ICML 2025 Papers

The Bradley-Terry (BT) model is widely practiced in reward modeling for reinforcement learning with human feedback (RLHF). Despite its effectiveness, reward models (RMs) trained with BT model loss as one-way classifiers are prone to over-optimization, losing generalizability to unseen inputs. In this paper, we study the cause of over-optimization and its downstream effects on the RLHF procedure, highlighting the importance of robustness in RMs. First, we show that the excessive dispersion of hidden state norms is the main source of over-optimization. Correspondingly, we propose batch-wise sum-to-zero regularization (BSR) that enforces reward sum for each batch to be zero-centered, constraining the rewards with abnormally large magnitudes. We assess the impact of BSR in improving robustness in RMs through four scenarios of over-optimization, where BSR consistently manifests better robustness on unseen inputs. Then, we compare the plain BT model and BSR on RLHF training and empirically show that robust RMs better align the policy to the gold preference model. Finally, we apply BSR to high-quality data and models, which surpasses state-of-the-art RMs in the 8B scale by adding more than 5\% in complex preference prediction tasks. By conducting RLOO training with 8B RMs, AlpacaEval 2.0, with reducing generation length by 40\% while adding a 7\% increase in win rate, further highlights that robustness in RMs induces robustness in RLHF training.

For question-answering (QA) tasks, in-context learning (ICL) enables language models (LMs) to generate responses without modifying their parameters by leveraging examples provided in the input. However, the effectiveness of ICL heavily depends on the availability of high-quality examples, which are often scarce due to data privacy constraints, annotation costs, and distribution disparities. A natural solution is to utilize examples stored on client devices, but existing approaches either require transmitting model parameters—incurring significant communication overhead—or fail to fully exploit local datasets, limiting their effectiveness. To address these challenges, we propose Federated In-Context Learning (Fed-ICL), a general framework that enhances ICL through an iterative, collaborative process. Fed-ICL progressively refines responses by leveraging multi-round interactions between clients and a central server, improving answer quality without the need to transmit model parameters. We establish theoretical guarantees for the convergence of Fed-ICL and conduct extensive experiments on standard QA benchmarks, demonstrating that our proposed approach achieves strong performance while maintaining low communication costs.

Recent advancements in reinforcement learning from human feedback have shown that utilizing fine-grained token-level reward models can substantially enhance the performance of Proximal Policy Optimization (PPO) in aligning large language models. However, it is challenging to leverage such token-level reward as guidance for Direct Preference Optimization (DPO), since DPO is formulated as a sequence-level bandit problem. To address this challenge, this work decomposes the sequence-level PPO into a sequence of token-level proximal policy optimization problems and then frames the problem of token-level PPO with token-level reward guidance, from which closed-form optimal token-level policy and the corresponding token-level reward can be derived. Using the obtained reward and Bradley-Terry model, this work establishes a framework of computable loss functions with token-level reward guidance for DPO, and proposes a practical reward guidance based on the induced DPO reward. This formulation enables different tokens to exhibit varying degrees of deviation from reference policy based on their respective rewards. Experiment results demonstrate that our method achieves substantial performance improvements over DPO, with win rate gains of up to 7.5 points on MT-Bench, 6.2 points on AlpacaEval 2, and 4.3 points on Arena-Hard. Code is available at https://github.com/dvlab-research/TGDPO.

Autoregressive Transformers rely on Key-Value (KV) caching to accelerate inference. However, the linear growth of the KV cache with context length leads to excessive memory consumption and bandwidth constraints. Existing methods drop distant tokens or compress states in a lossy manner, sacrificing accuracy by discarding vital context or introducing bias. We propose ${MorphKV}$, an inference-time technique that maintains a constant-sized KV cache while preserving accuracy. MorphKV balances long-range dependencies and local coherence during text generation. It eliminates early-token bias while retaining high-fidelity context by adaptively ranking tokens through correlation-aware selection. Unlike heuristic retention or lossy compression, MorphKV iteratively refines the KV cache via lightweight updates guided by attention patterns of recent tokens. This approach captures inter-token correlation with greater accuracy, which is crucial for tasks like content creation and code generation. Our studies on long-response tasks show 52.9\% memory savings and 18.2\% higher accuracy on average compared to state-of-the-art prior works, enabling efficient deployment.

In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

The multi-domain fine-tuning of large language models (LLMs) confronts a notorious trade-off among abilities across domains. Existing studies attribute this trade-off to the conflicts between samples rooted in inherent semantics. Recent approaches attempt to mitigate these conflicts through the empirical investigation or heuristic strategies. However, without a fundamental understanding of interactions between samples, they yield only marginal improvements, while incurring substantial trial-and-error costs. To address this challenge, we move beyond empirical studies by modeling interactions between samples as their influence on each other's loss, estimated using gradients. Intriguingly, we find that these interactions **evolve throughout training** rather than being purely determined by inherent semantics. Building on this insight, we propose **EV**olving **I**nteraction-guided **C**urriculum (**EVIC**), which iteratively selects samples that positively influence the overall dataset for training. By dynamically adapting the training curriculum to prioritize samples that contribute the most to the model training, EVIC effectively mitigates conflicts and improves the sample efficiency. Extensive experiments on a mixed dataset covering coding, math, and general tasks with several model architectures show that EVIC significantly outperforms all baselines across diverse capabilities.

Document-level relation extraction (RE) aims to extract comprehensive correlations between entities and relations from documents. Most of existing works conduct transfer learning on pre-trained language models (PLMs), which allows for richer contextual representation to improve the performance. However, such PLMs-based methods suffer from incorporating structural knowledge, such as entity-entity interactions. Moreover, current works struggle to infer the implicit relations between entities across different sentences, which results in poor prediction. To deal with the above issues, we propose a novel and effective framework, named DocKS-RAG, which introduces extra structural knowledge and semantic information to further enhance the performance of document-level RE. Specifically, we construct a Document-level Knowledge Graph from the observable documentation data to better capture the structural information between entities and relations. Then, a Sentence-level Semantic Retrieval-Augmented Generation mechanism is designed to consider the similarity in different sentences by retrieving the relevant contextual semantic information. Furthermore, we present a hybrid-prompt tuning method on large language models (LLMs) for specific document-level RE tasks. Finally, extensive experiments conducted on two benchmark datasets demonstrate that our proposed framework enhances all the metrics compared with state-of-the-art methods.

Despite the impressive capabilities of multimodal large language models (MLLMs) in vision-language tasks, they are prone to hallucinations in real-world scenarios. This paper investigates the hallucination phenomenon in MLLMs from the perspective of modality conflict. Unlike existing works focusing on the conflicts between model responses and inputs, we study the inherent conflicts in inputs from different modalities that place MLLMs in a dilemma and directly lead to hallucinations. We formally define the modality conflict and construct a dataset named Multimodal Modality Conflict (MMMC) to simulate this phenomenon in vision-language tasks. Three methods based on prompt engineering, supervised fine-tuning, and reinforcement learning are proposed to alleviate the hallucination caused by modality conflict. Extensive experiments are conducted on the MMMC dataset to analyze the merits and demerits of these methods. Our results show that the reinforcement learning method achieves the best performance in mitigating the hallucination under modality conflict, while the supervised fine-tuning method shows promising and stable performance. Our work sheds light on the unnoticed modality conflict that leads to hallucinations and provides more insights into the robustness of MLLMs.

Large Language Model (LLM) collaborative decoding techniques improve output quality by combining the outputs of multiple models at each generation step, but they incur high computational costs. In this paper, we introduce **Collaborative decoding via Speculation (CoS)**, a novel framework that accelerates collaborative decoding without compromising performance. Inspired by Speculative Decoding—where a small proposal model generates tokens sequentially, and a larger target model verifies them in parallel, our approach builds on two key insights: (1) the verification distribution can be the combined distribution of both the proposal and target models, and (2) alternating each model as the proposer and verifier can further enhance efficiency. We generalize this method to collaboration among *n* models and theoretically prove that CoS is never slower than standard collaborative decoding, typically achieving faster speed. Extensive experiments demonstrate CoS is **1.11x–2.23x** faster than standard collaborative decoding without compromising generation quality. Our code is available at https://github.com/Kamichanw/CoS/.

We introduce a new approach to systematically map features discovered by sparse autoencoder across consecutive layers of large language models, extending earlier work that examined inter-layer feature links. By using a data-free cosine similarity technique, we trace how specific features persist, transform, or first appear at each stage. This method yields granular flow graphs of feature evolution, enabling fine-grained interpretability and mechanistic insights into model computations. Crucially, we demonstrate how these cross-layer feature maps facilitate direct steering of model behavior by amplifying or suppressing chosen features, achieving targeted thematic control in text generation. Together, our findings highlight the utility of a causal, cross-layer interpretability framework that not only clarifies how features develop through forward passes but also provides new means for transparent manipulation of large language models.

Activation sparsity denotes the existence of substantial weakly-contributed neurons within feed-forward networks of large language models (LLMs), providing wide potential benefits such as computation acceleration. However, existing works lack thorough quantitative studies on this useful property, in terms of both its measurement and influential factors. In this paper, we address three underexplored research questions: (1) How can activation sparsity be measured more accurately? (2) How is activation sparsity affected by the model architecture and training process? (3) How can we build a more sparsely activated and efficient LLM? Specifically, we develop a generalizable and performance-friendly metric, named CETT-PPL-1\%, to measure activation sparsity. Based on CETT-PPL-1\%, we quantitatively study the influence of various factors and observe several important phenomena, such as the convergent power-law relationship between sparsity and training data amount, the higher competence of ReLU activation than mainstream SiLU activation, the potential sparsity merit of a small width-depth ratio, and the scale insensitivity of activation sparsity. Finally, we provide implications for building sparse and effective LLMs, and demonstrate the reliability of our findings by training a 2.4B model with a sparsity ratio of 93.52\%, showing 4.1$\times$ speedup compared with its dense version. The codes and checkpoints are available at https://github.com/thunlp/SparsingLaw/.

Pruning is a widely used technique for compressing large neural networks that eliminates weights that have minimal impact on the model's performance. Current pruning methods, exemplified by magnitude pruning, assign an importance score to each weight based on its magnitude and remove weights with scores below a certain threshold. Nonetheless, these methods often create a gap between the original dense and the pruned sparse model, potentially impairing performance. Especially when the sparsity ratio is high, the gap becomes more pronounced. To mitigate this issue, we introduce a method to bridge the gap left by pruning by utilizing a low-rank approximation of the difference between the dense and sparse matrices. Our method entails the iterative refinement of the sparse weight matrix augmented by a low-rank adjustment. This technique captures and retains the essential information often lost during pruning, thereby improving the performance of the pruned model. Furthermore, we offer a comprehensive theoretical analysis of our approach, emphasizing its convergence properties and establishing a solid basis for its efficacy. Experimental results on LLaMa models validate its effectiveness on large language models across various pruning techniques and sparsity levels. Our method shows significant improvements: at 50\% sparsity, it reduces perplexity by 53.9\% compared to conventional magnitude pruning on LLaMa-7B. Furthermore, to achieve a specific performance target, our approach enables an 8.6\% reduction in model parameters while maintaining a sparsity ratio of about 50\%.

High-quality data plays a critical role in the pretraining and fine-tuning of large language models (LLMs), even determining their performance ceiling to some degree. Consequently, numerous data selection methods have been proposed to identify subsets of data that can effectively and efficiently enhance model performance. However, most of these methods focus on general data selection and tend to overlook the specific nuances of domain-related data. In this paper, we introduce MASS, a Mathematical data Selection framework using the Skill graph for pretraining LLMs in the mathematical reasoning domain. By taking into account the unique characteristics of mathematics and reasoning, we construct a skill graph that captures the mathematical skills and their interrelations from a reference dataset. This skill graph guides us in assigning quality scores to the target dataset, enabling us to select the top-ranked subset which is further used to pretrain LLMs. Experimental results demonstrate the efficiency and effectiveness of MASS across different model sizes (1B and 7B) and pretraining datasets (web data and synthetic data). Specifically, in terms of efficiency, models trained on subsets selected by MASS can achieve similar performance to models trained on the original datasets, with a significant reduction in the number of trained tokens - ranging from 50\% to 70\% fewer tokens. In terms of effectiveness, when trained on the same amount of tokens, models trained on the data selected by MASS outperform those trained on the original datasets by 3.3\% to 5.9\%. These results underscore the potential of MASS to improve both the efficiency and effectiveness of pretraining LLMs.

Increasing diversity in language models is a challenging yet essential objective. A common approach is to raise the decoding temperature. In this work, we investigate this approach through a simplistic yet common case to provide insights into why decreasing temperature can improve quality (Precision), while increasing it often fails to boost coverage (Recall). Our analysis reveals that for a model to be effectively tunable through temperature adjustments, it must be trained toward coverage. To address this, we propose rethinking loss functions in language models by leveraging the Precision-Recall framework. Our results demonstrate that this approach achieves a substantially better trade-off between Precision and Recall than merely combining negative log-likelihood training with temperature scaling. These findings offer a pathway toward more versatile and robust language modeling techniques.

Large language models (LLMs) offer remarkable capabilities, yet their high inference costs restrict wider adoption. While increasing parameter counts improves accuracy, it also broadens the gap between state-of-the-art capabilities and practical deployability. We present **Puzzle**, a hardware-aware framework that accelerates the inference of LLMs while preserving their capabilities. Using neural architecture search (NAS) at a large-scale, Puzzle optimizes models with tens of billions of parameters. Our approach utilizes blockwise local knowledge distillation (BLD) for parallel architecture exploration and employs mixed-integer programming for precise constraint optimization. We showcase our framework’s impact via Llama-3.1-Nemotron-51B-Instruct (Nemotron-51B) and Llama-3.3-Nemotron-49B, two publicly available models derived from Llama-70B-Instruct. Both models achieve a 2.17x inference throughput speedup, fitting on a single NVIDIA H100 GPU while retaining 98.4% of the original model's benchmark accuracies. These are the most accurate models supporting single H100 GPU inference with large batch sizes, despite training on 45B tokens at most, far fewer than the 15T used to train Llama-70B. Lastly, we show that lightweight alignment on these derived models allows them to surpass the parent model in specific capabilities. Our work establishes that powerful LLM models can be optimized for efficient deployment with only negligible loss in quality, underscoring that inference performance, not parameter count alone, should guide model selection.

The residual stream acts as a memory bus where transformer layers both store and access features (Elhage et al., 2021). We consider changing the mechanism for retrieving and storing information in the residual stream, and replace the residual stream of the transformer with an outer product memory matrix (Kohonen, 1972, Anderson, 1972). We call this model the Residual Matrix Transformer (RMT). We find that the RMT enjoys a number of attractive properties: 1) the size of the residual stream can be scaled independently of compute and model size, improving performance, 2) the RMT can achieve the same loss as the transformer with 58% fewer FLOPS, 25% fewer parameters, and 41% fewer training tokens tokens, and 3) the RMT outperforms the transformer on downstream evaluations. We theoretically analyze the transformer and the RMT, and show that the RMT allows for more efficient scaling of the residual stream, as well as improved variance propagation properties.

Multi-agent frameworks can substantially boost the reasoning power of large language models (LLMs), but they typically incur heavy computational costs and lack convergence guarantees. To overcome these challenges, we recast multi-LLM coordination as an incomplete-information game and seek a Bayesian Nash equilibrium (BNE), in which each agent optimally responds to its probabilistic beliefs about the strategies of others. We introduce Efficient Coordination via Nash Equilibrium (ECON), a hierarchical reinforcement-learning paradigm that marries distributed reasoning with centralized final output. Under ECON, each LLM independently selects responses that maximize its expected reward, conditioned on its beliefs about co-agents, without requiring costly inter-agent exchanges. We mathematically prove that ECON attains a markedly tighter regret bound than non-equilibrium multi-agent schemes. Empirically, ECON outperforms existing multi-LLM approaches by 11.2% on average across six benchmarks spanning complex reasoning and planning tasks. Further experiments demonstrate ECON’s ability to flexibly incorporate additional models, confirming its scalability and paving the way toward larger, more powerful multi-LLM ensembles. The code is publicly available at: https://github.com/tmlr-group/ECON.

Existing methods for code generation use code snippets as seed data, restricting the complexity and diversity of the synthesized data. In this paper, we introduce a novel feature tree-based synthesis framework, which revolves around hierarchical code features derived from high-level abstractions of code. The feature tree is constructed from raw data and refined iteratively to increase the quantity and diversity of the extracted features, which captures and recognizes more complex patterns and relationships within the code. By adjusting the depth and breadth of the sampled subtrees, our framework provides precise control over the complexity of the generated code, enabling functionalities that range from function-level operations to multi-file scenarios. We fine-tuned widely-used base models to obtain EpiCoder series, achieving state-of-the-art performance on multiple benchmarks at both the function and file levels. In particular, empirical evidence indicates that our approach shows significant potential in the synthesizing of repository-level code data. Our code and data are publicly available.

Using large teacher models to guide the training of smaller student models has become the prevailing paradigm for efficient and effective learning. However, vocabulary mismatches between teacher and student language models pose significant challenges in language modeling, resulting in divergent token sequences and output distributions. To overcome these limitations, we propose Vocabulary-agnostic Teacher Guided Language Modeling (VocAgnoLM), a novel approach that bridges the gap caused by vocabulary mismatch through two key methods: (1) Token-level Lexical Alignment, which aligns token sequences across mismatched vocabularies, and (2) Teacher Guided Loss, which leverages the loss of teacher model to guide effective student training. We demonstrate its effectiveness in language modeling with 1B student model using various 7B teacher models with different vocabularies. Notably, with Qwen2.5-Math-Instruct, a teacher model sharing only about 6% of its vocabulary with TinyLlama, VocAgnoLM achieves a 46% performance improvement compared to naive continual pretraining. Furthermore, we demonstrate that VocAgnoLM consistently benefits from stronger teacher models, providing a robust solution to vocabulary mismatches in language modeling.