ICML 2025 Papers

Post-training compression reduces the computational and memory costs of large language models (LLMs), enabling resource-efficient deployment. However, existing compression benchmarks focus narrowly on language modeling (e.g., perplexity) and natural language understanding tasks (e.g., GLUE accuracy), ignoring the agentic capabilities—workflow, tool use/function call, long-context understanding and real-world application. We introduce the Agent Compression Benchmark (ACBench), the first comprehensive benchmark for evaluating how compression impacts LLMs' agentic abilities. ACBench spans (1) 12 tasks across 4 capabilities (e.g., WorfBench for workflow generation, Needle-in-Haystack for long-context retrieval), (2) 4-bit quantization (GPTQ, AWQ) and 50% pruning (Wanda, SparseGPT), and (3) 15 models, including small (Gemma-2B), standard (Qwen2.5-7B), and distilled reasoning LLMs (DeepSeek-R1-Distill). Our experiments reveal compression tradeoffs: 4-bit quantization preserves workflow generation and tool use (1%--3% drop) but degrades real-world application accuracy by 10%--15%. We introduce ERank, Top-k Ranking Correlation and Energy to systematize analysis. ACBench provides actionable insights for optimizing LLM compression in agentic scenarios, bridging the gap between algorithmic efficiency and real-world applicability.

Vision Language Models (VLMs) are impressive at visual question answering and image captioning. But they underperform on multi-step visual reasoning---even compared to LLMs on the same tasks presented in text form---giving rise to perceptions of *modality imbalance* or *brittleness*. Towards a systematic study of such issues, we introduce a synthetic framework for assessing the ability of VLMs to perform algorithmic visual reasoning, comprising three tasks: Table Readout, Grid Navigation, and Visual Analogy. Each has two levels of difficulty, SIMPLE and HARD, and even the SIMPLE versions are difficult for frontier VLMs. We propose strategies for training on the SIMPLE version of tasks that improve performance on the corresponding HARD task, i.e., simple-to-hard (S2H) generalization. This controlled setup, where each task also has an equivalent text-only version, allows a quantification of the modality imbalance and how it is impacted by training strategy. We show that 1) explicit image-to-text conversion is important in promoting S2H generalization on images, by transferring reasoning from text; 2) conversion can be internalized at test time. We also report results of mechanistic study of this phenomenon. We identify measures of gradient alignment that can identify training strategies that promote better S2H generalization. Ablations highlight the importance of chain-of-thought.

Despite recent advancements of fine-tuning large language models (LLMs) to facilitate agent tasks, parameter-efficient fine-tuning (PEFT) methodologies for agent remain largely unexplored. In this paper, we introduce three key strategies for PEFT in agent tasks: 1) Inspired by the increasingly dominant \textit{Reason+Action} paradigm, we first decompose the capabilities necessary for the agent tasks into three distinct roles: reasoner, executor, and summarizer. The reasoner is responsible for comprehending the user's query and determining the next role based on the execution trajectory. The executor is tasked with identifying the appropriate functions and parameters to invoke. The summarizer conveys the distilled information from conversations back to the user. 2) We then propose the Mixture-of-Roles (MoR) framework, which comprises three specialized Low-Rank Adaptation (LoRA) groups, each designated to fulfill a distinct role. By focusing on their respective specialized capabilities and engaging in collaborative interactions, these LoRAs collectively accomplish the agent task. 3) To effectively fine-tune the framework, we develop a multi-role data generation pipeline based on publicly available datasets, incorporating role-specific content completion and reliability verification. We conduct extensive experiments and thorough ablation studies on various LLMs and agent benchmarks, demonstrating the effectiveness of the proposed method. This project is publicly available at https://mor-agent.github.io

Despite the success of distillation in large language models (LLMs), most prior work applies identical loss functions to both teacher- and student-generated data. These strategies overlook the synergy between loss formulations and data types, leading to a suboptimal performance boost in student models. To address this, we propose DistiLLM-2, a contrastive approach that simultaneously increases the likelihood of teacher responses and decreases that of student responses by harnessing this synergy. Our extensive experiments show that DistiLLM-2 not only builds high-performing student models across a wide range of tasks, including instruction-following and code generation, but also supports diverse applications, such as preference alignment and vision-language extensions. These findings highlight the potential of a contrastive approach to enhance the efficacy of LLM distillation by effectively aligning teacher and student models across varied data types.

Despite the success of Instruction Tuning (IT) in training large language models (LLMs), such models often leverage spurious or biased features learnt from their training data and can become misaligned, leading to undesired behaviours. While existing techniques can steer model behaviour at inference-time, they are often post-hoc and do not embed steering as an intrinsic model feature. In this work, we introduce Focus Instruction Tuning (FIT), which trains LLMs to condition their responses by focusing on specific features whilst ignoring others, leading to different behaviours based on what features are specified. Across diverse benchmarks, we demonstrate that FIT: (i) successfully steers behaviour at inference time; (ii) increases robustness by amplifying core task signals and down-weighting spurious cues; (iii) mitigates social bias by suppressing demographic attributes; and (iv) generalises under distribution shifts and to previously unseen focus features. FIT therefore offers a lightweight, intrinsic mechanism for building more robust, fair, and easily controllable LLMs.

Reward models are widely used as proxies for human preferences when aligning or evaluating LLMs. However, reward models are black boxes, and it is often unclear what, exactly, they are actually rewarding. In this paper we develop Rewrite-based Attribute Treatment Estimator (RATE) as an effective method for measuring the sensitivity of a reward model to high-level attributes of responses, such as sentiment, helpfulness, or complexity. Importantly, RATE measures the *causal* effect of an attribute on the reward. RATE uses LLMs to rewrite responses to produce imperfect counterfactuals examples that can be used to measure causal effects. A key challenge is that these rewrites are imperfect in a manner that can induce substantial bias in the estimated sensitivity of the reward model to the attribute. The core idea of RATE is to adjust for this imperfect-rewrite effect by rewriting *twice*. We establish the validity of the RATE procedure and show empirically that it is an effective estimator.

We introduce SelfCite, a novel self-supervised approach that aligns LLMs to generate high-quality, fine-grained, sentence-level citations for the statements in their generated responses. Instead of only relying on costly and labor-intensive annotations, SelfCite leverages a reward signal provided by the LLM itself through *context ablation*: If a citation is necessary, removing the cited text from the context should prevent the same response; if sufficient, retaining the cited text alone should preserve the same response. This reward can guide the inference-time best-of-N sampling strategy to improve citation quality significantly, as well as be used in preference optimization to directly fine-tune the models for generating better citations. The effectiveness of SelfCite is demonstrated by increasing citation F1 up to 5.3 points on the LongBench-Cite benchmark across five long-form question answering tasks. The source code is available at https://github.com/facebookresearch/SelfCite.

Post-training of language models (LMs) increasingly relies on the following two stages: (i) knowledge distillation, where the LM is trained to imitate a larger teacher LM, and (ii) reinforcement learning from human feedback (RLHF), where the LM is aligned by optimizing a reward model. In the second RLHF stage, a well-known challenge is reward hacking, where the LM over-optimizes the reward model, leading to degraded performance on the true objective, in line with Goodhart's law. In this paper, we investigate whether a similar phenomenon, that we call teacher hacking, can occur during knowledge distillation. This could arise because the teacher LM is itself an imperfect approximation of the true distribution. To study this, we propose a controlled experimental setup involving: (i) an oracle LM representing the ground-truth distribution, (ii) a teacher LM distilled from the oracle, and (iii) a student LM distilled from the teacher. Our experiments reveal the following insights. When using a fixed offline dataset for distillation, teacher hacking occurs; moreover, we can detect it by observing when the optimization process deviates from polynomial convergence laws. In contrast, employing online data generation techniques effectively mitigates teacher hacking. More precisely, we identify data diversity as the key factor in preventing hacking. Overall, our findings provide a deeper understanding of the benefits and limitations of distillation for building robust LMs.

The alignment of large language models (LLMs) often assumes that using more clean data yields better outcomes, overlooking the match between model capacity and example difficulty. Challenging this, we propose a new principle: *Preference data vary in difficulty, and overly difficult examples hinder alignment, by exceeding the model's capacity*. Through systematic experimentation, we validate this principle with three key findings: (1) preference examples vary in difficulty, as evidenced by consistent learning orders across alignment runs; (2) overly difficult examples significantly degrade performance across four LLMs and two datasets; and (3) the capacity of a model dictates its threshold for handling difficult examples, underscoring a critical relationship between data selection and model capacity. Building on this principle, we introduce *Selective DPO*, which filters out overly difficult examples. This simple adjustment improves alignment performance by 9-16\% in win rates on the AlpacaEval 2 benchmark compared to the DPO baseline, surpassing a series of DPO variants with different algorithmic adjustments. These results together illuminate the importance of aligning data difficulty with model capacity, offering a transformative perspective for improving alignment strategies in LLMs. Code is available at https://github.com/glorgao/SelectiveDPO

Code Agent development is an extremely active research area, where a reliable performance metric is critical for tracking progress and guiding new developments. This demand is underscored by the meteoric rise in popularity of SWE-Bench -- a benchmark that challenges code agents to generate patches addressing GitHub issues given the full repository as context. The correctness of generated patches is then evaluated by executing a human-written test suite extracted from the repository after the issue's resolution. However, constructing benchmarks like SWE-Bench requires substantial manual effort to set up historically accurate execution environments for testing. Crucially, this severely limits the number of considered repositories, e.g., just 12 for SWE-Bench. Considering so few repositories, selected for their popularity runs the risk of leading to a distributional mismatch, i.e., the measured performance may not be representative of real-world scenarios running the riks of misguiding development efforts. In this work, we address this challenge and introduce SetUpAgent, a fully automated system capable of historically accurate dependency setup, test execution, and result parsing. Using SetUpAgent, we generate two new datasets: (i) SWEE-Bench an extended version of SWE-Bench encompassing hundreds of repositories, and (ii) SWA-Bench a benchmark focusing on applications rather than libraries. Comparing these datasets to SWE-Bench with respect to their characteristics and code agent performance, we find significant distributional differences, including lower issue description quality and detail level, higher fix complexity, and most importantly up to 60% lower agent success rates.

We examine diverging preferences in human-labeled preference datasets. We develop a taxonomy of disagreement sources spanning ten categories across four high-level classes and find that the majority of disagreements are due to factors such as task underspecification or response style. Our findings challenge a standard assumption in reward modeling methods that annotator disagreements can be attributed to simple noise. We then explore how these findings impact two areas of LLM development: reward modeling training and evaluation. In our experiments, we demonstrate how standard reward modeling (e.g., Bradley-Terry) and LLM-as-Judge evaluation methods fail to account for divergence between annotators. These findings highlight challenges in LLM evaluations, which are greatly influenced by divisive features like response style, and in developing pluralistically aligned LLMs. To address these issues, we develop methods for identifying diverging preferences to mitigate their influence in evaluations and during LLM training.

Large language model training with 8-bit floating point (FP8) formats promises significant efficiency improvements, but reduced numerical precision makes training challenging. It is currently possible to train in FP8 only if one is willing to tune various hyperparameters, reduce model scale, or accept the overhead of computing dynamic scale factors. We demonstrate simple, scalable FP8 training that requires no dynamic scaling factors or special hyperparameters, even at large model sizes. Our method, \textit{µnit Scaling (µS)}, also enables simple hyperparameter transfer across model widths, matched numerics across training and inference, and other desirable properties. µnit Scaling is straightforward to implement, consisting of a set of minimal interventions based on a first-principles analysis of transformer operations. We validate our method by training models with parameters ranging from 1B to 13B, performing all hidden linear layer computations in FP8. We achieve quality equal to higher-precision baselines while also training up to 33% faster.

We investigate an emerging setup in which a small, on-device language model (LM) with access to local data collaborates with a frontier, cloud-hosted LM to solve real-world tasks involving financial, medical, and scientific reasoning over long documents. *Can a local-remote collaboration reduce cloud inference costs while preserving quality?* First, we consider a naïve collaboration protocol, coined MINION, where the local and remote models simply chat back and forth. Because only the local model ingests the full context, this protocol reduces cloud costs by 30.4x, but recovers only 87% of the performance of the frontier model. We identify two key limitations of this protocol: the local model struggles to (1) follow the remote model's multi-step instructions and (2) reason over long contexts. Motivated by these observations, we propose MINIONS, a protocol in which the remote model decomposes the task into easier subtasks over shorter chunks of the document, that are executed locally in parallel. MINIONS reduces costs by 5.7× on average while recovering 97.9% of the remote-only performance. Our analysis reveals several key design choices that influence the trade-off between cost and performance in local-remote systems.

Understanding the decision-making processes of large language models is critical given their widespread applications. To achieve this, we aim to connect a formal mathematical framework—zigzag persistence from topological data analysis —with practical and easily applicable algorithms. Zigzag persistence is particularly effective for characterizing data as it dynamically transforms across model layers. Within this framework, we introduce topological descriptors that measure how topological features, $p$-dimensional holes, persist and evolve throughout the layers. Unlike methods that assess each layer individually and then aggregate the results, our approach directly tracks the full evolutionary path of these features. This offers a statistical perspective on how prompts are rearranged and their relative positions changed in the representation space, providing insights into the system’s operation as an integrated whole. To demonstrate the expressivity and applicability of our framework, we highlight how sensitive these descriptors are to different models and a variety of datasets. As a showcase application to a downstream task, we use zigzag persistence to establish a criterion for layer pruning, achieving results comparable to state-of-the-art methods while preserving the system-level perspective.

Vision-language instruction-tuning models have recently achieved significant performance improvements. In this work, we discover that large-scale 3D parallel training on those models leads to an imbalanced computation load across different devices. The vision and language parts are inherently heterogeneous: their data distribution and model architecture differ significantly, which affects distributed training efficiency. To address this issue, we rebalance the computational load from data, model, and memory perspectives, achieving more balanced computation across devices. Specifically, for the data, instances are grouped into new balanced mini-batches within and across devices. A search-based method is employed for the model to achieve a more balanced partitioning. For memory optimization, we adaptively adjust the re-computation strategy for each partition to utilize the available memory fully. These three perspectives are not independent but are closely connected, forming an omniverse balanced training framework. Extensive experiments are conducted to validate the effectiveness of our method. Compared with the open-source training code of InternVL-Chat, training time is reduced greatly, achieving about 1.8$\times$ speed-up. Our method's efficacy and generalizability are further validated across various models and datasets. Codes will be released at https://github.com/ModelTC/OmniBal.

Large language models (LLMs) excel at capturing global token dependencies via self-attention but face prohibitive compute and memory costs on lengthy inputs. While sub-quadratic methods (e.g., linear attention) can reduce these costs, they often degrade accuracy due to overemphasizing recent tokens. In this work, we first propose *dual-state linear attention* (**DSLA**), a novel design that maintains two specialized hidden states—one for preserving historical context and one for tracking recency—thereby mitigating the short-range bias typical of linear-attention architectures. To further balance efficiency and accuracy under dynamic workload conditions, we introduce DSLA-*Serve*, an online *adaptive distillation* framework that progressively replaces Transformer layers with DSLA layers at inference time, guided by a sensitivity-based layer ordering. DSLA-*Serve* uses a chained fine-tuning strategy to ensure that each newly converted DSLA layer remains consistent with previously replaced layers, preserving the overall quality. Extensive evaluations on commonsense reasoning, long-context QA, and text summarization demonstrate that DSLA-*Serve* yields **2.3×** faster inference than Llama2-7B and **3.0×** faster than the hybrid Zamba-7B, while retaining comparable performance across downstream tasks. Our ablation studies show that DSLA’s dual states capture both global and local dependencies, addressing the historical-token underrepresentation seen in prior linear attentions.

Large Language Models (LLMs) face significant deployment challenges due to their substantial resource requirements. While low-bit quantized weights can reduce memory usage and improve inference efficiency, current hardware lacks native support for mixed-precision General Matrix Multiplication (mpGEMM), resulting in inefficient dequantization-based implementations. Moreover, uniform quantization methods often fail to capture weight distributions adequately, leading to performance degradation. We propose GANQ (GPU-Adaptive Non-Uniform Quantization), a layer-wise post-training non-uniform quantization framework optimized for hardware-efficient lookup table-based mpGEMM. GANQ achieves superior quantization performance by utilizing a training-free, GPU-adaptive optimization algorithm to efficiently reduce layer-wise quantization errors. Extensive experiments demonstrate GANQ's ability to reduce the perplexity gap from the FP16 baseline compared to state-of-the-art methods for both 3-bit and 4-bit quantization. Furthermore, when deployed on a single NVIDIA RTX 4090 GPU, GANQ's quantized models achieve up to 2.57$\times$ speedup over the baseline, advancing memory and inference efficiency in LLM deployment.

We study memory-efficient optimization of neural networks (in particular language models) with *linear gradient transformations*, where the gradients are linearly mapped to a lower dimensional space than the full parameter space, thus saving memory required for gradient accumulation and optimizer state persistence. The model parameters are updated by first performing an optimization step in the lower dimensional space and then going back into the original parameter space via the linear map's transpose. We show that optimizing the model in this transformed space is equivalent to reparameterizing the original model through a *linear adapter* that additively modifies the model parameters, and then only optimizing the adapter's parameters. When the transformation is Kronecker-factored, this establishes an equivalence between GaLore and one-sided LoRA. We show that this duality between gradient transformations and adapter-based reparameterizations unifies existing approaches to memory-efficient training and suggests new techniques for improving training efficiency and memory use.

Large Language Model (LLM) agents equipped with external tools have become increasingly powerful for complex tasks such as web shopping, automated email replies, and financial trading. However, these advancements amplify the risks of adversarial attacks, especially when agents can access sensitive external functionalities. Nevertheless, manipulating LLM agents into performing targeted malicious actions or invoking specific tools remains challenging, as these agents extensively reason or plan before executing final actions. In this work, we present UDora, a unified red teaming framework designed for LLM agents that dynamically hijacks the agent's reasoning processes to compel malicious behavior. Specifically, UDora first generates the model’s reasoning trace for the given task, then automatically identifies optimal points within this trace to insert targeted perturbations. The resulting perturbed reasoning is then used as a surrogate response for optimization. By iteratively applying this process, the LLM agent will then be induced to undertake designated malicious actions or to invoke specific malicious tools. Our approach demonstrates superior effectiveness compared to existing methods across three LLM agent datasets. The code is available at https://github.com/AI-secure/UDora.