ICLR 2025 论文列表

Vehicle trajectory prediction is a crucial aspect of autonomous driving, which requires extensive trajectory data to train prediction models to understand the complex, varied, and unpredictable patterns of vehicular interactions. However, acquiring real-world data is expensive, so we advocate using Large Language Models (LLMs) to generate abundant and realistic trajectories of interacting vehicles efficiently. These models rely on textual descriptions of vehicle-to-vehicle interactions on a map to produce the trajectories. We introduce Trajectory-LLM (Traj-LLM), a new approach that takes brief descriptions of vehicular interactions as input and generates corresponding trajectories. Unlike language-based approaches that translate text directly to trajectories, Traj-LLM uses reasonable driving behaviors to align the vehicle trajectories with the text. This results in an "interaction-behavior-trajectory" translation process. We have also created a new dataset, Language-to-Trajectory (L2T), which includes 240K textual descriptions of vehicle interactions and behaviors, each paired with corresponding map topologies and vehicle trajectory segments. By leveraging the L2T dataset, Traj-LLM can adapt interactive trajectories to diverse map topologies. Furthermore, Traj-LLM generates additional data that enhances downstream prediction models, leading to consistent performance improvements across public benchmarks. The source code is released at https://github.com/TJU-IDVLab/Traj-LLM.

The development and evaluation of Large Language Models (LLMs) have largely focused on individual capabilities. However, this overlooks the intersection of multiple abilities across different types of expertise that are often required for real-world tasks, which we term **cross capabilities**. To systematically explore this concept, we first define seven core individual capabilities and then pair them to form seven common cross capabilities, each supported by a manually constructed taxonomy. Building on these definitions, we introduce *CrossEval*, a benchmark comprising 1,400 human-annotated prompts, with 100 prompts for each individual and cross capability. To ensure reliable evaluation, we involve expert annotators to assess 4,200 model responses, gathering 8,400 human ratings with detailed explanations to serve as reference examples. Our findings reveal that current LLMs consistently exhibit the ``Law of the Weakest Link,'' where cross-capability performance is significantly constrained by the weakest component. Across 58 cross-capability scores from 17 models, 38 scores are lower than all individual capabilities, while 20 fall between strong and weak, but closer to the weaker ability. These results highlight LLMs' underperformance in cross-capability tasks, emphasizing the need to identify and improve their weakest capabilities as a key research priority. The code, benchmarks, and evaluations are available on our [project website](https://www.llm-cross-capabilities.org).

The frequent retrieval of Key-Value (KV) cache data has emerged as a significant factor contributing to the inefficiency of the inference process in large language models. Previous research has demonstrated that a small subset of critical KV cache tokens largely influences attention outcomes, leading to methods that either employ fixed sparsity patterns or dynamically select critical tokens based on the query. While dynamic sparse patterns have proven to be more effective, they introduce significant computational overhead, as critical tokens must be reselected for each self-attention computation. In this paper, we reveal substantial similarities in KV cache token criticality across neighboring queries, layers, and heads. Motivated by this insight, we propose HShare, a hierarchical KV sharing framework. HShare facilitates the sharing of critical KV cache token indices across layers, heads, and queries, which significantly reduces the computational overhead associated with query-aware dynamic token sparsity. In addition, we introduce a greedy algorithm that dynamically determines the optimal layer-level and head-level sharing configuration for the decoding phase. We evaluate the effectiveness and efficiency of HShare across various tasks using three models: LLaMA2-7b, LLaMA3-70b, and Mistral-7b. Experimental results demonstrate that HShare achieves competitive accuracy with different sharing ratios, while delivering up to an $8.6\times$ speedup in self-attention operations and a $2.7\times$ improvement in end-to-end throughput compared with FlashAttention2 and GPT-fast respectively. The source code is publicly available at ~\url{https://github.com/wuhuaijin/HShare}.

In recent years, large language models (LLMs) have demonstrated significant success in performing varied natural language tasks such as language translation, question-answering, summarizing, fact-checking, etc. Despite LLMs’ impressive ability to generate human-like texts, LLMs are infamous for their inconsistent responses – a meaning-preserving change in the input query results in an inconsistent response and attributes to vulnerabilities of LLMs such as hallucination. Consequently, existing research focuses on simple paraphrasing-based consistency assessment of LLMs, and ignores complex queries that necessitate an even better understanding of logical reasoning by an LLM. *Our work therefore addresses the logical inconsistency of LLMs under complex logical queries with primitive logical operators, e.g., negation, conjunction, and disjunction.* As a test bed, we consider retrieval-augmented LLMs on a fact-checking task involving propositional logic queries from knowledge graphs (KGs). Our contributions are three-fold. **Benchmark:** We introduce three logical fact-checking datasets over KGs for community development towards logically consistent LLMs. **Assessment:** We propose consistency measures of LLMs on propositional logic queries and demonstrate that existing LLMs lack logical consistency, especially on complex queries. **Improvement:** We employ supervised fine-tuning to improve the logical consistency of LLMs on the complex fact-checking task with KG contexts. We have made our source code and benchmarks available.

Evaluating the reasoning abilities of large language models (LLMs) is challenging. Existing benchmarks often depend on static datasets, which are vulnerable to data contamination and may get saturated over time, or on binary live human feedback that conflates reasoning with other abilities. As the most prominent dynamic benchmark, Chatbot Arena evaluates open-ended questions in real-world settings, but lacks the granularity in assessing specific reasoning capabilities. We introduce GameArena, a dynamic benchmark designed to evaluate LLM reasoning capabilities through interactive gameplay with humans. GameArena consists of three games designed to test specific reasoning capabilities (e.g., deductive and inductive reasoning), while keeping participants entertained and engaged. We analyze the gaming data retrospectively to uncover the underlying reasoning processes of LLMs and measure their fine-grained reasoning capabilities. We collect over 2000 game sessions and provide detailed assessments of various reasoning capabilities for five state-of-the-art LLMs. Our user study with 100 participants suggests that GameArena improves user engagement compared to Chatbot Arena. For the first time, GameArena enables the collection of step-by-step LLM reasoning data in the wild.

Verifiable formal languages like Lean have profoundly impacted mathematical reasoning, particularly through the use of large language models (LLMs) for automated reasoning. A significant challenge in training LLMs for these formal languages is the lack of parallel datasets that align natural language with formal language proofs. To address this challenge, this paper introduces a novel framework for translating the Mathlib4 corpus (a unified library of mathematics in formal language Lean 4) into natural language. Building upon this, we employ a dual augmentation strategy that combines tactic-based and informal-based approaches, leveraging the Lean-jixia system, a Lean 4 analyzer. We present the results of this pipeline on Mathlib4 as Herald (Hierarchy and Retrieval-based Translated Lean Dataset). We also propose the Herald Translator, which is fine-tuned on Herald. Herald translator achieves a 96.7\% accuracy (Pass@128) on formalizing statements in the miniF2F-test and a 23.5\% accuracy on our internal graduate-level textbook dataset, outperforming InternLM2-Math-Plus-7B (73.0\% and 7.5\%) and TheoremLlama (50.1\% and 4.0\%). Furthermore, we propose a section-level translation framework for real-world applications. As a direct application of Herald translator, we have successfully translated a template section in the Stack project, marking a notable progress in the automatic formalization of graduate-level mathematical literature. Our model, along with the datasets, are open-sourced to the public.

Language is typically modelled with discrete sequences. However, the most successful approaches to language modelling, namely neural networks, are continuous and smooth function approximators. In this work, we show that Transformer-based language models implicitly learn to represent sentences as continuous-time functions defined over a continuous input space. This phenomenon occurs in most state-of-the-art Large Language Models (LLMs), including Llama2, Llama3, Phi3, Gemma, Gemma2, and Mistral, and suggests that LLMs reason about language in ways that fundamentally differ from humans. Our work formally extends Transformers to capture the nuances of time and space continuity in both input and output space. Our results challenge the traditional interpretation of how LLMs understand language, with several linguistic and engineering implications.

Large Language Models (LLMs) are increasingly deployed as chatbots, yet their ability to personalize responses to user preferences remains limited. We introduce PrefEval, a benchmark for evaluating LLMs' ability to infer, memorize and adhere to user preferences in long-context conversational setting. PrefEval comprises 3,000 manually curated user preference and query pairs spanning 20 topics. PrefEval contains user personalization or preference information in both explicit and implicit preference forms, and evaluates LLM performance using a generation and a classification task. With PrefEval, we have evaluated 10 open-sourced and proprietary LLMs in multi-session conversations with varying context lengths up to 100k tokens. We benchmark with various prompting, iterative feedback, and retrieval-augmented generation methods. Our benchmarking effort reveals that state-of-the-art LLMs face significant challenges in following users' preference during conversations. In particular, in zero-shot settings, preference following accuracy falls below 10\% at merely 10 turns (~3k tokens) across most evaluated models. Even with advanced prompting and retrieval methods, preference following still deteriorates in long-context conversations. Furthermore, we show that fine-tuning on PrefEval significantly improves performance. We believe PrefEval serves as a valuable resource for measuring, understanding, and enhancing LLMs' proactive preference following abilities, paving the way for personalized conversational agents.

Effective asynchronous planning, or the ability to efficiently reason and plan over states and actions that must happen in parallel or sequentially, is essential for agents that must account for time delays, reason over diverse long-horizon tasks, and collaborate with other agents. While large language model (LLM) agents show promise in high-level task planning, current benchmarks focus primarily on short-horizon tasks and do not evaluate such asynchronous planning capabilities. We introduce Robotouille, a challenging benchmark environment designed to test LLM agents' ability to handle long-horizon asynchronous scenarios. Our synchronous and asynchronous datasets capture increasingly complex planning challenges that go beyond existing benchmarks, requiring agents to manage over- lapping tasks and interruptions Our results show that ReAct (gpt-4o) achieves 47% on synchronous tasks but only 11% on asynchronous tasks, highlighting significant room for improvement. We further analyze failure modes, demonstrating the need for LLM agents to better incorporate long-horizon feedback and self-audit their reasoning during task execution.

The release of ChatGPT in November 2022 sparked an explosion of interest in post-training and an avalanche of new preference optimization (PO) methods. These methods claim superior alignment by virtue of better correspondence with human pairwise preferences, often measured by LLM-judges. In this work, we attempt to answer the following question -- do LLM-judge preferences translate to progress on other, more concrete metrics for alignment, and if not, why not? We define a concrete metric for alignment, and introduce SOS-Bench (Substance Outweighs Style Benchmark), the largest standardized, reproducible LLM meta-benchmark to date. We find that (1) LLM-judge preferences do not correlate with concrete measures of safety, world knowledge, and instruction following; (2) LLM-judges have powerful implicit biases, prioritizing style over factuality and safety; and (3) the supervised fine-tuning (SFT) stage of post-training has a large impact on alignment, with data scaling and prompt diversity as the driving factors.

Recent advancements in large language model (LLM)-powered agents have shown that collective intelligence can significantly outperform individual capabilities, largely attributed to the meticulously designed inter-agent communication topologies. Though impressive in performance, existing multi-agent pipelines inherently introduce substantial token overhead, as well as increased economic costs, which pose challenges for their large-scale deployments. In response to this challenge, we propose an economical, simple, and robust multi-agent communication framework, termed $\texttt{AgentPrune}$, which can seamlessly integrate into mainstream multi-agent systems and prunes redundant or even malicious communication messages. Technically, $\texttt{AgentPrune}$ is the first to identify and formally define the $\textit{Communication Redundancy}$ issue present in current LLM-based multi-agent pipelines, and efficiently performs one-shot pruning on the spatial-temporal message-passing graph, yielding a token-economic and high-performing communication topology. Extensive experiments across six benchmarks demonstrate that $\texttt{AgentPrune}$ $\textbf{(I)}$ achieves comparable results as state-of-the-art topologies at merely $\\$5.6$ cost compared to their $\\$43.7$, $\textbf{(II)}$ integrates seamlessly into existing multi-agent frameworks with $28.1\\%\sim72.8\\%\downarrow$ token reduction, and $\textbf{(III)}$ successfully defend against two types of agent-based adversarial attacks with $3.5\\%\sim10.8\\%\uparrow$ performance boost. The source code is available at \url{https://github.com/yanweiyue/AgentPrune}.

As LLM-as-a-Judge emerges as a new paradigm for assessing large language models (LLMs), concerns have been raised regarding the alignment, bias, and stability of LLM evaluators. While substantial work has focused on alignment and bias, little research has concentrated on the stability of LLM evaluators. In this paper, we conduct extensive experiments involving 9 widely used LLM evaluators across 2 different evaluation settings to investigate the uncertainty in model-based LLM evaluations. We pinpoint that LLM evaluators exhibit varying uncertainty based on model families and sizes. With careful comparative analyses, we find that employing special prompting strategies, whether during inference or post-training, can alleviate evaluation uncertainty to some extent. By utilizing uncertainty to enhance LLM's reliability and detection capability in Out-Of-Distribution (OOD) data, we further fine-tune an uncertainty-aware LLM evaluator named ConfiLM using a human-annotated fine-tuning set and assess ConfiLM's OOD evaluation ability on a manually designed test set sourced from the 2024 Olympics. Experimental results demonstrate that incorporating uncertainty as additional information during the fine-tuning phase can largely improve the model's evaluation performance in OOD scenarios. The code and data are released at: https://github.com/hasakiXie123/LLM-Evaluator-Uncertainty.

We present MM1.5, a new family of multimodal large language models (MLLMs) designed to enhance capabilities in text-rich image understanding, visual referring and grounding, and multi-image reasoning. Building upon the MM1 architecture, MM1.5 adopts a data-centric approach to model training, systematically exploring the impact of diverse data mixtures across the entire model training lifecycle. This includes high-quality OCR data and synthetic captions for continual pre-training, as well as an optimized visual instruction-tuning data mixture for supervised fine-tuning. Our models range from 1B to 30B parameters, encompassing both dense and mixture-of-experts (MoE) variants, and demonstrate that careful data curation and training strategies can yield strong performance even at small scales (1B and 3B). Additionally, we introduce two specialized variants: MM1.5-Video, designed for video understanding, and MM1.5-UI, tailored for mobile UI understanding. Through extensive empirical studies and ablations, we provide detailed insights into the training processes and decisions that inform our final designs, offering valuable guidance for future research in MLLM development.

Malicious content generated by large language models (LLMs) can pose varying degrees of harm. Although existing LLM-based moderators can detect harmful content, they struggle to assess risk levels and may miss lower-risk outputs. Accurate risk assessment allows platforms with different safety thresholds to tailor content filtering and rejection. In this paper, we introduce per-topic severity rubrics for 11 harmful topics and build BingoGuard, an LLM-based moderation system designed to predict both binary safety labels and severity levels. To address the lack of annotations on levels of severity, we propose a scalable generate-then-filter framework that first generates responses across different severity levels and then filters out low-quality responses. Using this framework, we create BingoGuardTrain, a training dataset with 54,897 examples covering a variety of topics, response severity, styles, and BingoGuardTest, a test set with 988 examples explicitly labeled based on our severity rubrics that enables fine-grained analysis on model behaviors on different severity levels. Our BingoGuard-8B, trained on BingoGuardTrain, achieves the state-of-the-art performance on several moderation benchmarks, including WildGuardTest and HarmBench, as well as BingoGuardTest, outperforming best public models, WildGuard, by 4.3\%. Our analysis demonstrates that incorporating severity levels into training significantly enhances detection performance and enables the model to effectively gauge the severity of harmful responses. Warning: this paper includes red-teaming examples that may be harmful in nature.

Large language models (LLMs) exhibit varying strengths and weaknesses across different tasks, prompting recent studies to explore the benefits of ensembling models to leverage their complementary advantages. However, existing LLM ensembling methods often overlook model compatibility and struggle with inefficient alignment of probabilities across the entire vocabulary. In this study, we empirically investigate the factors influencing ensemble performance, identifying model performance, vocabulary size, and response style as key determinants, revealing that compatibility among models is essential for effective ensembling. This analysis leads to the development of a simple yet effective model selection strategy that identifies compatible models. Additionally, we introduce the \textsc{Uni}on \textsc{T}op-$k$ \textsc{E}nsembling (\textsc{UniTE}), a novel approach that efficiently combines models by focusing on the union of the top-k tokens from each model, thereby avoiding the need for full vocabulary alignment and reducing computational overhead. Extensive evaluations across multiple benchmarks demonstrate that \textsc{UniTE} significantly enhances performance compared to existing methods, offering a more efficient framework for LLM ensembling.

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

Enhancing the adaptive capabilities of large language models is a critical pursuit in both research and application. Traditional fine-tuning methods require substantial data, computational resources, and specific capabilities, while in-context learning is limited by the need for appropriate demonstrations and efficient token usage. Inspired by the expression of in-context learned capabilities through task vectors and the concept of modular capability or knowledge, we propose ELICIT, a framework consisting of two modules designed to effectively store and reuse task vectors to enhance the diverse adaptive capabilities of models without additional training or inference tokens. Our comprehensive experiments and analysis demonstrate that our pipeline is highly transferable across different input formats, tasks, and model architectures. Externally storing and reusing vectors that represent in-context learned capabilities not only shows the potential to extract modular capabilities but also significantly enhances the performance, versatility, adaptability, and scalability of large language models, paving the way for more efficient and effective use of these models in a wide range of applications.

Optimizing the Key-Value (KV) cache of the Large Language Model (LLM) has been considered critical to saving the cost of inference. Most of the existing KV-cache compression algorithms attempted to sparsify the sequence of tokens by taking advantage of the different importance of tokens. However, most of these methods treat all layers equally, allocating the same KV budget to each layer. This approach is suboptimal, as some layers may be less sensitive to input tokens yet still receive the same budget as others. In this work, we found that by identifying the importance of attention layers, we could optimize the KV-cache jointly from two dimensions, i.e., sequence-wise and layer-wise. Based on our observations regarding layer-wise importance in inference, we propose \sys to precisely optimize the allocation of KV-cache budget among layers on-the-fly and then incorporate three representative sequence-wise algorithms to compress the KV-cache for each layer with its very own budget. Specifically, we first measure each layer's importance by calculating the cosine similarity of the input prompt differences before and after the self-attention layers. Based on this similarity, we then categorize the layers into two groups and adjust their KV budgets accordingly. By optimizing the KV-cache from both sequence's and layer's dimensions, \sys achieves around 30\% to 70\% of the memory reductions and up to 2.2 $\times$ of throughput improvements in a wide range of LLMs and benchmarks. The code is available at https://github.com/hetailang/SqueezeAttention.

Large Language Models (LLM) technology is rapidly advancing towards human- like dialogue. Values are fundamental drivers of human behavior, yet research on the values expressed in LLM-generated text remains limited. While prior work has begun to explore value ranking in LLMs, the crucial aspect of value correlation – the interrelationship and consistency between different values – has been largely un-examined. Drawing on established psychological theories of human value structure, this paper investigates whether LLMs exhibit human-like value correlations within a single session, reflecting a coherent “persona”. Our findings reveal that standard prompting methods fail to produce human-consistent value correlations. However, we demonstrate that a novel prompting strategy (referred to as "Value Anchoring"), significantly improves the alignment of LLM value correlations with human data. Furthermore, we analyze the mechanism by which Value Anchoring achieves this effect. These results not only deepen our understanding of value representation in LLMs but also introduce new methodologies for evaluating consistency and human-likeness in LLM responses, highlighting the importance of explicit value prompting for generating human-aligned outputs.