generation using a pre-trained language model generator oracle and a process verifier—which can decide whether a prefix can be extended to a string which satisfies the constraints of choice. We show that even in very simple settings, access to a verifier can render an intractable problem (information-theoretically or computationally) to a tractable one. In fact, we show even simple algorithms, like tokenwise rejection sampling, can enjoy significant benefits from access to a verifier. Empirically, we show that a natural modification of tokenwise rejection sampling, in which the sampler is allowed to "backtrack" (i.e., erase the final few generated tokens) has robust and substantive benefits over natural baselines (e.g. (blockwise) rejection sampling, nucleus sampling)—both in terms of computational efficiency, accuracy and diversity.

Evaluating in-the-wild coding capabilities of large language models (LLMs) is a challenging endeavor with no existing solution. We introduce Copilot Arena, a platform to collect user preferences through native integration into a developer's working environment. Copilot Arena comprises a novel interface for comparing pairs of model outputs, a sampling strategy to reduce experienced latency, and a prompting scheme to enable code completion functionality. Copilot Arena has served over 4.5 million suggestions from 10 models and collected over 11k pairwise judgements. Our results highlight the importance of model evaluations in integrated settings. We find that model rankings from Copilot Arena differ from those of existing evaluations, which we attribute to the unique distribution of data and tasks contained in Copilot Arena. We also identify novel insights into human preferences on code such as an observed consistency in user preference across programming languages yet significant variation in preference due to task category. We open-source Copilot Arena and release data to enable human-centric evaluations and improve understanding of coding assistants.

Uniform-state discrete diffusion models hold the promise of fast text generation due to their inherent ability to self-correct. However, they are typically outperformed by autoregressive models and masked diffusion models. In this work, we narrow this performance gap by leveraging a key insight: Uniform-state diffusion processes naturally emerge from an underlying Gaussian diffusion. Our method, Duo, transfers powerful techniques from Gaussian diffusion to improve both training and sampling. First, we introduce a curriculum learning strategy guided by the Gaussian process, **doubling training speed** by reducing variance. Models trained with curriculum learning surpass autoregressive models in zero-shot perplexity on 3 of 7 benchmarks. Second, we present Discrete Consistency Distillation, which adapts consistency distillation from the continuous to the discrete setting. This algorithm **unlocks few-step generation in diffusion language models** by accelerating sampling by two orders of magnitude. We provide the code and model checkpoints on the project page: http://s-sahoo.github.io/duo

As large language models (LLMs) play an increasingly important role in code generation, enhancing both correctness and efficiency has become crucial. Current methods primarily focus on correctness, often overlooking efficiency. To address this gap, we introduce SWIFTCODE to improve both aspects by fine-tuning LLMs on a high-quality dataset comprising correct and efficient code samples. Our methodology involves leveraging multiple LLMs to generate diverse candidate code solutions for various tasks across different programming languages. We then evaluate these solutions by directly measuring their execution time and memory usage through local execution. The code solution with the lowest execution time and memory consumption is selected as the final output for each task. Experimental results demonstrate significant improvements when fine-tuning with SWIFTCODE. For instance, Qwen2.5-Coder-7B-Instruct's pass@1 score increases from 44.8\% to 57.7\%, while the average execution time for correct tasks decreases by 48.4\%. SWIFTCODE offers a scalable and effective solution for advancing AI-driven code generation, benefiting both software development and computational problem-solving.

Large Language Models (LLMs) have exhibited exceptional performance in software engineering yet face challenges in adapting to continually evolving code knowledge, particularly the frequent updates of third-party library APIs. This limitation, rooted in the static pre-training datasets, often results in non-executable code or implementations with suboptimal safety and efficiency. To this end, we introduce CodeSync, a data engine to identify outdated code patterns and collect real-time code knowledge updates from Python third-party libraries. Building upon CodeSync, we develop CodeSyncBench, a comprehensive benchmark for assessing LLMs' ability to stay synchronized with code evolution, which covers real-world updates for 220 APIs from six Python libraries. Our benchmark offers 3,300 test cases spanning three evaluation tasks and an update-aware instruction tuning dataset of 2,200 training samples. Extensive experiments on 14 LLMs reveal that they struggle with dynamic code evolution, even with the support of advanced knowledge updating methods (e.g., DPO, ORPO, and SimPO). Our CodeSync lays a strong foundation for developing more effective and robust methods for real-time code knowledge updating in the future. The experimental code is available at: https://github.com/CGCL-codes/naturalcc/tree/main/examples/codesync.

Large Language Models (LLMs) have shown remarkable capabilities, yet ensuring their outputs conform to strict structural or grammatical constraints remains challenging, which is critical in function calls and domain-specific language (DSL) generation. Constrained decoding with context-free grammar is a flexible approach to guarantee LLMs' adherence to a specific format by dynamically building a token logits mask. However, creating this mask requires checking the validity of all tokens in the LLM vocabulary at every decoding step, which often incurs significant overheads in existing constrained decoding engines. To address this challenge, we propose $\textbf{ZapFormat}$, a novel $\textbf{dynamic pruning}$ strategy based on the Earley algorithm that identifies and eliminates invalid or redundant Earley states in real-time, significantly reducing memory occupation of the Earley algorithm's states. This further enables us to use a state cache to speed up structured generations on a large number of queries. We implemented ZapFormat in a new constrained decoding engine called Formatron which also incorporates existing optimizations. Through comprehensive experiments on structured generation tasks, including JSON generation, JSON Schema validation, and semantic parsing, we demonstrate that Formatron not only $\textbf{consistently maintains}$ high-precision compliant outputs but also achieves $\textbf{significant improvements}$ in inference speed up to 2x compared to state-of-the-art implementations. More importantly, Formatron is generally applicable across various LLM architectures. We release Formatron as open source at https://github.com/Dan-wanna-M/formatron.

We propose a data-driven approach to analyzing query complexity in Video Question Answering (VideoQA). Previous efforts in benchmark design have relied on human expertise to design challenging questions, yet we experimentally show that humans struggle to predict which questions are difficult for machine learning models. Our automatic approach leverages recent advances in code generation for visual question answering, using the complexity of generated code as a proxy for question difficulty. We demonstrate that this measure correlates significantly better with model performance than human estimates. To operationalize this insight, we propose an algorithm for estimating question complexity from code. It identifies fine-grained primitives that correlate with the hardest questions for any given set of models, making it easy to scale to new approaches in the future. Finally, to further illustrate the utility of our method, we extend it to automatically generate complex questions, constructing a new benchmark that is 1.9 times harder than the popular NExT-QA.

Automated architecture design methods, especially neural architecture search, have attracted increasing attention. However, these methods naturally need to evaluate numerous candidate architectures during the search process, thus computationally extensive and time-consuming. In this paper, we propose a prior knowledge guided neural architecture generation method to generate high-performance architectures without any search and evaluation process. Specifically, in order to identify valuable prior knowledge for architecture generation, we first quantify the contribution of each component within an architecture to its overall performance. Subsequently, a diffusion model guided by prior knowledge is presented, which can easily generate high-performance architectures for different computation tasks. Extensive experiments on new search spaces demonstrate that our method achieves superior accuracy over state-of-the-art methods. For example, we only need $0.004$ GPU Days to generate architecture with $76.1\%$ top-1 accuracy on ImageNet and $97.56\%$ on CIFAR-10. Furthermore, we can find competitive architecture for more unseen search spaces, such as TransNAS-Bench-101 and NATS-Bench, which demonstrates the broad applicability of the proposed method.

We consider molecule generation in 3D space using language models (LMs), which requires discrete tokenization of 3D molecular geometries. Although tokenization of molecular graphs exists, that for 3D geometries is largely unexplored. Here, we attempt to bridge this gap by proposing a novel method which converts molecular geometries into SE(3)-invariant 1D discrete sequences. Our method consists of canonical labeling and invariant spherical representation steps, which together maintain geometric and atomic fidelity in a format conducive to LMs. Our experiments show that, when coupled with our proposed method, various LMs excel in molecular geometry generation, especially in controlled generation tasks. Our code has been released as part of the AIRS library (https://github.com/divelab/AIRS/).

We consider the generative modeling of speech over multiple minutes, a requirement for long-form multimedia generation and audio-native voice assistants. However, textless spoken language models struggle to generate plausible speech past tens of seconds, due to high temporal resolution of speech tokens causing loss of coherence, architectural issues with long-sequence training or extrapolation, and memory costs at inference time. From these considerations we derive **SpeechSSM**, the first speech language model family to learn from and sample long-form spoken audio (e.g., 16 minutes of read or extemporaneous speech) in a single decoding session without text intermediates. SpeechSSMs leverage recent advances in linear-time sequence modeling to greatly surpass current Transformer spoken LMs in coherence and efficiency on multi-minute generations while still matching them at the utterance level. As we found current spoken language evaluations uninformative, especially in this new long-form setting, we also introduce: **LibriSpeech-Long**, a benchmark for long-form speech evaluation; new embedding-based and LLM-judged metrics; and quality measurements over length and time. Speech samples, the LibriSpeech-Long dataset, and any future code or model releases can be found at https://google.github.io/tacotron/publications/speechssm/.

The rapid advancement of code large language models (Code LLMs) underscores the critical need for effective and transparent benchmarking methods. However, current benchmarking predominantly relies on publicly available, human-created datasets. The widespread use of these static benchmark datasets makes the evaluation process particularly susceptible to data contamination—an unavoidable consequence of the extensive data collection processes employed during LLM training. Existing methods for addressing data contamination typically face significant limitations, including reliance on substantial human effort and difficulty in managing class imbalances. To overcome these challenges, we propose DyCodeEval, a novel benchmarking suite specifically designed to evaluate Code LLMs under realistic contamination scenarios. Given an initial seed programming problem, DyCodeEval utilizes multiple agents to systematically extract and modify contextual information without changing the core logic, generating semantically equivalent variations. We introduce a dynamic data generation method and conduct extensive empirical studies on two seed datasets involving 18 Code LLMs. The results demonstrate that DyCodeEval effectively assesses the reasoning capabilities of Code LLMs under contamination conditions while producing diverse problem variants, thereby ensuring robust and consistent benchmarking outcomes.

Software engineering activities frequently involve edits to existing code. However, contemporary code language models (LMs) lack the ability to handle diverse types of code-edit requirements. In this work, we attempt to overcome this shortcoming through (1) a novel synthetic data generation pipeline and (2) a robust model adaptation algorithm. Starting with seed code examples and diverse editing criteria, our pipeline generates high-quality samples comprising original and modified code, along with natural language instructions in different styles and verbosity. Today's code LMs come bundled with strong abilities, such as code generation and instruction following, which should not be lost due to fine-tuning. To ensure this, we propose a novel adaptation algorithm, SeleKT, that (a) leverages a dense gradient-based step to identify the weights that are most important for code editing, and (b) does a sparse projection onto the base model to avoid overfitting. Using our approach, we obtain a new series of models NextCoder (adapted from QwenCoder-2.5) that achieves strong results on five code-editing benchmarks, outperforming comparable size models and even several larger ones. We show the generality of our approach on two model families DeepSeekCoder and QwenCoder), compare against other fine-tuning approaches, and demonstrate robustness by showing retention of code generation and general problem-solving abilities post adaptation. We opensource the models, synthetic dataset, and implementation at http://aka.ms/nextcoder.

Curating data for instruction tuning is crucial for enhancing the performance of large language models (LLMs). This work aims to select training data for instruction tuning to improve the LLM performance on specific tasks. Existing methods often rely on next-token prediction (NTP) loss as a proxy for target task performance due to the non-differentiable nature of performance evaluation metrics. They select training data points that are most helpful in reducing validation loss. However, there is a discrepancy between minimizing NTP loss and maximizing performance (e.g., code pass rate in code generation). To remedy this, we introduce a novel Non-differentiable evaluation metric-based InfluenCe Estimation (NICE), which leverages the policy gradient to select the training data that improves the performance. Moreover, NICE can perform data selection in the absence of labels (ground-truth responses) when the evaluation metrics do not require labels (e.g., a reward model can output reward scores without supervision from labels). Experimental results show that our approach outperforms existing data selection baselines that use NTP loss in diverse and realistic scenarios. Notably, subsets selected by NICE often produce models that outperform those trained on the full dataset. Our code is available at https://github.com/JTWang2000/NICE.

Pruning has recently been widely adopted to reduce the parameter scale and improve the inference efficiency of Large Language Models (LLMs). Mainstream pruning techniques often rely on uniform layerwise pruning strategies, which can lead to severe performance degradation at high sparsity levels. Recognizing the varying contributions of different layers in LLMs, recent studies have shifted their focus toward non-uniform layerwise pruning. However, these approaches often rely on pre-defined values, which can result in suboptimal performance. To overcome these limitations, we propose a novel method called Dynamic Layerwise Pruning (DLP). This approach adaptively determines the relative importance of each layer by integrating model weights with input activation information, assigning pruning rates accordingly. Experimental results show that DLP effectively preserves model performance at high sparsity levels across multiple LLMs. Specifically, at 70% sparsity, DLP reduces the perplexity of LLaMA2-7B by 7.79 and improves the average accuracy by 2.7% compared to state-of-the-art methods. Moreover, DLP is compatible with various existing LLM compression techniques and can be seamlessly integrated into Parameter-Efficient Fine-Tuning (PEFT). We release the code\footnote{The code is available at: \url{https://github.com/ironartisan/DLP}.} to facilitate future research.

Speculative decoding (SPD) aims to accelerate the auto-regressive token generation process of a target Large Language Model (LLM). Some approaches employ a draft model with multiple heads to predict a sequence of future tokens, where each head handles a token in the sequence. The target LLM verifies the predicted sequence and accepts aligned tokens, enabling efficient multi-token generation. However, existing methods assume that all tokens within a sequence are equally important, employing identical head structures and relying on a single-generation paradigm, either serial or parallel. To this end, we theoretically demonstrate that initial tokens in the draft sequence are more important than later ones. Building on this insight, we propose Gumiho, a hybrid model combining serial and parallel heads. Specifically, given the critical importance of early tokens, we employ a sophisticated Transformer architecture for the early draft heads in a serial configuration to improve accuracy. For later tokens, we utilize multiple lightweight MLP heads operating in parallel to enhance efficiency. By allocating more advanced model structures and longer running times to the early heads, Gumiho achieves improved overall performance. The experimental results demonstrate that our method outperforms existing approaches, fully validating its effectiveness. Our code is available at https://github.com/AMD-AIG-AIMA/Gumiho.

Unsupervised domain adaptation (UDA) for time series data remains a critical challenge in deep learning, with traditional pseudo-labeling strategies failing to capture temporal patterns and channel-wise shifts between domains, producing sub-optimal pseudo labels. As such, we introduce TransPL, a novel approach that addresses these limitations by modeling the joint distribution $P(X,y)$ of the source domain through code transition matrices, where the codes are derived from vector quantization (VQ) of time series patches. Our method constructs class- and channel-wise code transition matrices from the source domain and employs Bayes' rule for target domain adaptation, generating pseudo-labels based on channel-wise weighted class-conditional likelihoods. TransPL offers three key advantages: explicit modeling of temporal transitions and channel-wise shifts between different domains, versatility towards different UDA scenarios (e.g., weakly-supervised UDA), and explainable pseudo-label generation. We validate TransPL's effectiveness through extensive analysis on four time series UDA benchmarks and confirm that it consistently outperforms state-of-the-art pseudo-labeling methods by a strong margin (6.1\% accuracy improvement, 4.9\% F1 improvement), while providing interpretable insights into the domain adaptation process through its learned code transition matrices.