COLM 2025 论文列表

Computer use agents automate digital tasks by directly interacting with graphical user interfaces (GUIs) on computers and mobile devices, offering significant potential to enhance human productivity by completing an open-ended space of user queries. However, current agents face significant challenges: imprecise grounding of GUI elements, difficulties with long-horizon task planning, and performance bottlenecks from relying on single generalist models for diverse cognitive tasks. To this end, we introduce Agent S2, a novel compositional framework that delegates cognitive responsibilities across various generalist and specialist models. We propose a novel Mixture-of-Grounding technique to achieve precise GUI localization and introduce Proactive Hierarchical Planning, dynamically refining action plans at multiple temporal scales in response to evolving observations. Evaluations demonstrate that Agent S2 establishes new state-of-the-art (SOTA) performance on three prominent computer use benchmarks. Specifically, Agent S2 achieves 18.9% and 32.7% relative improvements over leading baseline agents such as Claude Computer Use and UI-TARS on the OSWorld 15-step and 50-step evaluation. Moreover, Agent S2 generalizes effectively to other operating systems and applications, surpassing previous best methods by 52.8% on WindowsAgentArena and by 16.52% on AndroidWorld relatively. Code available at https://github.com/simular-ai/Agent-S.

Recent large language models (LLMs) achieve impressive performance in text generation but often fail to accurately attribute their outputs, undermining trust and verifiability. Moreover, existing attribution methods do not explain how and why models leverage the provided source documents to generate their final responses, limiting interpretability. Furthermore, current attributions fail to provide a reason as to how and why the model uses the context to arrive at the final output. To overcome these challenges, we introduce a modular generation framework, GenerationPrograms, inspired by recent advancements in executable ``code agent'' architectures. Unlike conventional generation methods that simultaneously generate outputs and attributions or rely on post-hoc attribution, GenerationPrograms decomposes the process into two distinct stages: first, creating an executable program plan composed of modular text operations (such as paraphrasing, compression, and fusion) explicitly tailored to the query, and second, executing these operations following the program's specified instructions to produce the final response. Empirical evaluations demonstrate that GenerationPrograms significantly improves attribution quality at both document-level and sentence-level granularity across two long-form question-answering tasks. We further demonstrate that GenerationPrograms can effectively function as a post-hoc attribution method, outperforming traditional techniques in recovering accurate attributions. In addition, the interpretable programs generated by GenerationPrograms enable localized refinement through modular-level improvements that further enhance overall attribution quality.

Can we simulate a sandbox society with generative agents to model human behavior, thereby reducing the over-reliance on real human trials for assessing public policies? In this work, we investigate the feasibility of simulating health-related decision-making, using vaccine hesitancy, defined as the delay in acceptance or refusal of vaccines despite the availability of vaccination services (MacDonald, 2015), as a case study. To this end, we introduce the VacSim framework with 100 generative agents powered by Large Language Models (LLMs). VacSim simulates vaccine policy outcomes with the following steps: 1) instantiate a population of agents with demographics based on census data; 2) connect the agents via a social network and model vaccine attitudes as a function of social dynamics and disease-related information; 3) design and evaluate various public health interventions aimed at mitigating vaccine hesitancy. To align with real-world results, we also introduce simulation warmup and attitude modulation to adjust agents' attitudes. We propose a series of evaluations to assess the reliability of various LLM simulations. Experiments indicate that models like Llama and Qwen can simulate aspects of human behavior but also highlight real-world alignment challenges, such as inconsistent responses with demographic profiles. This early exploration of LLM-driven simulations is not meant to serve as definitive policy guidance; instead, it serves as a call for action to examine social simulation for policy development.

To mitigate reward hacking from response verbosity, modern preference optimization methods are increasingly adopting length normalization (e.g., SimPO, ORPO, LN-DPO). While effective against this bias, we demonstrate that length normalization itself introduces a failure mode: the **URSLA shortcut**. Here models learn to satisfy the alignment objective by prematurely truncating low-quality responses rather than learning from their semantic content. To address this, we introduce **REFA**, a new alignment framework that proposes probabilistic control on a structural token that controls termination. Our core innovation is a new class of regularizers that operate directly on the probability of the End-of-Sequence (EOS) token, a previously unexploited control lever. This token-level intervention provides a principled solution to the URSLA shortcut, ensuring genuine quality improvements. Furthermore, it unlocks a versatile mechanism for managing the alignment-efficiency tradeoff, enabling practitioners to fine-tune models that adhere to specific token budgets. Empirically, REFA achieves a **60.29\%** win rate and a **52.17\%** length-controlled win rate on AlpacaEval2 with Llama-3-8B-Instruct, demonstrating the power of our token-level control paradigm.

Large language models (LLMs) have achieved impressive performance, leading to their widespread adoption as decision-support tools in resource-constrained contexts like hiring and admissions. There is, however, scientific consensus that AI systems can reflect and exacerbate societal biases, raising concerns about identity-based harm when used in critical social contexts. Prior work has laid a solid foundation for assessing bias in LLMs by evaluating demographic disparities in different language reasoning tasks. In this work, we extend single-axis fairness evaluations to examine intersectional bias, recognizing that when multiple axes of discrimination intersect, they create distinct patterns of disadvantage. We create a new benchmark called WinoIdentity by augmenting the WinoBias dataset with 25 demographic markers across 10 attributes, including age, nationality, and race, intersected with binary gender, yielding 245,700 prompts to evaluate 50 distinct bias patterns. Focusing on harms of omission due to underrepresentation, we investigate bias through the lens of uncertainty and propose a group (un)fairness metric called \emph{Coreference Confidence Disparity} which measures whether models are more or less confident for some intersectional identities than others. We evaluate five recently published LLMs and find confidence disparities as high as 40\% along various demographic attributes including body type, sexual orientation and socio-economic status, with models being most uncertain about doubly-disadvantaged identities in anti-stereotypical settings, such as when assigning transgender women to historically male-dominated occupations. Surprisingly, coreference confidence decreases even for hegemonic or privileged markers (e.g., 'White' or 'cisgender'), indicating that the recent impressive performance of LLMs is more likely due to memorization than logical reasoning. Notably, these are two independent failures in value alignment and validity that can compound to cause social harm.

Large Language Models (LLMs) demonstrate remarkable in-context learning capabilities but often struggle with complex, multi-step reasoning. Multi-Agent Debate (MAD) frameworks partially address these limitations by enabling iterative agent interactions. However, they neglect valuable historical insights by treating each new debate independently. In this paper, we propose Memory-Augmented MAD (MeMAD), a parameter-free memory-augmented MAD framework that systematically organizes and reuses past debate transcripts. MeMAD stores structured representations of successful and unsuccessful reasoning attempts enriched with self-reflections and peer feedback. It systematically retrieves them via semantic similarity at inference time to inform new reasoning tasks. Our experiments on challenging mathematical reasoning, scientific question answering, and language understanding benchmarks show that MeMAD achieves significant accuracy gains (up to 3.3\% over conventional MAD baselines) without parameter updates. Our findings underscore structured memory as a pivotal mechanism for achieving deeper and more reliable multi-agent reasoning in LLMs. Code is available in ~\url{https://github.com/LSHCoding/MeMAD}.

AI assistants interact with millions of real users everyday, imparting normative judgments that can have significant personal and societal impact—but little is known about what values guide these interactions in practice. To address this, we develop a method to empirically analyze values expressed in hundreds of thousands of real-world conversations with Claude models. We empirically discover and taxonomize 3,308 AI values, and study how model values and responses depend on context. We find that Claude expresses many professional and intellectual values, and typically supports prosocial human values while resisting values like "moral nihilism." While some values appear consistently (e.g. "professionalism"), most are highly context-dependent—"harm prevention" emerges when the model resists users, "historical accuracy" when discussing controversial events, "healthy boundaries" in relationship advice, and "human agency" in technology ethics discussions. By providing the first large-scale empirical mapping of AI values in deployment, this work creates a foundation for more grounded evaluation and design of values in increasingly influential AI systems.

Large language models (LLMs) are increasingly capable of simulating human behavior, offering cost-effective ways to estimate user responses during the early phases of survey design. While previous studies have examined whether models can reflect individual opinions or attitudes, we argue that a higher-order binding of virtual personas requires successfully approximating not only the opinions of a user as an identified member of a group, but also the nuanced ways in which that user perceives and evaluates those outside the group. In particular, faithfully simulating how humans perceive different social groups is critical for applying LLMs to various political science studies, including timely topics on polarization dynamics, inter-group conflict, and democratic backsliding. To this end, we propose a novel methodology for constructing virtual personas with synthetic user "backstories" generated as extended, multi-turn interview transcripts. Our generated backstories are longer, rich in detail, and consistent in authentically describing a singular individual, compared to previous methods. We show that virtual personas conditioned on our backstories closely replicate human response distributions (up to an 87% improvement as measured by Wasserstein Distance) and produce effect sizes that closely match those observed in the original studies. Altogether, our work extends the applicability of LLMs beyond estimating individual self-opinions, enabling their use in a broader range of human studies.

As large language models become increasingly capable at various tasks including writing, the need to generate unique and creative content arises. Although LLMs have the ability to generate text covering diverse topics, there is an overall sense of repetitiveness across texts that we aim to formalize. Such familiarity between documents is induced through the persistence of underlying discourse structures. However, existing similarity metrics dependent on lexical overlap and syntactic patterns are overly sensitive to volatility in content overlap, thus making them unsuitable for detecting $\textit{structural}$ similarities. We introduce an abstraction based on linguistics theories in Questions Under Discussion (QUD) and question semantics to help quantify differences in discourse progression. We then use this framework to build $\textbf{QUDsim}$, a similarity metric that can detect discursive parallels between documents. Using QUDsim, we find that LLMs often reuse discourse structures (more so than humans) to create seemingly new documents by simply swapping content. Furthermore, LLMs are not only repetitive and structurally uniform, but are also divergent from human authors in the types of structures they use.

Large Language Models (LLMs) have demonstrated promising capabilities to generate responses that simulate consistent personality traits. Despite the major attempts to analyze personality expression through output-based evaluations, little is known about how such traits are internally encoded within LLM parameters. In this paper, we introduce a layer-wise probing framework to systematically investigate the layer-wise capability of LLMs in simulating personality for responding. We conduct probing experiments on 11 open-source LLMs over the PersonalityEdit benchmark and find that LLMs predominantly simulate personality for responding in their middle and upper layers, with instruction-tuned models demonstrating a slightly clearer separation of personality traits. Furthermore, by interpreting the trained probing hyperplane as a layer-wise boundary for each personality category, we propose a layer-wise perturbation method to edit the personality expressed by LLMs during inference. Our results show that even when the prompt explicitly specifies a particular personality, our method can still successfully alter the response personality of LLMs. Interestingly, the difficulty of converting between certain personality traits varies substantially, which aligns with the representational distances in our probing experiments. Finally, we conduct a comprehensive MMLU benchmark evaluation and time overhead analysis, demonstrating that our proposed personality editing method incurs only minimal degradation in general capabilities while maintaining low training costs and acceptable inference latency. Our code is publicly available at https://github.com/universe-sky/probing-then-editing-personality.

Despite the success of text retrieval in many NLP tasks, code retrieval remains a largely underexplored area. Most text retrieval systems are tailored for natural language queries, often neglecting the specific challenges of retrieving code. This gap leaves existing models unable to effectively capture the diversity of programming languages and tasks across different domains, highlighting the need for more focused research in code retrieval. To address this, we introduce CodeXEmbed, a family of large-scale code embedding models ranging from 400M to 7B parameters. Our novel training pipeline unifies multiple programming languages and transforms various code-related tasks into a common retrieval framework, enhancing model generalizability and retrieval performance. Our 7B model achieves a new state-of-the-art (SOTA) in code retrieval, topping the CoIR Leaderboard. In addition to excelling in code retrieval, our models demonstrate competitive performance on the widely adopted BeIR text retrieval benchmark, offering versatility across domains. Experimental results demonstrate that improving retrieval performance significantly enhances end-to-end Retrieval-Augmented Generation (RAG) performance for code-related tasks.

Large language model (LLM) agents are increasingly employing retrieval-augmented generation (RAG) to improve the factuality of their responses. However, in practice, these systems often need to handle ambiguous user queries and potentially conflicting information from multiple sources while also suppressing inaccurate information from noisy or irrelevant documents. Prior work has generally studied and addressed these challenges in isolation, considering only one aspect at a time, such as handling ambiguity or robustness to noise and misinformation. We instead consider multiple factors simultaneously, proposing (i) RAMDocs (Retrieval with Ambiguity and Misinformation in Documents), a new dataset that simulates complex and realistic scenarios for conflicting evidence for a user query, including ambiguity, misinformation, and noise; and (ii) MADAM-RAG, a multi-agent approach in which LLM agents debate over the merits of an answer over multiple rounds, allowing an aggregator to collate responses corresponding to disambiguated entities while discarding misinformation and noise, thereby handling diverse sources of conflict jointly. We demonstrate the effectiveness of MADAM-RAG using both closed and open-source models on AmbigDocs – which requires presenting all valid answers for ambiguous queries – improving over strong RAG baselines by up to 11.40%, and on FaithEval – which requires suppressing misinformation – where we improve by up to 15.80% (absolute) with Llama3.3-70B-Instruct. Furthermore, we find that our proposed RAMDocs dataset poses a challenge for existing RAG baselines (the most performant Llama3.3-70B-Instruct only yields up to a 32.60 exact match score), as it requires handling conflicting information due to ambiguity, noise, and misinformation simultaneously. While MADAM-RAG begins to address these conflicting factors, our analysis indicates that a substantial gap remains, especially when increasing the level of imbalance in supporting evidence and misinformation.

Generation capabilities and language coverage of multilingual large language models (mLLMs) are advancing rapidly. However, evaluation practices for generative abilities of mLLMs are still lacking comprehensiveness, scientific rigor, and consistent adoption across research labs, which undermines their potential to meaningfully guide mLLM development. We draw parallels with machine translation (MT) evaluation, a field that faced similar challenges and has, over decades, developed transparent reporting standards and reliable evaluations for multilingual generative models. Through targeted experiments across key stages of the generative evaluation pipeline, we demonstrate how best practices from MT evaluation can deepen the understanding of quality differences between models. Additionally, we identify essential components for robust meta-evaluation of mLLMs, ensuring the evaluation methods themselves are rigorously assessed. We distill these insights into a checklist of actionable recommendations for mLLM research and development.

Multilingual vision-language models have made significant strides in image captioning, yet they still lag behind their English counterparts due to limited multilingual training data and costly large-scale model parameterization. Retrieval-augmented generation (RAG) offers a promising alternative by conditioning caption generation on retrieved examples in the target language, reducing the need for extensive multilingual training. However, multilingual RAG captioning models often depend on retrieved captions translated from English, which can introduce mismatches and linguistic biases relative to the source language. We introduce CONCAP, a multilingual image captioning model that integrates retrieved captions with image-specific concepts, enhancing the contextualization of the input image and grounding the captioning process across different languages. Experiments on the XM3600 dataset indicate that CONCAP enables strong performance on low- and mid-resource languages, with highly reduced data requirements. Our findings highlight the effectiveness of concept-aware retrieval augmentation in bridging multilingual performance gaps.

While the inconsistency of LLMs is not a novel topic, prior research has predominantly addressed two types of generative inconsistencies: i) Randomness Inconsistency: running the same LLM multiple trials, yielding varying responses; ii) Paraphrase Inconsistency: paraphrased prompts result in different responses from the same LLM. Randomness Inconsistency arises from the inherent randomness due to stochastic sampling in generative models, while Paraphrase Inconsistency is a consequence of the language modeling objectives, where paraphrased prompts alter the distribution of vocabulary logits. This research discovers Prompt-Reverse Inconsistency (PRIN), a new form of LLM self-inconsistency: given a question and a couple of LLM-generated answer candidates, the LLM often has conflicting responses when prompted "Which are correct answers?" and "Which are incorrect answers?". PRIN poses a big concern as it undermines the credibility of LLM-as-a-judge, and suggests a challenge for LLMs to adhere to basic logical rules. We conduct a series of experiments to investigate PRIN, examining the extent of PRIN across different LLMs, methods to mitigate it, potential applications, and its relationship with Randomness Inconsistency and Paraphrase Inconsistency. As the first study to explore PRIN, our findings offer valuable insights into the inner workings of LLMs and contribute to advancing trustworthy AI.

Unit tests (UTs) play an instrumental role in assessing code correctness as well as providing feedback to large language models (LLMs), motivating automated test generation. However, we uncover a trade-off between generating unit test inputs that reveal errors when given a faulty code and correctly predicting the unit test output without access to the gold solution. To address this trade-off, we propose UTGen, which teaches LLMs to generate unit test inputs that reveal errors along with their correct expected outputs based on task descriptions. Since model-generated tests can provide noisy signals (e.g., from incorrectly predicted outputs), we propose UTDebug that (i) scales UTGen via test-time compute to improve UT output prediction, and (ii) validates and backtracks edits based on multiple generated UTs to avoid overfitting, and helps LLMs debug effectively. We show that UTGen outperforms other LLM-based baselines by 7.59% based on a metric measuring the presence of both error-revealing UT inputs and correct UT outputs. When used with UTDebug, we find that feedback from UTGen's unit tests improves pass@1 accuracy of Qwen2.5 32B on HumanEvalFix and our own harder debugging split of MBPP+ by over 3.17% and 12.35% (respectively) over other LLM-based UT generation baselines. Moreover, we observe that feedback from Qwen2.5 32B-based UTGen model can enhance debugging with frontier LLMs like GPT-4o by 13.8%. Lastly, we demonstrate that UTGen is a better judge for code correctness, outperforming a state-of-the-art trained 8B reward model by 4.43% on HumanEval+ with best-of-10 sampling using Qwen2.5 7B.

Recent advances in video-large language models (Video-LLMs) have led to significant progress in video understanding. Current preference optimization methods often rely on proprietary APIs or ground-truth captions to generate preference data (i.e., pairs of model outputs ranked based on their quality or alignment with human judgment), which is then used to train models for video-language alignment. This approach is both costly and labor-intensive. To address this limitation, we introduce $\textbf{VideoSAVi}$ ($\underline{\textbf{S}}$elf-$\underline{\textbf{A}}$ligned $\underline{\textbf{Vi}}$deo Language Model), a self-training pipeline that enables Video-LLMs to reason over video content without external supervision. Our approach includes a self-critiquing mechanism that identifies reasoning errors in the model's initial responses and generates improved alternatives, creating preference pairs directly from video content. VideoSAVi then applies Direct Preference Optimization (DPO) to iteratively train the model using the preference data, thus enhancing its temporal and spatial reasoning for video understanding. Experiments show that VideoSAVi delivers significant improvements across multiple benchmarks, including a +4.2 percentage point gain on MVBench, +3.9 on PerceptionTest, and +6.8 on the challenging EgoSchema dataset compared to baseline models. Our model-agnostic approach is computationally efficient, requiring only 32 frames, offering a promising direction for self-aligned video understanding without reliance on external models or annotations.

Training of large language models (LLMs) is typically distributed across a large number of accelerators to reduce training time. Since internal states and parameter gradients need to be exchanged at each and every single gradient step, all devices need to be co-located using low-latency high-bandwidth communication links to support the required high volume of data exchange. Recently, algorithms like DiLoCo have relaxed the constraint that all devices need co-location: accelerators can be grouped into ``workers'', where synchronizations between workers need only occur infrequently. This in turn means that workers can afford being connected by lower bandwidth communication links without affecting learning quality. However, in these methods, communication across workers still requires the same peak bandwidth as before, as the synchronizations require all parameters to be exchanged across all workers. In this paper, we improve DiLoCo in three ways. First, we synchronize only subsets of parameters in sequence, rather than all at once, which greatly reduces peak bandwidth. Second, we allow workers to continue training while synchronizing, which decreases wall clock time. Third, we quantize the data exchanged by workers, which further reduces bandwith across workers. We show experimentally that by properly combining these modifications we can distribute training of billion-scale parameters and attain models of similar quality as before, but reducing required bandwidth by a factor of up to two orders of magnitude.

Large Audio Language Models (LALMs) have significantly advanced audio understanding but introduce critical security risks, particularly through audio jailbreaks. While prior work has focused on English-centric attacks, we expose a far more severe vulnerability: adversarial multilingual and multi-accent audio jailbreaks, where linguistic and acoustic variations dramatically amplify attack success. In this paper, we introduce Multi-AudioJail, the first systematic framework to exploit these vulnerabilities through (1) a novel dataset of adversarially perturbed multilingual/multi-accent audio jailbreaking prompts, and (2) a hierarchical evaluation pipeline revealing that how acoustic perturbations (e.g., reverberation, echo, and whisper effects) interacts with cross-lingual phonetics to cause jailbreak success rates (JSRs) to surge by up to +57.25 percentage points (e.g., reverberated Kenyan-accented attack on MERaLiON). Crucially, our work further reveals that multimodal LLMs are inherently more vulnerable than unimodal systems: attackers need only exploit the weakest link (e.g., non-English audio inputs) to compromise the entire model, which we empirically show by multilingual audio-only attacks achieving 3.1x higher success rates than text-only attacks. We plan to release our dataset to spur research into cross-modal defenses, urging the community to address this expanding attack surface in multimodality as LALMs evolve.